16.0 DEBLURRING⧉
A streetlamp smeared by camera shake and a point spread by defocus can both make a photograph look soft. Their reconstruction differs because the spatial response, exposure motion and depth dependence differ. Atmospheric turbulence adds changing wavefront distortion; short-exposure selection is one response, alongside optical correction and reconstruction.
16.0.1 Attenuated is not the same as destroyed⧉
A spatially invariant blur multiplies the image spectrum by a transfer function. Small responses make inversion sensitive to noise. A zero removes that component from the measurement, leaving its value to additional observations or assumptions. Both an ideal defocus disc and a uniform linear-motion exposure can have zeros. Real response and signal-to-noise ratio determine what detail is usable.
16.0.2 Two difficulties, stacked⧉
With a known operator, estimate the image using a likelihood and suitable regularization. With an unknown operator, also infer blur from the image, calibration or auxiliary measurements. Image and kernel can admit multiple explanations; an estimator must handle that ambiguity. Spatial variation, depth and occlusion can require a model beyond one convolution.
The unknown-kernel case has a particularly revealing ambiguity: the observed photograph is explained perfectly by calling it the sharp image and choosing a one-pixel impulse as the blur kernel. A useful method needs evidence or assumptions that favor another explanation. The choice between jointly estimating image and kernel and integrating over possible images is therefore consequential, not just a different implementation of the same objective. Chapter 16.4, Blind deblurring develops that distinction.
16.0.3 The arc of the chapters⧉
The known-kernel chapters develop Wiener filtering Wiener 1949, Richardson–Lucy (Richardson 1972; Lucy 1974) and image priors. Chapter 16.4, Blind deblurring and Chapter 16.5, Spatially varying blur address unknown and nonuniform operators. The defocus chapter connects calibration and dual-pixel cues to reconstruction. Learned methods, bursts and evaluation complete the part. Part 13, Denoising supplies the noise and prior foundations; Matching pixels across space and time supplies correspondence.
16.0.4 What to carry forward⧉
Preserve the capture and distinguish measured structure from inferred detail. Report the blur model, reference error, artifacts and cost. When an algorithm estimates a kernel, inspect that estimate as well as the final image.
Contents of this part
- 16.1 What blur is, and what deblurring can undo
- 16.1.1 Why naive inversion fails
- 16.1.2 The fix, stated once
- 16.1.3 Which space to deblur in
- 16.2 Non-blind deconvolution
- 16.3 Priors for deconvolution
- 16.4 Blind deblurring
- 16.5 Spatially varying blur
- 16.6 Defocus, and undoing what the lens softened
- 16.7 Learned deblurring
- 16.8 Video, bursts, and lucky imaging
- 16.9 Measuring a deblur
- 16.9.1 What went wrong, specifically
- 16.9.2 What to measure
- 16.9.3 On what data
- 16.9.4 What to report
- 16.10 Deblurring TLDR