Abstract
Recent years have witnessed remarkable achievements in perceptual imagerestoration (IR), creating an urgent demand for accurate image qualityassessment (IQA), which is essential for both performance comparison andalgorithm optimization. Unfortunately, the existing IQA metrics exhibitinherent weakness for IR task, particularly when distinguishing fine-grainedquality differences among restored images. To address this dilemma, wecontribute the first-of-its-kind fine-grained image quality assessment datasetfor image restoration, termed FGRestore, comprising 18,408 restored imagesacross six common IR tasks. Beyond conventional scalar quality scores,FGRestore was also annotated with 30,886 fine-grained pairwise preferences.Based on FGRestore, a comprehensive benchmark was conducted on the existing IQAmetrics, which reveal significant inconsistencies between score-based IQAevaluations and the fine-grained restoration quality. Motivated by thesefindings, we further propose FGResQ, a new IQA model specifically designed forimage restoration, which features both coarse-grained score regression andfine-grained quality ranking. Extensive experiments and comparisons demonstratethat FGResQ significantly outperforms state-of-the-art IQA metrics. Codes andmodel weights have been released in https://pxf0429.github.io/FGResQ/