Supervise-Assisted Self-Supervised Deep-Learning Method for Hyperspectral Image Restoration

Hyperspectral image (HSI) restoration is a chal-lenging research area, covering a variety of inverse problems.Previous works have shown the great success of deep learningin HSI restoration. However, facing the problem of distributiongaps between training HSIs and target HSI, those data-drivenmethods falter in delivering satisfactory outcomes for the targetHSIs. In addition, the degradation process of HSIs is usuallydisturbed by noise, which is not well taken into account inexisting restoration methods. The existence of noise furtherexacerbates the dissimilarities within the data, rendering itchallenging to attain desirable results without an appropri-ate learning approach. To track these issues, in this article,we propose a supervise-assisted self-supervised deep-learningmethod to restore noisy degraded HSIs. Initially, we facilitatethe restoration network to acquire a generalized prior throughsupervised learning from extensive training datasets. Then, theself-supervised learning stage is employed and utilizes the specificprior of the target HSI. Particularly, to restore clean HSIs duringthe self-supervised learning stage from noisy degraded HSIs,we introduce a noise-adaptive loss function that leverages innerstatistics of noisy degraded HSIs for restoration. The proposednoise-adaptive loss consists of Stein’s unbiased risk estimator(SURE) and total variation (TV) regularizer and fine-tunes thenetwork with the presence of noise. We demonstrate throughexperiments on different HSI tasks, including denoising, compres-sive sensing, super-resolution, and inpainting, that our methodoutperforms state-of-the-art methods on benchmarks underquantitative metrics and visual quality. The code is availableat https://github.com/ying-fu/SSDL-HSI.