Screen Image Quality Evaluation Model Based On Artificial Feature Fusion

With the rapid development of network technology,images displayed on consumer data terminals are usually not only natural images(NI),but composite images containing various computer-generated components such as natural images,text,and tables.An image generated by computer rendering is called Screen Content Image(SCI).According to many years of research,screen images have completely different image characteristics compared to natural images.The reason for this result is that screen images have a more complex structure than natural images.It is difficult to accurately predict the quality of screen images using traditional image quality evaluation models,because traditional image quality evaluation models are usually designed for natural images and do not achieve excellent results when predicting the quality of screen images with more complex structures.Therefore,it is of great significance to design a screen image quality evaluation model with excellent performance.In this paper,based on the different structural features of the different regions in the screen image,a convolutional neural network model(CNN-MAF)for screen image quality evaluation based on the fusion of multiple artificial features is proposed.The model first divides the screen image into two parts based on the index map based on the image activity,namely the text area and the graphics area;two convolutional neural networks are used to separate the two areas For quality prediction,the two CNNs have a similar architecture,but different artificial features are added for fusion;after obtaining the prediction scores of the two parts respectively,an adaptive weighting strategy based on image activity is used to obtain the prediction of the screen image fraction.Experimental results on two mainstream screen image databases show CNN-MAF has excellent performance and is more in line with the subjective evaluation results than most image quality evaluation models.The model innovations presented in this article are reflected in:(1)Use the image activity as the basis for distinguishing the screen image difference areas,cut the image according to the index map generated by the image activity,and cut the screen image area into several image blocks of the same size,the size is suitable for the CNN model learning.(2)The designed end-to-end CNN model takes the image block as input,and the prediction score of the image block as output,and uses two CNNs to train and predict the two parts of the image block.(3)The two CNNs add different features extracted from the human body respectively,and integrate with deep learning features to improve the prediction effect.(4)According to the image activity,design an adaptive weighting method,weighted fusion of the quality map to obtain the final prediction score.