← Contents · Index →

Glossary

673 terms. Each links to the chapter where it is defined; the companion to the index (where terms are discussed) and to the symbol list (Notations).

3-chip / 3-CCD camera — A color-capture design that splits incoming light with a prism onto three separate sensors, one per primary, giving full-resolution color with no interpolation at the cost of bulk and alignment. (3.12 Demosaicking)

35mm-equivalent focal length — A focal length expressed as the value giving the same field of view on full-frame, comparing differently sized sensors. (22.12 EXIF and image metadata)

Abbe number — A glass's dispersion measure (high means low dispersion). (9.1 Aberrations and optical challenges)

accommodation — The eye's focusing mechanism, in which the lens changes shape to focus objects at different distances. (2.8 Human (and animal) vision and color)

ACES (Academy Color Encoding System) — A color-management framework defining common working and interchange spaces so footage from different cameras grades and intercuts predictably. (3.3 Point operations)

achromat — A cemented crown-plus-flint doublet, the minimum system that corrects chromatic aberration. (9.2 Aberrations correction)

achromat condition — The relation under which a doublet's dispersion-weighted powers cancel, bringing two wavelengths to a common focus. (9.2 Aberrations correction)

active contour (snake) — An elastic curve dropped near an object that relaxes onto its boundary by minimizing an energy balancing smoothness, attraction to edges, and user constraints. (5.8 Seam optimization)

Adaptive Homogeneity-Directed (AHD) demosaicking — A high-quality demosaicker that reconstructs the image both horizontally and vertically and, per pixel, keeps whichever direction yields the more homogeneous local color neighborhood. (3.12 Demosaicking)

adaptive sharpening — Sharpening whose boost is gated by local context — strong in textured detail, backed off in smooth regions and at extreme edges — to avoid amplifying noise and producing halos. (3.6 Neighborhood operations and convolution)

affine bilateral grid (HDRnet) — A network that predicts a per-cell affine color/tone transform on a low-res image, then slices and applies it at full resolution for fast edge-aware enhancement. (5.2 Bilateral filtering)

affinity — a similarity weight between pixels (in space and/or value) that decides how much they influence each other; the core of edge-preserving filtering. (Bilateral filtering)

afterimage — A ghost image of the opposite color seen after staring at a colored patch, from opponent-channel adaptation. (2.8 Human (and animal) vision and color)

airlight — The bright, roughly constant scattered atmospheric light added to each pixel in proportion to 1−t in the haze model. (8.4 Dehazing)

Airy disk — the bright central disk (plus faint rings) a point source forms through a perfect circular aperture; its size (∝ λ·N) sets the diffraction limit. (Wave effects, diffraction, and the diffraction limit)

aliasing — high frequencies masquerading as low ones when a signal is sampled below the Nyquist rate; seen as moiré/jaggies, prevented by prefiltering before downsampling. (Linearity, Fourier, aliasing and deblurring)

alpha blending (feathering) — Compositing overlapping frames as a distance-to-boundary weighted average, which softens a seam but ghosts misaligned detail. (10.7 Blending)

alpha channel — A fourth per-pixel channel encoding opacity/coverage (RGBA), carrying the premultiplied-vs-straight pitfall and the common "PNG loads as four channels" trap. (3.1 Image representation)

alpha-expansion — An approximate solver for multi-label graph-cut energies that iterates a sequence of binary graph cuts. (5.8 Seam optimization)

alpha matte — The per-pixel foreground coverage α in the compositing equation C=αF+(1−α)B, fractional at soft boundaries. (8.8 Compositing, segmentation and matting)

anamorphic lens — A lens of cylindrical elements that squeezes the image horizontally to fit a wide scene on a narrow frame, de-squeezed later. (9.5 Special optics)

anisotropic diffusion (Perona–Malik) — Edge-stopping smoothing that diffuses freely where neighbors agree but halts across large gradients; the bilateral filter is its large-step form. (5.2 Bilateral filtering)

aperture — the lens opening that admits light; its diameter (via the f-number) trades light for depth of field and diffraction. (Thin lens optics)

aperture problem — The fact that from a single pixel only the motion component perpendicular to the local edge is recoverable. (7.6 Optical flow)

apochromat — A lens (ED or fluorite glass) that corrects chromatic aberration at three or more wavelengths. (9.2 Aberrations correction)

apodization — Shaping the aperture's transmission with a radially graded filter so the defocus disk has a soft rim. (9.5 Special optics)

APP1 marker segment — The stretch of a JPEG file, skipped by the image decoder, where the EXIF metadata payload is stored. (22.12 EXIF and image metadata)

ARAP (as-rigid-as-possible) deformation — A free-form warp regularizer keeping each local neighborhood close to a rotation-plus-translation so handles bend shapes without stretching. (6.1 Warping and resampling)

aspheric element — A surface departing from a sphere by a designed polynomial, killing spherical aberration with fewer elements. (9.2 Aberrations correction)

AsShotNeutral — The DNG tag recording the camera's own white-balance guess at capture. (22.13 DNG, the Digital Negative)

astigmatism — An off-axis aberration in which radial and tangential detail focus at different depths. (9.1 Aberrations and optical challenges)

auto-exposure (AE) — The camera's automatic choice of aperture, shutter, and ISO to place the scene's representative tone near middle grey. (3.13 Auto-exposure and auto white balance)

autofocus (contrast-detection, CDAF) — Autofocus maximizing high-frequency energy over lens position, accurate but blind to direction so it must hunt. (9.6 Focus, autofocus)

automatic white balance (AWB) — The camera's automatic estimate of the scene illuminant from the image itself, so it can be divided out. (3.13 Auto-exposure and auto white balance)

autoregressive image model — A generative model producing image tokens one at a time, the same next-token recipe as a language model. (4.5 Generative AI and diffusion)

B-frame — A video frame predicted bidirectionally from a past and a future reference, the smallest type but needing out-of-order decoding. (12.2 Video compression and motion compensation)

back-button focus — A handling technique decoupling autofocus from the shutter button, triggering focus with a thumb button. (2.10 Photography and camera 101)

back focal length — The distance from the last glass surface to the sensor, distinct from the effective focal length. (9.4 A short bestiary of classic designs)

backpropagation — The algorithm that computes the gradients used to fit a network's parameters during training. (4.3 Machine learning)

Bahtinov mask — An aperture grid throwing diffraction spikes that cross symmetrically only at perfect focus. (9.6 Focus, autofocus)

banding — visible steps in a smooth gradient from too-coarse quantization (too few code values / bit depth). (File formats and compression)

basis — A choice of coordinate system for a vector space; changing basis re-expresses the same vector in new coordinates. (22.1 Refreshers)

Bayer mosaic — the RGGB color-filter array over a sensor; each photosite sees one color, and demosaicking reconstructs full RGB per pixel. (Image formation → sensors)

Bayes' rule — The relation posterior ∝ likelihood × prior, giving what to believe about a scene after a noisy measurement. (22.1 Refreshers)

Beier–Neely field warping — A free-form warp driven by user-drawn before/after line-segment pairs, combined by a distance- and length-weighted average. (6.1 Warping and resampling)

benchmark — A standard public test set on which a method's quality is measured. (22.14 Datasets)

Bessel's correction — Dividing a stack's summed squared residuals by N−1 to get an unbiased per-pixel noise variance. (10.1 Denoising by averaging)

bias–variance trade-off — In denoising, the tension whereby a wider averaging filter cuts noise more but blurs detail more, and a narrower one does the reverse. (3.11 Denoising)

bicubic — Reconstruction by convolution with a fixed piecewise-cubic kernel (four samples wide), the standard editor default, with small negative side-lobes that add mild built-in sharpening. (3.8 Resampling)

bidirectional similarity — An image-completion objective requiring the result to be both coherent and complete with respect to source patches. (8.7 Patch match)

bilateral filter — an edge-preserving blur that weights neighbors by both spatial and value distance, so it smooths within regions but not across edges. (Bilateral filtering)

bilateral grid — A representation lifting each pixel into (x,y,I) so the bilateral filter becomes a plain separable 3-D Gaussian blur (splat, blur, slice). (5.2 Bilateral filtering)

bilinear interpolation — 2-D reconstruction done as two 1-D linear interpolations (one per axis), equivalently convolution with a tent kernel. (3.8 Resampling)

blind deblurring — Recovering both the sharp image and the unknown blur kernel from one blurry photo, broken by a natural-image prior. (8.3 Blind deblurring)

block matching — The per-block correspondence search in a codec that records a motion vector to the best-matching reference patch. (12.2 Video compression and motion compensation)

block-matching and 3-D filtering (BM3D) — A patch-based denoiser that groups similar patches into a stack and filters them jointly in a transform domain; long the benchmark for single-image denoising. (3.11 Denoising)

bokeh — The aesthetic shape and brightness profile of out-of-focus blur, the image of the aperture stop. (9.7 Depth of field)

boundary leakage — Spurious broadband energy (a cross in the spectrum) injected by the DFT's cyclic wrap-around where an image's opposite edges disagree; suppressed by windowing. (3.7 Linearity, Fourier, aliasing and deblurring)

box filter — The crudest blur kernel, giving every pixel in a square window equal weight; cheap but directional and prone to faint ringing. (3.6 Neighborhood operations and convolution)

BRDF (bidirectional reflectance distribution function) — for light arriving from one direction, how much a surface scatters toward each outgoing direction; the material's directional signature. (Light and physics)

brightness — An additive tonal offset that adds a constant to every value, lifting true black off the floor and fogging the shadows. (3.3 Point operations)

brightness constancy — The assumption that a scene point keeps the same intensity as it moves between frames, the basis of optical flow. (7.6 Optical flow)

Brown–Conrady model — A polynomial radial-distortion model, barrel for k1<0 and pincushion for k1>0. (9.1 Aberrations and optical challenges)

bullet-time — A capture rig of many synchronized cameras so one frozen instant can be viewed from a sweep of angles. (17.1 Human factors and the art of photography)

bundle adjustment — Jointly optimizing all cameras' parameters to minimize total reprojection error so a panorama loop closes. (10.8 Bells and whistles)

burst photography — Capturing many short identical frames, then aligning and robustly merging them to recover SNR, range, or resolution. (10.4 Application to cell phones HDR+ and burst imaging)

burst super-resolution — Reconstruction-based super-resolution fusing sub-pixel-shifted frames so hand-tremor offsets interleave into a denser sampling grid. (8.2 Super-resolution and image priors)

C2PA / content credentials — A cryptographic standard for signing an image's origin and edit history so a viewer can verify it. (17.2 Ethics of computational photography)

calibration frames (bias, dark, flat) — Reference frames that remove fixed-pattern errors averaging cannot touch. (10.1 Denoising by averaging)

camera obscura — A darkened box with a small hole that projects an inverted image of the world — the ancestor of every camera. (1.1 From Digital to Computational Photography)

camera path — The trajectory the camera swung through over a clip, the signal that stabilization smooths. (12.4 Video stabilization and rolling-shutter correction)

cardinal points — The reference planes/points from which a thick compound lens behaves like an ideal thin lens. (9.4 A short bestiary of classic designs)

cat's-eye bokeh — Off-axis defocus disks clipped to a lens shape by optical vignetting, producing the swirly look. (9.7 Depth of field)

catadioptric system — A folded mirror-and-lens design, achromatic by construction, whose central obstruction gives doughnut bokeh. (9.5 Special optics)

catchlight — The small reflected highlight in a subject's eye that makes it look alive. (17.1 Human factors and the art of photography)

Catmull–Rom — An interpolating cubic resampling kernel that passes through the samples with pronounced negative lobes, looking sharp but prone to ringing. (3.8 Resampling)

center-surround — A receptive-field organization subtracting a weighted neighbor average from the center, responding to local contrast. (2.8 Human (and animal) vision and color)

chain rule — The rule for differentiating a composition, the engine behind backpropagation. (22.1 Refreshers)

chroma keying — Matting made well-posed by shooting against a known uniform saturated background so B is known everywhere. (8.8 Compositing, segmentation and matting)

chroma subsampling — Storing the two chroma channels at reduced resolution (e.g. 4:2:0) while keeping luma full, exploiting the eye's low color acuity. (3.14 File formats and compression)

chromatic aberration — A color error where the wavelength-dependent index gives a lens different focal lengths for blue and red. (9.1 Aberrations and optical challenges)

