Face Recognition Based On Deep Learning

With the continuous development of computer vision,face recognition technology is a very widely used biometric technology.With the rapid change of artificial intelligence technology,the deep learning method relying on big data has brought new development to the field of face recognition,especially after the acceleration of GPU,the implementation speed of deep learning is faster.In the convolution neural network,the deep learning feature extraction is replaced by the convolution neural network.For the face image data,it is not necessary to build the eyes,nose and mouth of the face.A large amount of data is input into the convolution neural network,and the weights in the convolution neural network are trained automatically to extract the features that should be extracted,so as to minimize the loss function.Face detection is the first step of face recognition,because the face images in this paper are in different light?different face angle?different exposure?different posture and occlusion.In this paper,we use the inherent correlation between MTCNN detection and calibration to improve the performance under the multi task framework of deep cascade.Using three-level joint structure and convolutional neural network algorithm to detect and roughly locate the key points of human face is a processing method from coarse-grained to fine-grained.At the same time,according to the face detection task and face classification task to assist the detection of key points of human face,it has good robustness to the changes of illumination and posture.The change of light and attitude in the natural environment brings some challenges to the recognition accuracy.How to improve the recognition accuracy in the natural environment better.To solve this problem,based on the deep learning theory of face recognition,this paper proposes a weighted strategy of maximum pooling and mean pooling for FaceNet model,and uses Softmax Loss and TriHard Loss to carry out joint optimization of multiple losses,achieving higher recognition and good accuracy in LFW data set,and improving the efficiency of face recognition.The main contents and contributions of this paper are as follows:1?Based on the implementation of FaceNet in TensorFlow deep learning framework,an improved weighted strategy of maximum pooling and mean pooling is proposed to optimize the FaceNet model.In this paper,the pooled weighting strategy,Softmax Loss and TriHard Loss are combined to optimize the multiple losses.The improved facenet model has higher recognition accuracy.2?Based on the above optimization model,the real-time face detection and recognition system is implemented by Pyqt5.It includes three modules of face image real-time acquisition and registration,face detection and real-time face recognition.It has good maneuverability and can meet the needs of daily face recognition.