Abstract: A support vector machine constructs an optimal hyperplane from a small set of samples near the boundary.This makes it sensitive to these specific samples and tends to result in machines either too complex with poor generalization ability or too imprecise with high training error, depending on the kernel parameters.SVM focuses on the samples near the boundary in training time, and those samples intermixed in another class are usually no good to improve the classifier's performance, instead they may greatly increase the burden of computation.In order to improve the generalization ability we present an improved SVM:BS-SVM.It first classifies the training set.According to the distance between the sample and the mean sample, the training sample is classified three classes:good sample, poor sample and boundary sample, then trains the SVM with boundary sample.Experimental results show that BS-SVM is better than SVM in speed and accuracy of classification.