This paper proposes a novel weapon recognition approach based on the mixture of Support Vector Machine (SVM) and Bayes classifiers, thus presenting a robust way to assemble different approaches of recognition. Two kinds of features are used separately. The first kind of features is Sift feature which is dealt by SVM classifier. The second kind of features is obtained by calculating each image's color histogram. Due to the independence of each color feature, we use normal Bayes classifier to generate image's probability of containing a gun. To mix these two features, we assume two weights for the results probabilities of these two classifiers and exploit logistic regression analysis to build a linear model. Weight is valued by ROC curve. Experimental results show that the proposed approach maintains high recognition rate with a relatively low wrong-recognition rate.