Research On 2D To 3D Video Algorithms Based On Depth Map

Nowadays, 3D films are becoming more and more popular among people due to their higher stereoscopic realism over the conventional 2D ones. 3D has been regarded as the next revolution for entertainment, culture and consumer electronics industry. However, the development of 3D is constrained due to the lack of available stereoscopic 3D contents. The cost of making 3D films is still very high and it is also time consuming. Since there exists tremendous amount of media data in 2D format, the research on 2D-3D conversion can make it effective and very efficient. Accordingly, the 2D-3D conversion method is becoming more and more popular.Based on the above content, this thesis deeply researches the key algorithms on 2D to 3D conversion, which mainly includes the extraction of depth cues, the generation of depth map, Depth Image Based Rendering(DIBR) and so on.This thesis introduces the related theory of the 2D-3D conversion. Binocular depth cues contain binocular disparity, motion parallax, eye converge. While the relative size of object, shading, texture, color factors constitute the monocular depth cues. We can obtain the depth map by extracting depth cues. In addition, it introduces a variety of display technologies including naked display technology and visual aids display technology.A hybrid depth-generation algorithm based on motion information and linear perspective is analyzed. Moving objects can attract the most observers, so people have a shorter sensor distance from it. We define it as the foreground objects. Linear perspective can detect the vanishing point and vanishing line in order to solve the problem of depth map. Generally speaking, the vanishing point is the farthest point, we assign it 0 as the depth value. The other points are assigned the depth value according to the distance between the point and the vanishing point.What this thesis achieves is that the depth map generation algorithm based on geometric scene categorization and that an optimization to the depth map. Firstly, 12 kinds of global depth contour model are introduced. Secondly, the appropriate models are selected according to different images. Then, the depth value is assigned according to the depth assumption. In order to generate a better quality of 3D video, it is indispensable to optimize the depth map. Experimental results show that through the depth map optimization algorithm we can get the depth map with better quality.An improved DIBR algorithm is proposed. To generate virtual stereoscopic images from original 2D image and intermediate image, the holes filling and morphological filter will need more computing time and higher hardware conditions. The thesis proposes a new warping method based on the Non-Hole-Filling DIBR that combines hole-filling and pre-processing into one single step. Experimental results show that the image holes are reduced by this method and a better quality of the image is achieved.