Single Image Reflection Separation via Visual State Space Model and Feature Interactions

In this paper, we propose a novel single image reflection separation (SIRS) approach based on the Visual State Space Model (VSSM) and feature interaction mechanism. In particular, it is a two-stage cascaded network and consists of a Shallow Separation Network (SSNet) and a Deep Separation Network (DSNet). Given a mixed image, SSNet is used to extract multi-scale image features and achieve the initial transmission and reflection feature components. These initial separated feature components are then fed into the DSNet to perform a more fine-grained and complete component separation, and finally achieve the transmission image and reflection image. In both networks, an Exclusive and Complementary Feature Interaction Module (ECFIM) is employed to carry out the feature interactions between different feature components. Moreover, for the same feature component of DSNet, the visual state space model (VSSM) is utilized to capture and fuse the multi-scale and multi-channel global image spatial information. Experimental results show that our proposed approach generates high-quality transmission and reflection images, outperforms state-of-the-art methods on multiple real-world benchmark datasets.