Research On Methods Of Image Quality Comprehensive Assessment Based On Visual Characteristics

Images are vulnerable to noise pollution and exhibit distortion in the applications of sampling,compression,transmission,reconstruction,storage,etc.The studies of image quality assessment(IQA)is of great significance.Not only the accuracy and the monotonicity of the current objective IQA methods have a relatively low level in comparison with the subjective results of human visual system(HVS),but also the current indexes could not reflect the reliability and efficiency of various methods.To solve these problems,the dissertation integrates the principles and characteristics of HVS into IQA method.Objective methods based on visual perception for IQA are proposed.The performance of the proposed methods is analyzed by objective experiments.The main results are as the following:To overcome problem of the deficiency of the reliability and assessment efficiency in the existing objective IQA,an assessment efficiency method is proposed.Firstly,the reliability and time efficiency indexes are designed to comprehensively evaluate the performance of various methods.Secondly,evaluation criteria of correlation coefficient,error evaluation,reliability and time efficiency are unified by the normalized processing.Finally,an algorithm of assessment efficiency is proposed with pooling of objective IQA results for different distortion types,different distortion degree and different databases.The results indicate that the proposed algorithm is consistent with the results in the existing literatures.To improve the accuracy,the monotonicity and the stability of the pixel assessment methods,two novel objective methods on the visual perception of gray level are proposed.Integrating local Gaussian weighted algorithm modeling visual resolution and contrast information of images into the PSNR method,a structural peak signal to noise ratio(SPSNR)is proposed.The experimental results show that the SPSNR improves the assessment performance of the PSNR for different distortion types of images.The upper limit of the root mean square error(RMSE),the lower limit of Pearson Linear Correlation Coefficient(PLCC)and the lower limit of Spearman Rank Order Correlation Coefficient(SROCC)are improved by 25.55%,5.48% and 8.79%,respectively.By integrating local Gaussian weight and multi-scale sampling into information fidelity criterion(IFC),a multi-scale information fidelity criterion(MSIFC)method is proposed.The experimental results show that the assessment efficiency of the MSIFC is improved by 18.42% compared to the IFC.Furthermore,by integrating visual multi-channel characteristics into the pixel domain assessment,the performance of the PSNR and the SSIM on the above visual evaluation is discussed.Two objective methods named visual peak signal to noise ratio(VPSNR)and visual structural similarity(VSSIM)are proposed.Via the integration of Gaussian physiological photosensitive,spacial foveal and frequency contrast sensitive function(CSF),the paper presents a series of algorithms for local assessment,local and global pooling on the VPSNR and the VSSIM.The experimental results show that both the accuracy and the monotonicity of VPSNR and VSSIM for low quality images is increased by 10.85% in comparison with other similar methods.To reduce the gap between IQA based on image feature and subjective IQA,characteristics of feature assessment based on visual multi-channel are discussed.Three representative methods named visual gradient structure metric(VGSM),visual singular value decomposition(VSVD)and visual histogram(VHIST)are presented.By integration of Gaussian physiological photosensitive,spacial foveal and frequency contrast sensitive function(CSF),the paper presents a series of algorithms for local assessment,local and global pooling for the VGSM,the VSVD and the VHIST.The experimental results show that the presented visual feature assessment methods are superior to similar methods.The upper limit of the RMSE based on different distortion types for the VGSM is decreased by 10.00% in comparison with the GSM.The assessment efficiency for the VSVD is increased by 14.55% in comparison with the SVD.The PLCC and the SROCC for the VHIST on low quality images is increased by 72.56% in comparison with the similar method.To solve the problem that joint assessment methods based on image multi feature do not consider the principle of the HVS,four visual pooling methods are proposed.Firstly,the visual threshold iteratively adaptive(VTIA)algorithm is designed,and then an objective method named visual saliency adaptive pooling(VSAP)is proposed based on integrating the VTIA distortion assessment into the VSSIM similarity assessment.Secondly,based on the complementary of assessment algorithm and image visual features,three representative methods named visual gradient-structure pooled low-order-moment distribution(VGPLD),visual singular-values-energy pooled low-order-moments Distribution(VSPLD)and visual hist-statistics pooled gradient structure(VHPG)are proposed,respectively.Fast pooling algorithms for the above three methods are respectively designed based on regression function and experimental training.The experimental results show that assessment efficiency is improved by 14.50% for VSAP,VGPLD and VSPLD,7.35% for VHPG,compared to multi-feature assessment methods.