Multi-scale representation for image deraining with state space model

With the rapid development of deep learning, significant progress has been made in image deraining techniques, with current methods primarily based on Convolutional Neural Network (CNN) and Transformer architectures. However, these approaches often face limitations due to the fixed receptive fields of CNN and the computational complexity of Transformer. Moreover, most existing image deraining methods adopt a single-input single-output network architecture, which struggles with multi-scale feature representation, particularly in capturing both global and local information, leading to potential information loss. To address these issues, this paper proposes a multi-scale visual state space model for image deraining, aiming to improve deraining performance and image restoration quality by combining multi-scale networks with state space model. Specifically, we design a multi-scale Mamba block that models global features with linear computational complexity, and we develop an efficient multi-scale 2D scanning strategy that uses geometric transformations to apply different numbers of scanning directions at various scales, thereby better extracting feature information at each scale. Additionally, we introduce a Frequency Feature Enhancement Module to capture local feature information, and a Gated Feature Fusion Module to adaptively aggregate complementary features across scales, further enhancing image restoration quality and deraining performance. Experimental results demonstrate that our method achieves superior deraining performance on multiple public benchmark datasets, outperforming the current state-of-the-art methods, while significantly improving efficiency and maintaining low computational cost.