The major task of the video coding and video segmentation technologies for multimedia applications is how to improve the coding and segmentation efficiency. However, because of the limitation of CPU (Central Processing Unit) and other reasons, most of video coding and segmentation technologies can not satisfy the real-time demanding of multimedia applications. This thesis is to solve these problems by using GPU (Graphical Processing Unit). But GPU only supports integer computation and the precision of pixel shader is very low, this thesis tries to solve this difficulty, and several new algorithms are proposed by combining GPU and CPU. According to the properties of GPU, our algorithms are directed by two bases: 1) Floating operations are replaced by integer operations, or integer operations on CPU are transferred to GPU; 2) CPU is processed pixel values in image processing, while GPU is processed the whole image surface. The main contributions of the thesis are as following.: The first part proposes a motion estimation algorithm based on the Variable Block-Size Quad-tree segmentation algorithm. The algorithm first obtains different regions by analyzing the object character and image edge character, and then gets motion vectors of different regions. Experimental results show the proposed algorithm not only agrees on the human visual system, but also the whole running speed and quality of the predicted image are better than those of full search algorithm and Variable Block-Size Quad-tree segmentation algorithm. The second part will discuss how to complete Discrete Cosine Transform (DCT) by GPU. Because the operation precision of GPU is very low, we improve traditional DCT algorithm and propose an Inter Discrete Cosine Transform algorithm by lifting scheming and rounding operation. First we introduce the theory fundament of integer DCT, and then propose one dimension DCT and two dimensions DCT algorithms. The third part is to propose a video segmentation algorithm combining GPU and CPU. Firstly, we introduce how to select appropriate color space, and then discuss the time-domain segmentation algorithm based on object motion estimation. In the end, we propose a spatial-domain algorithm based on Markov Random Field and Graph model. Experimental results show the proposed technology not only has the small error probability but also has high coding efficiency.
修改评论