CIE — The standards body whose perceptual lightness L* underlies notions like "middle gray.". (2.10 Photography and camera 101)

CinemaDNG — The motion-picture variant of DNG: a per-frame raw sequence for video. (22.13 DNG, the Digital Negative)

circle of confusion (c) — the blur disk a defocused point forms on the sensor; the acceptable size of it sets depth of field. (Thin lens optics)

classifier-free guidance (CFG) — A diffusion sampling knob that extrapolates along the difference between conditioned and unconditioned predictions, trading diversity for prompt adherence. (4.5 Generative AI and diffusion)

clipping — The loss of tones pushed past 0 or 1, which snap to pure black or white and cannot be recovered. (3.3 Point operations)

closed-form matting — Natural-image matting assuming a local color-line model to eliminate F and B, leaving a quadratic energy in α solved via the matting Laplacian. (8.8 Compositing, segmentation and matting)

coarse-to-fine alignment — Estimating motion cheaply at the top of an image pyramid then refining down, avoiding local minima. (10.2 Image alignment)

coarse-to-fine warping — A pyramid scheme that rescues sub-pixel flow for large motion by estimating on coarse levels then refining the residual. (7.6 Optical flow)

coded aperture — A patterned mask replacing the round aperture so the out-of-focus blur becomes invertible and encodes depth. (1.1 From Digital to Computational Photography)

coded (flutter) shutter — A shutter chopped open and closed during exposure so the motion-blur kernel becomes cleanly deconvolvable. (12.1 Motion blur and temporal sampling)

color blindness — A condition where one cone type is missing, so spectra project onto two numbers and many colors become confusable. (2.8 Human (and animal) vision and color)

color constancy — Perceiving an object's intrinsic surface color despite changes in the illuminant, by discounting the light. (2.8 Human (and animal) vision and color)

color difference (R−G, B−G) — The cross-channel difference signals that, being nearly constant across edges, can be interpolated naively in green-based demosaicking and added back to green. (3.12 Demosaicking)

color filter array (CFA) — The regular grid of colored filters laid over a sensor so each photosite measures one color; the Bayer mosaic is the common case. (3.12 Demosaicking)

color fringing — Spurious colored speckles straddling a sharp edge in naive demosaicking, caused by the R, G, B channels being sampled at different positions. (3.12 Demosaicking)

color grading — The motion-picture craft of shaping a film's look using lift/gamma/gain wheels, primary/secondary corrections, and shipped looks. (3.3 Point operations)

color rendering index (CRI) — A score of how faithfully a light source renders test colors versus a reference of the same color temperature. (3.13 Auto-exposure and auto white balance)

color space / primaries — the primaries (sRGB, Adobe RGB, Display P3, …) that fix what RGB numbers actually mean as colors. (Color technology)

color temperature — the temperature (in kelvin) of the blackbody whose glow matches a light's color; low K = warm/orange, high K = cool/blue. (Light and physics)

color transfer — A color-matching method (Reinhard et al. 2001) that matches per-channel mean and standard deviation in the decorrelated lαβ space. (3.4 Histograms)

Color2Gray — A grayscale conversion preserving isoluminant contrast by solving for the gray image whose gradients match signed CIELAB differences. (8.3 Blind deblurring)

colorization (scribble-based) — Propagating user color scribbles across a gray image by an affinity-weighted least-squares solve that stops color at edges. (8.3 Blind deblurring)

ColorMatrix — A DNG tag encoding the transform between the camera's raw RGB and standard XYZ color space. (22.13 DNG, the Digital Negative)

coma — An off-axis aberration smearing a point into a one-sided comet shape. (9.1 Aberrations and optical challenges)

complex steerable pyramid — A multiscale decomposition into oriented band-pass sub-bands whose local phase encodes position, used in phase-based magnification. (12.3 Video magnification)

compositing equation — The model C=αF+(1−α)B; read forward it is compositing, read backward it is matting. (8.8 Compositing, segmentation and matting)

composition — The arrangement of elements within the frame; deciding what to include, exclude, and where to place the subject. (17.1 Human factors and the art of photography)

computational photography — Inserting arbitrary computation between the photons of a scene and the final picture. (1.1 From Digital to Computational Photography)

cone — One of three retinal photoreceptor types (L, M, S) working in daylight, the basis of color vision. (2.8 Human (and animal) vision and color)

conjugate gradient (CG) — A matrix-free iterative solver for symmetric positive-definite systems that picks non-interfering conjugate directions, converging in tens of steps. (4.1 Linear Inverse Problems and Regression)

conservative vector field — A field that is the exact gradient of some image (curl-free, integrable), the case admitting an exact integration and the membrane shortcut. (5.1 Poisson image editing)

content-aware fill — An editing operation that erases an object and fills the hole with plausibly matching content. (2.10 Photography and camera 101)

content-preserving warp — A spatially-varying, as-rigid-as-possible warp that stabilizes footage with genuine 3-D parallax. (12.4 Video stabilization and rolling-shutter correction)

content vs style — The neural-style split where content is what deep features encode and style is the location-discarded correlations among them. (8.5 Style transfer)

context encoders — The first CNN inpainter, an encoder–decoder trained with reconstruction plus adversarial loss, producing blurry semantic fills. (8.6 Inpainting, texture synthesis)

contrast (point operation) — A steepening of the tone curve about a fixed mid-grey pivot, pushing values away from (or toward) the pivot by a gain factor. (3.3 Point operations)

contrast-detection autofocus — Autofocus driving the lens until image sharpness is maximal, accurate but must hunt. (2.10 Photography and camera 101)

contrast sensitivity function (CSF) — The eye's sensitivity as a function of spatial frequency, the perceptual fact JPEG quantization and VDP metrics build on. (3.7 Linearity, Fourier, aliasing and deblurring)

ControlNet — A trainable side branch that injects a spatial hint (edges, pose, depth) into a frozen diffusion model so the output obeys both hint and prompt. (4.5 Generative AI and diffusion)

convexity — A function being bowl-shaped with no false bottoms, so a zero gradient pins down the global minimum. (22.1 Refreshers)

convolution — sliding a kernel over an image, replacing each pixel by a weighted sum of its neighbors; linear shift-invariant filtering. (Neighborhood operations and convolution)

convolution kernel — The small table of weights that defines a convolution; picking the weights picks the operation. (3.6 Neighborhood operations and convolution)

convolution theorem — The identity F{I∗g} = ηĝ — convolution in space is multiplication frequency by frequency in the Fourier domain. (3.7 Linearity, Fourier, aliasing and deblurring)

convolutional neural network (CNN) — A deep network that slides learnable filters over an image, the bread-and-butter architecture for learned image operators. (4.4 Deep learning)

Cooke triplet — A positive–negative–positive design, the minimum with enough parameters to correct all primary Seidel and chromatic aberrations. (9.4 A short bestiary of classic designs)

coring — Pyramid/wavelet denoising that clamps small band-pass coefficients to zero and keeps the large ones, exploiting natural-image sparsity. (3.9 Linear pyramids and wavelets)

correlation — The convolution-like operation without the kernel flip, natural for template matching but lacking convolution's clean algebra. (3.6 Neighborhood operations and convolution)

CoTracker — A learned point tracker that tracks many points jointly (a transformer over a window) for global consistency and occlusion handling. (7.2 Feature tracking)

covariant derivative (healing brush) — The multiplicative, log-domain form of gradient-domain cloning that transports ratios so a patch matches the destination's illumination. (5.1 Poisson image editing)

critical flicker-fusion frequency — The flicker rate above which a flickering source fuses into a steady one. (2.8 Human (and animal) vision and color)

cross-dissolve — The linear blend of two images' colors at fixed locations; correct for color but ghosts when features are misaligned. (6.3 Morphing)

cross / joint bilateral filter — A bilateral filter that reads its edge weight from a separate guide image while averaging the target (e.g. flash/no-flash). (5.2 Bilateral filtering)

cubic B-spline — The smoothest cubic-family kernel, all-positive and approximating rather than interpolating, so it visibly blurs. (3.8 Resampling)

cumulative distribution function (CDF) — The cumulative histogram rising monotonically 0→1, which doubles as the histogram-equalization tone curve. (3.4 Histograms)

damped least squares — The Levenberg–Marquardt update used by lens-design software, blending a least-squares and a gradient step. (9.3 Lens optimization)

dark channel prior — The observation that a haze-free outdoor patch has some channel near zero somewhere; haze lifts it, reading out transmission. (8.4 Dehazing)

data term — In a graph-cut energy, the unary cost measuring how well a pixel fits each label on its own. (5.8 Seam optimization)

dataset — A curated collection of images (often labeled or paired) that a learned method is trained and evaluated on. (22.14 Datasets)

DateTimeOriginal — The EXIF timestamp of when a photo was taken, as opposed to when it was written or scanned. (22.12 EXIF and image metadata)

DCF — The standard governing how cameras name files and folders on a card. (22.12 EXIF and image metadata)

DCP (DNG camera profile) — A DNG profile adding hue, saturation, and tone adjustments on top of the color matrices. (22.13 DNG, the Digital Negative)

DDPM objective — The diffusion training loss that trains a network to predict the Gaussian noise added at a random level. (4.5 Generative AI and diffusion)

Debevec–Malik calibration — Recovering a camera's log response curve and per-pixel radiances jointly by least squares from multiple exposures. (10.3 HDR merging)

deconvolution — The inverse problem of undoing a blur (a convolution) to recover the sharp image. (3.7 Linearity, Fourier, aliasing and deblurring)

deepfake — A synthesized or manipulated image/video at the extreme of the manipulation continuum. (17.2 Ethics of computational photography)

deghosting — Routing a stitch seam (or weighting a merge) so a moving object comes entirely from one frame. (10.8 Bells and whistles)

degrees-of-freedom ladder — The ordered family of parametric warps (translation→rigid→similarity→affine→projective), each adding freedom and giving up a preserved property. (6.1 Warping and resampling)

dehazing — Recovering clear scene radiance by inverting the scattering model I=Jt+A(1−t), with the dark-channel prior supplying transmission. (8.4 Dehazing)

demosaicking — reconstructing full per-pixel RGB from the Bayer mosaic; best done edge-directed and green-first. (Demosaicking)

denoiser-as-prior (PnP / RED) — The idea that a powerful denoiser implicitly encodes a prior on natural images and can be plugged into any inverse-problem solver as the regularizer. (4.5 Generative AI and diffusion)

denoising — estimating the clean image from a noisy one, trading noise reduction against detail; must respect the noise model (affine variance, clipping bias). (Denoising)

depth from defocus — Estimating per-pixel blur radius from a few differently-focused shots and converting it to depth. (9.6 Focus, autofocus)

depth-from-defocus AF — Autofocus reading defocus direction and amount from how a known blur changes between two frames. (2.10 Photography and camera 101)

depth from focus — Sweeping focus and finding, per pixel, the focus setting at which it was sharpest, which maps to depth. (9.6 Focus, autofocus)

depth of field — the range of object distances whose blur stays within the acceptable circle of confusion. (Thin lens optics)

dichromatic reflection model — The model that glossy-surface light is a diffuse term plus a specular (illuminant-colored) term, forming an L-shaped RGB cluster. (8.9 Illumination related effects in a single image)

