The Research Of Single-Sample Face Recognition Based On Neural Network

In this age of intelligence science and technology, automatic recognition, especially human face recognition has received more and more attention. Human face recognition technology employed computer vision, image processing and pattern recognition to detect human face, extract and classify human face features. Based on the extracted features, the human faces are classified and the identity is verified. Only one image is the sample for human face images obtained from some special places. Single registered sample greatly reduces the face recognition accuracy. In this article, a human face recognition method based on nerve network is proposed.It is necessary to preprocess human face images due to the differences of background, size and direction of the human faces obtained. First is to normalize images of human faces. The purpose is to normalize human faces into gray images with the same size to exclude the interference factors. In this article, the Haar features are extracted by modified Adaboost algorithm. A multi-scale strategy is used to detect the different size human faces in the same image. Based on the result of face detect, then to finish the detection of human eyes, which is done by two-directional Gabor filtering method combined with morphology processing. The detection area is reduced little by little, and finally tests the human eyes precise position by the sample match method.In the phase of human face recognition, wavelet transform is executed to the human face image after preprocessing. Based on the features after the integration of high and low frequency information, the low frequency information of the registered image and the high frequency image of unknown image are integrated. By computing the Euclidean distances between these two images and take it as the input feature to the nerve network to classify. In this article, the traditional nerve network is improved. A network is designed for each of the human faces to classify. Activation functions and nodes of nerve cells of the input layer, hidden layer and output layer are designed respectively. After the experiments on various image databases, the face detection rate of FERET databases is 96.07%, the face recognition of FERET databases is 90.53%. The face detection rate of self-made databases is 99.09%, the face recognition of FERET databases is 98%.From the result it is found that the classifier designed in this article is robust for the detection and recognition of human faces with perspective angles, ornaments and in different sizes.