Blind Image Quality Assessment for a Single Image From Text-to-Image Synthesis

Abstract

A fundamental bottleneck in text-to-image synthesis is that there are rarely subjective quality evaluation metrics for a single generated image. To address this issue, this paper proposed a procedure to evaluate the single generated image, which includes a specific dataset named multiple metrics quality assessment for birds(MMQA Birds) and a learning model named blind generated image evaluator(BGIE). The motivation of our proposal is twofold. On the one hand, subjective image quality evaluation is a human perceptual task; Therefore, it tends to be a process of supervised learning. To the best of our knowledge, there are not any datasets for this study. Thus, we handle this problem via designing a specific dataset. On the other hand, we observed that the spatial content of generated image attracts more attention when humans judge its quality; According to this finding, an efficient machine-learning model that combines both pixel-level features and spatial features is proposed. Extensive experiments manifest our method can solve this problem to some extent. In the generated image dataset, BGIE surpasses the state-of-art NSS-based method by 6.3% in PLCC and SRCC. In practice, we further discuss the rationality of the MMQA Birds dataset and the application of BGIE. It proves that both in subjective and objective aspects, our method achieves convincing results.

Keywords

• Generated image quality assessment
• generative adversarial networks
• image quality evaluation dataset