Face Recognition System Based On Hopfield Neural Network

The face recognition is an important topic in the pattern recognition research area. In the past decades, the research of the human face recognition was mainly limited to theoretical fields. From 80's to the early 90's, the face recognition technology has gradually become a hot issue in the application aspect. A complete automatic face recognition system consists of three main parts: detecting the human face from the image, extracting the face feature and recognizing face. In the light of previous literature on face-detecting and recognition, this paper attempts to establish a complete face recognition system and to conduct the discussion on the algorithms of the three parts and write relevant programs. At the face detection stage, this paper adopts the method of ellipse model of complexion detection. This method have a good effect on the image in natural light condition, but is not ideal for the image from the camera, as a result of the illumination influence, This paper overcomes this defect by carrying on the dynamic adjustment to the ellipse model parameter , and gives out the experimental contrast to the images under the different illumination conditions. In the pretreatment aspect, this paper mainly focuses on normalization of the human face image. The human face region detected from the image may have some differences in the size, direction of the face and the grayscale resulting from the deferent resolution of the camera, the deferent distance between the human face and the camera and the influence of illumination change. These differences will be able to have the direct influence on final recognition result of the human face. The method for automatically locating human eyes in face image has been used to find the eyes. And then according to angle between the line of the center of two eyes and the horizontal position, the face image is rotated. Because size of the image which has been rotated is bigger than the original one, the method of two-time detection, namely human face rough detection and refine detection, has been used to resolve the question above. By the bilinear interpolation, the face image is normalized in the same size (100 100). In the aspect of the feature extraction and face recognition, K-L transform is taken to extract the overall human face image, and the Hopfield Neural network is used in the face recognition. "Energy function" of Hopfield neural network has the characteristic of reducing gradually and tending to be steady state of balance in the end. Furthermore, once the network is established, it can automatically moves. The recognition is divided into two stages: the training stage and the recognizing stage. At the training stage, the network model is established by taking the target vector the same as the input one. At the face recognition stage, the need-to-be recognized face features are put into the network to operate. When the network stands in balance, the output vector is compared with the vector in database, and the face with the smallest distance is the sample image. Finally, the min-distance classifier and the BP neural network separately have been carried on the human face recognition. The result of the experiment indicates that the Hopfield neural network has a higher recognition rate and better stability on the face recognition of the small sample collection than the other two methods.