SARANIQA: Self-Attention Restorative Adversarial Network for No-Reference Image Quality Assessment

Siyu Xue,Hong Hou,Kangle Hui

DOI: https://doi.org/10.1109/ccisp51026.2020.9273475

2020-01-01

Abstract:In this work, we propose a self-attention restorative adversarial net (SARAN), a GAN-based model for no-reference image quality assessment (NR-IQA). Consistent with the human visual system, we add the self-attention module and improve the model structure to enhance the intensity of the against learning, the generated reference image for evaluation also could be more reliable. SARAN is composed of three parts: a generative network, a discriminative network and a quality evaluation network. The generative network performs feature extraction and reconstruction of the input distored image patches, while the discriminative network distinguishes the reconstructed patches from the pristine distortion-free patches. In order to capture the long range dependence which is beneficial to enhance the key texture and structure features of an image, the self-attention module is added into network. After restoration, the perceived distance between the restored and distorted patches is defined as appreciation of restoration (AoR). In the quality evaluation network, features vectors are concatenated from the distorted and restored patches, AoR. Eventually, the quality score of an input image is estimated by weighted sum of the patches scores. Experimental results are carried out on the LIVE and TID2013 datasets, SARAN is superior to the mainstream methods in the no-reference image quality assessment and has good performance for different distortion types. Meanwhile, it is more consistent with the human visual system (HVS) for no-reference image quality assessment.