| 摘 要: 随着遥感对地观测技术的发展，面向需要隐私保护的遥感数据，传统的集中式训练面临挑战。为打破数据孤岛，利用多源异构数据提升目标检测模型的泛化能力，本文提出一种名为 FedGS 的联邦学习框架，首次将梯度投影策略与旋转目标检测结合，解决遥感异构数据的联邦学习检测难题，区别于通用的联邦目标检测研究。并将轻量化共享卷积检测头用于改进目标检测算法,计算量降低 13.6%、参数量降低 8.3%。首先，部署基于改进yolon11-obb的本地检测器，通过组归一化与共享卷积参数降低边缘通信与计算开销；其次，提出异构梯度投影策略。实验结果表明，在异构场景下，FedGS的检测精度 mAP@50相比经典的 FedAvg、FedOpt、FedProx 提升了2.0%~9.67%、2.92%~10.67%、12.95%~15.64%,在类别分布不均衡的客户端中提升效果显著。 |
| 关键词: 联邦学习 遥感解译 目标检测 异构数据 |
|
中图分类号:
文献标识码:
|
|
| Heterogeneous Remote Sensing Object Detection Algorithm for Federated LearningWang H L1a , Xu Z Y1b? |
|
wang honglin1, xu zhouyue2
|
1.Nanjing University of Information Science and Technology School of Artificial Intelligence (School of Future Technology);2.Nanjing University of Information Science and Technology School of Computer Science
|
| Abstract: With the advancement of remote sensing earth observation technology, traditional centralized training is confronted with significant challenges when dealing with remote sensing data requiring privacy protection. To break down data silos and leverage multi-source heterogeneous data to enhance the generalization capability of object detection models, this paper proposes a federated learning framework named FedGS. For the first time, it integrates the gradient projection strategy with rotated object detection, addressing the dilemma of federated learning-based detection for heterogeneous remote sensing data and distinguishing itself from general federated object detection research. A lightweight shared convolutional detection head is adopted to improve the object detection algorithm, which reduces the computational complexity by 13.6% and the number of parameters by 8.3%. First, a local detector based on the improved YOLO11-OBB is deployed, which cuts down edge communication and computational overhead through group normalization and shared convolution parameters. Second, a heterogeneous gradient projection strategy is put forward. Experimental results demonstrate that in heterogeneous scenarios, the detection accuracy (mAP@50) of FedGS has improved by 2.0%–9.67%, 2.92%–10.67%, and 12.95%–15.64% compared with the classic FedAvg, FedOpt, and FedProx algorithms, respectively, with remarkable enhancement effects observed on clients with imbalanced category distribution. |
| Keywords: federated learning remote sensing interpretation object detection heterogeneous data |
