The Study On Evaluation Method Of Stereo Image Quality

Compared with the traditional two-dimensional(2D) images, three-dimensional(3D) images provide depth information, which can provide immersive viewing experience for viewers. It is a difficult problem to provide high-quality images to the viewers. However, current stereoscopic displays suffer from crosstalk that is detrimental to image quality, depth quality, and visual comfort. Though the existing crosstalk measurement methods have good correlation with the mean of subjective scores(MOS), they require using of both the original images and the crosstalk images. We have to synthesis crosstalk images before predicting crosstalk perception values. This process is very impractical. Therefore, in this paper, we propose a new method for predicting crosstalk perception values, which only use the original images, rather than both the original images and the ghosted images.The proposed metric(pdlcV) is based on the color information of the original stereoscopic images. Firstly, we extract disparity map, color difference map and color difference contrast map from original image pairs. Then we use those maps to construct two new metrics(dispcV and dlog cV). The metric dispcV considers the effect of disparity map and color difference map, while dlog cV addresses the influence of color difference contrast map. The prediction performace is measured on the testing data by Pearson correlation(PCor), Spearman correlation(SCor) and the root mean square error(RMSE) between the predicted values and the MOSs. Experimental results show that the new metrics, dispcV and dlog cV, achieve better evaluation performance than previous methods, which indicates that color information is one considerable factor to achieve high accuracy of crosstalk visible prediction. Besides, the metric pdlcV is made by combining dlog cV and dispcV together, to achieve a higher correlation with the perceived subject crosstalk scores.Furthermore, we construct a new dataset using various natural image scenes, which contains rich color information and depth structure. Then we evaluate the performance of the proposed methods using the real world images of our new dataset and LiyuanXing ‘s crosstalk stereoscopic dataset[15]. The results show that our new metric has better performce.