«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年08期

页码:

189-196

栏目:

新型计算应用系统

出版日期:

2024-08-10

文章信息/Info

Title:

A YOLO-based Federated Learning Network for Container Lock Pin Classification

文章编号:

1673-629X(2024)08-0189-08

作者:

王漫1; 李培剀1; 2; 熊勇2*

1. 上海第二工业大学 计算机与人工智能学院,上海 201209; 2. 中国科学院 上海微系统与信息技术研究所,上海 201899

Author(s):

WANG Man1; LI Pei-kai1; 2; XIONG Yong2*

1. School of Computer and Artificial Intelligence,Shanghai Polytechnic University,Shanghai 201209,China; 2. Shanghai Institute of Microsystem and Information Technology,Chinese Academy of Sciences,Shanghai 201899,China

关键词:

锁销图像分类; 联邦学习; 神经网络; 分布式架构; 早停

Keywords:

lock pin image classification; federated learning; neural network; distributed architecture; early stopping

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0144

摘要:

集装箱锁销自动拆卸任务是实现无人码头的最后一个技术瓶颈,锁销图像分类实时结果用于后继机械手启用适 配的拆卸装置和动作程序,是自动化拆卸任务的关键环节。 丰富多样的锁销数据集有利于保障锁销分类神经网络的鲁棒性。 然而,由于锁销使用具有商业敏感性、图像网络传输开销大等原因,锁销使用方如码头往往不将其拥有的锁销图像数据与他方共享,而是各自迭代自己的神经网络模型。 这导致各使用方的自有模型对其之前少用的锁销的分类准确率较低,极易在拆卸时引发故障,从而影响无人化操作效率。 由于传统的中心式学习无法解决上述问题,提出了一种基于YOLO 的锁销分类联邦学习网络。 首先,在 YOLOv8 的基础上建立锁销分类神经网络;其次,基于 Flower 框架建立分布式的联邦学习架构,并提出了改进的 FedAvg-mAP 算法,提高了聚合后的全局模型的性能。 此外,在本地模型训练阶段引入早停策略,加速了全局模型的收敛,且使收敛过程更加平滑。实验表明,提出的基于 YOLO 的锁销分类联邦学习网络能在图像数据不共享的前提下实现传统的中心式学习的功能,且改进的 FedAvg-mAP 算法相比传统算法具有更好的性能。

Abstract:

The task of automated container lock pin disassembly is the last technical bottleneck in realizing unmanned docks. The real-time results of lock pin image classification are used for succeeding mechanical arms to initiate the adapted disassembly device and action program,which is the key link in the automated disassembly task. A rich and diverse lock pin dataset is conducive to ensuring the robustness of the lock pin classification neural network. However,due to the commercial sensitivity of lock pin usage and excessive trans-mission cost of images,lock pin users such as docks do not usually share their own lock pin image data with others,but iterate their own neural network models. This leads to the fact that each user's own model is less accurate in the classification of previously rare lock pins,which is highly susceptible to failure and has an impact on the efficiency of unmanned operations. Since traditional center-based learning cannot solve the problem above,we propose a YOLO-based federated learning network for lock pin classification. Firstly,a lock pin classification neural network is established on the basis of YOLOv8. Secondly, a distributed federated learning architecture is established based on Flower framework, and an improved algorithm, which is called FedAvg - mAP, is proposed to improve the performance of the aggregated global model. In addition,an early stopping strategy is introduced in the local model training phase,which accelerates the convergence of the global model and makes the convergence process smoother. Experiments show that the YOLO-based federated learning architecture for lock pin classification proposed can realize the function of traditional centralized learning without image data sharing,and the improved FedAvg-mAP algorithm has better performance over the traditional algorithm.

常用功能

更新日期/Last Update: 2024-08-10