LWNAN: Lightweight Nonlocal Attention Network for Pharmaceutical Snapshot Spectral Computational Imaging

Pharmaceutical products exhibit different imaging effects at different wavelengths of light, which can guide pharmaceutical researchers to choose suitable raw materials for the development and production of new drugs. We propose a light-weight non-local prior approach to address the issues of poor spectral imaging quality and slow speed in specific scenarios such as medicine, in existing Coded Aperture Snapshot Spectral Imager (CASSI). We adopt a Light-Weight Non-local Attention Network (LWNAN) as a plugged-in denoiser, which can achieve high-speed and high-quality spectral image reconstruction on private hyperspectral medicinal dataset. The non-local attention module is designed with light-weight architecture, which reduces the number of FLOating-Point operations (FLOPs) by 60% and helps to quickly deploy imaging applications in different scenarios. In addition, this method is not limited by imaging models and can be applied to different spectral imaging systems. This paper selects two mainstream spectral imaging models for demonstration and comparison, and the imaging effect is 2.34 dB higher in PSNR and 4.5% better in SSIM than the existing SOTA level. We also provide a hyperspectral Chinese herbal medicine dataset.