Residual Pixel Attention Network for Spectral Reconstruction From RGB Images

Hao Peng, Xiaomei Chen, Jie Zhao; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 486-487

In recent years, hyperspectral reconstruction based on RGB imaging has made significant progress of deep learning, which greatly improves the accuracy of the reconstructed hyperspectral images. In this paper, we proposed a convolution neural network of the hyperspectral reconstruction from a single RGB image, called Residual Pixel Attention Network (RPAN). Specifically, we proposed a Pixel Attention (PA) module, which was applied to each pixel of all feature maps, to adaptively rescale pixel-wise features in all feature maps. The RPAN was trained on the hyperspectral dataset provided by NTIRE 2020 Spectral Reconstruction Challenge and compared with previous state-of-the-art method HSCNN+. The results showed our RPAN network had achieved superior performance in terms of MRAE and RMSE.