Difference-of-Gaussians (DoG) — A band-pass operator responding at intensity changes; thresholded it yields cartoon ink lines. (8.10 Non-photorealistic rendering)

differentiable image pipeline — An image pipeline made fast and differentiable so its parameters can be tuned by gradient descent or used as a network layer. (8.4 Dehazing)

diffraction grating — A periodic array of slits that splits light into orders, dispersing white light into a spectrum. (9.5 Special optics)

diffraction limit — the resolution bound a finite aperture imposes (the Airy disk, size ∝ λ·N); a larger aperture diffracts less. (Wave effects, diffraction, and the diffraction limit)

diffractive optical element — A finely-patterned element bending light by diffraction with dispersion opposite to glass, cancelling chromatic aberration. (9.5 Special optics)

diffuse / Lambertian — a matte surface that scatters light equally in all directions, so it looks equally bright from any view; the light it receives falls off as cos θ. (Light and physics)

diffusion inpainting (RePaint) — Hole-filling that runs a pretrained diffusion model, pinning the known region to real pixels at every step (PnP posterior sampling). (8.6 Inpainting, texture synthesis)

diffusion model — The dominant generative model that generates an image by denoising pure noise a little at a time, the universal denoiser-as-prior run in a loop. (4.5 Generative AI and diffusion)

digital intermediate — A film finished entirely as a digital color grade. (3.3 Point operations)

Dijkstra's algorithm — The shortest-path algorithm used by intelligent scissors to compute the least-cost edge-following path from a seed. (5.8 Seam optimization)

Dirichlet boundary condition — Pinning the boundary of a paste region to the destination's pixel values, the constraint that lets a Poisson clone absorb a level mismatch. (5.1 Poisson image editing)

discrete cosine transform (DCT) — A Fourier relative that re-expresses an 8×8 image block as 64 spatial-frequency cosine coefficients; JPEG's transform stage. (3.14 File formats and compression)

discrete Fourier transform (DFT) — The sampled Fourier transform of an N-sample signal, one coefficient per frequency index. (3.7 Linearity, Fourier, aliasing and deblurring)

disocclusion — A region uncovered by a moving foreground object, visible in only one frame — the chief difficulty in frame interpolation. (12.5 Frame interpolation and slow-motion synthesis)

disparity — The shift of a scene point's image position between two viewpoints, inverting to depth. (2.7 Multiple view geometry)

disparity map — A per-pixel image of disparity produced by stereo matching, equivalent to a depth map. (2.7 Multiple view geometry)

dispersion — the variation of refractive index with wavelength; why a prism splits white light and lenses show chromatic aberration. (Light and physics)

distortion (radial) — An aberration that keeps points sharp but varies magnification with radius, bowing straight lines (barrel/pincushion). (9.1 Aberrations and optical challenges)

divergence — The net outflow of a vector field; the divergence of the guidance field is the right-hand side of the Poisson equation. (5.1 Poisson image editing)

division model (Fitzgibbon) — A rational radial-distortion model capturing strong wide-angle distortion with few parameters and inverting cleanly. (9.1 Aberrations and optical challenges)

DNG (Digital Negative) — Adobe's open, documented, archival raw container built on TIFF/EP, a vendor-neutral stand-in for proprietary raw. (22.13 DNG, the Digital Negative)

dodging and burning — Local darkroom edits that brighten (dodge) or darken (burn) part of a print; the analog ancestor of masked local point operations. (3.3 Point operations)

DOL-HDR — A time-multiplexed on-chip HDR scheme reading long and short exposures and combining them, at the cost of ghosting. (10.3 HDR merging)

dolly zoom — Moving the camera while zooming to hold the subject the same size, so only the background swells (the "Vertigo" effect). (2.10 Photography and camera 101)

domain (spatial) operation — an operation where an output pixel is pulled from a different input location: a geometric warp / resampling. (Point operations; Warping and resampling)

double-Gauss (Planar) — A near-symmetric six-or-seven-element design whose symmetry about the stop cancels odd aberrations. (9.4 A short bestiary of classic designs)

DreamSim — A learned perceptual metric tuned to mid-level holistic similarity (layout, pose, content). (4.4 Deep learning)

drift (tracking) — The accumulation of per-frame sub-pixel tracking errors until the window slides off the true feature. (7.2 Feature tracking)

dual conversion gain — An on-chip HDR scheme reading each pixel at two readout gains and merging them. (10.3 HDR merging)

dual-pixel autofocus — On-sensor PDAF splitting every photosite into two half-photodiodes under one microlens. (9.6 Focus, autofocus)

dynamic programming (seam) — Filling a cumulative-energy table then back-tracing to recover the minimum-energy one-pixel-per-row seam. (5.8 Seam optimization)

dynamic range — ratio of the brightest recordable signal (sensor saturation / full-well) to the noise floor; motivates HDR. (Noise)

edge-directed interpolation — Demosaicking that, per pixel, picks the direction whose neighbors agree most and averages only along it. (3.12 Demosaicking)

edge-preserving filter — A blur that weights neighbors by value similarity as well as spatial distance so it smooths within regions but not across edges. (3.6 Neighborhood operations and convolution)

edge-spread function — The profile of a blurred step edge; differentiating it recovers the line-spread function. (3.6 Neighborhood operations and convolution)

effective focal length — The system focal length measured from the rear principal plane. (9.4 A short bestiary of classic designs)

Efros–Leung texture synthesis — A non-parametric method that grows an image one pixel at a time by copying from the patch whose neighborhood best matches the synthesized surroundings. (4.5 Generative AI and diffusion)

eigenvector — A direction a matrix only scales, never rotates; its scale factor is the eigenvalue. (22.1 Refreshers)

eight-point algorithm — The method for estimating the fundamental or essential matrix from noisy point matches. (2.7 Multiple view geometry)

electronic image stabilization — Crop-and-warp stabilization aligning frames after capture; cannot remove within-frame blur. (9.9 Optical stabilization)

electronic viewfinder (EVF) — A viewfinder showing the sensor's live feed, WYSIWYG with exposure and depth-of-field preview. (2.10 Photography and camera 101)

elliptical weighted average (EWA) — An anisotropic prefilter that approximates the true elliptical back-projected footprint of an output pixel under a non-uniform transform. (3.8 Resampling)

entrance pupil — The image of the aperture stop seen from the object side, which sets the f-number and is the no-parallax point. (9.4 A short bestiary of classic designs)

entropy — The average surprise of a source, equal to the minimum bits per symbol to encode it; low entropy means compressible. (22.1 Refreshers)

entropy coding — The lossless final JPEG stage that packs the mostly-zero quantized coefficients via run-lengths plus Huffman codes. (3.14 File formats and compression)

epipolar line — The line in one image to which a point's match is constrained, the intersection of the epipolar plane with the image plane. (2.7 Multiple view geometry)

epipolar plane — The plane spanned by a back-projected ray and the second camera's center, containing both centers and the 3-D point. (2.7 Multiple view geometry)

epitomes — A learned miniature patch model fit so every patch of an image matches somewhere inside it, used for denoising, inpainting, texture. (8.6 Inpainting, texture synthesis)

essential matrix — The 3×3 matrix encoding the epipolar constraint in calibrated coordinates, assembled from relative pose. (2.7 Multiple view geometry)

Eulerian — The "watch fixed locations" view of motion, recording each fixed pixel's time series; the basis of video magnification. (12.1 Motion blur and temporal sampling)

Eulerian video magnification — Pulling out tiny temporal variations by band-passing and amplifying each fixed pixel's time series, with no motion estimation. (12.3 Video magnification)

exemplar inpainting (Criminisi) — Object removal that copies best-matching patches in a priority order driven by confidence and an edge term, continuing structure first. (8.6 Inpainting, texture synthesis)

exiftool — The reference command-line tool for reading and writing image metadata, fluent in vendor MakerNote dialects. (22.12 EXIF and image metadata)

exit pupil — The image of the aperture stop seen from the image side; out-of-focus highlights are its projected image. (9.4 A short bestiary of classic designs)

expand (pyramid operator) — Upsample by two then blur — the approximate inverse of reduce. (3.9 Linear pyramids and wavelets)

expose to the right (ETTR) — Pushing exposure as bright as possible without clipping highlights, then pulling back in software. (2.10 Photography and camera 101)

exposure (H) — the total light a sensor patch collects: irradiance × time. (Radiometry; Exposure settings)

exposure bracketing — Shooting several frames at different exposures so their reliable bands tile the scene's radiance range for HDR. (10.3 HDR merging)

exposure compensation — A ±EV override biasing the meter's mid-gray target to correct high- and low-key metering errors. (2.10 Photography and camera 101)

exposure fusion — Blending an exposure bracket directly in a Laplacian pyramid by per-pixel quality scores, with no radiance map. (10.3 HDR merging)

exposure triangle — The interchangeable trio of shutter, aperture, and ISO, each in stops, that sets how bright a photo is. (2.10 Photography and camera 101)

exposure value (EV) — A log₂ unit of light; one stop equals +1 EV. (2.10 Photography and camera 101)

f-number (N) — focal length ÷ aperture diameter (N = f/D); sets both exposure and depth of field. (Thin lens optics)

fast Fourier transform (FFT) — The O(N log N) algorithm computing the DFT coefficients, making spectral methods practical. (3.7 Linearity, Fourier, aliasing and deblurring)

feature matching — The detect–describe–match–robust-fit pipeline that finds corresponding points between images. (10.6 Automatic panorama stitching and feature matching)

feature tracking — Following a sparse set of distinctive points across many frames, the Lagrangian complement to dense optical flow. (7.2 Feature tracking)

feed-forward stylization — Training a network once per style to map any content image to its stylized version in a single pass via a perceptual loss. (8.5 Style transfer)

FFT solver — The fastest Poisson/deblurring solver, valid for a circulant convolution, collapsing the solve to one transform, a per-frequency divide, and an inverse transform. (4.1 Linear Inverse Problems and Regression)

field curvature — The aberration where sharp points lie on a curved surface rather than the flat sensor. (9.1 Aberrations and optical challenges)

fill flash — Adding flash below ambient to open shadows in harsh or backlit light — dynamic-range management. (2.10 Photography and camera 101)

FILM (frame interpolation) — A learned interpolator using a scale-agnostic shared-weight feature pyramid and a perceptual loss for large motion. (12.5 Frame interpolation and slow-motion synthesis)

finite difference — A discrete approximation of a derivative — one pixel minus its neighbor — large exactly at edges. (22.1 Refreshers)

finite differences — Approximating image gradients with pixel-difference kernels, composed into the 5-point Laplacian that discretizes the Poisson equation. (5.1 Poisson image editing)

fisheye lens — A lens that abandons rectilinear projection for a mapping compressing a 180°+ field. (9.5 Special optics)

fixation — A pause where the high-resolution fovea is held on a region, between saccades. (2.8 Human (and animal) vision and color)

flare — Unwanted light from a bright source reflecting through the lens, producing streaks and washed-out regions. (9.8 Glare suppression)

flash / no-flash photography — Shooting a dim scene with and without flash and denoising the ambient image using the clean flash image's edges as a cross-bilateral guide. (5.2 Bilateral filtering)

fluorite (and ED glass) — Special low-dispersion materials with anomalous dispersion used in apochromats. (9.2 Aberrations correction)

focal length (f) — the distance behind a thin lens at which parallel rays converge; with sensor size it sets the field of view. (Thin lens optics)

focal stacking — Capturing frames focused at stepped distances and keeping the sharpest pixels for an all-in-focus image. (10.10 Focal stacks and depth of field extension)

focus breathing — The small magnification change as focus is racked, which forces a focus stack to be registered. (9.7 Depth of field)

focus-by-wire — Focusing in which the ring is electronically decoupled from the glass, signalling a motor. (9.6 Focus, autofocus)

focus peaking — A manual-focus aid highlighting the high-contrast, in-focus edges in live view. (2.10 Photography and camera 101)

forensics — Detecting that an image has been altered from its statistics and physics. (17.2 Ethics of computational photography)

forward mapping — Pushing each input pixel to its output destination under a transform — intuitive but broken, leaving gaps. (3.8 Resampling)

forward operator — The matrix A representing a known blur (or degradation), so the measurement is y=Ax and recovery is its inversion. (4.1 Linear Inverse Problems and Regression)

forward warping — Input-driven transport pushing each input pixel to its destination, producing holes and pile-ups. (6.1 Warping and resampling)

ForwardMatrix — A DNG calibration tag encoding part of the transform from camera raw RGB to standard XYZ. (22.13 DNG, the Digital Negative)

