摘要
为提高化纤丝生产加工中出现的断线和疵点类缺陷的检测精度,对Faster RCNN算法进行改进。首先,在主干特征提取网络上加入可变形卷积模型,以提高网络对不同缺陷特征的适应性;其次,采用递归特征金字塔(Recursive Feature Pyramid,RFP)结构替换原特征金字塔网络(Feature Pyramid Network,FPN),进行二次特征提取;最后,改进损失函数,采用Rank&Sort Loss(RS Loss)函数替代原分类损失函数,解决化纤丝2类缺陷样本量相差较大问题。对比实验后得出,改进后的方法训练得到的mAP值为84.7%,较初始模型提高了4.3%,可以满足实际生产加工中对化纤丝缺陷的智能检测要求。
An improved Faster RCNN algorithm was proposed to detect the defects in production and processing of chemical fiber silk.Firstly,a deformable convolution model was added to the backbone feature extraction network to improve the adaptability of the network to different defect features.Secondly,the Feature Pyramid Network(FPN)was replaced by Recursive Feature Pyramid(RFP)structure to extract features twice.Finally,the RS Loss function was improved,and RS Loss function was used to replace the original classification loss function to solve the problem caused by imbalanced sample categories.Experiment result shows that the mAP value calculated by the improved model is 84.7%,which is 4.3%higher than original Faster RCNN model.The improved model can meet the requirements of intelligent detection on chemical fiber defects in practical production and processing.
作者
郭磊
王洋
靳正轩
陈朝新
陈江义
沈鹏
GUO Lei;WANG Yang;JIN Zhengxuan;CHEN Chaoxin;CHEN Jiangyi;SHEN Peng(College of Mechanical and Power Engineering,Zhengzhou University,Zhengzhou,Henan 450001,China;Henan Qice Electronics Technology Co.,Ltd.,Zhengzhou,Henan 450001,China)
出处
《毛纺科技》
CAS
北大核心
2023年第3期74-79,共6页
Wool Textile Journal
基金
三束材料改性教育部重点实验室开放课题(KF1801)。
作者简介
第一作者:郭磊,硕士生,主要研究方向为机器视觉及图像处理,E-mail:goulie0903@163.com;通信作者:沈鹏,讲师,博士,主要研究方向为机械制造及其自动化、机器视觉等,E-mail:shenpengmtr@163.com。
