Font Size: a A A

Shape Recognition Algorithm Based On Fusion Of Global And Local Properties

Posted on:2017-03-31Degree:MasterType:Thesis
Country:ChinaCandidate:X GuoFull Text:PDF
GTID:2308330482995699Subject:Software engineering
Abstract/Summary:PDF Full Text Request
As the basic research topic of computer vision, shape recognition has a very wide range of applications,for example, target recognition, character recognition, medical image analysis, robot navigation, etc. The difficulty of shape recognition is that the shape from the same class, even after the geometric transformation(rotation, translation and scale transformation) or nonlinear deformation(noise, distortion, occlusion, etc.),can still be easily identified by the human eyes, but it’s a challenge for the computer. Therefore, a good shape recognition algorithm has become the key means to accurately identify the shape for the computer. Such an algorithm should be able to ignore the deformation difference from the same type of shape, but also can distinguish different types of shape. A complete shape recognition process consists of two steps. Firstly, to describe the shape feature, a model which can effectively describe the shape is established. This model which represents the shape will be involved in computing for the algorithm. Secondly, the shape matching algorithm, when compared to the two shapes, is how to describe the differences between them. The shape recognition algorithm is divided into the global algorithm and local algorithm according to the different ways of shape feature model. Although the classical global shape recognition algorithm is efficient, it is not good enough to deal with the nonlinear deformation without enough local details. The local shape recognition algorithm has a good retrieval rate, but the discriminability is still to be improved and will produce error matching with the noise. So people try to merge the global and local features together to make up for each other’s deficiencies.In this paper, we mainly study some classical recognition algorithms based on global and local shape features, and propose a novel feature-point classification method. Based on this, we further classify these points, taking different shape recognition algorithms for different kinds of points. The matching results are fused to make full use of the advantages of global and local shape feature descriptors. The main contents of this paper are as follows:(1) Summarize a comprehensive review for the current research in the field of shape recognition, make a brief description of research status and describe the characteristics of various types of algorithms, then introduce Blurred Shape Model based on global features and shape context algorithm base on local features. And analyze the deficiencies of the above two methods, laying the foundation for the next step of the study.(2) In this paper, we add a matching training session in the preprocessing stage of the traditional shape recognition algorithm, training the class-feature points of each shape in the database. These class-feature points form the skeleton of the shape, which represents the unique characteristics distinguished from other categories. Then in the retrieval phase, according to the matching results between the retrieved shapes and the shapes of the database, the feature points are further classified into three categories: matching class-feature points, non-matching points and deformation feature points. Through the above two steps, we complete the classification strategy.(3) The matching class-feature points will be used as the positive samples, and the non-matching points will be used as the negative samples to participate in the global feature information. Deformation feature points, which are used as local feature information, are involved in the fusion algorithm because of the local information and the deformation details.(4) In our experiments, Blurred Shape Model and shape context algorithm are applied to the two kinds of feature information, and the results are fused to generate the final fusion cost. This cost will represent the similarity of the two shapes, the smaller the value, the more similar and vice versa. The experiments show that the proposed framework can effectively combine different algorithms to achieve the shape recognition and get better results.
Keywords/Search Tags:Shape recognition, Blurred Shape Model, Shape Context, Feature Fusion
PDF Full Text Request
Related items