Fourier transform — representing a signal as a sum of sinusoids; diagonalizes convolution and underlies sampling, aliasing, and deblurring. (Linearity, Fourier, aliasing and deblurring)

fovea — The tiny central retinal pit of densely packed cones and highest acuity. (2.8 Human (and animal) vision and color)

Foveon sensor — A depth-multiplexing color sensor with stacked color layers per pixel, exploiting wavelength-dependent silicon penetration. (3.12 Demosaicking)

frame averaging — Averaging many repeated noisy captures so independent noise shrinks like 1/√n while signal survives. (5.4 NL-means (non-local means))

frame interpolation — Synthesizing in-between frames at a fractional time by estimating optical flow and warping both real frames. (12.5 Frame interpolation and slow-motion synthesis)

frame rate — The temporal sampling rate of a video; its half sets the temporal Nyquist limit. (12.1 Motion blur and temporal sampling)

Fresnel lens — A lens slicing a smooth profile into annular zones collapsed to thin sawtooth rings. (9.5 Special optics)

Fresnel reflectance — The ~4% per uncoated surface fraction of light reflected at an air–glass interface, the seed of flare and ghosting. (9.1 Aberrations and optical challenges)

full-reference metric — An image metric that scores a candidate against a known-good reference image. (3.10 Image metrics)

full-well capacity — the most charge a photosite can hold before saturating (clipping); the top of the dynamic range. (Noise)

fundamental matrix — The 3×3 matrix encoding the epipolar constraint in uncalibrated pixel coordinates. (2.7 Multiple view geometry)

gain (ISO) — amplification of the sensor signal; raising ISO brightens the image but amplifies noise (a steeper shot-noise slope). (Noise)

gamma encoding — storing V = L^(1/γ) so limited bits favor the shadows; decode with L = V^γ. (Linear vs gamma vs log encoding)

gated/partial convolution — Inpainting convolutions that renormalize or learn a per-pixel mask gate so hole pixels do not poison outputs. (8.6 Inpainting, texture synthesis)

Gaussian blur — The workhorse blur whose weights fall off as exp(−r²/2σ²), smooth and isotropic, truncated at ~3σ and renormalized. (3.6 Neighborhood operations and convolution)

Gaussian pyramid — A low-pass multiresolution tower built by repeatedly blurring and downsampling. (3.9 Linear pyramids and wavelets)

generative adversarial network (GAN) — A generative model that trains a generator against a discriminator to produce sharp, plausible images. (4.4 Deep learning)

generative model — A learned prior over natural images you can draw fresh samples from, rather than only score. (4.5 Generative AI and diffusion)

ghost — A discrete, often aperture-shaped bright blob from a double reflection between two lens surfaces. (9.8 Glare suppression)

global illumination — the multiply-bounced light in a scene (surface to surface), beyond the first bounce; makes real scenes hard to analyze. (Light and physics)

global shutter — A sensor exposing every pixel at once, avoiding rolling-shutter skew at higher cost. (12.4 Video stabilization and rolling-shutter correction)

golden hour — The hour after sunrise or before sunset when the sun is low, warm, and directional. (17.1 Human factors and the art of photography)

GOP (group of pictures) — The repeating arrangement of I, P, and B frames, trading random access against compression. (12.2 Video compression and motion compensation)

GrabCut — Graph-cut segmentation driven by a single loose rectangle, alternating a graph cut with re-estimation of color models. (5.8 Seam optimization)

gradient descent (GD) — The simplest matrix-free solver, stepping downhill along the negative gradient; correct but slow because it zig-zags. (4.1 Linear Inverse Problems and Regression)

gradient domain — Editing an image's gradient field rather than its pixels, then reconstructing the image whose gradients best match. (5.1 Poisson image editing)

gradient reversal — An artifact where a smooth ramp returns with its slope locally reversed near a strong edge, from the bilateral's range quantization. (5.7 Guided image filtering)

Gram matrix — The feature co-occurrence matrix summed over spatial positions, discarding location to capture style as a texture statistic. (8.5 Style transfer)

graph cut — Posing a binary object/background labeling as an energy whose global minimum is found exactly by min-cut/max-flow on a graph. (5.8 Seam optimization)

graph-cut segmentation — A hard binary foreground/background labelling found exactly by min-cut/max-flow on a grid graph with a data plus edge-aware smoothness term. (8.8 Compositing, segmentation and matting)

GraphCut textures — Graph cut applied to texture synthesis and compositing, routing the seam where two sources agree best. (5.8 Seam optimization)

gray card — An 18%-reflectance reference standing in for average scene reflectance and perceptual middle gray. (2.10 Photography and camera 101)

green-based demosaicking — Reconstructing green first, then interpolating the smooth color differences and adding green back, killing fringing. (3.12 Demosaicking)

grey-world assumption — An auto-white-balance estimator that assumes the scene averages to grey and scales channels so their means agree. (3.13 Auto-exposure and auto white balance)

GRIN (gradient-index) — An element whose refractive index varies continuously so flat-faced slabs focus by bending rays internally. (9.5 Special optics)

ground truth — The known correct answer a benchmark holds out to score a method against. (22.14 Datasets)

guidance field — The target gradient field a gradient-domain edit prescribes; changing only this field yields the whole family of Poisson edits. (5.1 Poisson image editing)

guided filter — An O(N) edge-preserving filter whose output is, in each window, the best least-squares line in a guidance image, with no gradient reversal. (5.7 Guided image filtering)

halftoning — Reproducing continuous tone with discrete marks (ordered dithering, error diffusion), exploiting the eye's spatial averaging. (8.10 Non-photorealistic rendering)

hallucination — Plausible detail invented by a learned or generative prior that was never measured. (4.3 Machine learning)

harmonic function — The smoothest fill of a region consistent with its boundary, the solution of Laplace's equation. (5.1 Poisson image editing)

harmonization — Adjusting a pasted cutout's color, illumination, contrast, and grain to match the new background. (8.8 Compositing, segmentation and matting)

Harris corner detector — A keypoint detector using the structure tensor's eigenvalues, large only where both gradient directions vary. (10.6 Automatic panorama stitching and feature matching)

Hays–Efros scene completion — Data-driven hole filling that retrieves a scene-matching photo, cuts a region along a graph-cut seam, and Poisson-blends it in. (8.6 Inpainting, texture synthesis)

HDR (high dynamic range) — capturing or representing a scene's full contrast, typically by merging multiple exposures, then tone-mapping for display. (HDR merging)

HDR merging — Estimating one floating-point radiance map per pixel as a reliability-weighted average of differently-exposed observations. (10.3 HDR merging)

HDR-VDP — A visible-differences predictor calibrated in physical luminance to predict visible differences across HDR imagery. (3.10 Image metrics)

healing brush — A retouching tool that copies a source region's texture/gradients while re-matching the destination's color and illumination. (8.6 Inpainting, texture synthesis)

heliography — The earliest photographic process, exposing a bitumen plate for days (Niépce, 1826/27). (1.1 From Digital to Computational Photography)

Hessian — The matrix of second partial derivatives, encoding curvature and distinguishing minima, maxima, and saddles. (22.1 Refreshers)

high-speed sync — A flash mode pulsing rapidly as the shutter slit travels, allowing flash above the X-sync speed. (2.10 Photography and camera 101)

hinge rule — The companion to Scheimpflug: the line about which the tilted plane of sharp focus pivots as you refocus. (9.5 Special optics)

histogram — the distribution of pixel values; the basis of contrast stretching and histogram equalization. (Histograms)

histogram equalization — Using the image's own CDF as a tone curve (L_out = CDF(L_in)) to flatten the histogram and maximize global contrast. (3.4 Histograms)

histogram matching — Reshaping an image's histogram to a target distribution by composing CDFs, with equalization the uniform-target special case. (3.4 Histograms)

homogeneous coordinates — Representing a 2D point by three numbers up to scale, so perspective is one matrix multiply. (10.5 Manual panorama stitching)

homography (projective transform) — The 8-DOF 3×3 projective warp that maps between any two images of a plane, preserving straight lines. (6.1 Warping and resampling)

Horn–Schunck — Optical flow with a global smoothness prior, minimizing brightness-constancy plus a penalty on flow variation. (7.6 Optical flow)

Hunt effect — The perceptual increase in colorfulness with luminance. (2.8 Human (and animal) vision and color)

hyperfocal distance — the focus distance that renders everything from half of it to infinity acceptably sharp. (Thin lens optics)

hyperlapse — Time-compressed first-person footage made watchable by reconstructing and smoothing the 3-D camera path. (12.6 Video editing)

hyperstereo — Widening the stereo baseline to exaggerate perceived depth (hypostereo shrinks it). (2.7 Multiple view geometry)

I-frame — A video frame coded alone with no temporal prediction, a self-contained random-access point. (12.2 Video compression and motion compensation)

ICC profile — An embedded description of an image's color space. (22.12 EXIF and image metadata)

ill-conditioning — The property of a blur operator that crushes high frequencies, so its inverse amplifies noise. (4.1 Linear Inverse Problems and Regression)

ill-posed problem — An inverse problem where many scenes explain the same measurement or noise is amplified, curable only with a prior. (22.1 Refreshers)

illuminant — a light source's spectrum S(λ); the color reaching the eye is illuminant × reflectance. (Light and physics)

Image Analogies — A pre-deep image translation method that learns a filter from a single example pair and applies it by patch matching. (4.4 Deep learning)

image averaging (1/√N law) — Averaging N independent zero-mean frames leaves the signal while noise std falls as 1/√N. (10.1 Denoising by averaging)

image gradient — The vector ∇I = (Ix, Iy) of partial derivatives; its magnitude is edge strength and its angle edge orientation. (3.6 Neighborhood operations and convolution)

image morphing — A smooth transition between two images by warping both to a common intermediate shape then cross-dissolving. (6.3 Morphing)

image quilting — Texture synthesis that lays down overlapping source blocks and stitches each along a minimum-error boundary. (8.6 Inpainting, texture synthesis)

image registration — Finding the transform that makes corresponding scene points land on the same pixel across frames. (10.2 Image alignment)

image retargeting — Content-aware resizing that changes a photo's aspect ratio while preserving important content. (8.7 Patch match)

image signal processor (ISP) — The fixed pipeline of stages a camera runs to turn raw sensor data into a viewable picture. (3.15 Recap ISP and non-destructive editing)

image stabilization — A system sensing camera shake with a gyroscope and countering it for slower hand-held shutter speeds. (2.10 Photography and camera 101)

image-to-image network — A deep net, typically a U-Net, that maps a degraded or partial image to a restored one. (4.4 Deep learning)

impulse (delta) — The discrete delta (one at the origin, zero elsewhere); convolving it with a kernel returns the kernel. (3.6 Neighborhood operations and convolution)

in-body image stabilization (IBIS) — Sensor-shift stabilization moving the sensor on a magnetic stage, working with any lens. (9.9 Optical stabilization)

inertial measurement unit (IMU) — A camera's accelerometer-plus-gyroscope package sensing its own motion, driving stabilization and alignment. (2.10 Photography and camera 101)

influence function — The derivative of a robust error norm, which rises then returns to zero so a neighbor across an edge is treated as an outlier. (5.2 Bilateral filtering)

inner product — The operation measuring how much two vectors align — "how much of this is in that.". (22.1 Refreshers)

inpainting — Filling a masked region with no measured data with a plausible, seamless completion drawn from a prior. (8.6 Inpainting, texture synthesis)

