Research On Edge Computing Technology Oriented To Object Detection Execution Efficiency Optimization

With the advent of the Internet of Everything era,the number of Internet of Things(Io T)devices with detection capabilities has exploded.Based on this,massive amounts of real-time data are generated at the edge of the network.The traditional cloud computing model has been unable to efficiently meet the needs of Io T devices.At this time,edge computing has become an emerging computing model with the characteristics of low latency,low bandwidth cost and high security.Because the current mainstream object detection models based on deep neural networks(DNN)have the characteristics of high storage and high computational capacity,it is impossible to directly use edge devices with limited storage and computing resources to complete computing tasks.Therefore,how to use the edge computing mode to complete the object detection task is an urgent problem to be solved.In many practical scenarios,new data samples and new data categories are gradually generated and updated.It is certainly feasible to train new models for these new data and categories independently,but the deployment of models will bring very large computing resource overhead to resource-constrained edge devices.Therefore,for scenarios where data and categories are gradually generated and updated and hardware resources are limited,how to achieve an increase in model capacity while maintaining or reducing the model size is a great challenge currently faced.In addition,greatly reducing the size of the model will cause a greater loss of accuracy.Therefore,in order to avoid compressing the model,a cloud server can be used to complete collaborative computing.In this scenario,how to accelerate the inference of the model while maintaining the accuracy of the model as much as possible is an urgent problem to be solved.In view of the above challenges,this paper studies edge computing technology for optimization of object detection execution efficiency.The main work and innovations are as follows:(1)For scenarios where data and categories are gradually generated and updated and hardware resources are limited,this paper proposes an incremental learning(ILMIL)method based on knowledge distillation of multiple intermediate layers.First of all,this paper proposes the MFRRK distillation knowledge containing multiple intermediate layer feature maps,which contains a lot of old knowledge.Secondly,use MFRRK distillation knowledge to complete the incremental training of the existing model,avoiding the resource overhead caused by training and using multiple independent models.Later,in order to further reduce the complexity of the model,the pruning operation can be performed before the knowledge distillation to compress the existing model.The experimental results show that when the existing model is incremented,the method in this paper has an effective accuracy improvement in the new class,the old class and the total class;when incrementing the compression model,at least 1.5%accuracy is lost when the compression ratio reaches 46%.(2)Aiming at scenarios with extremely limited hardware resources and high requirements for model accuracy,this paper proposes an edge-cloud collaborative inference(IDC)method based on intermediate data compression.Firstly,in order to reduce the network transmission delay of the output data at the model division point,a data compression method based on area compression and quantization is proposed,which can achieve differential data compression for the object area.Secondly,this paper divides the model in units of layers.In order to obtain the optimal division points under the premise of meeting actual needs,a method of selecting division points based on accuracy and delay is proposed.The experimental results show that the method in this paper achieves a speedup of 5.2 with a loss of only 5%;the method in this paper can obtain the optimal division point under the premise of meeting the actual demand.