Abstract: Abstract: To address the issues of overfitting and generalization in No-reference Image Quality assessment （NR-IQA） models based on convolutional neural networks under limited datasets and complex distortion conditions， this paper utilizes meta-learning to acquire shared experiential knowledge across different tasks， thereby enhancing the model’s generalization to unknown tasks. Using DenseNet as the backbone network to extract image features， we achieve comprehensive deep supervision， improving the information flow and gradients within the network， and reducing overfitting on small sample training tasks. Additionally， a multi-head self-attention mechanism is incorporated， allowing the network to capture diverse feature information and long-range dependencies of the global image from different subspaces， enhancing the model’s learning capability. A bi-level gradient optimization method from the support set to the query set is employed to train the quality prior model on various known distortion tasks， optimizing the subsequent gradient descent process of the model parameters. Fine-tuning is performed on the target NR-IQA task， where the model can quickly adapt to unknown distortion tasks under appropriately initialized parameters. Performance and generalization tests were conducted on the authentically distorted IQA dataset LIVEC and the synthetically distorted IQA dataset KADID-10K， where the SROCC values reached 0.834 and 0.831， respectively. The results indicate that the proposed model has better learning ability and generalization compared to traditional algorithms.

Key words:
Key words: convolutional neural networks, no-reference image quality assessment, meta-learning, DenseNet, attention mechanism