In this thesis, we present a robust face detection algorithm in cluttered images. It is a system based on a novel feature extraction method that uses wavelet-based sub- kernel principal component analysis (SKPCA) method. Usually, facial features such as eyebrows, eyes, nose and mouth are distinguished characteristics of human face, and that their global structure distribution is a guarantee of the existence of a face. The proposed method firstly transfers gray value images into a Haar wavelet spatial frequency domain to extract visually plausible features of the shape and interior detail of faces, Following this step, the constructed wavelet coefficients are mapped onto a higher dimensional feature space using SKPCA. By this way the structure distribution of laces can be captured. Theoretically, SKPCA is an extension of kernel principal component analysis algorithm (KPCA). In the process of SKPCA, a simple and effective subset choosing method is used to solve the computational and storage problems of KPCA. Via wavelet-based SKPCA feature extraction method, representative features of face images that include both local facial feature detail and global structure characteristic are constructed. Finally, a relatively simple support vector machine (SVM) is adopted as classifier.