Deep Multi-scale Recursive Residual Attention Network for Spectral Super Resolution

Abstract

The Spectral super-resolution methods based on deep learning have achieved promising performance when the network models are well trained, most of which require sufficient training dataset. However, capturing enough paired color image and hyperspectral images is expensive or complex under existing conditions. Thus, the available training dataset is still small to some extent, which is prone to over-fitting as the network becomes deeper. To solve this problem, a novel method based on deep network is proposed for spectral super-resolution, which is named as multi-scale recursive residual attention network. Without increasing network parameters, the recursive learning can increase the depth and receptive field of network to strengthen nonlinear modeling capabilities. In addition, multi-scale residual attention block is proposed to exploit contextual information in different scale and adjust channel-wise features adaptively. Experiments on public hyperspectral datasets demonstrates that our method can achieve better reconstruction effects compared with other spectral reconstruction methods when there is small training dataset.