To overcome the shortage of SVM and Dempster’s method, a new method of probabilistic outputs based on multi-class classification SVM is presented. The classification results are used as BPA functions, an improved evidence combination method is presented: the similarities between evidences are used as an approach judging whether conflicts exist, if there are more than 3 evidences as well as conflicts, the Mahalanobis Distance algorithm is used to calculate the distance between each evidence and the other groups of evidences so as to obtain the evidences’ weight coefficients. By means of these coefficients, BPA functions are transformed, and the Dempster’s method is used for the combination. Then, the probabilistic outputs of multi-class classification SVM are taken as BPA functions, the improved evidence combination method is used to fulfill the combination. Simulation results show that the output’s error rate is reduced, as well as the quantity of information is increased.