Algorithm Research Of2D-To-3D Video Image

3D technology is widely used,3D video has broad prospects, and3D ecological environment is good, but the limited existing3D video resources can’t meet people’s consumption demand.2D to3D technology is a breakthrough in solving this problem.2D to3D technology can effectively convert a large number of existing2D video resources and alleviate the current shortage of3D video resources. Also, the entire conversion process can be done automatically by the system with faster conversion speed and lower conversion costs.In this thesis, the project work is carried out from the research and improvement of2D to3D key algorithm, which includes depth information generation algorithm coalesced with motion depth cue, virtual viewpoint synthesis algorithm based on asymmetric edge self-adaptive filtering, layered void repair algorithms, etc. And on this basis, the improved algorithm is proposed. The general idea is to deeply analyze and detailed study every critical step of the2D to3D system process. First, start from2D monocular image free of depth information, obtain the depth information through depth information generation algorithm; then, composite the preprocessed depth information with the corresponding original2D image into voids virtual viewpoint image; finally, obtain a binocular image by filling voids of void virtual viewpoint image.This paper realized the combination of2D video features and characteristics of human visual motion cues based on the depth information generation algorithm, including algorithm theory, implementation steps and so on. Finally, through the video sequence algorithm test results show that the algorithm based on the depth information campaign trail generation algorithm can generate better prospect depth information campaign, but the disadvantages are also obvious algorithm for static scenes information processing less effective.In this thesis, a bilateral filter based asymmetric edge self-adaptive filtering algorithm is improved and implemented. It can preserve the edge information while smoothing the depth image. The asymmetry is reflected in that the pixels within the scope of image filtering nuclear are divided into three regions in vertical, horizontal, and diagonal directions. Different region has different weight. Then the filtered depth image and2D image are virtual viewpoint synthesized, In this thesis, using the relative position relations of people eyes, the DIBR method is simplified, so that can the right eye diagram can be achieved easier.In this thesis, layered void repair algorithm is implemented. First, down-sampling the virtual viewpoint synthesized image （referred to as Ro） with voids, until there is no voids in the down-sampling image （referred to as RN）; then, obtain EN-1by up-sampling RN, and obtain image FN_by using the pixels in EN-1to fill in voids pixels in RN₁, repeat this process until getting the image Fo, which is the repaired image. The foreground edge of the resulting image after layered voids repairing exists fuzzy problem. After analyzing, the problem is due to that when filling voids, foreground and background information are equal viewed. Therefore, in this thesis, layered depth repair algorithm is improved, and the depth self-adaptive layered voids repair algorithm is proposed. Firstly, correct the color image according to the different depth information of objects in the image, then, down-sampling the color image, the remaining processes are the same with the layered void repair algorithm, at last, the repaired image with better effect can be obtained.