Bp Neural Network-based Face Recognition Technology

It's well known that the technology of human face recognition has become a hot topicin pattern recognition field. Though a lot of progress has been made by many researchersthese years, many key problems still have to be solved in order to popularize theapplication of face recognition because of the complexity of face recognition.The background, development and main methods of face recognition are introducedfirstly in this paper, then a face recognition method which is based on wavelet transform,KL transform and BP neural networks is used in the paper. Here the face feature extractionincludes wavelet transform and KL transform. Moreover, the recognition classifier is BPneural networks. The simulation testing in the paper holds good recognition rate.The main work in this paper is introduced as follows:1. In the preprocessing stage, every face image is adjusted, located and standardizedto get the image of the same size: 32×32. After that, the influences on the face recognitioncaused by difference of scale and illumination intension are eliminated.2. The pre-processed images are dealed with wavelet transform in order to get the lowfrequency images. The low frequency images, which are compressed from 32×32 to 8×8,contain the most information and energy of the face images.3. The KL transform, which compress the data based on the rule of the least squaredifference, is a widely used method in feature extraction. The KL transform in this paperstarts with the collectivity spread matrix of the low frequency images from the wavelettransform to gain the eigenvector of the face images.4. In the training of the BP neural networks, the training sample set and the measuringsample set are not overlapped each other. Further, completed training and batch trainingstrategies are used to train the BP neural networks and the improved back propagationalgorithm: the flexible BP algorithm is adopted here.5. In the face recognition, the training sample set and the measuring sample set areseparately recognized to measure the BP neural networks which are trained in the forestage. The result of simulation testing, which is analysed and compared with other facerecognition methods, indicates that the method in this paper is feasible.