Abstract
No-reference image quality assessment (NR-IQA) quantifies image distortion. It plays an important role in computer vision. Distorted images vary greatly in content. Many existing methods tend to fuse content information with quality prediction. However, they often overlook human visual perception. To address this issue, we propose an attention-guided content-aware NR-IQA method. It combines meta-learning with image content understanding. The approach uses refined deep semantic features for quality evaluation. First, we train a meta-model on a baseline network. This improves sensitivity to diverse distortions. Second, we insert an attention module into the meta-model. This captures global information and highlights important regions. We also fuse multi-level semantic features. This enables a comprehensive description of both local and global distortions. Finally, we reduce feature dimensions and learn weights to predict the quality score. Extensive experiments show that our method achieves results closer to human perception. It effectively focuses on regions of interest during feature extraction.