Face Recognition Method Based On Tengine Edge Intelligence

In recent years,with the advent of the Internet of Everything,artificial intelligence has developed rapidly in the fields of smart home,smart security,smart voice,and unmanned driving.The demand for computing power of artificial intelligence is increasing.There will be a large delay in the transmission of data generated by edge terminals to cloud servers.The main processing and data storage need to be placed at the edge nodes of the network.In this case,edge intelligence comes out.Edge intelligence is edge computing with machine learning and advanced network functions.It is necessary to conduct experimental research to efficiently run deep learning models on terminal devices with limited resources.Face recognition technology,as a popular technology of artificial intelligence,is one of the important application scenarios of edge intelligence in the future.This paper analyzes the face recognition method on the edge intelligent platform,completes the model training on the cloud server,and optimizes and converts the generated model,removes the redundant calculation part and converts it into a model that Tengine can directly call,and then Based on the Tengine inference framework,the converted model is completed on the edge intelligent platform to complete the inference part of the model.An efficient face recognition method is designed and implemented.The main work includes:1)Use the EAIDK-610 development board to build the software and hardware of the edge intelligent platform.In order to compare the differences between the different frameworks of the edge intelligent platform,the performance of the Tengine inference framework was tested on the built platform.By comparing the time of running the same model with the Caffe framework,the model was effectively optimized.2)For the effectiveness of mobile device resources,lightweight neural network Mobile Face Nets is adopted,which effectively improves the speed and accuracy of face recognition.The deep separable convolution method is adopted,and the amount of parameters and calculations are greatly reduced under the condition of little loss of precision.The separable convolution is used to replace the average pooling layer,which improves the situation of poor network performance.Finally,the Arc Face loss function is used to make the different categories more discriminative,prompting the model to learn more in-depth features to improve face recognition performance.And on the cloud server,the face recognition training process is realized through the Pytorch framework.3)Optimize the model generated by the cloud server training and convert it into a Tengine model that can be deployed on the edge intelligent platform.Then directly call the Tengine interface to read the model and create a graph to complete the inference part of the model.Finally,the software is integrated on the development board to realize,and the performance indicators are analyzed.Contrasted with other ways deployed on the edge intelligent platform,this paper effectively separates the model training and inference part of CNN.The model achieves better recognition effect and the performance index of face recognition is also ideal.Therefore,face recognition ways can be deployed on the edge intelligent platform based on the Tengine inference framework.The reasoning part of edge intelligence based on the Tengine inference framework can also solve the vast majority of artificial intelligence application scenarios.With a smaller model and faster speed,from the perspective of edge computing to enable artificial intelligence,it effectively solves artificial intelligence "The last mile problem" of efficient equipment deployment.