SSCE-Net: a spatio-spectral co-enhancement network for compressive sensing reconstruction of hyperspectral images
TL;DR
Experimental results demonstrate that under various low-sampling-rate settings, the proposed SSCE-Net method achieves superior reconstruction performance compared to existing methods on datasets such as CAVE and KAIST, thereby validating its effectiveness and robustness in low-sampling-rate HSI reconstruction tasks.
Abstract
Hyperspectral images (HSI) contain rich spectral-spatial information and are of great value in remote sensing interpretation and detailed scene analysis; however, their high-dimensional nature poses significant challenges for storage and transmission. To address the challenge of high-fidelity reconstruction of hyperspectral images under low-sampling-rate conditions, we propose a Spatial-Spectral Co-enhancement Network (SSCE-Net). The proposed method is designed for hyperspectral image compressive sensing reconstruction. It employs an end-to-end learning framework consisting of a learnable compressive sampling encoder and a collaborative enhancement decoder. By combining local texture modeling with long-range dependency modeling, the proposed network achieves high-quality hyperspectral reconstruction at different sampling rates. Experimental results demonstrate that under various low-sampling-rate settings, our method achieves superior reconstruction performance compared to existing methods on datasets such as CAVE and KAIST. It achieves higher PSNR and lower SAM and RMSE, thereby validating its effectiveness and robustness in low-sampling-rate HSI reconstruction tasks.