FDANet: A frequency-domain attention network for super-resolution image quality assessment

Abstract

No-reference super-resolution image quality assessment (NR-SRIQA) is still a challenging task due to complicated degradation factors and the lack of reference high-resolution (HR) images. In this paper, we propose a novel NR-SRIQA framework that elaborately designs a frequency domain attention lightweight network called FDANet to predict the quality of SR images. Firstly the frequency domain attention module is used to extract the frequency domain salient map from the divided image patches, and then a frequency domain attention lightweight network is used as the backbone to learn the features in SR images to improve the performance of the network. Compared with the spatial domain attention module, the proposed frequency domain attention module is conducive to learn quality-aware features to quantify the quality of SR images. By integrating the deep learning-based features, the proposed NR-SRIQA model facilitates more compelling quality prediction but requires obviously less computational cost in contrast to compared deep networks for NR-SRIQA. Experimental results verify the superior performance of our method on two benchmark data-bases in terms of both quantitative evaluations and thorough analysis on the model deployment.