Most image deblurring algorithms are built on a simple mathematical promise: a blurry image is just a sharp image convolved with a blur kernel, plus some Gaussian noise ($B = K * I + N$). This linear model works beautifully in well-lit conditions, but it falls apart the moment a camera faces a night scene with bright streetlamps or specular reflections.
The problem is sensor saturation. When light intensity exceeds the camera sensor's maximum capacity, the pixel values "clip" at a ceiling (usually 255 in 8-bit images). This clipping is a non-linear process that destroys the mathematical relationship between the sharp scene and the blurred result. If you apply standard linear deconvolution to these clipped regions, the math tries to "undo" a process that didn't happen linearly, resulting in the ugly, ripple-like distortions known as ringing artifacts.
The Linearity Gap and Ringing Artifacts
In low-light photography, the dynamic range of the scene often dwarfs the dynamic range of the sensor. As Zhihong Zhang and colleagues note in their work on INFWIDE, this saturation introduces non-linearities that break fundamental deconvolution models.
When an algorithm assumes a linear relationship that isn't there, it encounters "outliers"—pixels that don't fit the expected blur pattern. Jinshan Pan has shown that these outliers can lead the optimization process toward a "trivial solution," where the algorithm incorrectly concludes there is no blur at all, or produces significant mathematical instability. The most visible symptom is the halo or ringing artifact: ghostly ripples that spread from bright light sources across the image, obscuring fine details in the shadows.
Screening and Masking Outliers
One of the most effective ways to handle saturation is to simply stop the algorithm from "looking" at the problematic pixels during the most sensitive part of the process: kernel estimation.
Jinshan Pan and Jiangxin Dong proposed frameworks that identify and mask saturated pixels, treating them as missing data rather than valid evidence for the blur kernel. Similarly, Meina Zhang introduced a pixel screening strategy that uses Bayesian estimation to exclude "unfavorable" pixels—those caused by saturation or heavy ringing—from the iteration process.
A more gradual approach is self-paced learning. Dong Gong and Qinfeng Shi developed a self-paced kernel estimation framework. The intuition is to start by estimating the blur kernel using only the most "reliable" (non-saturated) pixels. Once a rough kernel is defined, the algorithm gradually incorporates more complex regions, preventing the non-linearities of clipped pixels from poisoning the initial estimation.
Turning Streaks into Cues
While many researchers treat saturated highlights as a nuisance to be masked, others see them as a unique opportunity. In low-light scenes, a bright point source (like a distant star or a small LED) effectively draws the blur kernel directly onto the sensor as a "light streak."
Sunghyun Cho and Ming-Hsuan Yang pioneered the idea of explicitly modeling these streaks as blurred point light sources. By incorporating the camera's non-linear response function and the clipping process into the model, they turned saturated highlights from outliers into the primary cue for recovering the sharp image. Instead of failing because of the "ceiling," the algorithm uses the shape of the clipped region to understand the motion of the camera.
Restoration in the Irradiance Domain
Another path to robustness involves changing the space in which the deblurring happens. Standard images are stored in a non-linear "gamma-corrected" space designed for human viewing, which further complicates the deconvolution math.
Xianyong Fang and colleagues proposed deblurring in the irradiance domain. By first estimating the Camera Response Function (CRF), they map pixel values back to the actual physical light intensity (irradiance) that hit the sensor. In this domain, the relationship between the sharp and blurry scene is more strictly linear, making it easier to identify and handle saturated blocks using dual-threshold detection.
Deep Learning for Non-Linear Restoration
Recent work has shifted toward neural networks that can learn to handle these "rule-breaking" pixels without explicit manual masking.
Networks like NBDN (Night Blurry Images Deblurring Network) learn to identify clipped regions and adaptively tune their restoration parameters to avoid ringing. Others, like the Deep Richardson-Lucy approach by Liang Chen, replace traditional hand-tuned clipping functions with learned "latent maps." These maps allow the network to hallucinate the missing information in saturated areas while suppressing the artifacts that usually arise from the loss of data.
By moving away from the rigid assumptions of linear convolution, this research direction has transformed low-light deblurring from a fragile mathematical exercise into a robust tool capable of handling the messy, high-contrast reality of night-time photography.