No-Reference Image Quality Assessment Algorithm Based on Semi-Supervised Learning

This paper proposes a no-reference image quality evaluation algorithm based on semi-supervised learning and dual-branch network training to realize self-supervised learning in image quality evaluation. Specifically, it is a training process with two branches in which a small number of hand-labeled data samples are used for supervised learning in one branch. Self-supervised learning is performed in the other branch to assist the former in training the same feature extractor; the self-supervised learning part adopts several traditional full-reference methods to jointly label the training samples with soft labels. Extensive experiments are conducted on six public image databases. The results show that the proposed algorithm outperforms most current methods on the synthetic distorted image datasets and has a good generalization performance on the real distorted image datasets. The predicted results of the proposed algorithm are consistent with human subjective perception performance.