Quality assessment for super resolution image via feature optimization

Objectives: Most of the existing no-referenced (NR) image quality assessment (IQA) methods need to extract lots of image features and then predict the image quality. However, only a small part of these features plays a role in IQA models while the other makes the model inefficient. Therefore, a NR IQA framework based on feature optimization is proposed and applied it to the field of super-resolution (SR) IQA task. Methods: The features associated with the SR images are first collected and aggregated. Secondly, a variety of advanced feature selection algorithms are used to sort the feature sets and get the importance matrix of features. Then the linear correlation between the number of features and Pearson linear correlation coefficient (PLCC) is studied according to the importance matrix of features to determine the optimal number of features and the optimal feature selection algorithm. Results: In this paper, the image quality scores predicted by the model based on the aggregation feature are highly consistent with the human subjective scores. Using the proposed feature optimization framework, we can effectively reduce the number of features and make the model performance better. Conclusion: In this paper, feature engineering is used to optimize the performance and efficiency of existing NR IQA algorithms. Experimental results show that the image quality can be accurately predicted by only a small part of image features.