Low-level Features Based Image Quality Assessment Methods

With the rapid development of information and communication technology,end-users propose higher requirements for high-quality experience.Image Quality Assessment(IQA)aims to design an objective IQA model,which is consistent with Human Visual System(HVS)perception.Full Reference(FR)is the basis of other IQA method research,which evaluates the quality by comparing the differences or similarities between the reference image and the distorted image.In recent years,No Reference(NR)IQA has attracted more and more attention,the purpose is to develop a quantitative method for automatically and accurately estimating the perceived image quality without any prior information of the reference image.In view of this,we propose the FR and NR IQA methods respectively,which are based on extracting low-level features of images and combining machine learning methods.The main works of this paper are as follows:(1)A color Full Reference(FR)IQA method is proposed.First,four different types of low-level feature maps are extracted,Structural Contrast Index(SCI),gradient,Local Binary Pattern(LBP),and chroma,which are used to characterize different feature attributes of the image.Second,each features are processed separately by using different feature pooling strategies,and a set of similar feature vectors are formed as the detector of image quality.Then use the Extreme Learning Machine(ELM)to establish a regression model and map the feature vector into an objective quality score.Finally,extensive experiments performed on five benchmark IQA databases demonstrate that this method can effectively improve the accuracy of IQA index,and the evaluation results are competition than state-of-the-art FR IQA metrics.(2)A novel no-reference IQA is proposed,namely Two Low-level Feature Distributions(TLLFD)based method for NR IQA.Different from the deep learning method,TLLFD evaluates image quality by extracting two types of complementary low-level feature distributions and combining shallow machine learning as a regression model.Thus it has few parameters,simple model,high efficiency.Firstly,the sign feature and amplitude feature are calculated by generalized LBP,respectively,and the two feature weight histogram distribution coefficients are used to describe the texture change of the distorted image.Secondly,we use weibull distribution to fit the parameters of gradient map to represent the structural change of the distorted images.Furthermore,support vector regression is adopted to model the complex nonlinear relationship between feature space and quality measure.Finally,numerical tests are performed on LIVE,CISQ,MICT and TID2008 standard databases for five different distortion categories JPEG2000(JP2K),JPEG,White Noise(WN),Gaussian Blur(GB)and Fast Fading(FF).The experimental results indicate that TLLFD method achieves superior evaluation performance and generalization ability for image quality prediction as compared to state-of-the-art full-reference,no-reference and even deep learning IQA methods.