Improving No-Reference Image Quality Assessment with Limited Data Via a Semi-Supervised Regression GAN

Image quality assessment (IQA) is often considered a small-sample problem in the field of computer vision due to the limited size of available quality evaluation datasets, which has become a major bottleneck restricting the development of IQA methods. To address this issue, this paper proposes a Semi-supervised Regression Generative Adversarial Network for no-reference Image Quality Assessment (SRGAN-IQA). The proposed SRGAN-IQA integrates generative adversarial networks with regression training to form an end-to-end learning framework. By comparing the features of authentically captured (real-world) and distorted images, the discriminator can more effectively learn the quality degradation characteristics of distorted images. SRGAN-IQA exhibits strong generalization ability: even when trained on only half of the labeled images in a dataset, it achieves competitive and consistent cross-validation results across multiple standard IQA datasets. During training, a large number of unlabeled samples are simultaneously fed into the discriminator, guiding it to align the feature distributions between labeled and unlabeled images. This semi-supervised adversarial learning process enables the model to extract robust quality-aware representations even under limited annotation. Moreover, the model delivers promising evaluation results on both artificially and authentically distorted datasets. Two hypotheses are proposed and verified through contrast experiments, demonstrating that the method can effectively reduce reliance on labeled data, thereby offering significant potential for industrial applications.