Research And Implementation Of Image Identification Method Based On Active Shape Models

The research on Image recognition technology is trying to possess computers of intelligence which is similar to the humans, as well as the abilities to discriminate or extract meaningful goals and the attributes and features of objects in the images in the presence of noise, clutter and occlusion. Studies on image recognition technology, which have been widely applied in the area of medical, industrial, electronic commerce, communication and military and other fields, are of great significance. Image recognition methods can be roughly divided into three types: the first type is methods completely based on the information of image texture, and the second type includes methods based on models of the given target objects, and the third type consists of methods called hybrid which combinations of the above two type of methods.Methods completely based on the information of image texture, are designed depending on the specific features of the pixels in the image(such as the intensities of the pixels, the histogram, the gradient and texture characteristic of the image, etc.). Such methods usually produce results inconsistent in some degree with intuitive visual sense of the humans. Premise of methods based on models is the prior shape models which are obtained from the training set in accordance with the interested objects. They are effective methods to extract targets presenting in fixed shapes in images.However, the appearance of some objects in the real life is often changed. For example, the shapes of body tissues and organs of different individuals, or of the same one but in different period of time vary.In this case, the traditional rigid models show up serious limitations.Though many methods based on flexible models and deformable templates have been put forward, but most of those methods exist a common problem, that they sacrifice the specification of the models to accommodate shape changes, so there is a lack of robustness in image presentation. The Active Shape Models(ASM) overcomes this drawback by capturing natural changes of shapes of a specific category of objects, and being used to search in the presence of a particular structure of the image at the same time. It is a robust method to identify the shapes of deformable objects. But the traditional ASM search performs blindly iterative caculation of the local texture features of all the image pixels to match the control points, this is a serious time consuming process. To overcome the shortages of the above two methods of image recognition, scholars put forward the third type method of combination.Inspired by the idea of combination method, we combined an edge detection method which is based on the information of image texture, with ASM which is based on the models of the target objects, to design a new image search strategy,namely Boundary Neighborhood based Multi-Resolution Framework Active Shape Model, BN-MRASM for short, so as to gain both the robustness and efficiency through out the algorithm. First of all, on the basis of the traditional ASM with Multi Resolution Framework algorithm, we combine image edge detection in purpose of improving the efficiency of image search, and then expand the boundary pixels to a boundary neighborhood to enhance the robustness of the algorithm.Eventually we propose Multiresolution Active Shape Model method based on Boundary Neighborhood, BN- MRASM. This method is mainly divided into three stages: First, the training of model. In this stage, we should evaluate the mean shape of the object and modes of variation of the points of the shape from the sample training set, namely the object’s point distribution model(PDM), and figure out the local texture features of the landmarks of the models in the training set. Second, searching initialization. In this stage, we will construct the multi-resolution framework structure of a new coming image to be search, and then apply arbitrary edge detection method on images of each level of resolution, furthermore expand each boundary obtained from the last step to narrow belts. Third, image search. Utilize PDM and the boundary belt to perform multi-resolution ASM search in the image.We show experimence of three ways of image search of traditional ASM, MRASM based on multi-resolution framework, and BN – MRASM running on the same data set. Results show that the new method has an obviously higher execution efficiency than the classical method when searching in high resolution images, without significant effect on search acuracy.Main point of works in this paper is as follows:(1) The main types of image recognition methods and several representative identification algorithms are summarized.(2) Briefly describes the classical method based on image information, k-means clustering and its improved algorithm; The main ideas of the ASM algorithm is introduced and its advantages are put forward. Detailed implementation of its training and application are described, and image multiresolution framework is presented.(3) The shortage of the traditional ASM algorithm is pointed out and the efficiency enhanced combination method based on edge detection algorithm is proposed.(4) Experimented on traditional ASM algorithm,MRASM and the proposed algorithm,and their results are shown. In the last we analyzed all these methods on aspects of acuracy of the searching result and computational costs.