Scene-identity Face Matching Model Research In Embedded Equipment And Unconstrained Environments

With the help of AI technology based on deep leaning,face recognition has been successfully commercialized.However,most of the current face recognition systems work under the circumstance of limited environment and require users to cooperation actively,which are not effective for face recognition under occlusion.In addition,the current face models with excellent performance are large convolutional neural networks with slow running speed,large amount of computation and large number of parameters,which can't meet the requirements of some scenarios.So we study the scene-identity face matching model in an embedded and unconstrained environment.Specifically,this paper is based on the scene-identity face matching scenario,focusing on the face occlusion problem,using deep learning technology,based on the embedded computing device,to research the scene-identity face matching model from the three aspects of data,algorithm and application.The main research works are listed as follows:(1)Construct the dataset of partial occlusion scene-identity face matching.In view of the insufficient quantity and poor quality of existing partial occlusion datasets,on the basis of scene-identity face matching dataset,a partial occlusion simulation algorithm based on key point detection was designed to simulate the face occlusion of actual scene,automatically generate a scene-identity face matching dataset with partial occlusion.The comparison between the scene-identity face matching dataset and the partial occlusion scene-identity face matching dataset shows that the constructed partial occlusion dataset is effective and reliable.(2)Propose the partial occlusion scene-identity face matching algorithm.Aiming at the partial occlusion problem,we propose a weighted average pooling layer network structure.The weight matrix is used to improve the effective information,suppress the redundant information,and obtain effective performance improvement.Facing the sceneidentity face matching scene,we use two branch networks with the same structure and different parameters to process id and life photos respectively,and map the face image to the euclidean space of the same dimension for measurement learning,which obtained a huge performance gain.At FAR=0.01%,TAR gets a performance gain of about 10%.(3)Propose a neural network construction algorithm based on multi-mode convolution kernel.Aiming at the performance loss of the model in the process of transplantation to embedded device,the multi-branch network based on the multi-mode convolution kernel is built using the characteristics of the batch normalization layer to enhance the performance of the model without losing the speed of model inference.Experimental results show that the algorithm can obtain an additional performance gain of about 0.6%.(4)Design and implement a scene-identity face matching system based on embedded computing device.For the problem of limited computing resources of embedded devices,we deploy the lightweight convolution neural network on the embedded edge computing equipment,use the Tensor RT to reduce the inference time,build the scene-identity face matching system,and realize the automatic matching function.