摘要
钢材表面缺陷检测任务中,YOLO将目标检测转换为对位置信息的回归问题,实现高帧率实时检测,但对小目标缺陷定位精度有所欠缺。针对该问题,以YOLOv5s架构为基础,首先,在模型输入端设定动态尺度训练范式,提高小目标缺陷训练精度;其次,设计STD-CA模块利用图像转换技术,避免下采样过程中分辨率的降低,导致小目标缺陷特征信息的丢失,并引导特征提取能力,降低无关背景特征关注度,进一步提高模型小目标缺陷检测精度。结果表明,在NEU-DET数据集中,改进后模型在保证检测速度保持在54 frame/s的同时,平均精度均值达到86.6%,较YOLOv5s提高17.6%,对小目标缺陷定位更加准确,目前优于其他深度学习钢材实时检测模型。
In the task of steel surface defect detection,YOLO transforms object detection into a regression problem for position information,achieving high frame rate real-time detection.However,it lacks accuracy in locating small target defects.To address this issue,based on the YOLOv5s architecture,we first set a dynamic scale training paradigm at the model input end to improve the training accuracy for small target defects.Secondly,we designed the STD-CA module to use image transformation techniques to avoid the loss of small target defect feature information due to the decrease in resolution during downsampling.This module also guides feature extraction capability,reduces the attention of irrelevant background features,and further improves the accuracy of small target defect detection in the model.The results show that,on the NEU-DET dataset,the improved model maintains a detection speed of 54 frames/s while achieving an average precision of 86.6%,which is 17.6%higher than YOLOv5s.The model locates small target defects more accurately and currently outperforms other deep learning-based real-time detection models for stee.
作者
朱传军
梁泽启
付强
张超勇
刘荣光
ZHU Chuanjun;LIANG Zeqi;FU Qiang;ZHANG Chaoyong;LIU Rongguang(School of Mechanical Engineering,Hubei University of Technology,Wuhan 430068,China;State Key Lab of Digital Manufacturing Equipment&Technology,Huazhong University of Science and Technology,Wuhan 430074,China)
出处
《组合机床与自动化加工技术》
北大核心
2023年第11期133-137,共5页
Modular Machine Tool & Automatic Manufacturing Technique
基金
国家自然科学基金国际(地区)合作与交流项目(51861165202)
广东省重点领域研发计划项目(2019B090921001)。