| YOLOv10 的轻量化改进及其在联邦学习中的应用 |
| 点此下载全文 |
| 引用本文：杜孟林,胡小明,白双杰.YOLOv10 的轻量化改进及其在联邦学习中的应用[J].上海第二工业大学(中文版),2026,43(1):73-80 |
| 摘要点击次数: 312 |
| 全文下载次数: 19 |
|
|
| 中文摘要:针对车联网 (Internet of Vehicle, IoV) 环境下的实时目标检测中资源有限且隐私易泄露的问题, 本研究提出了一种联邦学习赋能的轻量化 YOLOv10 框架 (FL-YOLOv10)。该框架通过在 YOLOv10 中引入稀疏感知 ECA 模块与深度可分离重构的轻量化 CBAM 注意力机制模块, 针对性优化数据异构场景的特征交互能力; 使用 FedAvgM作为联邦聚合方案, 结合 YOLO 模块中针对边缘设备算力受限与数据异构性的改进, 显著降低了通信负载; 采用RSA/AES 混合加密实现前向保密, 确保隐私数据安全。在 KITTI 与 nuImages 数据集上的实验结果表明, 改进后的YOLOv10n 模型在 mAP@0.5 和 mAP@[0.5 : 0.95] 上分别达到 85.1% 和 62.5%, 比原版 YOLOv10n 模型分别提升了2.1% 和 2.5%; 精确率和召回率均提升 1.5%; 模型体积从 5.8 × 106 下降至 3.5 × 106。此外, 结合联邦学习框架后的FL-YOLOv10 相比集中式学习仅下降 0.5% 的精度, 但通信量减少 90%, 同时能有效解决隐私数据泄露的问题。 |
| 中文关键词:目标检测 联邦学习 YOLOv10 轻量化 |
| |
| Research on a Federated Learning-Based Lightweight YOLOv10 Object Detection Framework for Vehicular Networks |
|
|
| Abstract:To address the challenges of limited resources and privacy leakage in real-time object detection for Internet of Vehicle (IoV), this study proposes a federated learning-empowered lightweight YOLOv10 framework (FL-YOLOv10). The framework introduces a sparsity-aware ECA module and a lightweight CBAM attention mechanism with depthwise separable reconstruction into YOLOv10, specifically optimizing feature interaction capabilities for data-heterogeneous scenarios. FedAvgM is adopted as the federated aggregation strategy, combined with YOLO’s architectural improvements tailored for edge devices’ computational constraints and data heterogeneity, significantly reducing communication overhead. A hybrid RSA/AES encryption scheme ensures forward secrecy to protect private data. Experimental results on the KITTI and nuImages datasets demonstrate that the enhanced YOLOv10n model achieves 85.1% in mAP@0.5 and 62.5% in mAP@[0.5 : 0.95], outperforming the original YOLOv10n by 2.1% and 2.5%, respectively. Both precision and recall improve by 1.5%, while the parameter count decreases from 5.8 × 106 to 3.5 × 106. Furthermore, when integrated into the federated learning framework, FL-YOLOv10 incurs only a 0.5% accuracy drop compared to centralized training but reduces communication costs by 90% and effectively mitigates privacy leakage risks. |
| keywords:object detection federated learning YOLOv10 lightweight |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|