Study On Stereo Image Quality Evaluation Method Based On Reference

With the rapid development of Internet and multimedia technology,3D video andimage is increasingly walking into the life of people and gives people a strong visualfeast. But with the popularity of stereo videos and images, some problems appearconstantly. In order to balance the contradiction between the limited storage space andimage compression ratio, image quality assessment methods are taken into account. It isvery important for the whole system of the stereo video images to evaluate the qualityof the stereoscopic image accurately.Image quality evaluation method mainly divided into two categories: subjectiveevaluation and objective evaluation methods. The most commonly used subjectiveevaluation methods are Mean Opinion Score(MOS) and Difference Mean OpinionScore(DMOS). However, subjective methods have the problem of compleximplementation and poor stability. Objective methods have the advantages of stable,accurate, efficient results, and are highly consistent with the subjective evaluation, sothey are widely used. Representatives of the classic quality objective evaluationmethods are Peak Signal Noise Ratio (PSNR), Structural Similarity Index Method(SSIM) and the Gradient based Structural Similarity Index Method (GSSIM) Method.This thesis mainly focuses on the three-dimensional characteristic analysis, visualattention model establishment and the evaluation algorithms based on reference, andgives the in-depth study on them. A large number of relevant literatures in English andChinese were read and a lot of experiments were done. To enhance the precision ofPSNR evaluation method and improve the performance of the image quality evaluationmethod based on reference respectively, we propose an improved PSNR evaluationmethod of three-dimensional images in accordance with human visual perceptioncharacteristics and the image quality assessment method with reference based on visualattention model.The main content of this paper is as follows:(1) Sometimes the commonly used evaluation measure PSNR can not reflect thevisual perception characteristics of human eyes for three-dimensional reconstructionimages well, that is, the subjective and objective quality evaluations are not consistentin which the human subjective evaluation value is higher, while the objective evaluationis with low value. Therefore, based on the stereo parallax characteristics of thethree-dimensional image, this paper puts forward a method based on weighted averagefiltering, membership function and sliding window for combination optimal selection and gives a good solution, which makes the value of the improved PSNR of image canreflect the human eye perception well, and achieves consistency between the objectiveevaluation PSNR and the human visual perception. The improved PSNR method ismore in line with the human visual characteristics than the original PSNR method. Thusit can be more accurate and more convenient to evaluate the quality of the reconstructed3D image which brings more extensive application. Experimental results show that thevalue of evaluating three-dimensional image using the proposed PSNR method whichconforms to the human eye visual perception characteristics, can achieve the goal ofconsistency with subjective evaluation.(2)The key of image quality evaluation method is whether it can obtain accurateanalysis and research for human visual system to establish visual attention modelconforming to the human visual system at the true sense. In order to further enhance theprecise performance of the image quality assessment methods, and the consistency ofthe subjective and objective image quality evaluation, an image quality assessmentmethod with reference based on visual attention model is proposed in this paper. Theproposed algorithm firstly analyses the HVS characteristics, secondly establishes thehuman visual attention model based on salient model, finally combines the currentclassic PSNR, SSIM, and GSSIM methods to evaluate images. Experimental resultsshow that the evaluation results of the proposed algorithm is better than those of usingthe original algorithms. The proposed method can be widely used for many kinds ofimages, and it has strong commonality.In this paper, the algorithms are implemented by mixed programming of Clanguage and MATLAB. We use the images "Tsukuba" and "Venus" provided byUniversity of Tsukuba in Japan, image "corridor" provided by the University of Bonn inGermany, and image "Parkmeter" provided by University of Carnegie Mellon in theUnited States as stereo test images in our experiments. Image database (Release two)provided by the University of Texas at Austin and the Laboratory for Image and VideoEngineering are also used as test images. Comparison and analysis are given betweenthe algorithms of literatures and the proposed algorithms. The experimental results showthat the performance of the algorithms in this paper is better than that of most literatures,and it can evaluate objective image quality better.