Research On Deployment And Optimization Technology Of Deep Learning Algorithm Based On Edge Computing Platform

With the popularization of various artificial intelligence applications on mobile devices,a large number of new requirements pose new challenges to the deployment of edge computing platform applications.Due to the limited computing resources of edge devices,many AI applications cannot be implemented because it is difficult to meet real-time requirements.How to solve the contradiction between the algorithm requirements and the computing power of edge devices has become a key issue in edge applications.In view of the above problems,this thesis conducts research on the deployment and optimization of deep learning algorithms on edge computing platforms,mainly including the following work:A general software framework that can be used for algorithm deployment of machine vision inspection systems is studied.On the one hand,the framework can complete the rapid deployment of algorithms on the platform,and on the other hand,it can perform system resource scheduling for multi-task scenarios.The framework should have good stability,scalability,and maintainability.The software structure is friendly to developers,and it is easy to further expand and develop according to actual system requirements.Based on the channel pruning optimization and the Tensor RT framework,combined with the hardware characteristics of the Jetson AGX Xavier platform,a set of deep learning algorithm inference optimization schemes are proposed.First,the deep learning algorithm model is pruned through the channel,and then Tensor RT is used to complete the INT8 accuracy calibration.Combining CUDA’s storage management and communication optimization methods,and Tensor RT layer fusion and other optimizations,the inference engine is obtained for practical applications.Facing the needs of the intelligent road sweeper project,based on the above research contents,the system implementation of the intelligent road garbage identification system is completed,and the deployment and optimization of the road garbage detection algorithm on the platform is completed according to the proposed inference optimization strategy.The execution speed and performance of the optimized algorithm can meet the needs of the project.The test results show that the road garbage detection algorithm model reduces the model size and increases the inference speed at the cost of a slight decrease in accuracy.The amount of model parameters is reduced to 8.7% of the original,and the inference speed is increased from the original 97 ms to 24 ms,which proves that the inference optimization strategy proposed in this thesis can have better results in practical applications.