The Segmentation Of Deformable Object Image Base On Active Shape Model And Image Invariant Features

Since the deformable objects exist broadly in reality life the processing of deformable object image is involved in many image processing and computer vision tasks, such as face recognition, medical image processing, in which the segmentation of deformable object image is a basic and key step. However, the traditional segmentation methods can not obtain the accurate results as the parts of deformable objects may be locally deformed. Therefore, it is necessary for effectively segmenting the deformable object images to take the character of deformable object image into account.The Active Shape Model is one of the most efficient methods, which is widely used for the segmentation of deformable object images. It is a kind of statistic model based on the prior knowledge of the training set. First of all, the landmarks of object shape are marked manually. The variance of shape is then extracted by iterative matching steps and a point distribution model is built up by employing Primary Component Analysis. The next is that the gray information around the landmarks is sampled by which a gray model is constructed. At last, the model can be used for segmentation and fitting of images. Since the Active Shape Model models the local change of deformable object, promised segmentation results can be obtained. However, there are still some problems that need to be solved, for example, the automatic initialization of the start position of model, the robustness of segment and the quantification of the segment results. In this paper, we aimed at the former two problems and proposed some effective methods.During the investigation, the authors gained the following achievements:1) In this paper, an automatic initialization strategy is developed by combining SIFT features. According to the accurate matching, the high-repeatable and high-robust SIFT features are selected from training set. The key point model is then built. The relationship model is next constructed in accordance with the relationship between the key point model and point distribution model. When the original position of model is initialized, the whole position of PDM is firstly located by KPM and then the precise position of each landmark is obtained by RM. By the proposed method, the initial position of PDM can be precisely located, by which the ultimate accuracy of segment is enhanced.2) During the development of the initializing strategy, a novel refining method of the matching result of SIFT features is also presented. By improving the original shape context, the affine invariant shape context is proposed firstly, by which an effective and accurate refining method is developed. This method can refine the matching result and eliminate the missing matching to a large extent.3) This paper proposed a novel distortion detection and amendment method which is used for detecting and correcting the distortion during the iterative segmentation processing of ASM. Through this method, the robustness and accuracy of result is largely increased.