Shape Based Image Contour Grouping And Object Detection Technic

Silhouette is a key factor for understanding the vision system. Usually, people can simply recognize an object from its silhouette. It makes the vision system much easier to understand objects based on the silhouette. The objects from one category could be very similar to each other in spite of the fact that they have different brightness, color, texture, etc. Recognition based on the silhouette largely reduces the intra-class variance for . The computer need much more training samples and computation when the brightness, color and texture information are involved. Moreover, objects from different categories may have similar brightness, color and texture information which produces a lot of trouble for the computer to understand them. As regard of the importance of the silhouette, our research concentrates on shape (silhouette) extraction, representation and recognition based on the shape information.First, it is very important to partition an image to regions of interests (ROI) and extract the contour of the objects. We propose a closed contour extraction algorithm based on particle filter which can efficiently extract the silhouettes of an object, and it is especially suitable for medical images. It has very comparable performance with Active Contour Models like Snake and Level Set, and enjoys the advantage of fast processing speed. In the field of medical imaging, the very popular algorithm Active Contour Models (for example: snakes and level set) faces the problem of high complexity, and some simple methods like watershed do not guarantee the precision of segmentation, thus our algorithm takes the advantage in the application that both the fast processing and precision are required.Then, given the silhouette of an object, i.e. a shape, it is another important issue of computer vision to understand and represent the shape. We propose a novel shape descriptor which has excellent property of invariant of scaling and rotation, it also has good capacity of handling the deformation of non-rigid objects, furthermore, compared with other peer methods, our algorithm enjoys the advantage of handling 3D and higher dimensional point matching problem.Finally, Recognition is the most critical technic in computer vision to recognize the objects from an image. This thesis proposes an edge-linking technic which can extract the sequential contour segments for the convenience of computer understanding. A Multiscale Random Field (MSRF) model is proposed subsequently and utilized to handle partial silhouette missing problem, we improve Relaxation Labeling algorithm to fit our optimization function. At last, a particle filter framework is proposed to integrate our shape descriptor with the edge linking technic, and applied to shape based object recognition to achieve the state of the art performance on ETHZ dataset.