Research On Depth Information Mining And Evaluation Algorithm Of Auto-stereo Image

Depth image is an effective 3D data representation method. The auto-stereo image processing system based on the depth image provides flexible adaptability on network and custom display, so it brings the new opportunity and challenge in the area of stereo image processing. In the system, depth image generation and evaluation are so important that they restrict the post-processing ability of the whole system. This paper carries out deep and detailed theoretical and experimental work on the depth image generation and evaluation of the auto-stereo image.Firstly,this paper proposes a new auto-stereo image processing system based on the depth information. It only generates the depth image flows for view 1 and N, and designs the scalable coding algorithm to transmit the specific video streams and depth information based on the network quality and custom feedback. In the custom display, it designs the virtual-view rendering algorithm to generate the virtual views based on the custom demands. Both the theoretical analyses and experimental results indicate that this system realizes the effective transmission of the auto-stereo images and is compatible with the 2D television system.Secondly, this paper proposes the depth information generation algorithm with the fractional-pixel precision. According to the specific characteristics of the left and right images, it proposes the half-phase DCT interpolation algorithm to estimate the fractional-pixel information. This paper designs experiments to compare the quality of disparity maps with the integer-pixel precision, half-pixel precision and quarter-pixel precision. It concludes that the disparity maps with half-pixel precision reflect more details of the whole image, so they have better quality than the others. Besides, it designs experiments to compare the half-phase DCT interpolation algorithm with the common interpolation algorithms. Experimental results indicate that the half-phase DCT interpolation algorithm is more suitable to be integrated into the stereo matching algorithm.Thirdly, this paper proposes the digital total variation model to post-process the initial disparity map. According to different characteristics of the left and right disparity maps, it designs different discrete matrices for both of them. The experimental results reflect that this post-processing algorithm eliminates the"black hole"phenomenon in the textureless areas, doesn't blur the edges, and improves the depth map quality.Finally, this paper newly proposes a depth information evaluation index based on Neyman-Pearson judgment criterion. This index is deduced from the Gaussian disparity distribution model. The experimental results reflect that this objective index can estimate the disparity map quality correctly.