Studies On Subjective And Objective Consistency Of Image Quality Assessment

With the rapid development of digital image and communication technologies, image quality has become an important index to measure image processing systems like image acquisition, transmission, compression, de-noising and restoration, etc. Image quality assessment has also become one of the hot research fields of image engineering. It is of great importance to optimize system, evaluate algorithm and improve image quality. Image quality mainly includes two aspects:image fidelity and image intelligibility, and there are two categorizations of image quality assessment methods:subjective and objective ones. Unfortunately engineers in the field of image engineering are confused by the inconsistency of subjective and objective image quality assessment methods, and how to improve the consistency of subjective and objective methods has been raised to the agenda, as the12th Five Year pre-research projects of high resolution major project put forward an important task: raise the consistency of subjective and objective methods to more than95%. How to make the subjective image quality assessment results meet the objective results better, and how to improve the consistency of them, have become an urgent problem.In order to improve the consistency of subjective and objective image quality assessment methods, the model of the human visual system is well studied first, and different image quality assessment methods are proposed by combining with different models and different characteristics of human visual system.Image quality assessment method based on Prewitt magnitude is proposed in this article. Based on Multichannel effect of human visual system, HSV color space is introduced into image quality assessment. HSV color space encapsulates information about a color in terms that are more natural and intuitive to humans. Researches on visual psychology confirmed that edges and contours play decisive role in the process of human cognitive scene. Images are transformed into HSV color space first, Prewitt magnitude is then utilized to extract the edge and contour information which are more sensitive to human visual system, and regional mutual information is used to calculate the similarity of corresponding channels of reference image and test image. Thus image quality assessment method called PMRMI is obtained. Experimental results demonstrate that PMRMI is more constant with human visual system than PSNR and SSIM.Image quality assessment method based on Contourlet transform is proposed in this article. Since human visual system has multi-scale, multi-directional characteristics, and it has different sensitive degree to different spatial frequency information, the Contourlet transform is introduced into image quality assessment because it can get texture information and direction information of images. First, visual sensitivity coefficients of different sub-bands and different directions are obtained via Contourlet transform, the coefficients that cannot be perceived by human visual system are combined by just noticeable difference model, and regional mutual information is used to calculate the similarity of corresponding sub-bands and corresponding directions of reference image and test image. Finally, weighted by contrast sensitivity function model, image quality evaluation algorithm is obtained which is called CRSIM. Experimental results demonstrate that the proposed algorithm can achieve better consistency with the subjective evaluations.Multi-level image quality assessment method is proposed in this article. Researches on visual psychology confirmed that human visual system is more sensitive to image information in low frequency than high frequency. Inspired by this, gradient information of image is segmented into different levels. Since the coefficients of Riesz transform can reflect the overall spatial structure information of the image, the coefficients of different levels are obtained by Riesz transform. Regional mutual information is utilized to calculate the similarity of the corresponding levels of spatial structure information. The weighting factors of different levels are obtained by analyzing the performance of the proposed method on different image quality assessment databases. This method is called MLSIM (in the form of2LSIM and3LSIM).These proposed methods are tested on one or more public image quality assessment databases, and five commonly used metrics like SROCC, KROCC, CC, RMSE, MAE are utilized to evaluate the performance of the proposed methods and the results demonstrate that these methods are more constant with the subjective evaluations, and can reflect the characteristics of human visual system much better. At last, multispectral images after compressed are tested to verify the correctness of the methods proposed in this article.