16.10 Deblurring TLDR⧉
- Prevent blur when possible: improve focus, shorten exposure for motion, stabilize the camera, or capture a burst. Stabilization does not freeze a moving subject.
- With a calibrated kernel, start with regularized deconvolution. Match the noise model and boundaries; Richardson–Lucy Richardson 1972 is for a suitable nonnegative Poisson-count model, not arbitrary background-subtracted RGB (Chapter 16.2, Non-blind deconvolution).
- With unknown or spatially varying blur, test a suitable learned model or a physical estimator using available motion, depth or calibration data. No method guarantees the identity of unresolved text or face detail. Keep the original.
- Do not equate sharpening with recovery. Weak transfer responses amplify noise; common zeros leave information unmeasured. Check ringing, texture loss and changes to small structures, not just apparent crispness.
- With several frames, compare selection with fusion. Different blurs can supply complementary information; lucky imaging discards photons to select favorable atmospheric states. Judge the final stack at the same capture budget (Chapter 16.8, Video, bursts, and lucky imaging).
- Evaluate the complete output on representative blur and noise. If a kernel is estimated, inspect it and test its effect through a controlled non-blind solver. Report reference preparation, runtime and any generative processing (Chapter 16.9, Measuring a deblur).
A blur spreads each scene point according to a point-spread function. In the spatially invariant model, the corresponding transfer function tells us how each spatial frequency is scaled. Dividing by a small response amplifies noise; dividing by zero cannot recover a component the measurement did not record. Regularization limits that amplification by introducing assumptions. Both defocus and motion kernels can have zeros. Additional observations help when they supply complementary responses (Chapter 16.1, What blur is, and what deblurring can undo).
With a known kernel, compare a simple regularized baseline before using a more elaborate reconstruction. Wiener filtering Wiener 1949 balances inversion against the relative signal and noise power. Richardson–Lucy is derived from a nonnegative Poisson observation model; the forward and adjoint operators, boundary handling, and any background term must agree with that model. More iterations can fit noise, so stopping is part of the reconstruction choice. Neither method makes a wrong kernel harmless (Chapter 16.2, Non-blind deconvolution).
The prior determines which structures survive uncertain inversion. A strong smoothness penalty can remove the texture we hoped to recover. Sparse-gradient and total-variation penalties allow strong edges while suppressing small fluctuations, with characteristic biases such as piecewise-constant regions. A denoiser inside an iterative reconstruction supplies a more flexible image model, but its behavior and convergence need to be checked with the chosen data term (Chapter 16.3, Priors for deconvolution, Chapter 13.11, Denoising as a universal prior).
Blind deblurring must estimate both the image and the blur. A delta kernel paired with the observed blurry image already explains the data, so the data fit alone does not select a useful answer. The image prior, kernel constraints, and estimator all matter. Joint maximum a posteriori estimation can favor this trivial explanation under common formulations; marginalizing over images changes how candidate kernels are assessed. Inspect the estimated kernel as well as the final image (Chapter 16.4, Blind deblurring).
Choose the physical model from the cause. Camera rotation can smear different parts of the image differently; depth-dependent defocus needs spatially varying blur; independently moving objects and occlusion add further complications. A single global convolution can then fail even with an excellent solver. Pose trajectories, depth cues, multiple captures, and learned models provide different ways to constrain the problem. Dual-pixel observations supply a defocus cue under a calibrated model, not an unconditional depth measurement (Chapter 16.5, Spatially varying blur, Chapter 16.6, Defocus, and undoing what the lens softened).
Learned deblurring can handle complex image statistics and trained blur distributions, but its output still needs inspection. Check fine text, repeated patterns, and faces where a plausible replacement would be misleading. An unfamiliar camera pipeline or motion distribution can change the result; a successful synthetic benchmark does not establish performance for every real exposure (Chapter 16.7, Learned deblurring).
With multiple frames, compare selection, registration and fusion, and joint reconstruction. Lucky imaging selects favorable atmospheric states, losing some photons to gain sharper observations; turbulence can also require local registration and restoration. Differently blurred frames may contain complementary frequency information, provided the model accounts for their relationship. Treating a burst as an unordered set does not remove the need to handle scene motion (Chapter 16.8, Video, bursts, and lucky imaging).
Evaluate kernel estimation with a controlled non-blind solver when possible, so a stronger image prior does not hide a poorer kernel. Report boundary treatment, reference alignment, noise, and the capture budget. Ringing can make a sharpness score improve while making the photograph worse. Look at bright boundaries and small structures in addition to the aggregate error (Chapter 16.9, Measuring a deblur).