Surface morphology segmentation and evaluation of diamond lapping pad based on improved Mask R-CNN
-
摘要: 金刚石磨盘广泛应用于各类硬脆材料的磨削加工中，磨盘表面形态对加工工件质量与磨盘磨削性能有着直接的影响。为对磨盘表面形态进行检测，提出基于改进Mask R-CNN模型的分割方法对磨盘表面图像中的磨粒、气孔进行识别与分割，并对模型进行训练与验证。结果表明：使用该方法能够实现磨盘表面图像中磨粒、气孔的识别与分割，平均准确率为78.2%。为验证该方法分割的磨粒、气孔与实际结果的差异，提出目标数量识别准确率、目标分割面积准确率、目标位置误差3个参数来评价分割效果，结果表明：磨粒、气孔的数量识别准确率分别为82.1%与93.4%，分割面积准确率分别为89.9%与95.3%，位置误差分别为3.80%与2.80%，证明该方法有效。
-
关键词:
- 磨盘检测 /
- 深度学习 /
- 改进Mask R-CNN模型 /
- 分割评价
Abstract:Objectives The surface morphology of diamond lapping pads has a direct impact on the lapping quality of hard and brittle materials such as sapphire and silicon carbide. Detecting and controlling the surface morphology of diamond lapping pads is a crucial step in improving the lapping quality. Collecting surface images of diamond lapping pads, which contain numerous tiny abrasive particles, pores, and complex background textures, makes it challenging to conduct quantitative detection of their surface morphology. An improved Mask R-CNN model and three parameter evaluation indicators, namely the target number recognition accuracy, target segmentation area accuracy, and target position error, are utilized to explore the segmentation performance and effect of the proposed model. Methods Based on the Mask R-CNN model, the dilated convolution method is adopted to improve the feature extraction network in the model's backbone. The ResNet50, which serves as the feature extraction network in the backbone, is divided into five stages: Input stem and stages 1 to 4. The network structures of the Input stem, stages 1, 2, and 4 are kept unchanged. Dilated convolution is introduced in stage 3, and each residual block in stage 3 is improved into a residual block using dilated convolution to expand the receptive field, enhance the model's ability to extract deep semantic features of abrasive particles and pore targets of smaller scales on the surface of the lapping pad, and improve segmentation performance for abrasive grain and pore targets of different scales. The loss function and mean average precision (mAP) of Mask R-CNN are used to comprehensively reflect the performance of the model. For evaluation of the segmentation effect, three parameters, namely target number recognition accuracy, target segmentation area accuracy, and target position error, are proposed. These are mainly calculated based on the number, area, and center of abrasive particles and pores, and evaluating diamond abrasive particles and pores separately to assess the overall surface morphology of the lapping pad. Results Through training and verification of the improved Mask R-CNN model, results show that this method can realize the recognition and segmentation of diamond abrasive particles and pores in the surface images of the lapping pad, achieving an mAP of 78.2%. By comparing the surface images of the lapping pad with the model segmentation images, there is no significant difference between the number of diamond abrasive particles and pores obtained by the improved Mask R-CNN model and the actual numbers, indicating that this method effectively recognizes and segments diamond abrasive particles and pores. Comparing the model's segmentation results with manually annotated results and calculating the three evaluation indicators, the recognition accuracies for the number of diamond abrasive particles and pores are 82.1% and 93.4%, respectively. This is due to the complex background of the lapping pad's surface images and unclear contrast between abrasive particles, pores, and binders, which can cause some missed or false detections when using this method. For successfully identified diamond abrasive grain and pore targets, the segmentation area accuracies are 