Research Of Image Contour Extraction Method Based On SNAKES Model

Active contour model, is a hot topic in the research of computer vision. It brings higher-level knowledges into the low level tasks. The method is a challenge to Marr's computational vision theory, where the vision system is divided into three bottom-up levels with information transfering in a single direction. Active contour model can be used in a series of vision tasks such as contour extraction, stereo matching, motion tracking, 3D reconstruction and so on. Snakes model, that's parametrical active contour model, is a kind of active contour model. This dissertation focuses on the object contour extraction by Snakes model.First, specific deduction of the numerical implementation of traditional snakes model is given. Then the according flow chart is presented. Combined with the simulations and experiments, the limitations of the traditional snakes model are analysed.GVF snakes model, makes an external force field of lager capture range over traditional snakes model. But the computation of GVF is expensive. In this dissertation, in terms of solving two PDEs, finite differece method together with the SOR method is implemented to compute GVF with higher converging speed. Using B splines to describe contours, the number of control points is reduced, so that the computation complexity is decreased. GVF B Snakes model is implemented by combing GVF Snakes with B Snakes, and it retains the advantages of both models, which is demonstrated by experiments. A new conjugate gradient method is applied in the snake deforming process. The new method shows higher speed of convergence in the experiments. An energy term, which is a constraint of smoothness for curves, is involved in the usual B snakes model to improve the capacity of resisting disturbance. And the flow chart of GVF B Snakes model is presented.In the end, the image acquisition system is set up and the software named Snakes is given. The experiments show that the Snakes model can be used to extract the object contours from real images acquired by the system effectively.