Research On Face Recognition And Its Application

Face recognition is one of the important biological characteristics. It is paid much more attention now as the technology of face recognition is direct, friendly and convenient There are still some open problems in face recognition because it is influenced by illumination, pose, expression, overlap and age span. And it is far away from the practical application. These problems are such as: Facial feature location is very important to face recognition, and how to ensure the right location; Robustly recognition of faces can get accurate recognition in these conditions: small and/or noisy images, images acquired years apart and outdoor acquisition: light and pose etc. How to get the robust method; what's the limit on the number of faces that can be distinguished; what's the principal and optimal way to arbitrate and combine local features and global features. It has warmly attracted many researchers to effectively solve these problems. It is surveyed about the methods of face detection and recognition. The technologies of face detection are usually about: knowledge -based methods, feature invariant approaches, template matching methods and appearance -based methods. The technology of facial feature extraction and face recognition are about: face recognition based on geometry features and template matching, eigenfaces-based recognition, low frequency subimage-based recognition, neural network based recognition and SVM-based recognition. On basis of these methods, this paper discusses that face detection, facial feature extraction face recognition and its application in color and gray image with complex background. The followings are the main research contents in my paper: (1) Face detection in complex background. When detecting, we locate the rough face region in the image and validate it using template model matching. We use projection slope of difference image to enhance the robustness to the background in static gray image. At the same time the method that we use K-L model in color image location and increase the searching speed combining agent model is proposed. The former improves the method of difference. The speed and veracity will be effected as there are much more edges when the searched object moving. The more corrected location will be gotten when the redundant edges are filtered by projection slope. The latter has more fitness. Face Locating by the skin color is almost not affected by the complex and dynamic background and can locate some faces. (2) It introduces wavelet translation and compression and the technology is used to facial feature extraction and vector data's dimension decrease. A main problem is related to the pattern recognition is so-called dimension disaster. The classed system needs the minimal feature data. Wavelet transform is widely used in image processing as its good limitation of space-time. There are 3 methods using wavelet in facial feature extraction and data dimension decrease: the low frequent subimage at some level; the weighted sum of the fourth part of a transformed image; the facial feature data extracted by EZW and entropy encoding. (3) It introduces the adaptive resonance theory (ART) and applies it to face recognition. The characteristics of ART are compatibility to same class samples and easy to change the number of the sample's number. It can be trained online. The cost time is short when recognizing a face. The parameters are adjusted properly to fit face recognition during training. When using to recognize a face, the facial feature data is input the ART, the most similar class to the input sample one is found by the competed mechanism. When adding a new sample, the network needn't retrain and only add an output node, also adjust the corresponding weights. (4) When we use a face as a probe searching in the large database, the searching time is very long. In order to improve the searching speed, the face database should be divided into some subclasses. The range of searching is reduced and the speed of searching is effectively increased. We mainly introduce K-mean clustering and cluster...