InstructPix2Pix — A diffusion model that edits an image in place conditioned on an input image and a natural-language instruction. (4.5 Generative AI and diffusion)

intelligent scissors (live-wire) — An interactive selection tool where a least-cost path on a gradient cost graph snaps to the nearest strong edge. (5.8 Seam optimization)

interleaved (HWC) layout — A memory layout where a pixel's channels sit side by side (RGB RGB …); NumPy's default. (3.1 Image representation)

intraocular lens — An artificial lens implanted to replace the clouded crystalline lens in cataract surgery. (9.10 The eye as an optical instrument)

intrinsic-image decomposition — Splitting one photo into a reflectance layer and a shading layer, I=R·S; under-determined and requiring a prior. (8.9 Illumination related effects in a single image)

intrinsic images — The decomposition of an image into time-invariant reflectance and time-varying shading, I=R·S. (10.12 Intrinsic images with time lapse)

inverse filtering — Deblurring by dividing each frequency of the blurred image by the kernel's spectrum; exact on paper but unstable. (3.7 Linearity, Fourier, aliasing and deblurring)

inverse mapping — Looping over output pixels and applying the inverse transform to reach back into the input — the gap-free way to resample. (3.8 Resampling)

inverse-variance weighting — Combining observations weighted by the inverse of each one's noise variance, the minimum-variance merge. (10.3 HDR merging)

inverse warping — Output-driven transport that, for each output pixel, applies the inverse map and resamples, with no holes or double-writes. (6.1 Warping and resampling)

IPTC — The older press-and-captioning metadata standard carrying caption, byline, credit, keywords, and location. (22.12 EXIF and image metadata)

irradiance (E) — power per unit area arriving on a surface (the sensor); the integral of incoming radiance over the collecting solid angle. (Radiometry)

ISO — Electronic gain applied to the sensor signal after capture, brightening a dim exposure while amplifying noise. (2.10 Photography and camera 101)

isophote — A level-line of constant intensity perpendicular to the gradient; PDE inpainting diffuses along isophotes to continue edges. (8.6 Inpainting, texture synthesis)

Jacobian — The matrix of all first partial derivatives of a vector-to-vector map, its best local linear approximation. (22.1 Refreshers)

JPEG — The lossy, perception-based image format: opponent color, chroma subsampling, 8×8 DCT, CSF-shaped quantization, and entropy coding. (3.14 File formats and compression)

keystoning — Converging verticals when a camera is tilted off a plane, undone by a single rectifying homography. (6.7 Perspective distortion and its correction)

KLT (Kanade–Lucas–Tomasi) tracker — Sparse feature tracking applying the Lucas–Kanade patch solve to one window and threading it greedily through the video. (7.2 Feature tracking)

Lagrangian — The "follow the particles" view of motion, the frame of optical flow and feature tracking. (12.1 Motion blur and temporal sampling)

Lanczos — A sharp resampling kernel formed by windowing the ideal sinc, close to ideal but prone to edge ringing. (3.8 Resampling)

Laplace's equation — The source-free Poisson equation ∇²f=0, whose pinned-boundary solution is the smoothest membrane interpolating the boundary. (5.1 Poisson image editing)

Laplacian pyramid — A band-pass multiresolution decomposition, each level the detail one scale adds; an exact, invertible encoder/decoder. (3.9 Linear pyramids and wavelets)

latent diffusion — Running diffusion in a compact autoencoder latent space instead of pixels, which made high-res text-to-image practical (Stable Diffusion). (4.5 Generative AI and diffusion)

lateral inhibition — Retinal wiring where each cell subtracts a weighted neighbor average, enhancing edges and producing Mach bands. (2.8 Human (and animal) vision and color)

leaf shutter — A shutter built into the lens that opens fully at any speed and so syncs flash at all shutter speeds. (2.10 Photography and camera 101)

learned operator — Replacing a hand-designed operator with a function fit to example pairs, learning the prior from data (L8). (4.3 Machine learning)

learned perceptual image patch similarity (LPIPS) — A learned metric measuring image distance in a deep network's feature space. (3.10 Image metrics)

learning rate — The step size in gradient descent; too large overshoots, too small crawls. (22.1 Refreshers)

least squares — Recasting inversion as a regression finding the image whose blur best reproduces the photo, minimizing the squared residual. (4.1 Linear Inverse Problems and Regression)

lens hood — A tube or petal that blocks off-axis light from a source outside the frame before it can scatter. (9.8 Glare suppression)

lens-profile correction — Playing a measured per-lens profile in reverse in the ISP to undo distortion, vignetting, and lateral CA. (9.2 Aberrations correction)

levels — The tonal tool that maps a chosen black point to 0 and white point to 1, with a mid-grey gamma to bend the midtones. (3.3 Point operations)

light field — The 4-D bundle of rays (preserving direction) that light-field cameras capture, enabling refocus and depth after capture. (1.1 From Digital to Computational Photography)

light painting — A long-exposure effect in which a light is moved through a dark scene, accumulating a luminous trail. (17.1 Human factors and the art of photography)

linear DNG — A DNG storing already-demosaicked data, a full RGB triple at every pixel. (22.13 DNG, the Digital Negative)

linear interpolation — Reconstructing a value between two samples as their distance-weighted blend, the straight line joining consecutive samples. (3.8 Resampling)

linear inverse problem — Recovering an unknown image from a measurement produced by a known linear operator. (4.1 Linear Inverse Problems and Regression)

linear light — pixel values proportional to scene radiance; the space in which light adds and scales, so white balance, blur, and HDR are computed there. (Linear vs gamma vs log encoding)

linear map — A matrix operation y=Ax that respects sums and scaling; blur, color conversion, and resampling are all linear maps. (22.1 Refreshers)

linear shift-invariant (LSI) filter — An operation that is both linear and shift-invariant; convolution is precisely the LSI operation. (3.6 Neighborhood operations and convolution)

local Laplacian filters — Halo-free edge-aware detail/tone manipulation that reshapes Laplacian-pyramid coefficients with a per-output-pixel curve. (5.5 Local Laplacian filters)

local tone mapping — A tone map whose compression curve varies spatially, compressing a large-scale layer while keeping local contrast. (3.5 Tone mapping)

locally adaptive regression kernel (LARK) — The steering-kernel form of the affinity, fitting a local model with an anisotropic kernel stretched along edges. (5.3 Locally adaptive regression kernel (LARK))

log encoding — A value encoding where equal light ratios become equal steps; the workhorse space for tone mapping. (3.5 Tone mapping)

look-up table (LUT) — A precomputed table of a point operation's output per input value, turning any tone curve into one index per pixel. (3.3 Point operations)

loss function — The per-example penalty summed over a dataset to define the training objective. (4.3 Machine learning)

lossless compression — Compression that returns bit-for-bit identical pixels by squeezing out statistical redundancy (PNG). (3.14 File formats and compression)

lossy compression — Compression that discards perceptually irrelevant detail for much smaller files (JPEG, HEIC, WebP, AVIF). (3.14 File formats and compression)

low-pass filter — A filter that passes low frequencies and attenuates high ones; blur is a low-pass filter. (3.7 Linearity, Fourier, aliasing and deblurring)

Lucas–Kanade — Optical flow with a local constant-flow assumption over a window, an over-determined least-squares solve whose matrix is the structure tensor. (7.6 Optical flow)

lucky imaging — Selecting only the sharpest few percent of frames to beat atmospheric blur in planetary astrophotography. (10.1 Denoising by averaging)

luma (Y′) — a weighted sum of gamma-encoded RGB (Rec. 601: 0.299 R + 0.587 G + 0.114 B); a coding convenience (YUV/YCbCr), not the same as luminance. (Linear vs gamma vs log encoding)

luminance (Y) — linear, perceptual brightness; a weighted sum of linear RGB (CIE Y). (Color technology)

luminance / chrominance — the brightness vs color components of an image; detail lives in luminance, color is smooth, which licenses chroma subsampling. (Demosaicking; Color technology)

Mach bands — Illusory bright and dark bands at a luminance ramp's edges, produced by center-surround overshoot. (2.8 Human (and animal) vision and color)

macro photography — Imaging at magnification ≥ 1, where depth of field becomes razor-thin and flat-field correction matters. (9.5 Special optics)

macroblock — A block into which a frame is partitioned for independent motion-compensated prediction. (12.2 Video compression and motion compensation)

MakerNotes — The manufacturer's private EXIF annex holding vendor-specific data like shutter count. (22.12 EXIF and image metadata)

Malvar–He–Cutler demosaicking — A fast, non-iterative high-quality linear demosaicker using a single fixed 5×5 filter per pixel type. (3.12 Demosaicking)

MAP estimation — Taking the most probable scene under Bayes' rule; in log form it becomes the data-term-plus-prior loss. (22.1 Refreshers)

Markov random field (MRF) — The prior behind the graph-cut energy — a per-pixel data term plus a pairwise smoothness term. (5.8 Seam optimization)

matrix (evaluative) metering — Smart metering adding brightness, color, contrast, and distance cues to guess the photographer's intent. (2.10 Photography and camera 101)

matting — The under-determined inverse of recovering α, F, and B from a single observed pixel, made tractable by a trimap and a prior. (8.8 Compositing, segmentation and matting)

matting Laplacian — The affinity-weighted sparse matrix whose energy minimization extracts a soft alpha matte from scribbles. (5.6 Edge-preserving optimization — colorization)

maze / labyrinth artifacts — Residual false color in demosaicking on fine high-frequency color texture. (3.12 Demosaicking)

mean squared error (MSE) — The mean of squared pixel differences between two images; cheap but blind to where errors are and what structure they destroy. (3.10 Image metrics)

median filter — A neighborhood filter replacing each pixel by the median of its window, excellent against salt-and-pepper noise. (3.11 Denoising)

median-threshold bitmap (MTB) — A brightness-robust alignment comparing frames thresholded at their median, used to register an HDR bracket. (10.3 HDR merging)

membrane — The smooth correction surface a Poisson clone adds to spread the boundary discrepancy across the interior. (5.1 Poisson image editing)

merit function — The single scalar a lens optimizer minimizes — weighted squared ray errors plus constraint penalties. (9.3 Lens optimization)

mesh / triangulation warp — A piecewise-affine free-form warp that warps each triangle by the affine fixed by its vertices; fast but only C0. (6.1 Warping and resampling)

mesopic — The dim regime where rods and cones work together, with desaturation and the Purkinje shift. (2.8 Human (and animal) vision and color)

metalens — A flat film tiled with sub-wavelength nanostructures imposing a chosen phase delay, reproducing a lens with no thickness. (9.5 Special optics)

metamerism — when two different spectra produce the same three (RGB) responses — a match under one light that breaks under another. (Light and physics; Color)

metering — The camera's measurement of scene light (spot, center-weighted, evaluative) rendered to an 18%-gray assumption. (2.10 Photography and camera 101)

Michelson contrast — A normalized, level-independent contrast measure, (Imax−Imin)/(Imax+Imin). (2.8 Human (and animal) vision and color)

microscope objective — A high-NA, apochromatic, flat-field lens; infinity-corrected versions send parallel light to a tube lens. (9.5 Special optics)

min-cut / max-flow — The algorithm that finds the globally optimal binary boundary by pushing maximum flow from a source to a sink through a pixel graph. (5.8 Seam optimization)

mipmap — A precomputed pyramid of progressively blurred, downsampled copies for cheap isotropic prefiltering during resampling. (3.8 Resampling)

mirrorless — A camera design with the SLR mirror box removed, the sensor feeding an EVF directly. (2.10 Photography and camera 101)

Mitchell–Netravali — The balanced default cubic resampling kernel, chosen to minimize combined blur, ringing, and anisotropy. (3.8 Resampling)

mixing gradients — A Poisson-cloning variant that keeps, at each pixel, whichever gradient is stronger, letting background texture show through. (5.1 Poisson image editing)

mode collapse — A GAN failure where the generator covers only part of the data distribution; diffusion avoids it by construction. (4.5 Generative AI and diffusion)

modulation transfer function (MTF) — The lens's spatial-frequency response, whose near-zeros bound how far deconvolution can recover detail. (9.2 Aberrations correction)

moiré — False low-frequency bands that appear when fine detail is sampled too coarsely without a prefilter. (3.7 Linearity, Fourier, aliasing and deblurring)

monocular depth — Estimating a dense relative depth map from a single RGB image, driven by a learned scene prior (MiDaS, Depth Anything). (4.4 Deep learning)

monotone spline — An interpolating spline held non-decreasing so a tone curve can never invert tones. (3.3 Point operations)

mosaic DNG — A DNG storing the original color-filter-array samples straight off the sensor. (22.13 DNG, the Digital Negative)

motion blur — The streak recorded when a feature moves across the frame during exposure, a convolution in time. (12.1 Motion blur and temporal sampling)

motion compensation — Predicting each block from a displaced reference patch and coding only the residual — "optical flow on a budget.". (12.2 Video compression and motion compensation)

