HyperMixNet: Hyperspectral Image Reconstruction with Deep Mixed Network from a Snapshot Measurement

Many hyperspectral imaging systems resort to computational photography technique for capturing spectral information of the dynamic world in recent decades of years. Therein, Coded aperture snapshot spectral imaging encodes the 3D hyperspectral image as a 2D compressive image (snapshot) and then employs an inverse optimization algorithm embedded in the imaging system to reconstruct the underlying HSI. This study proposes a novel HyperMixNet to reconstruct an underlying HSI from the single snapshot image. Specifically, to reduce the size of the reconstruction model for being handy embedded in the real imaging system, we integrate the MixConv block instead of the conventional convolutional layers in our proposed HyperMixNet, which can not only greatly decrease the network parameter amount but also learn multi-level context for more representative feature extraction. Simultaneously, we employ a mixed spatial and spectral convolutional module to effectively learn the spatial structure and spectral attribute for more robust HSI reconstruction. We further design a mixed loss function for network training, which incorporates not only spatial fidelity but also spectral fidelity aiming at recovering the hyperspectral signature with small spectral distortion. Experimental results on three benchmark HSI datasets validate that our proposed method outperforms the state-of-the-art methods in quantitative values, visual effect, and reconstruction model scale.