The Theory And Application Of Structural Invariance In Pattern Recognition

Image Recognition is one of the most important part of Pattern Recognition, in which traditional Pattern Recognition theories are widely used. For example, the theory of moment invariants, with the properties that they are stable in linear transformation, work well in rigidity objects recognition. Meanwhile, the idea by combine Haar-like feature and AdaBoost algorithm became one of the most useful method in face recognition. Generally, facing different objects, the effects of various methods may be sky or ground, and the choice of feature is the key factor. From the viewpoint of feature selection, the traditional methods can be divided into two categories: a) project by entire data; b) project by local data. These two methods have advantages and disadvantages: the former is the statistical information of the object as a whole, which works well for a fixed shape and less noise object; and the latter, by focusing on local feature, has a strong robustness in recognition. By analysis, both of them have disadvantages. The first class method relies of the stability of the overall data, and the other method d is dependent on the characteristics of local differences of object. Thus, we propose a idea of recognizing by structural feature, and also build models for experiment.Specific work is as follows:(1) Propose a recognition method by structural feature. We divide objects to several parts before training, and then train each part by conventional method, then do the second training by combine each part, finally we find the classifier which can recognise by structure of objects. The part of first training of this paper,we respectively use two methods, Moment Invariants with Support Vector Machine algorithm and Haar-Like feature with AdaBoost algorithm. The second training is to combine the results of the first work.(2) Propose the method of automatically dividing region.