A Federated Learning-Based Object Detection System with Edge–Cloud Collaboration

Abstract

With the development of smart agriculture, field object detection has become a key component in improving management efficiency. This paper proposes and implements a cloud-edge collaborative federated learning system for object detection, using a field weed dataset as a case study. The system employs a federated learning framework to enable parallel training across multiple edge devices and integrates the YOLOv11 model to perform collaborative training on several Jetson Nano devices. To enhance global model performance and convergence speed, only model parameters that meet predefined accuracy thresholds and pass a target filtering mechanism are uploaded to the cloud for aggregation. On the client side, a PyQt5-based graphical user interface is developed to support inference on images, videos, and real-time camera feeds.In the experimental evaluation, we compared federated training across multiple edge devices with centralized training on a single device. The results show that federated learning outperforms centralized training across several key performance metrics. Specifically, precision improved from 77.89% to 82.17%, recall increased from 72.25% to 73.56%, and mean average precision (mAP) rose from 77.82% to 81.45%. These findings demonstrate that federated learning can significantly enhance model accuracy and generalization while keeping data localized.