| With the development of the Internet of Things,many new application scenarios such as smart home,auto-driving and smart medical emerge gradually.The traditional centralized data processing mode centered on cloud computing can not meet the new requirements of delay,energy consumption,security,etc.Therefore,a new computing model of "end device-edge-cloud" collaboration has proposed;Convolution neural network has achieved great success in the new application scenarios due to its excellent performance,but its excellent performance depends on a large number of parameters and computation.In the edge scenario,the resources of computing and storage are limited.Convolution neural network is difficult to deploy directly on end and edge devices.Therefore,it is important to reduce the amount of parameters and calculation of the convolution neural network to meet the real-time inference requirements while guaranteeing its performance,and to promote the large-scale application of the convolution neural network on the collaboration of "end device-edge-cloud".In this paper,the "end device-edge-cloud" collaborative convolution neural network is optimized from two aspects.On the one hand,convolution neural network is complex and difficult to deploy.From the perspective of network lightweight,this paper optimizes the structure of convolutional neural network and improves the deployment efficiency;On the other hand,convolution neural network has large input data and difficult to calculate.From the perspective of computational efficiency,this paper optimizes data preprocessing and improves data compression rate.The specific research contents are as follows:(1)To solve the problem that network structure is complex and difficult to deploy,this paper presents a "Triple-partition" convolution neural network partition mechanism based on model performance evaluation.Firstly,the traditional convolution neural network is divided into three lightweight subnetworks,and the early exit mechanism is used to set the split points to provide efficient dynamic computing services for the convolution neural network.Secondly,in order to select the optimal network partition point,the EntropyTopsis comprehensive evaluation model is introduced to evaluate the network performance.Based on this,the "Triple partition" network is constructed to improve the efficiency of computing services and deployment.In this paper,classical convolution neural networks(AlexNet,ResNet)are used to explore the performance of the model on public datasets.The results show that the trained "Triple partition" convolution neural network can be well deployed in resourceconstrained environments,and it can improve the inference speed while guaranteeing high accuracy of the model,reducing end-to-end delay to the original 1/3.A good balance has been achieved between inference performance and resource consumption.(2)To solve the problem that the input data is large and difficult to calculate,this paper presents a data preprocessing mechanism of convolution neural network based on DCT-Attention module.Firstly,the DCT algorithm is used to convert the conventional spatial domain data into more compact frequency domain data,which simplifies the preprocessing steps by using the feature of block parallel computing.Secondly,using the energy aggregation property of DCT algorithm,the Attention module is introduced to filter out important frequency domain and suppress noise,thereby greatly reducing the size of input data and achieving the purpose of data compression.In this paper,classical convolution neural networks(AlexNet,ResNet)are used to explore the performance of the model on public datasets.The experimental results show that when the input data of the "Triple partition" network is compressed to only 1/3 of the original through the DCT-Attention module,the accuracy of the model is almost lossless. |
