In this paper, we propose a new multi-classification algorithm based on the non-parallel plane support vector machine (SVM). In the approach, data points of each class are proximal to one of nonparallel planes, and at the same time, are far from the other categories to certain extent. This leads to solve convex quadratic optimization problems which the number is the same as the varieties of category. Optimization problem for each is less than the size of the quadratic programming problem of standard SVM. We also induce the kernel method into our algorithm to solve the non-linear problems. Experimental results show that the proposed method which compared to the current multi-classification methods, not only in the overall accuracy rate but also in specific categories of accuracy, plays a good performance.