Boosting No-Reference Image Quality Assessment Using Handcrafted LBP Features

Aiming at the problem that the existing no-reference image quality assessment methods are mainly based on deep neural network for feature learning, and fail to effectively use the hand-crafted features containing domain expertise, a no-reference image quality assessment model based on hand-crafted local binary pattern (LBP) feature enhancement was proposed. The performance of the no-reference image quality assessment model was improved by introducing the LBP feature. Firstly, the corresponding LBP map was extracted from the input RGB image. Secondly, the dual-branch network was used to learn the feature representation of RGB image and LBP image respectively, and the deep features related to perception were extracted. Then, the dual-branch feature representation was adaptively integrated through the weighted feature fusion module. Finally, the fused features were fed into the regression module to predict the quality score of the image. The experimental results on six public datasets show that the overall performance of the proposed method is better than that of the comparison methods, and the model has good generalization ability. Compared with the baseline model without introducing LBP features, the performance of the model after introducing the manual feature strategy has been greatly improved, especially on the LIVE, CSIQ and BID datasets, compared with the baseline model, the SRCC and PLCC indicators are improved by about 5% to 8%, which verifies the effectiveness of the strategy of improving the performance of the model through manual features.