Visual Salience Measure Of Shape Parts And Shape Coding

Modern multimedia information processing requires that computers can stimulate the functions of the human brain, including mass image and video data retrieval, classification, recognition, edit, interaction and storage. The shape, served as the key feature in object recognition, plays an indispensible role in these applications. Thus, how to make the shape computing and storage accurately and efficiently becomes the focus of this dissertation.The diversity of object poses and the changes of viewpoints result in the myriad changes of shape appearances. Moreover, the human eyes’ regioselectivity and multi-scale characteristics can easily lead to the difference between the shape measure and the visual experiences. Not to mention that the shape acquisitions often introduce quantification and segmentation noise. These all brings great challenges of the shape computing and storage.To meet these challenges, the part-based and the contour-based shape representation theory and applications are investigated in this dissertation. The visual salience of shape parts are measured, the distortion measurement and the encoding structure of the opera-tional rate-distortion (ORD) shape coding framework are improved, and the dual distortion regularization strategy of this framework is presented. These proposals are applied to shape parsing, fingertip detection and hand gesture recognition, which lead to promising results. The detailed works of this dissertation are provided as follows.1) A visual salience measure of shape parts is proposed. Firstly, measuring the visual salience of shape parts is simplified into measuring the visual salience of triangles. Then, the triangle salience is modeled as a function with respect to the total length, length ratio and turning angle. Using the completeness, independence and visualization of these three factors, the final metric of the triangle salience are derived. Experiments show that this measure outperforms the existing ones, and can significantly improve the accuracy of shape parsing and hand gesture recognition.2) An accurate distortion measurement using the analytical method is proposed. First-ly, the contour point distortion is defined as the shortest distance from the parametric curve to that contour point. Then, the parametric equation is derived using the perpendicular rela- tionship between the shortest distance line segment and the tangent line, and this equation is resolved using the analytical method. In order to reduce the execution time, a hybrid admissible distortion checking algorithm is proposed. Experiments show that this measure can accurately measure the actual distortion with the computational complexity the same as the lowest among the existing measurements.3) Two arbitrary directional edge encoding structures are proposed. Firstly, the digital plane is partitioned into eight or sixteen sectors. Then, the approximate edge is represented by a sector number, a short component and a long component. Experiments show that these structures can save a host of polygon vertices, so they can improve the coding efficiency as well as make the approximate polygon more compact.4) A dual distortion regularization-based ORD shape coding method is proposed. An (?)2-norm distortion regularization term is added to the original ORD framework, in order to find a better tradeoff between the edge rate and the edge distortion. The range of regular-ization parameter is provided to avoid the boundary noise influence, and the final solution to this combinatorial optimization problem is obtained using a shortest path algorithm for a weighted directed acyclic graph. Experiments show that this method can not only reduce edge distortions without loss of coding efficiency, but also be robust against boundary noise and therefore applicable to fingertip detection and hand gesture recognition.The main contributions of this dissertation are twofold. Firstly, the visual salience measure of shape parts make it possible to compare the visual saliences among different shape parts. Secondly, the accurate distortion measurement using the analytical method, two arbitrary direction edge encoding structures and the dual distortion regularization strategy make the ORD shape coding results more accurate, compact and efficient. These all lay a good foundation for further shape-based object recognition.