SSDSNet: a dual-stream state-space network for efficient image denoising with long-range dependency modeling

Abstract In recent years, advancements in deep learning architectures such as Convolutional Neural Networks (CNNs) and Transformers have significantly enhanced the field of image denoising. Nevertheless, current denoising techniques often encounter challenges stemming from inherent local inductive biases or the quadratic computational demands of their designs. Recently, Selective Structured State Space Models, particularly exemplified by Mamba, have demonstrated substantial promise in capturing long-range dependencies with linear computational complexity. Despite their potential, these models have not been extensively explored for low-level computer vision applications. In this paper, we present the Dual-Stream State-Space Module (DSSM), a simple yet powerful model tailored for image denoising tasks. The DSSM integrates convolutional operations with the Mamba model, thereby enhancing the ability to extract both local features and global contextual information. This dual-stream framework effectively utilizes the structured nature of local image patches while adeptly modeling long-range interactions, resulting in feature representations optimized for denoising. Comprehensive experiments validate the efficacy of our approach. Notably, on the SIDD dataset, our model matches the performance of leading state-of-the-art methods, maintaining a global receptive field with comparable computational efficiency.