Motion-JPEG — Coding every video frame as an independent JPEG, ignoring temporal redundancy. (12.2 Video compression and motion compensation)

motion vector — The per-block displacement to its best match in the reference frame, the codec's coarse correspondence. (12.2 Video compression and motion compensation)

moving least squares (MLS) deformation — A shape-preserving free-form warp making each local neighborhood transform as close to rigid/similarity as possible under handle dragging. (6.1 Warping and resampling)

MS-SSIM — A multi-scale variant of SSIM that evaluates structural similarity across several resolutions. (3.10 Image metrics)

multiband blending — Blending each frequency band over a transition matched to its scale, so every band gets its own width. (10.7 Blending)

multigrid — A coarse-to-fine solver that fixes low-frequency error on coarse grids and prolongs corrections back up in a V-cycle. (4.1 Linear Inverse Problems and Regression)

Naka–Rushton curve — A saturating S-shaped response to intensity whose half-saturation point slides with adaptation. (2.8 Human (and animal) vision and color)

natural vignetting — The cos⁴θ corner darkening from pixel-irradiance geometry alone, distinct from optical vignetting. (2.4 Image measurements as integrals)

nearest-neighbour field (NNF) — A per-output-pixel offset field to the most similar source patch, computed by PatchMatch and Shift-Map. (8.7 Patch match)

negative — The developed film image, dark where the scene was bright, from which positive prints are made. (2.10 Photography and camera 101)

neighborhood operation — an operation where an output pixel is a function of a window of input pixels (e.g. convolution, blur). (Neighborhood operations and convolution)

Neumann boundary condition — A zero-normal-derivative border condition used when integrating a guidance field over a whole image with no destination to pin. (5.1 Poisson image editing)

neural style transfer (Gatys) — Stylization by optimizing an image to match a content image's deep features and a style image's Gram matrices simultaneously. (8.5 Style transfer)

neutral-density filter — A uniform gray filter cutting light by a chosen number of stops with no color shift. (2.10 Photography and camera 101)

no-reference metric — A metric that scores a single image's quality with no reference, using a built-in model of undegraded photographs. (3.10 Image metrics)

nodal points — Cardinal points where a ray crosses the axis undeviated in angle. (9.4 A short bestiary of classic designs)

non-blind deconvolution — Deblurring when the blur kernel is known, versus blind deblurring where the kernel is unknown. (3.7 Linearity, Fourier, aliasing and deblurring)

non-destructive editing — The editor model that never alters the raw and stores edits as a recipe of pipeline parameters, recomputing the preview on demand. (3.15 Recap ISP and non-destructive editing)

non-linear editing — Assembling random-access clips on a timeline, the edit being a non-destructive list of references with in/out points. (12.6 Video editing)

non-local means (NLM) — A denoiser that averages a pixel with the centers of similar patches found anywhere in the image. (3.11 Denoising)

non-photorealistic rendering (NPR) — Deliberately discarding information to turn a photo into a depiction (painting, sketch, cartoon). (8.10 Non-photorealistic rendering)

norm — A vector's length. (22.1 Refreshers)

normal equations — The linear system AᵀA x = Aᵀy obtained by zeroing the least-squares gradient. (4.1 Linear Inverse Problems and Regression)

normal flow — The only locally recoverable flow component, along the gradient; the along-edge component is undetermined. (7.6 Optical flow)

normalized cross-correlation (NCC) — A matching score invariant to per-frame gain and offset, maximized at the true alignment. (10.2 Image alignment)

normalized cuts — A spectral, seedless segmentation dividing cut weight by each side's total connection, relaxing to an eigenproblem. (5.8 Seam optimization)

null space — The set of scene components a forward operator throws away — exactly what the measurement cannot see. (22.1 Refreshers)

Nyquist rate — twice the highest frequency present; sampling below it causes aliasing. (Linearity, Fourier, aliasing and deblurring)

on-chip HDR — Producing a high-dynamic-range signal inside the sensor by multiplexing range in time, space, or gain. (10.3 HDR merging)

on-sensor PDAF — Phase-detection autofocus moved onto the imaging sensor via masked or dual pixels. (2.10 Photography and camera 101)

opcode (DNG) — A deferred correction instruction a DNG carries for the raw processor to apply at decode time. (22.13 DNG, the Digital Negative)

opponent color — the brightness / red–green / blue–yellow encoding used by vision (and by JPEG); the basis for coarse chroma. (Human (and animal) vision and color)

opponent recoding — The retinal re-mixing of L/M/S cone signals into light–dark, red–green, and blue–yellow channels. (2.8 Human (and animal) vision and color)

opsin — A light-sensitive photopigment protein; the set an animal carries defines its color vision. (2.8 Human (and animal) vision and color)

optical flow — The dense per-pixel 2-D motion field between two frames, the automatic dense answer to correspondence. (7.6 Optical flow)

optical-flow constraint equation — The linearized brightness-constancy equation Ix·u + Iy·v + It = 0, one scalar equation per pixel. (7.6 Optical flow)

optical image stabilization (OIS) — Lens-shift stabilization where a floating group, driven from a gyro, steers the image to stay still. (9.9 Optical stabilization)

optical low-pass filter (OLPF) — A birefringent layer that slightly blurs the image before the sensor samples it, trading sharpness for fewer aliasing artifacts. (3.12 Demosaicking)

optical viewfinder — A viewfinder showing the real world directly through a mirror-and-prism — zero lag but not the captured exposure. (2.10 Photography and camera 101)

optical vignetting — Corner darkening from the barrel and rims clipping the off-axis pupil, worst wide open. (9.1 Aberrations and optical challenges)

optimal transport — The exact N-dimensional method matching a full joint color distribution. (3.4 Histograms)

optimization — Finding the image or parameters that best explain measurements by minimizing data-fit plus prior. (22.1 Refreshers)

orientation tag — The EXIF flag recording how the camera was held, instructing a viewer to rotate the image upright. (22.12 EXIF and image metadata)

orthonormal basis — A basis of mutually perpendicular unit vectors in which coordinates are just inner products and energy is preserved. (22.1 Refreshers)

overfitting — A model fitting the noise rather than the signal; what the prior is for. (4.1 Linear Inverse Problems and Regression)

P-frame — A video frame predicted forward from a previous decoded frame, far smaller than an I-frame. (12.2 Video compression and motion compensation)

panorama — One wide image assembled by registering and stitching several overlapping frames. (10.5 Manual panorama stitching)

PatchMatch — A randomized algorithm computing a near-exact nearest-neighbour field in milliseconds via random init, propagation, and random search. (8.7 Patch match)

PDE / diffusion inpainting (Bertalmío) — Weak-prior hole filling that propagates smoothness along isophotes; great for thin scratches, blurry on large holes. (8.6 Inpainting, texture synthesis)

peak signal-to-noise ratio (PSNR) — A logarithmic re-expression of MSE in decibels; convenient but inheriting MSE's perceptual blind spots. (3.10 Image metrics)

perception–distortion tradeoff — The formal limit that fidelity and perceptual realism cannot both be maximized past a frontier. (8.2 Super-resolution and image priors)

periscope (folded-optics) design — A smartphone telephoto bending the optical axis ~90° with a prism so a long lens fits a thin body. (9.5 Special optics)

phase (Fourier) — The component of a spectrum saying where each wave sits; it, not the magnitude, carries a scene's structure. (3.7 Linearity, Fourier, aliasing and deblurring)

phase-based magnification — Video magnification amplifying the band-passed local phase of a complex steerable pyramid. (12.3 Video magnification)

phase correlation — Recovering a whole translation in one shot via the inverse FFT of the normalized cross-power spectrum. (10.2 Image alignment)

phase-detection autofocus — Autofocus comparing two sub-aperture views to read defocus direction and amount in one shot. (2.10 Photography and camera 101)

photoelectric effect — A photon freeing an electron, the conversion that lets a photosite count photons as electrons. (1.1 From Digital to Computational Photography)

photon counting noise — Poisson noise from light arriving as discrete photons, with variance equal to mean. (22.1 Refreshers)

photopic — The daylight regime in which cones dominate, giving full color and sharp foveal vision. (2.8 Human (and animal) vision and color)

photosite — a single light-collecting well on the sensor; accumulates photoelectrons during the exposure. (Sensors — photosites, CCD vs CMOS)

pinhole camera — A lensless camera in which a small hole forms a real but dim inverted image. (2.8 Human (and animal) vision and color)

PIPs (Particle Video Revisited) — A learned tracker that tracks one point over a window, jointly solving location and appearance and predicting visibility. (7.2 Feature tracking)

Pix2Pix — A conditional GAN that learns to map one image domain to another from paired examples. (4.4 Deep learning)

pixel-shift — A multi-shot mode micro-shifting the sensor between frames and merging for higher resolution and full color. (22.15 A camera-feature wish list)

planar (CHW) layout — A memory layout storing all of one channel contiguously then the next; the book's C++ default and PyTorch's expectation. (3.1 Image representation)

plenoptic function — the radiance of every ray filling space, parameterized by 3D position, 2D direction, wavelength, and time (7-D); any imaging operation samples a portion of it. (The plenoptic function)

Plug-and-Play (PnP) priors — A splitting solver in which the prior step is any denoiser dropped into the proximal slot, decoupling physics from prior. (8.2 Super-resolution and image priors)

PNG — The book's lossless, easy-to-read save/load format, supporting an alpha channel. (3.14 File formats and compression)

point (range) operation — an operation where an output pixel depends only on the input at the same location (a value remap): brightness, gamma, levels. (Point operations)

point spread function (PSF) — How an imaging system spreads a single point of light — equivalently the kernel and impulse response of the blur. (3.6 Neighborhood operations and convolution)

Poisson (gradient-domain) blending — Seamless compositing that pastes a region's gradients and re-derives boundary color from the background. (8.8 Compositing, segmentation and matting)

Poisson distribution — The distribution governing photon counts, with variance equal to mean. (22.1 Refreshers)

Poisson equation — The linear system ∇²f = div v that reconstructs the image whose gradients best match a guidance field. (5.1 Poisson image editing)

polarization — the orientation of a light wave's oscillation; the basis of polarizing filters that cut glare and darken skies. (Light and physics)

posterior — What we believe about a scene after seeing a measurement, the output of Bayes' rule. (22.1 Refreshers)

posterior sampling — Solving an inverse problem with a generative prior by drawing from p(x|y) ∝ p(y|x)p(x), alternating a diffusion step with a data-fit step. (4.5 Generative AI and diffusion)

preconditioning — Applying a cheap approximate inverse each CG iteration to re-round the elongated bowl, slashing the iteration count. (4.1 Linear Inverse Problems and Regression)

prefilter (anti-aliasing) — The low-pass blur applied before downsampling to remove frequencies above the output grid's Nyquist. (3.7 Linearity, Fourier, aliasing and deblurring)

presbyopia — The age-related failure of accommodation as the crystalline lens stiffens. (9.10 The eye as an optical instrument)

principle of univariance — A single cone confounds wavelength with intensity, reporting one number, so only comparing cones separates color from brightness. (2.8 Human (and animal) vision and color)

prior (regularizer) — A penalty R(x) added to the data-fit term that prefers plausible images, stabilizing the otherwise unstable inverse. (4.1 Linear Inverse Problems and Regression)

priority modes — Exposure modes (aperture/shutter priority) where the photographer sets one control and the camera picks its partner. (2.10 Photography and camera 101)

progressive (varifocal) lens — A spectacle lens whose power increases continuously top to bottom with no line. (9.10 The eye as an optical instrument)

projection (vector) — A vector's shadow onto a direction; in an orthonormal basis every coordinate is such a projection. (22.1 Refreshers)

Prokudin-Gorskii — The temporal-multiplexing color method shooting three filtered plates in sequence; exact for static scenes only. (3.12 Demosaicking)

