Boosting One-Shot Spectral Super-Resolution Using Transfer Learning

Though deep learning based spectral super-resolution (SSR) methods have state-of-the-art performances, most previous deep spectral super-resolution approaches require extensive paired RGB images and hyperspectral images (HSIs) for well-fitting learning. However, in real cases, the cost of generating such paired images is too prohibitive to collect sufficient training samples. To solve this problem, we investigated one-shot SSR in a target domain. To avoid over-fitting, we introduced knowledge from a source domain to guide the one-shot SSR in the target domain and use the idea of spectral unmixing to remove the interference of different spectral characteristics, with which we proposed a spectral-unmixing inspired deep SSR framework. Experimental results on three benchmark SSR datasets showed the effectiveness of the proposed method.