Perceptual Salience Structure Extraction Based On Adaptive Tensor Voting

With the development of computer vision area, lots of new graphical image forms are gradually becoming the research objects of computer vision. At the same time, the requirements for processing outcome are also increasingly demanding. In such circumstances, several new algorithms for perceptual salience structure extraction are emerging, among which there is an outstanding algorithm, called Tensor Voting.In this dissertation, accompanied by the design and implementation scheme, we propose an improved adaptive Tensor Voting,which is based on the in-depth analysis and research on the Tensor Voting framework for perceptual salience structure extraction. We apply this framework to extract perceptual salience structure from 2-D point cloud, and the preferable outcome was obtained. The contributions are as follows:(1) Proposing Adaptive Tensor Voting Algorithms, which are combined with point cloud density and image texture, to improve and enrich the Tensor Voting theoretical framework.(2) Based on the in-depth analysis and research on point cloud density, we propose a denseness feature, which can reflect the degree of local aggregation. With this feature, the problem of averaging the overall density of point cloud, which always accompanies the traditional methods for extracting point cloud density, can be solved. Meanwhile, combining denseness with uniform blocking density and variance, we suggest an approach for estimating stochastic distribution and homogeneity of the point cloud, which fills in the gap of denseness research area.(3) Design and Implement an adaptive Tensor Voting which combined with point cloud density and stochastic estimation, while the existing algorithms are usually designed only for a special type of point cloud. This approach firstly implement a process called pre-discriminant for point cloud, which can eliminate the possibility of extracting pseudo-entity structure characteristics by distinguishing different types of point cloud, selecting those which have perceptual salience structure for extraction and excluding the others. This increases the accuracy of the Tensor Voting algorithm. Furthermore, through adaptively getting voting field scale parameter, we realize an adaptive Tensor Voting which improves the extraction accuracy.(4) Proposing an adaptive Tensor Voting combined with image texture. By adaptively combining scale parameter of voting field with extracted image texture spectrum, we realize an adaptive Tensor Voting which makes the voting process adapt itself with image texture spectrum. This improves the accuracy of Tensor Voting in Extracting curves from gray image.As an outstanding algorithm for computer vision, researches and improvement on Tensor Voting will also be inspiring for research of other visual structure extraction algorithm.