Research On Embedded Face Recognition Technology Based On Lightweight CNN

With the continuous improvement of social informatization,face recognition technology is widely used in the fields of access control system,public security,consumer entertainment and so on.Face recognition algorithms can be divided into two categories:traditional methods and deep learning methods.Currently,mainstream face recognition algorithms are based on deep learning methods.Although the deep face recognition system has good detection effect and strong robustness,the algorithm is often deployed on the cloud server in practical application due to its complex system model and large computational load.In recent years,various kinds of lightweight convolutional neural networks have been continuously proposed,bringing new ideas to the application of face recognition system.Under the above background,this thesis designs a lightweight face recognition systembased on the lightweight convolutional neural network,which is deployed on an embedded platform with low cost,low power consumption and low computational power.A complete face recognition system can be divided into two parts:face detection and recognition.In the face detection part,this thes i s designs a lightweight face detection algorithm,which is based on end-to-end object detection algorithm.Through the design of lightweight main network to extract multi-scale features,the multi-scale feature fusion is realized by using the feature pyramid,and the simple detector with context enhancement is used for face detection.The anchor mechanism is used to generate and regression the face detection boxes,and the redundant information is filtered by non-maximum suppression to obtain the coordinates of face area and key points of faceIn the face recognition part,this thesis designs a lightweight face recognition algorithm.Designing a lightweight facial feature extraction network,the number of parameters and calculations of the algorithm are reduced;the network is trained with an improved loss function to improve the generalization ability of the facial feature extraction network;after extracting face features by using lightweight convolutional neural network,face recognition is completed according to the similarity of face features.Finally,based on the theoretical analysis and algorithm design,the Jetson Nano embedded platform is used as the application-processing platform and the algorithm is implemented and optimized on the embedded platform to realize the deployment of the face recognition system,and the real-time and robust performance of the system is verified.The face recognition system takes 37ms to perform an identity recognition.It can adapt to complex background and lighting changes,and has good robustness.Therefore,the embedded face recognition system based on lightweight convolutional neural network designed in this thes i s has good engineering application value.