Hyperspectral Image Compression Sensing Network With CNN–Transformer Mixture Architectures

The tasks of storing and transmitting hyperspectral images (HSIs) face challenges due to the extensive number of present bands. The recovery of HSIs at extremely low compression ratios represents a significant challenge in hyperspectral applications. To address these challenges, we propose a CNN-transformer mixture (CTM) architecture-based compressed sensing network (CTCSN). Its core feature enhancement backbone network incorporates a multilayer CTM module and is integrated with a spatial-spectral superresolution (SR) decoder module; this network aims to surmount the challenges posed by traditional compressed sensing (CS) technology in terms of reconstructing HSIs. Furthermore, to satisfy the receptive field requirements of strip data acquired from spaceborne push-broom hyperspectral sensors, we employ nonsquare convolutional layers and CTM-like blocks to devise a lightweight encoder. Overall, our approach, which is grounded in deep learning (DL), achieves significantly improved reconstruction quality and efficiency by merging the local modeling capabilities of a convolutional neural network (CNN) with the global information processing prowess of a transformer. Experimental results demonstrate that, compared to the recently proposed deep CS network (DCSN) and bit-level tensor data compression network (BTC-Net) (32-bit) methods, our method achieves an average SAM improvement of approximately 10% at a sampling rate of 1%, confirming its superiority in the field of CS for HSIs. The code is available at
https://github.com/ANIMZLS/CTCSN
.