Hyperspectral image super-resolution based on recurrent dense fusion network

With the development of optical imaging technology, more and more imaging techniques are being used in the field of cultural relics conservation and analysis. Among these, hyperspectral imaging has received much attention for its nondestructive, non-contact and ‘image-spectrum integration’ characteristics. However, due to the physical limitation of the imaging sensor, the spatial resolution of existing hyperspectral images is far less than that of RGB three-channel high-definition digital cameras, which greatly hinders the application of hyperspectral digitization of cultural relics. To address this problem, this paper proposes a hyperspectral image superresolution method based on a recurrent dense fusion network. By incorporating a fusion network with a recurrent neural network and a dense spatial-spectral attention module, the spatial and spectral features are fused. Finally, the hyperspectral image with high spatial and spectral resolution is obtained by the reconstruction module. The experimental results show that the proposed method can better improve the spatial resolution while maintaining the spectral resolution with a low number of parameters, both in terms of subjective vision and objective metrics.