Task Oriented Image Quality Assessment for Synthesized Images

Ke Xu, Qian Zhang, Fei Yang, Zhanghao Jiang, Boon-Giin Lee

Abstract: In this study, we propose a new general learning-based framework, named Task-Oriented Image Quality Assessment, for evaluating the performance of Reference-guided image synthesis (RIS) tasks. Our framework uniquely employs both source and target images to construct content- and style-encoded feature embeddings, and then evaluates the quality of the synthesized images by comparing their feature distances to those of the source and target images. We designed a two-branch network that embeds both content and style elements simultaneously. Our training process uses a style-level interpolation strategy to generate intermediate styled images for training, eliminating the need for human annotations. The quality score is calculated using a ratio-based distance that considers both the synthesized image from the source image and to the target image. Our method was evaluated using the HIDER dataset and RESIDE dataset, which provide subject scores for each image. The obtained result shows the efficiency of our method.