Fast non-iterative blind restoration of hyperspectral images with PSFs

Hyperspectral imaging is the basis for many data analysis techniques. As more application scenarios demand the portability and snapshot imaging capabilities of hyperspectral imaging systems, snapshot spectral imaging system based on diffracted rotation has gained increasing attention and development. However, in this system, the traditional iterative optimization-based unrolled network architectures require the assistance of point spread functions (PSFs) in the reconstruction process and are costly in terms of time and computational re-sources. Aiming to improve the quality of the reconstruction results, reduce the consumption of computational resources and time, and broaden the spectral range, we firstly construct a visible to near-infrared hyperspectral (VNH) dataset. Then we propose a convolutional neural network (CNN) based blind restoration network for hyperspectral images. In this work, we take Unet as the initial framework and propose using the spectral upsampling block, the weight adaptive residual (WAR) block, and the hybrid loss function to enhance the network performance. Experiments show that our proposed method can effectively reconstruct hyperspectral images from rotational diffraction blurred images, consuming less computational resources while retaining a competitive spatial resolution and spectral precision.