| 董俊杰1,2， 任志俊1,2*， 苏晨1,2， 王琨1,2.基于深度学习的渗碳齿轮金相图像分割算法[J].轻工机械,2024,42(5):66-73 |
| 基于深度学习的渗碳齿轮金相图像分割算法 |
| Carburized Gear Metallographic Image Segmentation Algorithm Based on Deep Learning |
| |
| DOI：10.3969/j.issn.1005 2895.2024.05.009 |
| 中文关键词: 渗碳齿轮 金相图 区域卷积神经网络RCNN Swin Transformer神经网络模型 图像分割 深度学习 |
| 英文关键词:〗carburized gear metallographic image RCNN（region convolutional neural network） Swin Transformer neural network image segmentation deep learning |
| 基金项目:江苏省科技支撑计划（工业）项目重点项目（BE2020006 5）；2023年度无锡市科学技术协会软科学研究课题（KX 23 B006）。 |
|
| 摘要点击次数: 22 |
| 全文下载次数: 12 |
| 中文摘要: |
| 针对渗碳齿轮的质检依赖人工观察判断显微组织图像而造成不确定性较大的问题，课题组基于掩膜区域卷积神经网络Mask RCNN（region based convolutional neural network）提出了STN Mask RCNN模型，对渗碳齿轮中的残余奥氏体与马氏体进行图像分割，将Mask RCNN的主干特征提取网络替换为Swin Transformer模块，引入了神经架构检索(neural architecture search，NAS)算法与特征金字塔网络(feature pyramid network，FPN)相结合的NAS FPN模块，并在Mask图像分割分支中加入卷积块注意力模块 (convolutional block attention module，CBAM)，最后与DeepLabV3[KG-*3]+模型和U Net模型进行对比实验，并进行消融实验分析每个模块与网络性能之间的关系。实验结果表明：课题组提出的模型对渗碳齿轮中残余奥氏体与马氏体的图形分割能力较强，平均像素精度(mean pixel accuracy，mPA)达90.64%，整体性能明显优于其他模型，且各个模块对于模型性能都有不同程度的提升。 |
| 英文摘要: |
| Aiming at the problem of significant uncertainty in the quality inspection of carburized gears caused by manual observation of microstructure tissue images, the STN Mask RCNN model based on Mask RCNN to segment residual austenite and martensite in carburized gears was proposed. The backbone feature extraction network of Mask RCNN was replaced with Swin Transformer, and the NAS FPN module combining FPN and neural retrieval algorithm were introduced, and the CBAM attention mechanism was added in the Mask image segmentation branch. Finally, compared the model with DeepLabV3+ and U Net models, and performed ablation experiments to analyze the relationship between each variable and network performance. The experiments show that the proposed model has strong segmentation capabilities for residual austenite and martensite in carburized gears, with an mean pixel accuracy of 90.64%. The overall performance is significantly better than other model structures, and each module contributes to the improvement of the model performance to varying degrees. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
