Research On The Algorithm Of Image Quality Evaluation Without Reference Based On Information Entropy And Deep Learning

Objective image quality evaluation methods can be divided into three types: full reference,semi-reference and non-reference,among which the non-reference model has gained more and more attention because it does not need any reference information.Deep convolutional neural network has very powerful modeling and analysis ability,which can represent abstract and distinguishing features.In recent years,it has been widely used in the quality evaluation of non-reference images.When using convolutional neural network to evaluate image quality,because there is not enough training data with labels in the general experimental database,researchers tend to divide the distorted images in the database into blocks to expand the data set,which will result in the imprecise determination of the mass fraction of each block.In view of the above problems,this paper proposes a convolutional network model(Image quality assessment convolution neural network,IQA-CNN_weight)based on information entropy,considering the influence of information entropy on image quality.By analyzing the relationship among the distortion types,distortion levels and image quality in the LIVE data set,it is found that there is a strong correlation between image acuity and information entropy.Based on this,this paper proposes two targeted convolution network training methods for different types of images based on the average information entropy of original images.When the entropy of the whole image is higher than the average entropy of the original image,the blocks with higher entropy are given high weight.The information entropy of each block is calculated as the weight of block importance to represent the degree of its influence on the quality of distorted images,and the loss function of convolutional neural network is adjusted based on the weight.When the information entropy of the whole image is lower than the average information entropy of the original image,the blocks with lower information entropy are given high weight,and the blocks with larger deviation from the label value are given higher weight.The blocks with higher evaluation error are compensated to reduce the gap between the evaluation value and the real value.The cross-validation results on the LIVE dataset and TID2008 dataset show that the proposed model can evaluate the quality of the distorted images well,and the evaluation results are closer to human visual perception.