89.9% and 95.3%, respectively, indicating small differences between segmented abrasive areas and actual areas, with a high degree of agreement, and good classification and segmentation performance by the model. By comparing the contours of diamond abrasive particles and pores obtained by model segmentation with the actual contours, the position errors are 3.80% and 2.80%, respectively, indicating small differences between the segmented and actual contours, and demonstrating good segmentation accuracy. Conclusions The dilated convolution method can effectively expand the receptive field and improve the ability to extract deep semantic features of targets at different scales. Therefore, based on the comparison between the segmentation images of the improved Mask R-CNN model and manually annotated images and the evaluation of the three indicators, the improved Mask R-CNN model demonstrates good segmentation performance for diamond abrasive particles and pores of different scales on the lapping pad surface, proving the effectiveness of the segmentation method. -
图 2 Mask R-CNN模型改进方法示意图
Figure 2. Schematic diagram of Mask R-CNN model improvement method
图 4 磨盘表面磨粒和气孔图像
Figure 4. Images of abrasive particles and pores on surface of lapping pad
图 8 磨盘表面图像与模型识别分割图像
Figure 8. Lapping pad surface image and model recognition segmentation image
图 9 人工标注图像与模型识别分割图像对比
Figure 9. Comparison of manually labeled image and model recognition segmentation image
图 10 人工标注图像与模型分割图像对比
Figure 10. Comparison of manually labered image and model segmentation image
图 11 2种方法分割磨粒数量对比与磨粒数量识别准确率
Figure 11. Comparison of number of abrasive particles segmented by two methods and accuracy of number of abrasive particles recognition
图 12 2种方法分割气孔数量对比与气孔数量识别准确率
Figure 12. Comparison of number of pores segmented by two methods and accuracy of recognition of number of pores
图 13 2种方法分割磨粒面积对比与磨粒分割面积准确率
Figure 13. Comparison of two methods to segment abrasive particle area and accuracy of abrasive particle segmentation area
图 14 2种方法分割气孔面积对比与气孔分割面积准确率
Figure 14. Comparison of pore area and accuracy of pore segmentation area by two methods
图 15 2种方法分割的磨粒、气孔形心图
Figure 15. Abrasive particle and pore centroids segmented by two methods
-
[1] 邓朝晖, 伍俏平, 张高峰, 等.新型砂轮研究进展及其展望 [J]. 中国机械工程, 2010, 21(21): 2632-2638, 2645.DENG Zhaohui, WU Qiaoping, ZHANG Gaofeng, et al. Recent advances and future perspectives in new type grinding wheels [J]. China Mechanical Engineering, 2010, 21(21): 2632-2638, 2645. [2] HADAD M, SHARBATI A. Thermal aspects of environmentally friendly-MQL grinding process [J]. Procedia CIRP,2016,40:509-515. doi: 10.1016/j.procir.2016.01.125 [3] MARKOPOULOS A P. Finite element method in machining processes [M]. London: Springer, 2013. [4] 徐西鹏, 黄辉, 胡中伟, 等. 磨粒工具的研究现状及发展趋势 [J]. 机械工程学报,2022,58(15):2-20. doi: 10.3901/JME.2022.15.002XU Xipeng, HUANG Hui, HU Zhongwei, et al. Development of abrasive tools: State-of-the-art and prospectives [J]. Journal of Mechanical Engineering,2022,58(15):2-20. doi: 10.3901/JME.2022.15.002 [5] 陈光军, 韩松鑫, 潘佳琦, 等. 切削加工表面粗糙度影响因素及预测建模综述 [J]. 机床与液压,2020,48(13):185-188. doi: 10.3969/j.issn.1001-3881.2020.13.040CHEN Guangjun, HAN Songxin, PAN Jiaqi, et al. Review on the influence factors and predictive modeling of surface roughness in machining [J]. Machine Tool & Hydraulics,2020,48(13):185-188. doi: 10.3969/j.issn.1001-3881.2020.13.040 [6] FAN K C, LEE M Z, MOU J I. On-line non-contact system for grinding wheel wear measurement [J]. The International Journal of Advanced Manufacturing Technology,2002,19(1):14-22. doi: 10.1007/PL00003964 [7] LACHANCE S, WARKENTIN A, BAUER R. Development of an automated system for measuring grinding wheel wear flats [J]. Journal of Manufacturing Systems,2003,22(2):130-135. doi: 10.1016/S0278-6125(03)90010-0 [8] 杨栖凤, 崔长彩, 黄国钦. 