SFM-IQA: A Hierarchical Model for Image Quality Assessment via Optimization-Based Meta Learning

Training data-driven IQA models is trapped in large size datasets, and when assessing authentic images with unknown distortion types, insufficient labeled data can lead to the decrease in the prediction accuracy. Therefore, a model named SFM-IQA that uses gradient optimization is proposed for authentic image assessment scenarios. A meta learning framework in this model is developed to acquire the prior knowledge shared by different kinds of distortions, so as to adapt to unknown distortions. A Swin hierarchical extractor is used to extract global and local quality-aware features to characterize distortions. In order to make full use of features at different scales to enhance the feature representation, a multi-scale fusion module is designed. On LIVEC and KonIQ-10k, the proposed model achieved SROCC of 0.894 and 0.933 respectively. The result demonstrates that the proposed model outperforms other state-of-art IQA metrics.