Dual-domain deep unfolding Transformer for spectral compressive imaging reconstruction

To solve the low spatial and/or temporal resolution problem which the conventional hyperspectral cameras often suffer from, spectral compressive imaging systems (SCI) have attracted more attention recently. By formulating the problem into a data fidelity module and a prior denoising module, the deep unfolding reconstruction method for SCI has demonstrate promising performance but still face problems: 1) In the prior denoising modules, feature reconstruction in the frequency domain is neglected. 2) In the unfolding multi-stage networks, there are still a lack of multi-level feature interaction and fusion between adjacent stages that can filter and learn the flow of beneficial information from the previous stage to the next stage. 3) The noise levels of prior denoising modules at different stages have not been effectively modeled. To solve these problems, in this paper, we proposed a dualdomain deep unfolding Transformer (DUT). Specifically, firstly, in the prior denoising module of DUT, a Dualdomain Self-Attention mechanism (DSA) is proposed to simultaneously reconstruct features in both image and frequency domains. Secondly, in the DUT multi-stage networks, a Gated-Dconv Stage Interaction mechanism (GDSI) is proposed to achieve effective multi-level feature interaction and fusion between adjacent stages, preventing the loss of beneficial features. This enables prior models in different stages to share parameters, substantially reducing the number of parameters in DUT. In addition, the proposed Noise Estimation block (NE) effectively models the noise levels of each stage, further improving the reconstruction accuracy. Extensive experiments show that our method significantly outperforms state-of-the-art methods on both simulated and real datasets, while requiring cheaper computational costs. Code and models are publicly available at: https://github. com/Vzhouhan/DUT.