Research On Lightweight Scalable Convolutional Neural Network Algorithm And Application For UAV

In emergency rescue scenarios,rescuers will use heterogeneous unmanned aerial vehicles(UAV)with different coverage,different equipment loads,and different volumes according to the rescue stage,the purpose of the rescue mission,the characteristics of the mission,the relevant areas of the mission,and the complexity of the mission.UAV clusters cooperate to complete large-scale emergency rescue missions.In recent years,with the development of deep learning and the improvement of hardware performance,it has become possible for UAVs to carry lightweight deep learning applications to perform emergency rescue missions.However,there are still the following challenges to uniformly deploy deep learning applications into heterogeneous UAV clusters:Firstly,in a heterogeneous UAV cluster,the computing resources of different UAVs are different.If a fixed-scale deep network model is used,the balance between real-time performance and accuracy of deep learning applications cannot be guaranteed.Even if a slimmable neural network is used,in airborne devices with limited computing resources,the accuracy of small-scale sub-networks in that network is still relatively low,which will affect the execution effect of the entire cluster task.Secondly,the complexity of input data is different in different scenarios.If a fixed-scale deep network model is used,the same resource and time consumption will be incurred for data of different complexity,and higher resource utilization and task execution efficiency cannot be achieved.Finally,there are many types of heterogeneous UAVs,and the process of deploying deep learning applications is cumbersome and complex,which is also a heavy burden on rescuers.This thesis is based on the emergency rescue scenario,with the ultimate goal of deploying a deep learning model in a heterogeneous UAV cluster and completing the emergency rescue task accurately and efficiently,studying the algorithm and application of lightweight scalable convolutional neural network for UAV.The main work and contributions of this thesis are as follows:(1)This thesis proposes a slimmable neural network adapted to the deployment of heterogeneous UAVs.Aiming at the problem of poor performance of small-scale subnetworks in the current slimmable neural network,this thesis proposes methods of inplaced feature map attention transfer and inplaced knowledge distillation to optimize the original network.Therefore,a slimmable neural network adapted to the deployment of heterogeneous UAVs is designed,and based on the mainstream lightweight network model structure,a lightweight slimmable network model adapted to the deployment of heterogeneous UAVs is constructed.Finally,through experiments,the classification performance of the small-scale sub-network in this model is better than the original model.(2)This thesis designs a lightweight scalable convolutional neural network model for UAVs.In order to allow the network to dynamically adjust its network depth according to the complexity of the input data while having variable width capability,this thesis proposes a layer-hopping module suitable for slimmable neural networks,and embeds it into the lightweight variablewidth network model constructed in(1)to design a lightweight scalable neural network model for UAVs.Through experiments,this thesis verifies that the model can adjust the network depth according to the complexity of the input data,reduce the overall calculation amount,and improve the efficiency of resource utilization and task execution.(3)This thesis designs a lightweight deep model deployment platform for heterogeneous UAV clusters.In order to make it easier for rescuers to deploy the lightweight and scalable convolutional neural network model of UAVs into heterogeneous UAV clusters,this thesis is based on containerization and container orchestration technologies,combined with the design ideas of cloud native and edge computing,designed a lightweight deep model deployment platform for heterogeneous UAV clusters.The platform uses the design principles of software engineering to conduct implementations,which can streamline,automate,and simplify the deployment of lightweight deep models,reduce the learning cost and operational burden of rescuers,and is of great significance for improving emergency response capabilities and emergency rescue efficiency.