PPD-IQA: A Bottom-Up and Top-Down Combined Approach for Blind Image Quality Assessment via Prototype-Prompted Disentangling

作者：Hui Wang, Zhongzhou Zhang, Zi-Yuan Yang, Guangcheng Wang, Yi Zhang · 发表于：IEEE Transactions on Emerging Topics in Computational Intelligence · 年份：2026 · DOI：10.1109/tetci.2025.3574579 · 被引用次数：1 · 研究领域：Computer Science

Blind Image Quality Assessment (BIQA) aims to predict the image quality automatically by mimicking the perception mechanism of the human vision system (HVS). As a perception task, BIQA should consider the synergy of bottom-up and top-down information. The bottom-up and top-down information are sensory information and task-relevant prior information, respectively. Specifically, we consider that quality perception derives from the synergy of bottom-up information obtained from vision sensory and the quality-shared prior as top-down information, which is learned by HVS during daily life. Recently, with the successful application of large pre-trained vision-language models like CLIP in prompting downstream tasks, a series of prompt-based BIQA methods have been proposed to obtain top-down information and synergize it with bottom-up information for the final quality-oriented features. However, directly deploying such models for BIQA is challenging since additional annotations and well-designed prompt templates are required, and a slight change in template wording could have a huge impact on performance. To avoid these challenges, we proposed a novel end-to-end method, dubbed PPD-IQA, which learns the top-down information from a proxy classification task without a pre-trained vision-language model by prototype learning. The learned top-down information is referred to as a group of quality prototypes, representing centers of samples with similar quality perception. Prompted by the qu...