电子测量与仪器学报

赵佰亭,吴俊东,贾晓芬.融合特征增强的轻量化罐道缺陷检测算法[J].电子测量与仪器学报,2023,37(6):159-168

融合特征增强的轻量化罐道缺陷检测算法

Lightweight guide defect detection algorithm based on feature enhancement

DOI：

英文关键词:guide defect feature fusion target detection YOLOv5s lightweight

基金项目:国家自然科学基金面上项目( 52174141)、安徽省自然科学基金面上项目( 2108085ME158)、安徽高校协同创新项目(GXXT-2020-54)、安徽省重点研究与开发计划项目(202104a07020005)资助

作者	单位
赵佰亭	1. 安徽理工大学电气与信息工程学院
吴俊东	1. 安徽理工大学电气与信息工程学院
贾晓芬	1. 安徽理工大学电气与信息工程学院，2. 安徽理工大学省部共建深部煤矿采动响应与灾害防控国家重点实验室

Author	Institution
Zhao Baiting	1. China Institute of Electrical and Information Engineering, Anhui University of Science and Technology
Wu Jundong	1. China Institute of Electrical and Information Engineering, Anhui University of Science and Technology
Jia Xiaofen	1. China Institute of Electrical and Information Engineering, Anhui University of Science and Technology，2. State Key Laboratory of Mining Response and Disaster Prevention and Control in Deep Coal Mines, Anhui University of Science and Technology

摘要点击次数: 1071

全文下载次数: 1160

中文摘要:

为了解决矿井复杂环境下,缺陷特征提取不充分问题,融合特征增强和级联注意力机制提出一种快速智能的罐道缺陷识别算法 RDM-YOLOv5,旨在解决人工巡检效率低的现状。首先,为了提高主干网络特征图信息表征能力,设计特征增强模块 RLKM,它通过重参数化大内核卷积增强主干网络对目标特征的提取能力,并且有效降低模型参数量;然后,经过主干网络提取到高低层级特征后,由设计的级联注意力机制 DCAM 进一步挖掘缺陷目标的深层语义信息,显著增强小目标的特征信息;最后,为提升检测精度的同时保障检测网络的轻量化,在特征增强网络中引入轻量级卷积 GSConv,在保持模型检测准确性的同时降低计算成本。实验结果表明,相较于 YOLOv5s,RDM-YOLOv5 的检测精度和速度分别提高 3. 7%、11. 4%,模型参数量减少 15. 4%。它能基本满足实际应用中精准识别和快速定位罐道表面缺陷的需求。

英文摘要:

In order to solve the problem of insufficient defect feature extraction in the complex mine environment, a fast and intelligent guide defect recognition algorithm RDM-YOLOv5 is proposed by integrating feature enhancement and cascading attention mechanism, aiming to solve the current situation of low efficiency of manual inspection. Firstly, in order to improve the information representation ability of the backbone network feature map, the feature enhancement module RLKM is designed, which enhances the backbone network’s ability to extract target features through reparameterized large kernel convolution, and effectively reduces the amount of model parameters. Then, after extracting the high-level and low-level features through the backbone network, under the action of the cascaded attention mechanism DCAM composed of the designed channel attention mechanism and coordinate attention mechanism, the deep semantic information of the defective target is further mined, and the feature information of the small target are significantly enhanced. Finally, in order to improve the detection accuracy while ensuring the lightweight of the detection network, a lightweight convolution GSConv is introduced into the feature enhancement network to reduce the computational cost while maintaining the accuracy of model detection. The experimental results show that compared with YOLOv5s, the detection accuracy and speed of RDM-YOLOv5 are increased by 3. 7% and 11. 4%, respectively, and the number of model parameters is reduced by 15. 4%. It can basically meet the needs of accurate identification and rapid location of guide surface defects in practical applications.

查看全文查看/发表评论下载PDF阅读器