Deep Dual Internal Learning for Hyperspectral Image Super-Resolution

Yongqing Sun, Hong Liu, Qiong Chang, Xianhua Han

Abstract: Super-resolution (SR) of Hyperspectral images (HSI) has emerged as a prominent research focus. While deep learning has shown remarkable improvements over traditional methods for HSI SR, these approaches typically require large, paired datasets of low-resolution (LR) and high-resolution (HR) HSIs for training, which are costly and difficult to obtain. In this study, we introduce a novel deep dual internal learning framework for HSI SR that does not rely on any external training dataset. Our method generates pseudo LR-HR pairs from the observed LR HSI and its downsampled version. We then employ a supervised internal learning approach to train an SR model specific to the image. Additionally, to address cross-scale differences, we use the observed LR HSI as a single training sample in an unsupervised internal learning phase, further refining the SR model. Our experiments on three benchmark datasets show that this dual internal learning framework outperforms existing unsupervised methods and achieves performance comparable to state-of-the-art supervised SR techniques.

External IDs:dblp:conf/mmm/SunLCH25