Quite recently the support vector machine (SVM) has shown a great potential in the area of automatic face detection. Generally the SVM based methods fall into two categories: component-based and whole face-based. However there exist some limitations to each category. In this paper we present a two-stage method using both SVM categories based on multiresolution wavelet decomposition (MWD). In the first stage, the whole face-based SVMs are used for coarse location of faces from small sub-images of low resolution. Then a set of component-based SVMs are applied to verify the extracted candidates in subsequent larger sub-images of higher resolutions. Experimental results show that this wavelet-SVM based method takes the advantage of the effectiveness of both categories of SVM-based methods and the computation efficiency.