An Asymmetric Filter Based 3D Synthesis Algorithm with Depth Region Classification

To reduce geometric distortion in three dimensional (3D) synthesized images and high computation time in the depth map preprocessing, this paper proposes a region classification based asymmetric filter for depth map preprocessing. First, the depth map can be classified into three types of region: depth continuous region, sharp depth transition regions with, and without vertical lines. To maintain 3D experience and reduce computation time, the depth value in the depth continuous region is left unchanged, such as the background region. For the sharp depth transition regions, we determine the presence or lack of vertical lines. Next, an adaptive Horizontal Gaussian filter is used to reduce rubbersheet holes generated in the region which does not contain vertical lines. Additionally, an adaptive Gaussian filter based on the vertical line direction is used to reduce geometric artifacts and further reduce the computation time compared to other asymmetric Gaussian filters used in the region which has vertical lines. The proposed algorithm can improve the quality of the Depth-image-based rendering (DIBR) virtual views by using depth sensors. According to the experiment result, the computation time is reduced by 90-95% compared with the asymmetric Gaussian filter. Moreover, the vertical line comparison and the skewness comparison show that geometric distortion is reduced obviously and the 3D experience is preserved successfully.