A Snake Model Analysis Based On Bayesian Theory And Applied Research

Active contour model is also called snakes. It was widely used in recent years as a tool for image segmentation. Active contour model (snakes) was first introduced by Kass in 1987, since then, it has been improved by many researchers. A quite successful model among them is that of Bayesian statistics-based shape models. The improvement of often in the two sides, one is to improve energy functions' styles, the other is to improve minimisation algorithms. The targets of these improved algorithms are to enhance matching precision, or to promote searching ability, or to accelerate searching speed. In this thesis, we will study an Affine-Invariant EigenSnake model (AIES), which improved the model's searching ability. Combined with a snake algorithm using growing phase analysis, we aimed to implement a new model with strong searching ability and high matching precision. On second thoughts, we studied geometrical shape expression and feature extraction methods based on snakes.In the beginning of this thesis, we firstly studied the principles of traditional active contour models, summarized characteristics of different active contour models. Then we placed emphasis on the theory and implement of Bayesian statistics-based shape model, and analyzed its details. Based on the energy definition of AIES, we mended the snake function and the greedy algorithm, which had been used in this model, to accelerate searching speed. In the coarse matching period (the growing phase), in order to accelerate searching speed, we only searched upside, underside, left, right and four diagonally positions according to the result of sobel edge detection. A growing phase analysis snake algorithm and the original computational solution were combined to make the snake growing algorithm much more simple, at the same time, do not reduce the model's matching ability.In the last part of this thesis, we studied shape expression and feature extraction methods based on snakes. Our target is to form a set of scheme from bottom edge detection of images to high level feature extraction. Firstly, we got a snake describing the object'scontour. Then we rebuilt the whole slick contour using cubic spline data interpolation based on the snake gained in the previous step. A signature is a 1-D functional representation of a boundary and may be generated in various ways. We adopted a simplest signature definition, which is to plot the distance from the center to the boundary as a function of angle. Towards human face feature extraction, we extracted signatures of the two eyes, mouth, and face boundary as local features, and extracted the center of the above four contours as global feature. The local and global features altogether formed the geometrical feature of human face. Our experiment showed this scheme was effective in human facial feature extraction.