Research On Blind Image Quality Evaluation Algorithms Based On Deep Learning

With the advent of the era of interconnection of all things,the quality of images plays a decisive role in the reliability and validity of images as information carriers.Therefore,the accurate evaluation of image quality is directly related to whether the image can be widely used as an information carrier.Most distorted images in practical applications have no corresponding distorted HD images.Therefore,the evaluation of blind image quality has more research significance.In the research of blind image quality evaluation,how to accurately extract features highly related to the degree of image distortion and how to accurately evaluate different types of distorted images by the algorithm has been a difficult problem.To solve these problems,this paper applies convolutional neural network in depth learning to blind image quality evaluation.In the training image quality blind evaluation regression network,a method based on local region confidence is proposed to make the network prediction results fit human subjective vision better.Based on the block expansion of database samples,the relationship between the statistical mean value of predicted mass fraction of different size image sub-blocks and subjective mass fraction of image is studied,and the optimal size of image sub-blocks is determined.The experimental results show that the mean of image sub-block mass fraction predicted by the statistical network as a method of image prediction quality can accurately fit the subjective perception of human eyes,and the network model in this paper has strong robustness and validity under different distortion types of images.In order to further fit the subjective consistency of human eyes,a prediction score calculation method based on the confidence of image block prediction fraction is proposed in this paper.This paper mainly counts the contrast of image sub-blocks and takes it as the confidence of sub-block prediction fraction.At the sametime,the confidence interval of sub-block prediction fraction is constructed according to the brightness value of image sub-blocks,and the sub-block prediction fraction with too high and too low brightness value is eliminated.Combining the sub-block quality prediction score obtained by the network model with the sub-block confidence,the predicted mass fraction which is more in line with the human eye observation model is calculated.The experimental results show that this method can improve the accuracy of the false prediction in most cases.