Research on No-reference Image Quality Assessment Algorithm Based on Generative Adversarial Networks

Abstract

Currently, with the massive generation and transmission of digital images, especially the rapid development of the Internet and mobile Internet industries where images are widely used, the study of reference-free image quality assessment has attracted much attention in academia and practical applications, and is a popular research direction in the field of computer vision. In response to the low performance of existing reference-free image quality assessment algorithms in the face of real distorted images, a reference-free image quality assessment algorithm based on generative adversarial networks is proposed. Firstly, the generator structure is changed, the U-Net structure is improved, and the channel attention mechanism SeNet structure is introduced to update the feature map after down sampling. Secondly, a feature similarity measurement system is incorporated and a dual discriminator structure is used to discriminate multiple groups of images. The FPN structure is combined in the feature extractor to produce a multi-scale feature representation. Experiments are conducted on KonIQ-10k dataset and LIVEC dataset, and the experimental results show that the algorithm exhibits good prediction accuracy as well as good generalization performance in the face of real distorted images.