16.7 Learned deblurring⧉
A moving person can produce different blur in different regions, with changing visibility during the exposure. A direct learned model can estimate a sharp image without first returning a kernel, using training pairs to learn these patterns.
Such scenes exceed a single-convolution model, but not every physical or classical model. Given geometry, motion and visibility, exposure integration can still be linear in scene radiance. The difficulty is inferring those latent quantities and content hidden from the chosen reference view.
16.7.1 Networks that skip the kernel⧉
The direct approach maps blurred to sharp without an explicit estimated blur operator, and two design choices recur because they suit this problem specifically.
Multi-scale processing. Blur has a size, and a kernel that spans thirty pixels is invisible to a network whose receptive field is ten. Coarse-to-fine architectures deblur a downsampled image, where the blur is small, and use that result to condition the next scale up, which is the same coarse-to-fine reasoning the classical blind methods use for kernel estimation.
Recurrence and weight sharing across scales, which is the same idea with fewer parameters, and which produced the strong early results (Nah, Kim & Lee 2017 (GoPro, multi-scale deblurring); Tao et al. 2018 (SRN-DeblurNet)).
Later architectures are the general restoration line rather than anything deblurring-specific: multi-stage designs (Zamir et al. 2021 (MPRNet)) and the simplified blocks of NAFNet (Chen et al. 2022 (NAFNet)), which provides a useful published baseline when its training domain and computation match the task.
Unrolled deconvolution retains an explicit operator and learns some update components (Chapter 13.7, Learned denoisers). Learned steps can include data updates, transforms and regularization, not only a prior. A finite unrolled network does not automatically enforce measurement consistency or inherit the original solver's convergence guarantee.
Adversarial methods such as DeblurGAN (Kupyn et al. 2018 (DeblurGAN); Kupyn et al. 2019 (DeblurGAN-v2)) optimize appearance-sensitive losses. Diffusion restoration uses conditional generation or an image prior with measurement guidance (Chapter 13.9, Generative denoising with diffusion models). Both can introduce incorrect sharp detail; inspect text and repeated patterns against a reference.
16.7.2 The datasets are the argument⧉
Supervised deblurring needs a target corresponding to a defined instant and view. The reference has its own exposure, blur and noise; neither synthetic nor captured pairs are automatically exact.
GoPro uses high-frame-rate video and averages consecutive frames to synthesize longer exposure (Nah, Kim & Lee 2017 (GoPro, multi-scale deblurring)). The authors provide both gamma-corrected and uncorrected construction variants; identify which is used. Discrete frames approximate continuous exposure, and frame rate alone does not specify each target's shutter duration. Residual target blur, clipping and the camera response remain modeling issues.
RealBlur (Rim et al. 2020 (RealBlur)) captures instead: a beam splitter feeds two cameras, one with a long exposure and one with a short one, triggered together, so the pair is real. The residual difficulties are geometric alignment between the two arms and the noise in the short exposure, which is what a short exposure costs.
Evaluate across the intended motion, noise and camera domains. A score on a synthetic averaged-frame set does not by itself establish performance on captured blur or a different response pipeline.
16.7.3 What learning did and did not change⧉
Learning can fit complex restoration mappings, including spatially varying motion and appearance statistics. It does not eliminate ambiguity or make physically modeled reconstruction irrelevant. Compare a known-operator solver when calibration is available, and a suitable learned model when its training conditions match the input. For a direct model, report the final image and failure cases rather than demanding a kernel it never estimates.