Dual-path fusion network based on transformer and local binary patterns for compressive sensing reconstruction

Abstract In recent years, deep-learning-based methods have made significant progress in the field of compressed sensing. However, most existing deep-learning-based solutions commonly encounter issues in image reconstruction due to inadequate interaction with image textures during the reconstruction process. In this study, we developed a dual-path fusion network that combined structural and textural information for image reconstruction. Guided by structural priors, we designed a new focus linear cross window transformer network that computes attention through parallel cross-SWs of different sizes, integrating both local and global structural information to enhance feature interaction. Meanwhile, a submodule based on local binary pattern is used to leverage image texture priors for learning texture features, thus providing rich texture information for image reconstruction. The features from these two distinct paths are adaptively fused, and the reconstruction results are optimized using an iterative thresholding algorithm. This method effectively combines the advantages of traditional algorithms and deep neural networks, taking advantage of the structural and textural prior information. Experiments demonstrated that the proposed method significantly improves the quality of image reconstruction and robustness to noise.