Individual/Joint Deblurring and Low-Light Image Enhancement in One Go via Unsupervised Deblurring Paradigm

Suiyi Zhao, Zhao Zhang, Jicong Fan, Yang Zhao, Shuicheng YAN

Primary Area: representation learning for 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.

Keywords: deblurring, low-light image enhancement, unsupervised learning, joint task processing

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: A novel strategy for joint deblurring and low-light image enhancement

Abstract: Image restoration and enhancement, e.g., image deblurring, deraining and low-light image enhancement (LLIE), aim to improve the visual quality according to the corrupted/low-quality observation. Deep learning-based methods have achieved remarkable results on these individual tasks, but it is still hard to tackle them together. While some attempts have been made to implement joint task processing, they inevitably lead to higher data cost and higher training cost. Moreover, these attempts are strictly limited by the data distribution, i.e., the distribution of the inference data is required to be as close as possible to the training data, otherwise the data cannot be used for inference. In this paper, we take the LLIE and deblurring task as the subjects of this study in an attempt to seek a novel solution to the joint task processing problem. Specifically, we tackle this kind of problem in an extraordinary manner, i.e., \textit{Individual/Joint \underline{D}eblurring and Low-Light Image \underline{E}nhancement in One Go \underline{v}ia \underline{U}nsupervised \underline{D}eblurring \underline{P}aradigm (DEvUDP)}, which integrates the noise self-regression and could avoid the limitations of aforementioned attempts. More specifically, a novel architecture with a transformation branch and a self-regression branch is elaborated, which only accepts unpaired blurry-sharp data as input to train the model; in this way, the pre-trained model can be surprisingly applied to both LLIE, deblurring and mixed degradation processing. Besides, we can choose to highlight perceptual performance or distortion performance of the model by configuring different components to the architecture. Extensive experiments have demonstrate the superiority of the method on different widely-used datasets.

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: 2538