Research On Image Quality Assessment Methods Via Machine Learning

Image quality assessment as a basic research in the field of image processing aims to design an objective evaluation algorithm that automatically evaluates image quality and achieves consistency with the subjective evaluation score.The images of the real scene inevitably suffer from distortion during imaging,transmission and processing.It is a challenging study to propose an image quality assessment method that can meet the needs of practical applications.The traditional assessment algorithms directly perform weighted pooling,so the importance of each feature cannot be reflected well.Based on this,some machine learning methods are continuously applied in the research of image quality assessment algorithms.This article mainly proposes two image quality assessment methods by combining the underlying image features and machine learning methods.(1)A random forest based spatial-frequency domain joint feature full reference color image quality evaluation method is proposed.Firstly,this method extracts the chroma and gradient features in the spatial domain,which are used to characterize the color information and spatial structure information of images.The texture detail information of the response of the log-Gabor filter bank and spatial frequency features are extracted in the frequency domain,which are used to be joint features.Then,random forest is implemented for learning the mapping relationship between the feature vector and the subjective opinion score to predict the objective quality score.Finally,experiments conducted on three standard databases,i.e.TID2008,TID2013,and CSIQ prove that the comprehensive evaluation performance by our method is better than the state-of-the-art full reference assessment algorithms,especially on TID2013 database,the Pearson linear correlation coefficient value can reach 0.9397.(2)A no-reference image quality assessment algorithm based on two machine learning methods is proposed.Firstly,this method extracts the local binary pattern on the gradient map.The gradient weighted histogram is used to describe the image structure and contrast information.Secondly,the natural scene statistics of an asymmetric generalized Gaussian distribution are extracted to simulate the sensitivity of the human visual system to the local laws of the image.Then,the random forest and support vector machine are used to establish the regression model.Finally,experiments are performed on natural images with four types of distortion(JPEG-2000,JPEG,WN,GB)on the TID2008 and CSIQ databases,and experiments are performed on all distorted images on the LIVE,CSIQ,TID2013,LIVEWC,CID2013 databases.Experimental results show that the method in this paper shows better evaluation performance on 5 databases,and has stronger competitiveness than other state-of-the-art no reference methods.Among them,the Pearson linear correlation coefficient on the LIVE database reaches 0.9596 and on the CID2013 database reaches 0.8471.