Real&Synthetic Dataset and the Linear Attention in Image Restoration

Yuzhen Du, Teng Hu, Jiangning Zhang, Ran Yi, Chengming Xu, Xiaobin Hu, Kai WU, Donghao Luo, Yabiao Wang, Lizhuang Ma

Keywords: Image Restoration, Vision-RWKV

Abstract: Image restoration (IR), which aims to recover high-quality images from degraded inputs, is a crucial task in modern image processing. Recent advancements in deep learning, particularly with Convolutional Neural Networks (CNNs) and Transformers, have significantly improved image restoration performance. However, existing methods lack a unified training benchmark that specifies the training iterations and configurations. Additionally, we construct an image complexity evaluation metric using the gray-level co-occurrence matrix (GLCM) and find that there exists a bias between the image complexity distributions of commonly used IR training and testing datasets, leading to suboptimal restoration results. Therefore, we construct a new large-scale IR dataset called ReSyn, that utilizes a novel image filtering method based on image complexity to achieve a balanced image complexity distribution, and contains both real and AIGC synthetic images. From the perspective of measuring the model's convergence ability and restoration capability, we construct a unified training standard that specifies the training iterations and configurations for image restoration models. Furthermore, we explore how to enhance the performance of transformer-based image restoration models based on linear attention mechanism. We propose RWKV-IR, a novel image restoration model that incorporates the linear complexity RWKV into the transformer-based image restoration structure, and enables both global and local receptive fields. Instead of directly integrating the Vision-RWKV into the transformer architecture, we replace the original Q-Shift in RWKV with a novel Depth-wise Convolution shift, which effectively models the local dependencies, and is further combined with Bi-directional attention to achieve both global and local aware linear attention. Moreover, we propose a Cross-Bi-WKV module that combines two Bi-WKV modules with different scanning orders to achieve a balanced attention for horizontal and vertical directions. Extensive experiments demonstrate the effectiveness and competitive performance of our RWKV-IR model.

Primary Area: datasets and benchmarks

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: 2816