摘要
深度学习的目标检测模型主要应用于自然场景,若将其直接应用于工业场景,由于其数据样本不足,有可能导致对领域目标特征提取不够充分,模型检测精度不理想。为提升磨粒缺陷的检测精度,以Cascade R-CNN模型作为基线,提出了一种二维手工特征与深度学习相融合的磨粒缺陷检测模型。采用双向多尺度特征融合模块BiCE-FPN,解决目标尺度不一问题;采用梯度调和分类损失GHMC,解决缺陷样本不均衡问题;通过融合磨粒缺陷的二维纹理、梯度手工特征与深度特征,解决磨粒缺陷样本不足造成的深度学习特征抽取不够充分问题。研究结果表明:双向多尺度特征融合BiCE-FPN、梯度调和分类损失GHMC策略和特征融合均对提升模型的检测精度起到了积极作用。
The object detection model of deep learning is mainly applied in natural scenarios.If it is directly applied to industrial scenarios,due to the insufficient data samples,it may lead to insufficient extraction of domain object features and unsatisfactory model detection accuracy.To improve the detection accuracy of wear debris defects,taking the Cascade R-CNN model as the baseline,a wear debris defects detection model integrating two-dimensional manual features and deep learning is proposed.The bidirectional multi-scale feature fusion module BiCE-FPN is adopted to solve the problem of inconsistent target scales.The gradient harmonic classification loss(GHMC)is adopted to solve the problem of imbalanced defect samples.By integrating the two-dimensional texture,gradient manual features and depth features of wear debris defects,the problem of insufficient deep learning feature extraction caused by insufficient wear debris defects samples is solved.The research results show that the bidirectional multi-scale feature fusion BiCE-FPN,the gradient harmonic classification loss GHMC strategy and feature all play a positive role in improving the detection accuracy of the model.
作者
侯志昌
杨灏瀛
汪红兵
Hou Zhichang;Yang Haoying;Wang Hongbing(Shanghai Anchor Science and Technology Co.,Ltd.,Shanghai 200433,China;School of Computer and Communication Engineering,University of Science and Technology Beijing,Beijing 100083,China)
出处
《河北冶金》
2025年第6期46-52,共7页
Hebei Metallurgy
关键词
深度学习
磨粒缺陷
手工设计特征
特征融合
检测精度
deep learning
wear debris defects
handcrafted features
feature fusion
detection accuracy
