Research On Transfer Learning Based Image Recognition And Object Detection In The Internet Of Things

With the development of artificial intelligence and the proliferation of the Internet of Things(IoT)devices,more and more intelligent applications are deployed in IoT systems.Computer vision applications have strict requirements on task performance and response delay,and the computing power and bandwidth resources of devices need to be comprehensively considered during deployment.Cloud-edge collaborative computing emerges as the times require.Collaborative computing can not only reduce the communication delay caused by massive data transmission but also break the bottleneck of the edge server’s computing power to meet the needs of different intelligent task deployment in IoT.However,under the cloud-edge collaboration framework,the edge server needs to perform feature extraction for multidomain and multi-task,and deploying independent models will increase the storage pressure on the edge.Transfer learning can complete knowledge transfer between different domains and tasks,enabling edge servers to use their prior knowledge to solve new problems,thereby saving storage overhead.This work was supported by the Beijing Natural Science Foundation under Grant No.4202049.Facing the need of deploying multiple intelligent tasks under the cloud-edge collaboration framework,this paper improves the generalization of the feature extraction network in the edge server based on transfer learning and proposes optimization algorithms for the multi-domain image recognition task and target detection task respectively to save the edge server.resource overhead.The main research contents of the paper are as follows:(1)Systematically sort out and summarize the related research on the technology involved in this paper.Firstly,the computer vision applications in the Internet of Things are summarized,the requirements for the deployment of various intelligent tasks in the Internet of Things are analyzed,and the model basis of image recognition and target detection tasks is summarized.Then,the cloud-edge collaboration framework in the Internet of Things and its development and challenges are summarized,and the research direction of this paper is clarified.Finally,the related concepts and metrics of transfer learning are summarized,the research status of different types of transfer learning algorithms is summarized,and the advantages and disadvantages of their application in the Internet of Things are analyzed,and the key issues that this paper focuses on in algorithm design are pointed out.(2)Aiming at the problem that the feature extraction ability deteriorates due to the different data distribution when the edge server processes the data of multiple domains,this paper proposes a crossdomain transfer algorithm based on knowledge distillation to realize the multi-domain image recognition task.Compared with existing methods,this paper designs a feature extraction network model that can share parameters,thereby reducing storage overhead.Then,based on the knowledge distillation algorithm,the feature extraction network is pruned and fine-tuned with the optimization goals of recognition accuracy,computational complexity,and transmission data volume,and the resource consumption of the edge server is further reduced on the premise of ensuring the task performance.Finally,an experimental analysis is performed to evaluate the advantages of the algorithm compared with other algorithms.(3)Aiming at the storage pressure faced by the edge server for feature extraction of different tasks,and the problem that the detection ability decreases when the target detection task label is insufficient,this paper proposes a target detection algorithm that jointly considers the cross-task transfer of shallow-deep semantic information,migrating from image recognition to object detection tasks with less memory overhead.First,a parameter sharing layer partitioning mechanism is proposed to migrate shallow features to save storage resources.Then,the semantic distance between tasks is measured based on the deep semantic relationship as a soft label loss to improve the detection performance of the target detection algorithm when the labels are sparse.Finally,compare with other algorithms for experimental analysis to evaluate the advantages of the algorithm in parameter storage and detection performance.