Visual-Language Model Knowledge Distillation Method for Image Quality Assessment
This work addresses efficiency and accuracy issues in IQA for computer vision applications, but it is incremental as it builds on existing CLIP-based methods.
The study tackled the problem of excessive parameter burden and insufficient local distortion identification in CLIP for Image Quality Assessment (IQA) by proposing a visual-language model knowledge distillation method, which significantly reduces model complexity and outperforms existing IQA methods on multiple datasets.
Image Quality Assessment (IQA) is a core task in computer vision. Multimodal methods based on vision-language models, such as CLIP, have demonstrated exceptional generalization capabilities in IQA tasks. To address the issues of excessive parameter burden and insufficient ability to identify local distorted features in CLIP for IQA, this study proposes a visual-language model knowledge distillation method aimed at guiding the training of models with architectural advantages using CLIP's IQA knowledge. First, quality-graded prompt templates were designed to guide CLIP to output quality scores. Then, CLIP is fine-tuned to enhance its capabilities in IQA tasks. Finally, a modality-adaptive knowledge distillation strategy is proposed to achieve guidance from the CLIP teacher model to the student model. Our experiments were conducted on multiple IQA datasets, and the results show that the proposed method significantly reduces model complexity while outperforming existing IQA methods, demonstrating strong potential for practical deployment.