Research On Key Technologies Of Face Detection And Recognition Based On Convolutional Neural Network

In recent years,with the continuous advancement of CNN and hardware equipment,face detection and face recognition have also been widely used in railway gate system ticket verification,mobile device login and identity authentication in the financial field.However,a series of problems have emerged in face detection and recognition in practical application:the face detection algorithm which under complex environmental factors has problems such as high false detection rate and missed detection rate,and low recognition accuracy of largeangle and large-pose faces.To address the above problems the main research work of this paper is as follows.First,in face detection,because of the poor performance of traditional linear interpolation feature upsampling,this paper proposes a feature super-resolution reconstruction algorithm based on CNN。And it is applied to the FPN of the face detection algorithm Retina Face to replace the original interpolation algorithm for feature upsampling.On data sets such as FDDB,the improved algorithm has significantly improved its detection accuracy compared to the original algorithm.Second,face recognition: Aiming at the complex and changeable angle and posture of the face to be detected in the real scene,and the conventional face recognition algorithm cannot meet the requirements of high-precision recognition in this environment,this paper proposes a three-dimensional like face recognition based on Siamese network and Res NetSE(Siam3DLike Face base on Res Net-SE)to address those challenges,siamese network is one of CNN.Siam3 DLike Face improves the accuracy of face recognition by extracting features and building a database of face in different postures and angles.Finally,experiments on face detection have proven the effectiveness of feature superresolution reconstruction and FR-FPN.At the same time,this paper has amplified data and face recognition experiments on the CAS-PEAL-R1 and other data sets,which proves that the average accuracy of Siam3 DLike Face base on Res Net-SE proposed in this paper has been improved compared to the original algorithm.