Research On Cloud-edge Collaborative Task Offloading Methods For Artificial Intelligence Image Recognition

The deep integration of cloud-edge collaboration and artificial intelligence(AI)promotes the further development of image recognition technology,which is widely used in AI applications in medical,marine and other fields.AI image recognition applications often adopt the cloud training and edge inference,in which a large amount of data transmission can lead to network bandwidth pressure and slow inference response.The cloud-edge collaborative computing enables a collaborative computing mode for cloudbased model training and edge inference,which effectively supports large-scale cloud computing tasks and real-time edge data processing and analysis to meet the requirements of image recognition applications.However,cloud-based model training and edge inference are separated for AI applications,which makes it difficult to apply and update the model online due to an unclosed loop of application model training and inference.And the edge inference resources are limited,which often faces the problem of insufficient computing power due to the growth of image data.The problem of low accuracy rate arises due to the deployed lightweight inference model.The unbalanced sample data may result in low aggregation accuracy because in cloud-edge collaborative federated learning.In this thesis,we study cloud-edge collaborative task offloading methods for artificial intelligence image recognition from three aspects: cloud-edge image training and inference integration,cloud-edge collaborative image inference and cloud-edge collaborative federated learning model aggregation based on the analysis of cloud-edge collaborative computing features,task offloading technologies and platform frameworks.The main research contents and innovations are as follows.(1)A task offloading method of image training and inference integration is proposed based on cloud-edge architecture.Firstly,a task offloading model with an integrated image training and inference is constructed based on cloud-edge architecture,in which the correlation between task offloading and resource allocation,and that among taskresource-data are described formally in the cloud-edge collaborative architecture.Then,the task offloading method based on cloud-edge architecture with integrated image training and inference is studied,and the cloud model training,model image production and distribution,and edge application deployment and inference are realized by applying the tachnologies of cloud-edge collaborative computing,container and YAML orchestration automatic deployment.Finally,a cloud-edge collaborative computing environment based on Kubernetes and Kubeedge is constructed to realize the integration of cloud-edge training and inference for image recognition applications.And the experiments are conducted to compare the results of edge inference and cloud inference.The experimental results show that the proposed method simplifies the model deployment process,greatly reduces the data transmission time by edge inference,decreases the image recognition task execution time,and improves the execution efficiency of image recognition model training and inference.(2)A resource-constrained task offloading method of cloud-edge collaborative image inference is proposed for the problems of limited resources and low inference recognition probability value of lightweight models at the edge node.The task migration and offloading strategy of cloud-edge collaborative inference with load overrun is constructed to ensure the image recognition inference efficiency by calculating the load overrun parameters of edge nodes to trigger the cloud-edge collaborative inference.The migration and offloading strategy of cloud-edge collaborative inference task with low inference recognition probability value is constructed to compare the inference probability value with the tolerance evaluation,in which images with low recognition quality or beyond the recognition range of lightweight models are uploaded to the cloud for collaborative inference to guarantee the image recognition accuracy.Finally,the method plug-in is developed and integrated into the cloud-edge collaborative inference framework built based on Kubernetes,Kubeedge,and Sedna.And the experimental analysis of cloud-edge collaborative inference is carried out in cancer pathology and marine fish recognition.The results show that the method solves the problems of task blocking,slow image recognition,and low recognition rate caused by insufficient arithmetic power of edge nodes by regulating cloud-edge task collaborative offloading and resource utilization,and improves the efficiency and accuracy rate of image recognition.(3)A cloud-edge collaborative federal learning model aggregation method is proposed for sample imbalance.The method starts from improving the fairness and effectiveness of model aggregation,addresses the problem of unbalanced sample distribution leading to large differences in global model aggregation contributions,integrates the accuracy and stability of federal learning local models for each sample feature recognition and the number of samples used for training as evaluation weights,and establishes a federal learning model aggregation method for sample imbalance to evaluate high-quality local models.The method increases the weight of contribution value of high-quality local models and reduce the contribution of sample imbalance and lowquality models to improve the accuracy of the global model.Finally,the method plug-in is developed and integrated in the cloud-edge collaborative computing environment.The experiments are conducted to compared this method with the classical federal average algorithm(Fed Avg)model aggregation under various sample imbalance scenarios.The experimental results show that the method reduces the impact of sample variability on the global model aggregation and improves the accuracy of the global model under multiple sample imbalance scenarios.