Towards Generalized No-Reference Image Quality Assessment Via Ascend Acceleration Learning
Year: 2026
Type: conference-paper
Abstract: The swift advancement of Vision-Language Pretraining (VLP) models has offered new insights into No-Reference Image Quality Assessment (NR-IQA) by exploiting the auxiliary visual priors embedded within these models to generate perceptual quality scores that closely match human evaluations. However, recent works reveal the limitations of this paradigm. Human-crafted prompt templates usually fail to achieve optimal performance and cannot be optimized online. Moreover, overlooking image contextual information prevents further performance improvements. Enlightened by the prompt learning technology and Ascend acceleration, this paper explores a new Contrastive Language-Image Pre-training (CLIP)-based NR-IQA paradigm, denoted as Multi-Component Prompt Learning (MCPL)-IQA, which is efficiently implemented on Ascend platforms. Instead of optimizing towards a fixed categorical prompt template, the proposed MCPL-IQA is characterized by learning a descriptive aggregation prompt that accounts for both task-specific and image contextual information, with its training and inference fully accelerated by Ascend. Moreover, we propose a specially designed Vision-Prompt Interaction (VPI) module that captures useful information from the multi-level CLIP image features based on the learned prompt, leveraging Ascend-optimized operators to contribute to perceptual quality prediction. Thus, the proposed MCPL-IQA can be directly adapted to any IQA datasets without human-crafted prompts, enjoying flexibility, interpretability, and seamless Ascend acceleration support. Experimental results across diverse datasets demonstrate that the proposed method achieves consistent performance improvements over mainstream state-of-the-art methods. Powered by Ascend computing acceleration, the designed framework yields considerable efficiency gains during both model training and inference. In addition, systematic ablation analyses confirm the validity of each key component in our architectural design. To facilitate academic research and reproducibility, we will release the Ascend-accelerated version of our source code to the public.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225624632