Unsupervised Hyperspectral Image Blind Super-Resolution via a Kernel Prior Network

Hyperspectral images typically suffer from low resolution due to inherent hardware constraints, complicating downstream tasks like detection, classification, and recognition. In our study, we introduce a kernel prior network (KPNet) designed to address the the hyperspectral image (HSI) blind super-resolution (BSR) challenge in an unsupervised fashion. Double-DIP, comprising two deep image priors (DIP), enables the simultaneous reconstruction of high-dimensional spectral data and kernel estimates via an untrained encoder-decoder network configuration in the parameter space. The KPNet creates a reversible connection between complex blur kernel distributions and simple hidden variable distributions. In addition, the proposed DIP can also estimate reliable blur kernels, which can be integrated into existing blind super-resolution techniques. Comprehensive experiments conducted on standard datasets demonstrate that our proposed approach markedly surpasses current techniques in performance.

I. Introduction

The challenge of upgrading hyperspectral images from low resolution to high resolution, referred to as hyperspectral image super-resolution, plays a crucial role in fields including detection [1], classification [2], and recognition. Hyperspectral images are three-dimensional datasets that capture the reflective properties of electromagnetic waves over hundreds of contiguous spectral bands with precise sensors [3], as shown in Fig. 1. These images consist of multiple grayscale layers, with each pixel signifying a spectral vector that indicates the reflection intensity at that position. The performance of hyperspectral image super-resolution faces challenges due to factors such as the extensive dimensionality of the data, inadequate image quality, and the difficulty in obtaining training samples [4].