Construction and Application of Quality Assessment Model of No-reference Images Two-Stream Convolutional Neural Network

Abstract

With the advancement of information and multimedia technologies, screen content images have been widely applied in multiple fields. However, in image transmission across devices, image distortion may occur due to various reasons. To solve this problem, feature extraction is carried out for the image area of the text area of the image, and the features of different areas are fused to construct a no reference image quality assessment model with areal feature fusion. The transfer learning method is introduced to build a model with the Two-Stream convolutional neural network. The test findings indicate that the Spearman rank correlation coefficients of the models based on areal feature fusion and Two-Stream convolutional neural network are 0.5263 and 0.9242 respectively, the Kendal rank correlation coefficients are 0.5745 and 0.8059 respectively, the Pearson linear correlation coefficients are 0.3852 and 0.9284 respectively, and the root-mean-square erro is 1.0012 and 7.8523 respectively, which are superior to other equivalent algorithms. Moreover, the accuracy of individual RGB branch network testing is 93.7%, the accuracy of individual gradient branch network testing is 95.2%, and the accuracy of Two-Stream network testing is 96.1%. The results show that the two algorithms constructed are effective in evaluating no-reference images and have high generalization ability.