Learning Texture Enhancement Prior with Deep Unfolding Network for Snapshot Compressive Imaging

Abstract

AbstractCoded Aperture Snapshot Spectral Imaging (CASSI) utilizes a two-dimensional (2D) detector to capture three-dimensional (3D) data, significantly reducing the acquisition cost of hyperspectral images. However, such an ill-posed problem desires a reliable decoding algorithm with a well-designed prior term. This paper proposes a decoding model with a learnable prior term for snapshot compressive imaging. We expand the inference obtained by Half Quadratic Splitting (HQS) to construct our Texture Enhancement Prior learning network, TEP-net. Considering the high-frequency information representing the texture can effectively enhance the reconstruction quality. We then propose the residual Shuffled Multi-spectral Channel Attention (Shuffled-MCA) module to learn information corresponding to different frequency components by introducing the Discrete Cosine Transform (DCT) bases. In order to overcome the drawbacks of grouping operations within the MCA module efficiently, we employ the channel shuffle operation instead of a channel-wise operation. Channel shuffle rearranges the channel descriptors, allowing for better extraction of channel correlations subsequently. The experimental results show that our method outperforms the existing state-of-the-art method in numerical indicators. At the same time, the visualization results also show our superior performance in texture enhancement.KeywordsCoded Aperture Snapshot Spectral Imaging (CASSI)Deep unfoldingResidual shuffled multi-spectral channel attentionTexture enhancementChannel shuffle