面向少样本及小目标环境下接触网零部件缺陷检测模型
于小四1韦宝泉2李泽文2邓芳明2
Defect Detection Model for Catenary Components in Environments with Few Samples and Small Targets
YU Xiaosi1WEI Baoquan2LI Zewen2DENG Fangming2
-
作者信息:1.中铁七局集团有限公司，450007，郑州
2.华东交通大学电气与自动化工程学院，330013，南昌
-
Affiliation:1.China Railway Seventh Bureau Group Co., Ltd., 450007, Zhengzhou, China
2.School of Electrical and Automation Engineering, East China Jiaotong University, 330013, Nanchang, China
-
关键词:
-
Key words:
-
DOI:10.16037/j.1007-869x.2024.12.015
-
中图分类号/CLCN:U225.4
-
栏目/Col:学术专论
摘要:
［目的］针对少样本条件下轨道交通接触网小体积零部件检测困难问题，提出一种融合生成对抗网络和深度分割模型的缺陷检测方法。［方法］介绍了检测系统；介绍了改进的DCGAN(深度卷积生成对抗网络)结构和改进的YOLACT(单阶段条件对抗性网络)模型，并基于实际缺陷数据集验证缺陷检测效果。［结果及结论］在样本扩充部分，改进的DCGAN模型能够通过引入若干量纲一化操作，增加高效通道注意力机制，提高生成样本的质量。在模型检测部分，采用改进的ResNeXt-FPN(残差网络与特征金字塔结合的深度卷积神经网络)结构代替原始YOLACT模型的主干网络，以充分表征目标多尺度特征。在掩码分支网络中引入CS(通道和空间)注意力机制，能够有效提高小体积零部件的检测精度。所提缺陷检测模型能够在复杂接触网图像中实现开口销钉的高精度检测，其检测精度和召回率分别高达88.63%和87.49%。相比于原始YOLACT模型，所提缺陷检测模型的综合性能提升约6.2%。
Abstracts:
［Objective］ Aiming at the difficulties of detecting small-volume parts of rail transit catenary under few-sample conditions, a defect detection method integrating generative adversarial networks and deep segmentation models is proposed. ［Method］ The composition of the detection system, the improved DCGAN (deep convolutional generative adversarial network) structure and the improved YOLACT (single-stage conditional adversarial network) model are introduced, and the defect detection effect is verified based on the actual defect datasets. ［Result & Conclusion］ In the sample expansion part, the improved DCGAN model can improve the quality of generated samples by introducing several normalization operations and adding an efficient channel attention mechanism. In the model detection part, the improved ResNeXt-FPN (deep convolutional neural network combining residual network and feature pyramid network) structure is used to replace the backbone network of the original YOLACT model, aiming to fully characterize the multi-scale features of the target. The introduction of CS (channel and space) attention mechanism in the mask branch network can effectively improve the detection accuracy of small-volume parts. The proposed defect detection model can achieve high-precision detection of cotter pins in complex contact network images, with a detection accuracy and recall rate of up to 88.63% and 87.49% respectively. Compared with the original YOLACT model, the comprehensive performance of the proposed defect detection model is improved by about 6.2%.
- 上一篇: 钢弹簧浮置板浸水条件下的减振效果
- 下一篇: 复杂工况下受电弓升弓碰撞特性
过刊浏览
资料下载
友情链接