Face Recognition Based On Convolutional Neural Network

In recent years,thanks to the development of face recognition technology,face gates,intelligent check-in system,intelligent monitoring and so on have gradually entered people's daily life,providing many conveniences for people's lives.Face recognition,as a means of identity authentication,has the advantages of convenience,non-perception,non-coercion and concurrency compared with iris recognition and fingerprint recognition.However,face is easily affected by external environment such as gesture,expression and occlusion,which brings many challenges to face recognition.Traditional face recognition algorithms use artificially designed features to extract features.It is often difficult to describe face features in an all-round and multi-level way,which leads to the difficulty in improving the recognition effect.Since 2012,with the proposal of Alexnet,deep learning has made great progress in the field of image detection and recognition,which greatly improves the accuracy of object recognition.In image aspect,deep learning uses convolutional neural network algorithm,which does not need artificial design features.It is attributed to the design of large-capacity models and rich tag data collection.Through a large number of image data and massive parameters,feature learning can be carried out automatically,and more distinguishing and realizable features can be obtained.In this paper,face detection and recognition based on convolutional neural network algorithm is carried out.The main work is as follows:1)In view of the vulnerability of face detection to multi-pose,illumination,occlusion,etc.,this paper takes MTCNN as the basic model.On this basis,BN initialization method and focal loss loss function are used to improve the training of the model,and the face detection data set is progressed on wider face.2)In the process of face recognition,because of the influence of illumination and expression,the recognition will also be disturbed.Therefore,it is necessary to minimize the intra-class distance of face features and maximize the inter-class distance of face features.Based on facenet algorithm,this paper compares the influence of different face soft intervals and similarity measure functions on recognition accuracy.After that,this paper designs a deep convolution neural network face recognition model with high recognition accuracy and achieves 99.15% recognition accuracy on LFW data sets.Finally,a lightweight convolution neural network model with less parameters suitable for mobile terminals is designed,and itsparameters are 1.03 M.In practical applications,different models can be selected according to the application scenarios to carry out face recognition tasks.