《基于svm的图像分割e》由会员分享，可在线阅读，更多相关《基于svm的图像分割e（3页珍藏版）》请在金锄头文库上搜索。
1、Medical image segmentation based on threshold SVM Chen Xiao-juan Information Engineering College NorthEast Dianli University(NEDU) JiLin ，JiLin，China cxj_ Li Dan Information Engineering College NorthEast Dianli University(NEDU) JiLin ，JiLin，China AbstractMedical image segmentation is a basic problem
2、 in medical image processing field and the key to the problem from processing to analyzing.Medical image segmentation based on SVM needs the category attribute of image training sample set in order to achieve the goal of image segmentation by machine learning.The method of obtaining sample set manua
3、lly involves heavy workload,what is worse,the accuracy entirely depends on the experience of operators.Therefore this paper proposes a second-order segmentation method based on threshold SVM.The experiment shows that the new method is feasible and its performance is nice,in which the hung is segment
4、ed from the chest X-ray film. Keywords- medical image segmentation,support vector machine, threshold algorithm I. INTRODUCTION Image segmentation is a basic problem in image processing field and the key to the procedure from processing to analyzing.Extracting object contour of medical images,obtaini
5、ng specific edge information can help doctors understand diseases more visually and play an important role especially in 3-d reconstruction for human vision. According to the particularity of medical images,segmentation methods based on region or edge are hard to achieve satisfactory segmentation re
6、sult whereas those based on learning can obtain relatively better segmentation results. Traditional learning classification methods are based on empirical risk minimization rather than expectation risk minimization, so there are serious defects in generalization1- 2.In 1995, Vapnik proposed a new ma
7、chine learning method which has the potential-Support Vector Machine3.SVM transforms the sample data to a high dimension feature space in which obtains the optimal hyperplane that minimizes the expected risk between the classes. When SVM is used in medical image segmentation,first it needs the categ
8、ory attribute of image training sample set,then obtains the category attribute of other samples by machine learning,furthermore achieves the goal of image segmentation.The method of obtaining training sample set manually involves heavy workload,what is worse,the accuracy entirely depends on the expe
9、rience of operators.Therefore this paper proposes a new method that segments approximately with threshold method at the beginning to get training sample set and then precisely and automatically with SVM.The experiment shows that the new method is feasible and its performance is nice,in which the lun
10、g is segmented from the chest X-ray film. II. ALGORITHM A. Threshold Segmentation Algorithm Threshold segmentation method includes global and local threshold method.Global threshold method is applied frequently in image processing,which uses fixed threshold in the whole image to segment it.Gray hist
11、ogram is the processing object in classic threshold selecting.Objects and background of medical images can be supposed in different grayscales and polluted by zero-mean Gaussion noise,so gray scale distribution curve of the image is considered formed by the superposition of two normal distribution f
12、unction22()21( )2x f xe =.For such images,in whose histogram two separate peaks will appear,the gray value of trough between two peaks can be selected as the segmentation threshold.This paper applies threshold method only to segment images approximately,so this threshold will be used as baseline thr
13、eshold baseand on this baseline threshold basis plus or minus the offset of the same sized (based+,based),in order to segment the source image approximately to background,object and mid-point set.Background and object point sets are training sample sets of SVM and the mid-point set is testing sample
14、 set which will be segmented precisely then. B. SVM Theory SVM is essentially a linear learning machine.For the input training sample set ( ,),1, 1, 1n iix yin xRy= +?, (1) the classification hyperplane equation is let to be ()0xb+=, (2) 978-1-4244-5316-0/10/$26.00 2010 IEEEthus the classification m
15、argin is 2 /.To maximize the margin,that is to minimize ,the optimal hyperplane problem is transformed to quadratic programming problem as follows, 1min ( )( ,)2 . . ()1,1,2,ist yxbil = +=?(3) Afer introduction of lagrange multiplier ,the dual problem is given by, 11111( )()2s.t. 001,2,maxnnniijijij iijniii iQy yK xxyin = =?，(4) According to Kuhn-Tucker rules,the optimal solution must satisfy () 10iiiyw xb+ =,1,2,in=?. (5) That is to say if the optima solution is * 12(,)Ti =?,1,2,in=?. (6) then *1niii iwy x= ， *1()nii
