Adaptive Reconstruction for Block-based Image Compressed Sensing via A Texture-gated Network

In the existing deep learning-based methods for block-based image compressed sensing (CS) reconstruction, blocks are reconstructed non-adaptively through one network with fixed capacities, which ignores block textures and restricts reconstruction quality. Intuitively, blocks with smooth textures are easier to reconstruct than those with complex textures, which means blocks with different textures should be adaptively reconstructed through networks with different capacities. To address this problem, this paper proposes a texture-gated network for block-based image CS adaptive reconstruction, dubbed AdapBCS-Net. Specifically, a gate module based on block textures and a gate loss are developed to learn the activations of layers of the backbone network. As a result, backbone networks with specific activations are equivalent to networks with specific capacities, which can adaptively reconstruct blocks. Simulation results demonstrate that AdapBCS-Net effectively improves reconstruction quality.