Abstarct: Feature extraction from facial images is used in various applications Over the past two decades, e.g. face detection, face recognition, facial exxpression and face reconstruction. Among these applications, Face recognition has a particular importance.Face recognition has problems such as poor quality of the input image, unanticipated coatings on the face, high volume databaxse and being only one picture per person. In this research, a new method is presented to feature extraction from facial images. The main purpose of this research is retrieval of probe image from a big databaxse. By increasing the size of the databaxse, the similarities between people increases and the separation capability decreases. The proposed method increases the distance between people in feature space by extracting appropriate features. This method is baxsed on properties of the human vision system and sequentially extracts features in top-down manner. For this purpose, spatial- frequency features are used. In this method, by applying concentric windows in different size on the facial image, the content of each window are mapped to frequency space. The change of frequency components in different windows forms the feature space of image. Then frequency component with high separation capability between face images is remained by appropriate filter. In the end, the final image is retrieved from databaxse by Euclidean distance criterion. In most existing facial recognition methods, desirable results are achived considering the small size of the databaxse or in big databaxse, the number of image per person increases. In this research only one image per person is considered in databaxse and used more subjects in databaxse against another methods. In this research the FERET databaxse is used. Recognition rate compared with the best current method in similar size of databaxse with 2% improvement, reached to 99%. By increasing databaxse size to 990 classes, 90/4% of recognition rate is achieved.