Research On Deep Learning Face Recognition Based On Convolutional Neural Networks

Image classification and recognition is a research focused area in the field of computer vision and pattern recognition.Face recognition technology is also an important research direction of image classification.It can be used to verify the identity of people through non-contact method,which is simple and convenient way to implement.Therefore,it is widely used in many fields.However,the traditional face recognition technology could only be used in laboratory scene.When testing the performance of the algorithm,the changes of the samples in the database in lighting,posture,expression have to be strictly controlled.and some images' features need intensive computation.Therefore,it has important research significance and application value to solve the field application problems of face recognition.In recent years,deep learning has achieved a good recognition effect in the field of image recognition.Therefore,after deep researching the CNN(LeNet-5 convolutional neural network)model,this paper puts forward some improved methods and measures.The main work of this essay is as follows:(1)By adding cascaded original LeNet-5 convolutional neural network layers,modifying and adjusting network layer parameters,an improved Lightened CNN(LCNN)network model is obtained,and a parameter initialization method is used to reduce the model's convergence time.In the end,the new model not only solves the problems that the original LeNet-5 model's training parameters are not easy to adjust and the training model's extraction features are insufficient,but also selects different feature matching algorithms in the feature matching process to perform matching test research in order to seek final recognition.The face model trained in CISIA face database was tested on AR,LFW and other databases,and the rate of recognition is 94.70%.(2)In order to further solve the disturbing factors of human face in complex natural environments,and to deal with pixel contamination and small occlusion of individual pictures,sparse filtering is added at the LCNN data input network layer to form a new SF-The LCNN network model adjusts network layer parameters,and performs Dropout technology improvement on the SF-LCNN network model to improve the convergence speed,enhances the network's feature extraction capabilities for more complex face images,and improves the recognition accuracy in AR,Test experiments were conducted on databases such as LFW,and the recognition rate reached 97.65%.(3)Finally,a real-time face recognition system based on camera surveillance video is designed and implemented based on the above model algorithm,which is applied in the Zeng Xianzi experimental building lobby scene.The function and proceeding of each module of the intelligent face recognition system are introduced in detail.The collected face database of the students is used to establish a face database,and this database is tested.The test achieves an accuracy of more than 97%.The face recognition system verifies the validity of this method and the practical application requirements of face recognition.