Blind Image Quality Assessment Based On Sparse Representation And Convolutional Neural Network

With the rapid development of multimedia technology and the rapid popularization of social networking tools,images as the source of visual information and the carrier of information communication are widely used.Image is inevitably distorted during acquisition,storage,transmission,accessing,and its quality directly affects the subjective feeling of people and the acquisition of information.So image quality assessment technology has become a hot research topic in the field of image processing.Objective evaluation of image quality means that the computer can automatically predict image quality by designing effective algorithms to imitate human visual perception.Among these methods,blind image quality assessment(BIQA)refers to the prediction of the perceived quality of arbitrary distortion type without pristine reference image.Most of the actual scenes are not available for reference images,and in addition to the kaleidoscope of reality image content,blind image quality assessment shows extremely high research significance and application value.This thesis focuses on three critical issues,i.e.image feature extraction model,quality prediction model and experimental verification,and goes deep into studying on blind image quality assessment technology from codebook learning and convolutional neural network,aming to improve the subjective and objective consistency.The main content of this dissertation is summarized as follows.The first part focus on BIQA via sparse representation which belongs to codebook learning.Firstly,we use salient local feature descriptors to build dictionary,effectively eliminating redundant information,enhancing expressive ability of the dictionary,so as to ensuring the accuracy of feature expression while reducing the size of dictionary.Secondly,using sparse representation coefficients to quantify distortion degree of the image,and constructing equal-energy sub-matrices from sparse representation coefficient matrix by singular value decomposition,in order to forming coefficient matrix group to enrich feature expression content.Then,using max-pooling and LI-norm depicts intensity and sparsity characteristics respectively.Finally,uniting these features combined with support vector regression to predict quality.Experiments on standard database of LIVE illustrate that our method can effectively enhance the expressive ability of image features,and improve the subjective and objective consistency with lower size dictionary.The second part concentrates on BIQA via convolutional neural network.Aiming at the problems of the feature representation susceptible to artificial combination,the feature extraction stage and the quality prediction stage not jointly optimized,and the decrease of stability and generalization ability dealing with small samples of high dimensional feature space in the upper part,we study the convolutional neural network based end-to-end model for BIQA.Not only using distortion type prediction task to assist quality prediction task,but also jointing learning local quality and local weights inside the network when traing.And uniting saliency as weight of image patch outside the network when testing,which making prediction more accurate.Experimental results shows that the establishment of a large-scale data set applied to convolutional neural network can contribute to improving stability and robustness dealing with small samples of high dimensional feature space,and achieving better subjective and objective consistency.What' more,the prediction accuracy of some scenes on CSIQ and TID2013 databases can go up with more than one percentage point.