Real-time Face Detection and Recognition Based on Deep Learning
Face recognition is one of the most important applications in video surveillance and computer vision. However, the conventional algorithms of face recognition are susceptible to multiple conditions, such as lighting, occlusion, viewing angle or camera rotation. Therefore, face recognition based on deep learning can greatly improve the recognition speed and compatible external interference. In this thesis, we use convolutional neural networks (ConvNets) for face recognition, the neural networks have the merits of end-to-end, sparse connection and weight sharing. The purpose of this thesis project is to identify the name of different people based on location of the detected box of a face. Then, we can obtain recognition results with different confidence under various distances. This thesis presents different methods with comparisons, namely, comparing the training results and the test results of different parameters under the same model, training results of the same test video under different models. We find that the recognition accuracy of this model is mainly affected by face proportion and the number of samples. If we get larger proportion of a face on screen, then we have higher recognition accuracy; if we obtain much greater number of samples, we can get higher recognition accuracy. In this work, we first collect sufficient samples as our dataset and use the suitable model embedded in the platform Google TensorFlow to complete the training and test. We collected five different faces and obtained 500 images on each face as training set, each of which can be cropped and rotated by using 50 different angles of the picture having a human face, of which 40 for training, 10 for verification. The use of neural networks for face recognition improves the speed of recognition. The contributions of this thesis are: (1) The use of elliptical markers can identify a human face including rotation and position. (2) The confidence of human face recognition is mainly affected by the proportion of face occupied on the screen.