Image restoration—the process of recovering a sharp, high-resolution image from a blurry or low-quality observation—is a classic "inverse problem." To solve it, we typically assume a forward model: the original image was smeared by a specific blur kernel (or Point Spread Function, PSF) and then corrupted by noise. If we know that kernel exactly, we can mathematically "undo" the blur.
However, research across decades reveals a persistent, frustrating reality: image restoration is hyper-sensitive to the accuracy of this kernel. Even a tiny discrepancy between the assumed blur and the actual physical process can lead to catastrophic failure.
The Fragility of the Forward Model
In a perfect world, "non-blind" deconvolution—where the blur kernel is provided—should be a solved problem. Yet, as S. Blondin and colleagues noted in their work on Extracting clean supernova spectra, even minor inaccuracies in modeling the spatial distribution of light can lead to disproportionately large errors. In astronomy, a small percentage error in the estimated blurring kernel can result in a total failure to restore the underlying signal, potentially skewing measurements of the universe's expansion.
This sensitivity isn't just a problem for telescopes. In consumer electronics, Keiichi Kondo demonstrated that in motion-blur restoration, the precise length of the smear is more critical than the signal-to-noise ratio. A single-pixel discrepancy in the estimated blur length causes the mean square error to spike and introduces "ringing" artifacts—ugly, oscillating ghosts that ripple out from sharp edges.
Why Small Errors Cause Big Artifacts
To understand why this happens, think of deconvolution as a form of division in the frequency domain. If your kernel estimate is slightly off, you are essentially dividing by the wrong numbers. In areas where the true kernel has low energy (high frequencies), a slight mismatch causes the restoration algorithm to over-amplify noise or create artificial patterns.
Anat Levin showed in Blind Motion Deblurring Using Image Statistics that this sensitivity necessitates extremely precise estimation. If the algorithm cannot distinguish between very similar blur levels, the resulting "ringing" can completely obscure structural details. This is why many traditional methods, such as those discussed by Takashi Shibata, struggle with "halos" and oscillations when the kernel isn't a perfect match.
Bridging the Gap with Deep Unfolding
Modern research has moved toward "blind" restoration, where the kernel and the image are estimated simultaneously. However, simply throwing a neural network at the problem doesn't automatically solve the sensitivity issue. Zhengxiong Luo and others found that decoupling kernel estimation from restoration often backfires; if the estimation module produces an imperfect kernel, the restoration module—trained on ground-truth data—won't know how to handle it.
The solution has been Deep Unfolding. By taking classical iterative optimization algorithms and "unfolding" them into neural network layers, researchers like jiawei zhang and Jinshan Pan have created models that respect the physical image formation process while remaining flexible. These networks, such as the Deep Alternating Network, iteratively refine both the kernel and the image, allowing the model to "correct" its own estimation errors before they propagate into artifacts.
Plug-and-Play and Internal Learning
Another powerful direction is the Plug-and-Play (PnP) framework. Instead of training a single model for a single blur, researchers like Kai Zhang developed methods like DPSR that can take any blur kernel as input at test time. These models use a pre-trained denoiser as a "prior" to keep the image looking natural, even when the kernel is complex.
To combat the "unseen kernel" problem, some researchers turn to the image itself. Raja Giryes and Tom Tirer introduced "internal learning," where a model fine-tunes itself on the specific textures of the test image. By ensuring the high-resolution output remains mathematically consistent with the low-resolution input—a concept also explored by Yuval Bahat in Explorable Super Resolution—these systems become far more robust to the "fragility" of the initial degradation model.
Ultimately, the field is moving away from the search for a "perfect" kernel and toward systems that can gracefully handle the uncertainty of the real world. Whether through meta-learning (Jae Woong Soh) or diffusion-based priors (Kai Zhang), the goal remains the same: preventing a single-pixel error from ruining a perfect picture.