ProRAW — Apple's computational-photography raw format, which is DNG under the hood. (22.13 DNG, the Digital Negative)

provenance — Cryptographically signing an image's origin and edit history so a viewer can check where it came from. (17.2 Ethics of computational photography)

Purkinje shift — The mesopic shift, as rods take over, making blues look relatively brighter and reds darker. (2.8 Human (and animal) vision and color)

pushbroom (line-scan) imaging — A hyperspectral architecture capturing one spatial line in all bands at once, sweeping to fill the frame. (10.11 Hyperspectral imaging color wheels)

pyramid (Gaussian / Laplacian) — a multi-scale image representation, local in both space and frequency; the workhorse for blending and multi-scale editing. (Linear pyramids and wavelets)

quad-Bayer — A high-megapixel CFA layout where a 2×2 block of same-color photosites shares one filter tile, requiring a remosaic step. (3.12 Demosaicking)

quantization — rounding continuous values to a finite set of codes; too few (low bit depth) produce visible banding. (File formats and compression)

radiance (L) — power per unit area per unit solid angle carried along a ray; conserved along a ray in vacuum. (Radiometry)

radiance map — A single floating-point estimate of scene radiance per pixel, the output of an HDR merge. (10.3 HDR merging)

RAFT — A learned optical-flow network built as a neuralized classical pipeline: learned features, an all-pairs correlation volume, and a recurrent update. (7.6 Optical flow)

random-dot stereogram — Julesz's stimulus yielding depth from disparity alone — "depth without objects.". (2.7 Multiple view geometry)

random variable — A quantity scattered around its true value, described by a distribution with mean and variance. (22.1 Refreshers)

RANSAC — A robust estimator fitting a model from random minimal samples, scoring by inlier consensus, and re-fitting on inliers. (10.6 Automatic panorama stitching and feature matching)

ratio test — Keeping a descriptor match only when the nearest/second-nearest distance ratio is small enough. (10.6 Automatic panorama stitching and feature matching)

raw file — The sensor's early data — a linear, one-color-per-pixel Bayer mosaic before demosaicking — preserving the most information. (3.14 File formats and compression)

ray-tracing (lens design) — Following many rays surface by surface through full Snell's law to evaluate a candidate lens. (9.3 Lens optimization)

re-detection — Periodically re-running feature detection to spawn fresh features in the gaps left by drift and occlusion. (7.2 Feature tracking)

read noise — a signal-independent noise floor from the readout electronics; the affine noise model's intercept. (Noise)

reconstruction vs hallucination — The axis distinguishing genuinely-measured recovered detail from invented detail synthesized by a learned prior. (8.2 Super-resolution and image priors)

rectification — A projective reprojection making a stereo pair appear parallel so matching searches only horizontal lines. (2.7 Multiple view geometry)

RED (Regularization by Denoising) — An explicit denoiser-as-prior whose regularizer gradient is the denoising residual x−D(x). (8.2 Super-resolution and image priors)

reduce (pyramid operator) — Blur with a small Gaussian then downsample by two, the building step of the Gaussian pyramid. (3.9 Linear pyramids and wavelets)

reflection removal — Un-summing a through-glass image into transmitted and reflected layers using defocus, ghosting, sparsity, or polarization cues. (8.9 Illumination related effects in a single image)

refraction — the bending of light crossing between media of different refractive index (Snell's law); how a lens focuses. (Light and physics; Thin lens optics)

refractive surgery (LASIK, PRK) — Reshaping the cornea with an excimer laser to dial in a new prescription. (9.10 The eye as an optical instrument)

regression — Fitting an unknown image to data by minimizing a squared discrepancy. (4.1 Linear Inverse Problems and Regression)

regularization — Adding a prior term weighted by λ to a loss, taming ill-posed inverse problems and noise amplification. (22.1 Refreshers)

Reinhard operator — The global tone curve L_out = L/(1+L), passing dark tones nearly unchanged and rolling highlights off without clipping. (3.5 Tone mapping)

relative pose — The rotation and translation of the second camera with respect to the first, recoverable up to scale. (2.7 Multiple view geometry)

ReLU — A common neuron nonlinearity (rectified linear unit) applied after the weighted sum. (22.1 Refreshers)

remosaic — The CFA-conversion step rearranging quad-Bayer samples into a standard RGGB Bayer pattern. (3.12 Demosaicking)

remote photoplethysmography — Recovering vital signs such as heart rate from faint pulse-driven color changes in ordinary video. (12.3 Video magnification)

resampling — Reconstructing a continuous signal from samples and re-sampling it on a new grid (resize, rotate, warp). (3.8 Resampling)

residual (codec) — The prediction error a codec actually codes, small wherever motion-compensated prediction was good. (12.2 Video compression and motion compensation)

retina — The thin neural sheet at the back of the eye holding rods and cones and pre-processing the image. (2.8 Human (and animal) vision and color)

Retinex — The intrinsic-image prior that reflectance changes are sharp and shading changes smooth, sorting log-gradients by a threshold. (8.9 Illumination related effects in a single image)

retrofocus (inverted telephoto) — A negative-front, positive-rear design lengthening the back-focal distance to clear an SLR mirror box. (9.4 A short bestiary of classic designs)

RMS spot radius — The root-mean-square radius of a spot diagram, plotted versus field angle to read a design's sharpness. (9.3 Lens optimization)

robust combination (sigma-clipping) — Combining a stack while iteratively rejecting samples beyond kσ then re-averaging. (10.1 Denoising by averaging)

robust statistics — Edge-preserving smoothing seen as estimation with a saturating penalty, so neighbors across an edge become outliers. (5.2 Bilateral filtering)

rod — A highly sensitive retinal photoreceptor for dim, colorless night vision, dominant in the periphery. (2.8 Human (and animal) vision and color)

rolling shutter — a sensor read out row-by-row rather than all at once, skewing or wobbling fast motion. (Sensors — photosites, CCD vs CMOS)

rolling-shutter correction — Rectifying row-by-row readout distortion by reprojecting each scanline from its per-row pose. (12.4 Video stabilization and rolling-shutter correction)

rule of thirds — A composition default placing the subject a third of the way into the frame. (17.1 Human factors and the art of photography)

S-curve (sigmoid) — A tone curve steep through the midtones and rolling off at the extremes, avoiding the hard clip of a straight contrast line. (3.3 Point operations)

saccade — A fast, ballistic eye movement re-aiming the fovea between fixations. (2.8 Human (and animal) vision and color)

saliency — The property of regions that draw the gaze (faces, motion, high contrast), feeding content-aware cropping. (2.8 Human (and animal) vision and color)

SAM (Segment Anything) — A promptable pretrained segmentation model returning clean masks from a point, box, or "everything" prompt. (8.8 Compositing, segmentation and matting)

saturation (point operation) — A color operation scaling each pixel's chrominance about its luminance; s=0 gives grayscale, s>1 pushes colors from grey. (3.3 Point operations)

scanpath — The record of where the eye lands over time, a sequence of fixations connected by saccades. (2.8 Human (and animal) vision and color)

Scheimpflug condition — With a tilted lens, the lens, sensor, and plane of sharp focus all meet along one line. (9.5 Special optics)

Scheimpflug / tilt-shift — View-camera optics that fix perspective at capture: shift selects an off-center image region, tilt sets the focus plane. (6.7 Perspective distortion and its correction)

score — The gradient of the log-density pointing toward more probable images, which Tweedie's formula ties to a Gaussian denoiser. (4.5 Generative AI and diffusion)

scotopic — The night regime in which only rods work, giving no color and poor acuity. (2.8 Human (and animal) vision and color)

seam carving — Content-aware resizing that removes the lowest-energy connected one-pixel-per-row seam found by dynamic programming. (5.8 Seam optimization)

Seidel aberrations — The five monochromatic aberrations — spherical, coma, astigmatism, field curvature, distortion. (9.1 Aberrations and optical challenges)

separable filter — A 2-D filter expressible as the outer product of two 1-D filters, applied as a row pass then a column pass. (3.6 Neighborhood operations and convolution)

shadow removal — Flattening a shading drop by detecting the shadow boundary, zeroing its gradients, and Poisson-reconstructing with a soft penumbra matte. (8.9 Illumination related effects in a single image)

shape-from-focus — Using a focal stack's per-pixel sharpest-frame index, which maps to focus distance, as a depth map. (10.10 Focal stacks and depth of field extension)

sharpening — Making an image crisper by amplifying how each pixel differs from its neighbors, classically the unsharp mask. (3.6 Neighborhood operations and convolution)

Shi–Tomasi (good features to track) — The selection rule keeping a feature only if the smaller eigenvalue of its structure tensor exceeds a threshold (a corner). (7.2 Feature tracking)

shift-invariant — The property that the same operation applies at every pixel, so shifting the input merely shifts the output. (3.6 Neighborhood operations and convolution)

Shift-Map editing — Posing image editing as a per-pixel shift labelling solved by graph cut, with a data term and an edge-placing seam term. (8.7 Patch match)

shift theorem — The Fourier fact that translating an image multiplies its spectrum by a phase ramp and changes nothing else. (10.2 Image alignment)

Shirley card — The light-skinned calibration reference historically used to tune film and exposure, a source of skin-tone bias. (17.2 Ethics of computational photography)

shot noise — Poisson noise from the discreteness of photons; its variance grows linearly with the signal, so its standard deviation grows as √signal. (Noise)

shutter angle — The cinematographic expression of exposure time as a fraction of the frame interval, 180° the natural film look. (12.1 Motion blur and temporal sampling)

shutter speed — The exposure-time control, entering exposure linearly and governing motion blur. (2.10 Photography and camera 101)

sidecar file — A separate file holding metadata or edit settings alongside an untouched raw rather than embedded. (22.12 EXIF and image metadata)

SIFT — A scale- and rotation-invariant detector (DoG extrema) plus a 128-D gradient-orientation-histogram descriptor. (10.6 Automatic panorama stitching and feature matching)

signal-to-noise ratio (SNR) — signal ÷ noise standard deviation; for shot noise SNR = √N, so it is worst in the shadows even though absolute noise is largest in the highlights. (Noise)

sinc — The ideal infinite-support reconstruction filter sin(πx)/(πx), optimal but unrealizable, so practical kernels approximate it. (3.7 Linearity, Fourier, aliasing and deblurring)

singular value decomposition — The factorization A=UΣVᵀ reading as rotate–scale–rotate, the basis of low-rank approximation, PCA, and the pseudo-inverse. (22.1 Refreshers)

slow-motion synthesis — Producing synthetic frames between real pairs via frame interpolation to slow footage smoothly. (12.5 Frame interpolation and slow-motion synthesis)

smoothness term — In a graph-cut energy, the pairwise penalty for neighboring pixels taking different labels, small across true edges. (5.8 Seam optimization)

snapshot spectral imaging — Capturing the whole hyperspectral cube in one exposure via a filter mosaic or coded compressive sensing. (10.11 Hyperspectral imaging color wheels)

soap-opera effect — The over-smooth, cheap-video look of heavily frame-rate-up-converted film. (12.5 Frame interpolation and slow-motion synthesis)

