Study Of Rendering-Quality-Oriented Depth Map Compressed Sampling

The three dimension(3D) video gains growing intrests for its vivid reflection of the scene. However, the huge number of video data brings new challenge to data capturing and compressing. Thus an efficient video data sampling and compressing scheme is in need. “Texture plus Depth” is a common way to describe 3D videos, and depth compressing plays a vital role in 3D video system. With high sparcity, depth map can be precessed with Compressed Sampling(CS), which shows benign performance in sparse signals. In this paper, different CS methods and their applications are anaylized, as well as the features of depth maps and their influence on viewpoint rendering quality. Based on that, CS method on depth map is studied.The major work includes:First, the features and applications of existing CS methods are analyzed and compared, and properties and compressing methods of depth maps are analyzed. Based on that, the sparse representation and filtering methods of depth map are studied. Results show that multiscale geometric analysis is the sparsest representation of the edge information in depth map, and median filtering during depth reconstruction helps increase the rendering quality.Second, an adaptive block-based depth map CS scheme in accordance with edge information is proposed. For loss of edge information during compressing precess reduces the final rendering quality, in the proposed method, depth map blocks are firstly classified according to their variance, which indicates edge information, then sampled adaptively. Depth maps are reconstructed using smoothed projection Landweber method with contourlet as its sparse expression. Experiments show that the proposed method can increase the rendering quality in low sampling rates by protecting edge information in depth maps.Last, an adaptive depth sequence CS scheme combining with temporal correlation is proposed. Within the ME/MC framework, the key frame in a GOP is adaptively CS to yield benign reconstructing quality, and non-key frames’ differences are estimated with their sampled values to decide whether to carry out ME/MC or reconstruct independently. Experiment results demonstrate that the proposed schem lead to better rendering quality than direct ME/MC model, especially for the scene with fast movements of objects.The study on adaptive CS on depth map combines CS theory and depth map’s own feature. The proposed methods has lower the computing complexity in the encoding end and better robustness compared to traditional video compressing techniques.