Dual-Branch Network for No-Reference Super-Resolution Image Quality Assessment

No-reference super-resolution image quality assessment (SR-IQA) has become an critical technique for optimizing SR algorithms, the key challenge is how to comprehensively learn visual related features of SR image. Existing methods ignore the context information and feature correlation. To tackle this problem, this letter proposes a dual-branch network for no-reference super-resolution image quality assessment (DBSRNet). First, dual-branch feature extraction module is designed, where residual network and receptive field block net are combined to learn multi-scale local features, stacked vision transformer blocks are utilized to learn global features. Then, correlations between dual-branch features are learned and fused based on self-attention mechanism structure, final predicted score is obtained by adaptive feature pooling strategy. Finally, experimental results show that DBSRNet significantly outperforms State-of-the-Art methods in terms of prediction accuracy on all SR-IQA datasets.