Deep Residual Attention Network for Spectral Image Super-Resolution

Zhan Shi, Chang Chen, Zhiwei Xiong, Dong Liu, Zheng-Jun Zha, Feng Wu; Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0-0


Spectral imaging sensors often suffer from low spatial resolution, as there exists an essential tradeoff between the spectral and spatial resolutions that can be simultaneously achieved, especially when the temporal resolution needs to be retained. In this paper, we propose a novel deep residual attention network for the spatial super-resolution (SR) of spectral images. The proposed method extends the classic residual network by 1) directly using the 3D low-resolution (LR) spectral image as input instead of upsampling the 2D bandwise images separately, and 2) integrating the channel attention mechanism into the residual network. These two operations fully exploit the correlations across both the spectral and spatial dimensions and greatly promote the performance of spectral image SR. In addition, for the scenario when stereo pairs of LR spectral and high-resolution (HR) RGB measurements are available, we design a fusion framework based on the proposed network. The spatial resolution of the spectral input is enhanced in one branch, while the spectral resolution of the RGB input is enhanced in the other. These two branches are then fused together through the attention mechanism again to reconstruct the final HR spectral image, which achieves further improvement compared to using the single LR spectral input. Experimental results demonstrate the superiority of the proposed method over plain residual networks, and our method is one of the winning solutions in the PIRM 2018 Spectral Super-resolution Challenge.

Related Material

author = {Shi, Zhan and Chen, Chang and Xiong, Zhiwei and Liu, Dong and Zha, Zheng-Jun and Wu, Feng},
title = {Deep Residual Attention Network for Spectral Image Super-Resolution},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV) Workshops},
month = {September},
year = {2018}