A Research On Key Technologies Of Face Recognition In Surveillance Environment

With the popularity of surveillance cameras in daily life,the information captured by the camera can provide more security for humans.Face recognition technology based on surveillance camera can play an important role in security,education,smart city,smart home and other fields.Face recognition in the monitoring environment is difficul,because there are problems such as blurred face images,small face size,insufficient illumination and so on.Aiming at the above problems,the main research contents include face recognition technology based on deep learning,face image preprocessing technology in monitoring environment,and face recognition method based on deep convolution network feature fusion.The specific contents are as follows.Firstly,this thesis briefly analyzes the face recognition technology based on traditional algorithms,and deeply studies the face recognition technology based on deep learning.It mainly includes basic knowledge of deep learning,face recognition processing flow,face detection knowledge,face alignment knowledge,and face verification knowledge.This thesis focuses on face detection of cascade convolutional neural networks and small-scale face,face alignment of cascade convolutional neural networks,and multi-task cascade convolutional neural networks,face verification of deep identity features and angle & cosine margin loss function.Secondly,in view of the main problems faced by face recognition in surveillance environment,this thesis solves the problem of small face image size by using fast superresolution convolutional neural network after analyzing and comparing multiple image super-resolution algorithms.Based on Retinex theory,this thesis designs an adaptive image enhancement algorithm based on HSV spatial domain and Retinex image enhancement algorithm based on different color space fusion,which solves the problem of blurred image recognition.The experimental analysis shows that image superresolution technology and image-enhanced face image preprocessing technology can improve the face recognition accuracy.Finally,based on the analysis of the traditional image feature extraction methods,this thesis proposes a face image feature extraction method based on convolutional neural network and data augmentation.In order to improve the correct rate of face recognition,the face features extracted by multiple convolutional neural networks are fused,and then face verification is performed by using support vector machine and nearest neighbor classification algorithm.Experiments show the classification algorithm based on support vector machine is more stable and reliable.In the long-distance face recognition,the correct rate of long-distance face recognition is further improved by using the close-up face image feature as a training set.