Support Vector Machines (SVM) are a kind of novel machine learning methods which have become the hotspot of machine learning because of their excellent learning performance. They also have successful applications in many fields, such as: face detection, handwriting digit recognition, text auto-categorization, etc. But as a new technique, SVM also have many shortcomings that need to be researched, including: sensitive to noise, have a limitation in the scale of training set, the shortcomings of training methods, incremental learning, and the combination with the prior knowledge, etc. The applications in many fields are limited because of these problems. In this paper, some of above problems are probed into the application in image processing area. And the corresponding solutions are given.The work including:(1) SVM based auto detection and segmentation of left ventricle MRI image The aim of the segmentation of the left ventricle MRI image is to find out theinner and the outer contour of the left ventricle, while for that the structure of the heart is very complex, meantime the tag lines may also decay, the contrast of the object and the background will also descend, which still make the segmentation more difficult. In fact, now, most of the segmentation methods for the tagged left ventricle are alternating, the typical one of them is the Active Contour Model (Snake Model) and it's deformation, which need to give a initial contour of the probably position of the object. Although this method work sound, it must be given a proper initial contour, so it still need much labor. Different with them, the method I proposed in this paper is an auto segmentation procedure based on SVM, and it is a realizable and effective method.(2) The improvement of the SVM training procedureWhen we training an object recognition system, the image with object always contains a background. For the background has nothing to do with the object recognition, it should not be learned. The problem is that the learning algorithm can not distinguish the object from the background, so they often be learned equally. This will result in the confusion and error when there is a new image with the same object but different background. A method is used in this paper to decrease this sort of generalization error. Through changing the training set, but not changing the training algorithm, to improve the insensitivity of the algorithm to the unimportant dimensions.(3) The improvement of the incremental learning algorithmIn this paper, the shortcomings of the traditional incremental learning procedure are pointed out, and the n corresponding resolving methods are advanced. The idea is proposed that those increased date, which near the separating hyperplane, is significant for the forming of the new hyperplane, whenever these date are classed by the former hyperplane to Test Error Set Berr or Test Right Set Bok. When theincreased data are learned, the data, which belong to Bok but near by the hyperplane, play an important role in the changing of the new hyperplane. And the change focuses in the local area of the increased data. Based on the above two ideas, a improved incremental learning procedure is advanced and tested. |