Quality Assessment of Synthetic Images (2025)

Quality assessment of synthetic images via spatial distortion recognition

Abstract　Generative adversarial networks (GANs) trained on limited data often produce images with spatial distortions. In this study, we propose an efficient method for evaluating the quality of GAN-generated images by training a quality assessment model using artificially distorted images created through random spatial transformations of real images. Experimental results demonstrate that the proposed method outperforms existing no-reference image quality assessment methods and produces evaluation scores that are more consistent with human judgments.

Authors　Tomoya Sawada, Marie Katsurai, Masashi Okubo

Publication Venue　Machine Vision and Applications

Overview of the Proposed Method