CLIP-DQA V2: Exploring CLIP for Dehazed Image Quality Assessment From a Fragment-Level Perspective

Contrastive Language-Image Pretraining (CLIP) models have demonstrated strong performance in blind dehazed image quality assessment (DQA), yet their efficiency remains a concern. In this letter, we introduce CLIP-DQA V2, which explores CLIP for efficient blind DQA from a fragment-level perspective. To effectively map fragments sampled from dehazed images to quality scores, CLIP-DQA V2 integrates two key components: (1) multi-modal prompt learning, which jointly optimizes CLIP’s image and text encoders for better alignment between fragments and quality-related text descriptions, and (2) a semantic consistency loss that alleviates the semantic degradation caused by fragment sampling. Experiments on two widely used benchmark datasets show that CLIP-DQA V2 reduces computational cost by nearly 45% compared to previous methods, while delivering more accurate quality predictions.