Research And Implementation On The Face Recognition Platform Based On Collaborative Edge Computing

In resent years,with the rapid development on the research of computer vison technology,especially with the breakthrough in the field of object recognition with deep convolutional neural networks,now computer vision and especially the face recognition,has been more and more widely used.Because of the computational complexity of convolutional neural network,it is difficult to run better network models on local machines with limited calculation power,especially for the embedded device.While it is also difficult for a single cloud device to achieve the best performance for the factors like the network bandwidth.To solve the problems above,a technology of face recognition based on collaborative edge computing is putted forward on this thesis.By deploying convolutional neural network on the embedded system and the cloud corner respectively,it is possible to make the face recongnition more fastly and accurately in different scenarios,especially in a complex environment like a large airport.The focus of this thesis is to build a residual network in the cloud,so that the face features can be quickly extracted and the face can be accurately identified.In addition,a lightweight convolutional network is designed on the embedded side to implement face recognition.And then all of the functional modules on both the cloud and the embedded side are completed.Finally,a strategy is designed for choosing a module between both sides to achieve a faster and more accurate face recognition platform.The main contents of this thesis include:(1)The techniques related to the neural network of deep learning is in-depth studied,including gradient descent algorithm,back propagation algorithm and the structure of convolutional neural network.A commonly used open source framwork Tensorflow and the edge computing idea is also explored.(2)A convolution model based on neural network and Siamese network is deployed on the cloud.The ability of the model to extracted face features is enhanced by adding a Siamese structure.A dataset named WebFace is used to train the model and optimize the parameters.At last,an accuracy of about 93.9%was obtained by validating the trained model using the LFW dataset.(3)A lightweight convolutional neural network structure named MobileNet was designed and deployed on the embedded platform,making it possible to run on an embedded devices with weak computing power.A recognition accuracy of about 88.3%in the LFW dataset was last obtained after training by FaceScrub dataset with a cost function of triplet loss.(4)The whole framework of the face recognition platform based on collaborative edge computing proposed in this thesis is designed and each functional modules of it are finally realized.After designing a selection strategy for choosing a platform,the feasibility of the platform is finally verified by using a collected face images from 10 students.