Sobel kernel — A 3×3 gradient kernel pairing a finite-difference derivative on one axis with 1-2-1 smoothing on the other. (3.6 Neighborhood operations and convolution)

soft-focus lens — A portrait lens deliberately leaving spherical aberration under-corrected, overlaying a glow on a sharp core. (9.5 Special optics)

spectral reflectance (ρ(λ)) — the fraction of light a surface reflects at each wavelength; intrinsic to the surface, independent of the light. (Light and physics)

spectral scanning — Filling the hyperspectral cube one band at a time with a filter wheel or tunable filter, requiring a still scene. (10.11 Hyperspectral imaging color wheels)

spectral theorem — A symmetric matrix's eigenvectors are orthogonal, so diagonalizing it gives a basis of pure per-axis scaling. (22.1 Refreshers)

spectrum — the distribution of light power across wavelengths; "color" upstream of the camera, before it is compressed to three numbers. (Light and physics)

specular — a shiny surface that reflects into a narrow lobe near the mirror direction, giving a view-dependent highlight. (Light and physics)

specular-highlight removal — Stripping a glossy highlight by projecting each pixel onto its diffuse color line in the dichromatic model. (8.9 Illumination related effects in a single image)

spherical aberration — The aberration where marginal rays focus nearer the lens than central rays, giving a glow, worst wide open. (9.1 Aberrations and optical challenges)

splat, blur, slice — The three-step bilateral-grid pipeline: accumulate pixels into a coarse 3-D grid, convolve, then read back by trilinear interpolation. (5.2 Bilateral filtering)

spline — A smooth curve stitched from low-degree polynomial pieces through control points; the fit used for the freehand tone curve. (3.3 Point operations)

spot diagram — A scatter of ray landing points for one field point and color — the ray-sampled estimate of the PSF. (9.3 Lens optimization)

steerable pyramid — A multiresolution decomposition splitting each band into oriented sub-bands, adding direction at the cost of redundancy. (3.9 Linear pyramids and wavelets)

steering kernel — LARK's anisotropic regression kernel that steers its size and orientation to the local image structure. (5.3 Locally adaptive regression kernel (LARK))

stereo matching — Finding each point's correspondence in the other view and reading off its disparity. (2.7 Multiple view geometry)

stereo / stereopsis — Recovering depth from two viewpoints, the visual system's fusion of two retinal images into 3-D. (2.7 Multiple view geometry)

stochastic gradient descent — Gradient descent using a random minibatch per step, the workhorse for training neural networks. (22.1 Refreshers)

stop — The photographic unit of exposure, one stop being a factor of two in light. (3.3 Point operations)

stopping down — Closing the iris to use only the well-behaved central aperture zone, reducing most aberrations. (9.2 Aberrations correction)

stroke-based rendering — Repainting a photo with discrete brush strokes coarse-to-fine, oriented perpendicular to the gradient. (8.10 Non-photorealistic rendering)

structural similarity (SSIM) — A metric comparing local windows on luminance, contrast, and structure, correlating with perceived quality far better than PSNR. (3.10 Image metrics)

structure from motion — Reconstructing both cameras and scene by chaining relative poses across many views. (2.7 Multiple view geometry)

structure tensor — The windowed covariance of local gradients whose eigenvectors give an edge's along/across directions and eigenvalues its strength. (5.3 Locally adaptive regression kernel (LARK))

sub-pixel alignment — Refining an integer shift to fractional precision by fitting a parabola to the score basin. (10.2 Image alignment)

submodularity — The condition on a graph-cut smoothness term under which binary min-cut yields the exact global optimum. (5.8 Seam optimization)

sum of squared differences (SSD) — A matching score minimized at the true shift, assuming brightness constancy. (10.2 Image alignment)

sunstars — Radiating spikes around a bright source, the Fourier transform of the polygonal diaphragm opening. (9.1 Aberrations and optical challenges)

super-resolution — Recovering an image meaningfully sharper than the input by inverting the blur-and-downsample model with a prior. (8.2 Super-resolution and image priors)

Super SloMo — A learned frame interpolator estimating bidirectional flow, refining intermediate flow, and predicting visibility. (12.5 Frame interpolation and slow-motion synthesis)

supervised learning — Fitting a parametric function to example input-output pairs by minimizing average loss. (22.1 Refreshers)

synthetic aperture — Combining separate apertures into one large virtual lens, as in the Event Horizon Telescope. (1.1 From Digital to Computational Photography)

synthetic data — Manufacturing supervised training pairs by simulating the degradation from clean targets. (4.3 Machine learning)

T-stop — An f-number corrected for transmittance, giving the actual exposure a lens delivers. (9.4 A short bestiary of classic designs)

teal-and-orange — A color grade pushing skin toward warm orange and everything else toward complementary teal. (3.3 Point operations)

telephoto — A positive-front, negative-rear design pushing the rear principal plane forward so the barrel is shorter than the focal length. (9.4 A short bestiary of classic designs)

temporal aliasing — Motion faster than half the frame rate folding to a false slow apparent motion — the wagon-wheel effect. (12.1 Motion blur and temporal sampling)

temporal redundancy — The near-identity of consecutive video frames that codecs exploit by coding only the difference. (12.2 Video compression and motion compensation)

tetrachromat — An animal with four cone types, seeing color dimensions humans lack. (2.8 Human (and animal) vision and color)

text encoder (CLIP) — A frozen model turning a prompt into embeddings that condition a diffusion denoiser via cross-attention. (4.5 Generative AI and diffusion)

texture synthesis — Growing more of a self-similar texture by copying from a source, pixel-at-a-time or by overlapping seam-cut blocks. (8.6 Inpainting, texture synthesis)

The Bitter Lesson — Sutton's argument that general methods scaling with data and compute eventually beat hand-built cleverness. (4.3 Machine learning)

thin-plate spline (TPS) — A radial-basis free-form warp that interpolates control displacements while minimizing bending energy. (6.1 Warping and resampling)

three-point lighting — A studio setup of key, fill, and rim lights that shapes the subject and separates it from the background. (17.1 Human factors and the art of photography)

through-the-lens (TTL) — Metering or flash metering performed through the taking lens, including a pre-flash. (2.10 Photography and camera 101)

TIFF/EP — The imaging variant of TIFF on which DNG is built. (22.13 DNG, the Digital Negative)

TIFF/IFD — The image file directory structure — a table of tagged fields — that holds EXIF and DNG metadata. (22.12 EXIF and image metadata)

tilt-shift lens — A lens that shifts (to control perspective) or tilts (to reorient the plane of sharp focus). (9.5 Special optics)

tolerancing — Budgeting manufacturing errors and predicting as-built performance and yield, often the real driver of lens cost. (9.3 Lens optimization)

tone curve — The most general point operation — a freehand input→output transfer curve through draggable control points. (3.3 Point operations)

tone mapping — compressing a high-dynamic-range image into a displayable range while preserving local contrast. (Tone mapping)

train/validation/test split — Partitioning data so a model is fit on one part, tuned on another, and honestly evaluated on a third. (22.1 Refreshers)

transcript-based editing — Editing a talking-head video by editing its time-aligned speech transcript. (12.6 Video editing)

transpose operator — The transpose of a blur, which is convolution by the flipped kernel; applied matrix-free like A. (4.1 Linear Inverse Problems and Regression)

triangulation — Intersecting two verified rays to recover a 3-D point from two views. (2.7 Multiple view geometry)

trichromacy — Three-cone color vision, in primates a re-evolved trait from a gene duplication. (2.8 Human (and animal) vision and color)

trimap — A three-way labelling (definite foreground, definite background, thin unknown band) that turns matting into a tractable solve. (8.8 Compositing, segmentation and matting)

Tweedie's formula — The identity expressing the optimal Gaussian denoiser's estimate in terms of the score, so a diffusion model is a learned score field. (4.5 Generative AI and diffusion)

two-image operation — An operation combining two same-size images pixel by pixel (add, subtract, multiply). (1.3 Basic intro to digital images)

two-scale blending — Blending the low band of each source smoothly and the high band winner-take-all, the two-level case of multiband. (10.7 Blending)

two-scale (base/detail) decomposition — Splitting an image into a large-scale base and a fine-scale detail layer so each can be treated differently. (5.2 Bilateral filtering)

two-scale style transfer — Transferring a model photo's look by splitting into base and detail layers and matching each separately. (8.5 Style transfer)

U-Net — An encoder–decoder convolutional architecture processing an image across scales, dominant for image-to-image tasks. (22.1 Refreshers)

unsharp mask — The standard sharpen, I + k·(I − blur(I)), adding back amplified high-frequency detail. (3.6 Neighborhood operations and convolution)

V-cycle — The down-and-up traversal of an image pyramid in multigrid that handles each error band where it is cheap. (4.1 Linear Inverse Problems and Regression)

vanishing point — The image of the point at infinity where world-parallel lines meet; the basis of automatic rectification. (6.7 Perspective distortion and its correction)

variational autoencoder (VAE) — A generative model that encodes to a latent, samples there, and decodes; its autoencoder defines latent diffusion's space. (4.5 Generative AI and diffusion)

vector space — The high-dimensional space in which an image, treated as a stacked column of pixels, is a single point. (22.1 Refreshers)

veiling glare — A low-frequency wash from countless weak reflections that lifts blacks and crushes contrast. (9.8 Glare suppression)

vibrance — A gentler saturation that boosts muted colors more than vivid ones and spares skin. (3.3 Point operations)

video magnification — The family of methods that pull out and amplify tiny temporal variations in video until visible. (12.3 Video magnification)

video textures — Turning a short clip of a repetitive phenomenon into an arbitrarily long seamless loop by finding low-cost frame transitions. (5.8 Seam optimization)

view morphing — A projectively-correct morph of two views that prewarps both images, interpolates, then postwarps, preserving shape. (6.3 Morphing)

vignetting — The dimming of an image toward the corners, from natural cos⁴ falloff and from optical clipping of the pupil. (9.1 Aberrations and optical challenges)

visibility mask — A per-frame weight that down-weights the frame where a pixel is occluded, enabling occlusion-aware interpolation. (12.5 Frame interpolation and slow-motion synthesis)

visible differences predictor (VDP) — A metric built on a human-vision model that outputs, per pixel, the probability a person would notice a difference. (3.10 Image metrics)

visual acuity — The finest line spacing the eye can resolve, the high-frequency tail of the contrast sensitivity function. (2.8 Human (and animal) vision and color)

visual microphone — Recovering audio from the micrometre-scale sound-induced vibrations of an object in high-speed video. (12.3 Video magnification)

von Kries adaptation — The white-balance model correcting each color channel by an independent gain (a diagonal transform). (3.13 Auto-exposure and auto white balance)

wagon-wheel effect — The temporal-aliasing demonstration where an undersampled spoked wheel appears to slow, stall, or spin backward. (12.1 Motion blur and temporal sampling)

Ward histogram-based tone mapping — A constrained equalization that caps the local slope of the CDF tone curve so contrast is never pushed past what the eye accepts. (3.4 Histograms)

warp (image warping) — A domain transform that moves where pixels live while leaving their colors alone, the shared primitive of the motion/warping part. (6.1 Warping and resampling)

wavefront coding — Engineering a lens's PSF (via a phase mask or coded aperture) so its blur inverts cleanly or encodes depth. (9.2 Aberrations correction)

wavelet — A multiresolution transform forming a non-redundant basis localized in both space and frequency, the natural choice for compression. (3.9 Linear pyramids and wavelets)

Weber's law — The just-noticeable intensity increment is a roughly constant fraction of the background, ΔI/I ≈ const. (2.8 Human (and animal) vision and color)

Weiss intrinsic-image estimator — Taking the per-pixel median over time of log-gradients to get the reflectance gradient field, then integrating it. (10.12 Intrinsic images with time lapse)

white balance — correcting a scene's color cast so neutrals read neutral; gray-world scales each channel to a common mean. (Color technology)

white-patch assumption — An auto-white-balance estimator that takes the brightest pixels to be a white highlight reflecting the illuminant. (3.13 Auto-exposure and auto white balance)

wide-angle portrait correction — A content-aware mesh warp applying a locally stereographic map over faces and original perspective elsewhere. (6.7 Perspective distortion and its correction)

Wiener filter — The optimal linear deblurring/denoising filter that restores signal where the blur left healthy content and declines to amplify noise. (3.7 Linearity, Fourier, aliasing and deblurring)

XMP — Adobe's XML-based metadata standard where edit settings, keywords, and ratings are stored. (22.12 EXIF and image metadata)

zippering — The alternating light/dark teeth along a sharp edge in naive demosaicking. (3.12 Demosaicking)

Zone System — Ansel Adams and Fred Archer's discipline carving the tonal range into eleven zones and previsualizing which scene tone lands in which. (3.3 Point operations)

zoom lens — A lens of variable focal length where a variator group changes magnification while a compensator holds the image plane. (9.4 A short bestiary of classic designs)