Improving Transformer-based Image Compressed Sensing via Filtering and Fusion

Compressed sensing (CS) technology has a wide range of application prospects and research value in many fields, which will have a positive impact on the development of digital signal processing and communication technology. Considering the actual storage of the observation matrix, the observation of images is usually processed in blocks, which leads to the problem of block effect and discontinuous diffusion caused by zero filling operations. To improve the visual quality of the latest Transformer-based CS reconstruction, we propose a joint scheme based on Transformer reconstruction module, the filtering module constructed by the BM3D method and image fusion module. Given a reconstructed image with some singular pixels form the reconstruction module, nonlocal filtering is effective in removing irregular points by using self-similarity between image blocks, but it may lead to over-smoothed images. Then, fusion module is used to extract the advantages of the outputs of both modules. Experimental results show that the singular pixels can be significantly reduced while obtaining slightly higher objective evaluations.