A Metric for Evaluating Image Quality Difference Perception Ability in Blind Image Quality Assessment Models

Blind Image Quality Assessment (BIQA) is a challenging research area essential for preprocessing and optimizing visual tasks like semantic recognition and image restoration. Due to the high cost of subjective experiments and the complexity of IQA data, BIQA remains a small-sample learning problem. Transfer learning methods have been introduced to address this. With advancements from VGG to Swin Transformer, new semantic backbones are continuously proposed. This paper is interested in whether the backbones with improving performance in semantic recognition also enhance predictive accuracy in BIQA tasks. However, comparative experiments showed that different semantic backbones, using appropriate pipelines, exhibit minimal differences in PLCC and SRCC, making it hard to identify the superior model. To resolve this, we propose a novel model comparison method, IQDP, based on the model falsification method. IQDP experiments revealed that models with similar accuracy can differ significantly in perceiving image quality differences, which traditional PLCC and SRCC struggle to capture. Based on this, we further implemented a new metric, Image Quality Difference Perception Ability, to supplement the traditional PLCC and SRCC, providing an effective means of identifying superior models.