Research On The Key Technology Of2D-to-3D Video Conversion

With the rapid development of computer image/video production and display technology,3D video is better than2D video compared with the feeling of the real three-dimensional that transferred to moviegoers. The expensive production costs of pure3D movie make2D-to-3D video conversion technology become the spotlight of movie producers. At present, the key step of2D-to-3D video conversion technology is how to extract the depth information of the original2D video images.The quality of the depth map directly affects the quality of the generated3D video.This paper aims to research on the depth extraction method based on linear perspective principle and the depth extraction method based on motion parallax principle.It’s committed to obtaining high quality of depth image.Several innovations are followed:1. According to the depth extraction method based on linear perspective principle, put forward a method that determines the line extraction threshold adaptively. The straight line extraction is more in line with the characteristics of the image itself; This paper also refines the applicable scene types.By this way,it can make the depth extraction method more practical, and also make ready for the depth assignment work. Present a priority depth assignment method based on the vanishing point and the intercept of the vanishing lines.The depth information obtained by this method is more in line with the real depth. The quality of the depth map is better.2. According to the depth extraction method based on motion parallax, put forward the Pyramid Lucas-Kanade optical flow method based on image segmentation to extract the depth information:Firstly, determine the number of layers by getting the maximum motion vector of the image. It can make up for lack of the lost of information caused by too many layers and the failure to meet the Lucas-Kanade optical flow caused by too few layers. Then, in each layer, use the segmentation information to remove the error pixels, making depth more accurate. Reduce the time complexity with almost unchanged experimental results by adaptive iteration. Lastly, deal with the depth map through the statistics of depth in each class to make the edge information more clear. This paper also puts forward a dual threshold image segmentation method based on image hue for the video of Kimono.