An improved support vector machine will be presented for solving the problem of the inefficiency of license plate recognition.License plate pretreated and location is the first step of the automatically license plate recognition,and we define one multiple kernel for each feature.The optimal weight with each kernel matrix in the combination is obtained through the semi-definite programming learning method.We solve multiple kernel learning with an improved semi-definite programming to reduce the time and space requirements.Last we build the model of license plate recognition.Experimental results show that our algorithm is fast and accurate.