Research On Depth Generation Methods For2D-3D Conversion

With the development of signal processing and display technologies, the first black and white TV has developed into color TV, and the standard format color TV has developed into high-definition format. Now, three dimensional television (3DTV) has become the new hot spot of research and applications for it can bring people immersive visual experience.Any3DTV system’s success depends largely on whether there is plenty of interesting3D content. Enough3D content can make people willing to buy a3D display or3DTV system services, especially in the early stage of3DTV popularity. However, recording a large number of new3D videos will cost a lot and take some time, which limits the development and popularization of3DTV. In this situation, converting existing2D videos to3D videos can enrich the3D content. Besides, it also allows people to watch classic videos in3D format.Depth map generation is a key issue of2D-3D conversion. Depth map generation methods can be divided into manual, automatic and semi-automatic three categories. Manual methods can generate depth maps with high accuracy, but consume a lot of time and manpower cost. The accuracy of the automatic methods is not high, but can be applied to real-time conversion systems. Semi-automatic methods can combine the advantages of the manual and automatic methods.Currently, there is not a depth map generation method can be applied to all scenes. Moreover, poor3D experiences caused by inaccurate depth maps may bring viewers eye strains and headaches. How to generate high-quality depth maps is the key issue in2D-3D conversion need to be addressed. Depth map generation methods can be not in real time, can be for a particular scene or involve manual interactive operations.This paper analyzes the basic principles of stereoscopic vision of the human eyes, including monocular and binocular stereo vision. On this basis, depth map generation methods of single images are researched and a depth generation method based on linear perspective and motion is proposed. The specific content is summarized as follows:We research and simulate a depth generation map based on edge information: images are segmented into different regions based on edge information; a depth value is assigned to each region according to the prior hypothesis; finally, depth maps are filtered by bilateral filter. For focused images:focused regions are extracted based on higher order statistics and based on wavelet transform, respectively; then, the extracted focused regions are took as foreground and remained defocused regions as background and different depth assignment methods are used for foreground and background; after depth fusion, the final depth map can be obtained.A depth generation method based on linear perspective and motion is proposed. First, the static scene background is reconstructed. Then, the coarse depth map of the background is extracted by linear perspective; stationary objects in the background are segmented; after assigning depth values to the segmented stationary objects and depth fusion, the depth map of background is obtained. After that, moving objects in each frame are segmented. Finally, depth values are assigned to moving objects; the depth maps of moving objects and background are integrated into one depth map.In sum, classifying videos into categories according to depth cues or scene content and designing specific depth map generation method for each type, we can get depth maps with higher quality. Such, the generated stereo videos will bring people more comfortable visual enjoyment.