Degradation learning for unsupervised hyperspectral image super-resolution based on generative adversarial network

Abstract

High-resolution is essential to achieve remarkable performance in many applications using hyperspectral images. However, the acquired hyperspectral images are often low-resolution due to the limitations of sensors. Recently, deep learning-based methods have been widely studied to address the super-resolution problem, and most super-resolution methods are learned with the simulated dataset. However, the pre-defined down-sampling function employed in the simulation is simple, thus the trained model often suffers from the poor generalization for real-world images. In this paper, we propose an unsupervised super-resolution approach that does not require the paired dataset. The method is conducted in two stages: unsupervised image generation and supervised image super-resolution. Specifically, the image generation model generates a low-resolution image through a generative adversarial network in an unsupervised manner. Then, the generated low-resolution images are used to train the super-resolution model with high-resolution images in a supervised manner. To explore the high correlation of hyperspectral image in spectral and spatial domain, group convolution, attention mechanism, and multi-level features are employed in the super-resolution model. Experiments on the public dataset CAVE and Harvard show that the proposed model provides outperforming super-resolution ability over the compared methods. Also, the results on images with unknown degradation show a promising generalization of the proposed model.