MULOR-Net: Multi-Branch Low-Rank Network for Spectral Compressive Imaging

The coded aperture snapshot spectral imager (CASSI) facilitates efficient hyperspectral image (HSI) acquisition owing to its low bandwidth requirement. The accurate reconstruction of the original HSI from its compressive measurement is crucial but consistently challenging. Over the years, self-supervised reconstruction based on deep image prior (DIP) has garnered wide popularity due to its ease of use and the absence of requiring training data. However, its performance is inferior to that of supervised reconstruction because of overfitting. To address this problem, this paper proposes a Multi-Branch Low-Rank Network (MULOR-Net) to break the performance bottleneck of DIP. Specifically, MULOR-Net employs a multi-branch architecture to learn different modalities of HSI, with each branch decomposed into a spatial abundance tensor and a spectral signature matrix learned by a spatial decoder and a spectral decoder, respectively. The proposed MULOR-Net is then incorporated into the alternating direction method of multipliers, and a double-stage training strategy is utilized to accelerate the algorithm. Extensive experiments on simulated and real datasets validate that the proposed method achieves the state-of-the-art performance with high accuracy, interpretability, and robustness against additive measurement noise. Notably, the proposed method achieves an average PSNR of 41.46 dB on the KAIST dataset, which is 1.31 dB higher than that of the second-best method.