Compressive sensing networks based on attention mechanism reconfiguration

The combination of deep learning and compressive sensing has brought new breakthroughs in the field of image and video processing, but how to design compressive sensing networks with good generalization ability and low computational complexity is still a great challenge. In this paper, we propose a multiscale compressive sensing network reconstructed based on the attention mechanism, where training a single model allows sampling and reconstruction of arbitrary sampling ratios. Initially, in the sampling phase, we employ multi-scale adaptive sampling within the wavelet domain. This method dynamically adjusts the sampling ratios of various image blocks to accommodate the varying complexities of different regions through a multi-scale mechanism, thereby enhancing data utilization. Next, we construct a deep reconstruction module based on the pyramid model, which realizes adaptive feature enhancement at different resolutions by applying the attention mechanism at different scales. We jointly optimize the sampling network and the reconstruction network, and the model obtained by training this network is able to adapt to arbitrary sampling ratios. Testing results across different datasets demonstrate that our proposed compressive sensing reconstruction network exhibits rapid operational speed while ensuring the high quality of image reconstruction.