The Models Of SVM And The Applications Of SVM To Image Segmentation

Support vector machine(SVM) is a new sort of recognizing technology. Based on the principle of structural risk minimization instead of the principle of experiential risk minimization, combining the techniques of statistical learning, machines learning and neural networks etc, support vector machines has good capability of generalization. Because of having self-contained theories and good experimental results, Support vector machines are coming researched by more and more researcher.Although, Support vector machines have some merits which discussed before, they have some shortcoming in practice, for examples: the bound of structural risk too relax, lacking adaptive capability in the process of learning, sensitive to outliers too, etc. For the seek of resolving the three problem, the following researchs are done by modifying the model of the present support vector machines:Firstly, the subsections linear learning model based on support vector machines is proposed, which is used to resolve the problem that the bound of structural risk of the present support vector machines is too relax. Based on the principle of structure risk minimization, which is both able to learning to data sets which are classed by inferior linear functions, and able to learning to data sets which are only classed by nonlinear functions with less structural risk. Secondly, a sort of support vector machine based on positive feedback is proposed, which is used to resolve the problem that the present support vector machines lack adaptive capability to data sets in the learning process. Data set are preprocessed using the model of L1-SVM, then, the support vector machine based on positive feedback is obtained by training the data set with adaptive kernel functions which are computed using the distance between support vectors. The sort of support vector machine has some adaptive capability. Experimental results showing, the sort of support vector machine has better capability of generalization than other models of statistical learning to some data set which distribute feature is not same in different sub-regions.Lastly, the support vector machine based on adjustive boundary is proposed, which is used to resolve the problem that the present support vector machines are sensitive to outliers too. The sort support vector machine imposes different penitentiary coefficient to samples with different training error. The sort support vector machine is robust by control the infection caused by outliers. Presently, the techniques of image processing are used in most domains. Especially, the techniques of image processing are used and developed in autocontrol, communication, scatheless test, resource reconnaissance, medical diagnosis, biological engineering very generally. An image processing system is based on image segmentation commonly. That the results of image segmentation be good or bad is very important to an image processing system commonly. We have not general image segmentation methods to segment complex image rightly. Which method is chosen, is based on the image features, experiences, and targets. Because of the performance of an image segmentation system depending on the features of images severely, the area of using of an image segmentation system is limited badly.Based on the applying research in the thesis, two image segmentation systems are obtained matching the need of road building and engineering building. The applying research are comprised by the following:Firstly, the method which be used to extract seed of image automatically is proposed by using the features of stone-earth images. Combining the method and region growing, an image segmentation system is obtained which is able to segment stone-earth images rightly. The system is compiled by using VC++. In the thesis, the system is also the base of the image segmentation system based on support vector machines. Secondly, support vector machines are used to segment complex images first one. Both the necessity using the methodology of statistical learning, and the feasib...