For the disadvantage of passive millimeter wave (PMMW) image, such as poor quality, obscure boundary and difficult identification, an improved FCM clustering algorithm is proposed by the subtractive clustering based on the traditional words model. Moreover, with the visual words extracted, the pistol, knives and explosives are briefly by the method of SIFT. Finally, experimental results show that the improved algorithm can be accurately classify the contraband. Furthermore, the average recognition rate can reach more than 90%. Compered with FCM clustering algorithm and K clustering performance, the improved FCM clustering algorithm is excellent.