Target Recognition And Detection Model Deployment Research Based On Edge Intelligence

In recent years,the application of object recognition and detection models in artificial intelligence has become more and more important for extracting effective information from massive data,but the excellent ability of the model requires abundant computing resources and storage resources.At present,the target recognition and detection model is more urgently needed to be deployed and applied on edge devices,but the problems of insufficient computing power,limited storage resources,and certain requirements for power consumption of edge devices limit the direct deployment of applications of this type of model on edge devices.Therefore,how to efficiently deploy object recognition and detection models at the edge end is a very worthy problem.Based on deep learning technology,this thesis takes the model compression method in edge intelligence as the starting point,and discusses the localization deployment strategy of target recognition and detection model at the edge end.The main research contents of this thesis are as follows:(1)The complex network and redundant parameters of the target recognition and detection model make the volume of the model more and more bloated,and also increase the difficulty of deploying such models to edge devices.This method compresses the model volume by clipping the convolution channel of the model and compresses the model volume while retaining the target recognition and detection ability of the model,which successfully reduces the model volume and reduces the dependence of the model on computing resources and storage resources.(2)In order to accelerate the deployment speed of the model and reduce the power consumption of edge devices,this thesis quantifies the model after pruning,and proposes a model quantization method based on DETVM.This method uses low-precision data to represent the high-precision floating-point parameters in the model through the conversion of data types,which further compresses the volume of the model,improves the speed of model deployment to edge devices,and reduces the power consumption of edge devices running the model.(3)This thesis introduces the artificial intelligence accelerator SKTPU,which provides dedicated convolutional neural network processing functions for a wide range of edge devices.At the same time,the accelerator is equipped on the domestic edge terminal Loongson Education Pie to achieve the purpose of collaborative reasoning and computing power support,and further reduce the dependence of the model processed by(1)and(2)on the computing resources and storage resources of the edge terminal.