Semi-supervised Learning with Conditional GANs for Blind Generated Image Quality Assessment

Evaluating the quality of images generated by generative adversarial networks (GANs) is still an open problem. Metrics such as Inception Score(IS) and Fréchet Inception Distance (FID) are limited in evaluating a single image, making trouble for researchers’ results presentation and practical application. In this context, an end-to-end image quality assessment (IQA) neural network shows excellent promise for a single generated image quality evaluation. However, generated image datasets with quality labels are too rare to train an efficient model. To handle this problem, this paper proposes a semi-supervised learning strategy to evaluate the quality of a single generated image. Firstly, a conditional GAN (CGAN) is employed to produce large numbers of generated-image samples, while the input conditions are regarded as the quality label. Secondly, these samples are fed into an image quality regression neural network to train a raw quality assessment model. Finally, a small number of labeled samples are used to fine-tune the model. In the experiments, this paper utilizes FID to prove our method’s efficiency indirectly. The value of FID decreased by 3.32 on average after we removed 40% of low-quality images. It shows that our method can not only reasonably evaluate the result of the overall generated image but also accurately evaluate the single generated image.