Reduced-reference Perceptual Discrepancy Learning for Image Restoration Quality Assessment

Leida Li, Bo Hu, Yipo Huang, Hancheng Zhu

Abstract: Image restoration has been receiving extensive attention owing to its widespread applications, and numerous restoration algorithms have been proposed. However, how to accurately evaluate the performances of image restoration algorithms remains largely unexplored. Current image restoration quality metrics make predictions solely based on the restored images without making full use of the original degraded image, which we believe also provides valuable information. For image restoration quality assessment, accurate measurement of the perceptual discrepancy between the degraded image and the restored images is crucial. Motivated by this, this paper presents a perceptual discrepancy learning (PDL) framework for image restoration quality assessment, where the original degraded image is utilized as reduced-reference to achieve reliable predictions. First, a large-scale paired image quality database with weakly annotated labels is built, based on which a prior quality model is trained using Siamese network. Then, based on the prior model, the degraded-restored image pairs (DRIPs) are further used to train the perceptual discrepancy prediction model in an end-to-end manner. Finally, the performances of image restoration algorithms can be obtained based on the predicted relative perceptual discrepancy (RPD) values directly. Experimental results on four image restoration quality databases demonstrate the advantage of the proposed metric over the state-of-the-arts.