Haste Makes Waste: Teaching Image Restoration to Learn Distributions from Pixels to Patterns

Enxuan Gu, Hongwei Ge, Chen Zhao, Yong Guo

Keywords: Image Restoration, Low-level Vision, Training Strategy

Abstract:

In this paper, we revisit the image restoration (IR) task and propose a new training strategy that models the IR problem as a distribution mapping challenge from two perspectives, i.e., (1) the intra-pixel regression and (2) the inter-pixel interaction. At the beginning of optimization, due to the pattern distribution involving a group of pixels within a neighborhood, it is not very easy for the model to capture such multi-pixel distribution mapping. A more optimal solution would be firstly teaching the model to learn a relatively simple yet important distribution w.r.t the pixel-by-pixel mapping between the degraded/clean pixels, as warming up. By doing so, the learned distribution is served as a prior, regarded as an injection of a kind of inductive bias into the model's whole optimization procedure. Subsequently, as conventional, the model is shifted to focus on the mapping distribution of the cross-pixel patterns, which ensures the consistency and fidelity of the image patterns. The final learned mapping is a joint distribution, which transfers the knowledge from the pixel distributions to the pattern ones. Experimental results indicate that under the compact and elegant training paradigm, the newly learned joint distribution is closer to the ideal one and yields a stronger representation ability, to circumvent the dilemma of the difficulty for existing methods to learn the patterns mapping distribution between degraded/clean images right off the bat.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4346