Global Feature Learning Based No-Reference Image Quality Assessment for Super Resolution

Assessing the quality of super resolution (SR) image is one of the most significant issues for SR, existing no-reference quality assessment methods ignore contextual relationships among different regions of SR images. Therefore, this paper proposes a global feature learning based no-reference quality assessment method for SR image. First, stacked vision transformers are applied to extract the global features and to learn contextual relationships. Then, adaptive weight strategy is designed to pool the global features and to predict the quality score, which could further improve the prediction accuracy. Finally, experimental results indicate that the proposed method outperforms state-of-the-art methods in terms of prediction accuracy on benchmark datasets.