Research On Machine Learning Based Image Quality Assessment

With the development of science and technology,the information era has come.As the most widely used information medium,the digital image is closely related to our daily life.The quality of an image represents its core value which determines the quality of image service to a great extent.However,in the processing procedures such as capture,storage,compression,transmission and display,inevitable loss of perceptual quality caused by the image system itself or external environmental factors always lowers the quality of viewing experience.In order to efficiently assess the image quality,a perfect Image Quality Assessment(IQA)system is highly in demand.Since the subjective IQA is not only unstable and expensive but also cannot be performed in real time system,researchers have spent much effort on objective IQA which is able to evaluate the image accurately,efficiently and in real time.Given the fact that most objective IQA algorithms are not highly correlated with human Mean Opinion Score(MOS),this paper proposed two new objective IQA algorithms based on machine learning.On the one hand,we proposed an image distortion based full reference IQA.A Discrete Cosine Transform(DCT)domain feature based K-Nearest Neighbor(KNN)classifier is designed to predict the distortion type of the image.Then a multiple linear regression based model,a fusion of PNSR,SSIM,JND and VIF,is generated to assess the image quality according to distortion type.In this way,it can make full use of advantages of every classic algorithm.Extensive experimentation on TID2008 and LIVE database shows that for each distortion separately as well as across distortion types,the proposed scheme provides a superior performance to PSNR,SSIM and JND.On the other hand,combined with the advanced achievements of deep learning,we proposed a Convolutional Neural Network(CNN)based no reference IQA.Compared with traditional machine learning based algorithms,the CNN can excavate the inherent law in the data through training.Independence of artificial image features shows the superiority of deep learning which may lead a trend of image quality assessment in the future.