金刚石砂轮表面二维形貌全场测量和分析 [J]. 华侨大学学报(自然科学版),2018,39(4):479-484. doi: 10.11830/ISSN.1000-5013.201711013YANG Qifeng, CUI Changcai, HUANG Guoqin. Measurement and analysis of two-dimensional surface topography of whole grinding wheel [J]. Journal of Huaqiao University(Natural Science),2018,39(4):479-484. doi: 10.11830/ISSN.1000-5013.201711013 [9] 刘明宇, 佃松宜. 基于机器视觉的金刚线表面质量检测 [J]. 四川大学学报(自然科学版),2020,57(5):920-926. doi: 10.3969/j.issn.0490-6756.2020.05.015LIU Mingyu, DIAN Songyi. Surface quality detection of diamond wire based on machine vision [J]. Journal of Sichuan University(Natural Science Edition),2020,57(5):920-926. doi: 10.3969/j.issn.0490-6756.2020.05.015 [10] 赵玉康, 毕文波, 葛培琪. 电镀金刚石线锯表面磨粒分布密度的多相机视觉检测 [J]. 金刚石与磨料磨具工程,2021,41(2):64-68. doi: 10.13394/j.cnki.jgszz.2021.2.0011ZHAO Yukang, BI Wenbo, GE Peiqi. Multi-camera visual inspection of abrasives distribution density on electroplated diamond wire saw surface [J]. Diamond & Abrasives Engineering,2021,41(2):64-68. doi: 10.13394/j.cnki.jgszz.2021.2.0011 [11] YOU F Y, ZHOU W, WANG X, et al. Systematic monitoring and evaluating the wear of alumina wheel when grinding the workpiece of Cr12 [J]. Complexity,2021(1):6665043. doi: 10.1155/2021/6665043 [12] 李弘扬, 方从富. 基于K-Means聚类与凸包检测的金刚石磨粒分割与评价 [J]. 金刚石与磨料磨具工程,2023,43(2):188-195. doi: 10.13394/j.cnki.jgszz.2022.0099LI Hongyang, FANG Congfu. Segmentation and evaluation of diamond abrasive grains based on K-Means clustering and convex hull detection [J]. Diamond & Abrasives Engineering,2023,43(2):188-195. doi: 10.13394/j.cnki.jgszz.2022.0099 [13] 胡伟栋, 王占奎, 董彦辉, 等. 基于深度学习的固结磨料研磨垫表面形态表征 [J]. 金刚石与磨料磨具工程,2022,42(2):186-192. doi: 10.13394/j.cnki.jgszz.2021.0096HU Weidong, WANG Zhankui, DONG Yanhui, et al. Surface morphology characterization of fixed abrasive lapping pad based on deep learning [J]. Diamond & Abrasives Engineering,2022,42(2):186-192. doi: 10.13394/j.cnki.jgszz.2021.0096 [14] 谢运鸿, 孙钊, 丁志丹, 等. 基于Mask R-CNN和迁移学习的无人机遥感影像杉木单木树冠提取 [J]. 北京林业大学学报,2024,46(3):153-166. doi: 10.12171/j.1000-1522.20210343XIE Yunhong, SUN Zhao, DING Zhidan, et al. UAV remote sensing image extraction of single tree crown of Chinese fir based on Mask R-CNN and transfer learning [J]. Journal of Beijing Forestry University,2024,46(3):153-166. doi: 10.12171/j.1000-1522.20210343 [15] 张丽媛, 赵海蓉, 何巍, 等. 融合全局-局部注意模块的Mask R-CNN膝关节囊肿检测方法 [J]. 图学学报,2023,44(6):1183-1190. doi: 10.11996/JG.j.2095-302X.2023061183ZHANG Liyuan, ZHAO Hairong, HE Wei, et al. Knee cysts detection algorithm based on Mask R-CNN integrating global-local attention module [J]. Journal of Graphics,2023,44(6):1183-1190. doi: 10.11996/JG.j.2095-302X.2023061183 [16] 杨智宏, 贺石中, 冯伟, 等. 基于Mask R-CNN网络的磨损颗粒智能识别与应用 [J]. 摩擦学学报,2021,41(1):105-114. doi: 10.16078/j.tribology.2020020YANG Zhihong, HE Shizhong, FENG Wei, et al. Intelligentidentification of wear particles based on Mask R-CNN network and application [J]. Tribology,2021,41(1):105-114. doi: 10.16078/j.tribology.2020020 [17] 肖潇. 基于深度学习的遥感图像处理系统的设计与实现 [D]. 北京: 北京邮电大学, 2019.XIAO Xiao. Design and implementation of remote semsing image processing system based on deep learning [D]. Beijing: Beijing University of Posts and Telecommunications, 2019. [18] AMIRGAN B, ERENER A. Semantic segmentation of satellite images with different building types using deep learning methods [J]. Remote Sensing Applications: Society and Environment,2024,34:101176. doi: 10.1016/j.rsase.2024.101176 [19] JIA P Y, CHEN C, ZHANG D L, et al. Semantic segmentation of deep learning remote sensing images based on band combination principle: Application in urban planning and land use [J]. Computer Communications,2024,217:97-106. doi: 10.1016/j.comcom.2024.01.032 -
下载:
点击查看大图
计量
- 文章访问数: 557
- HTML全文浏览量: 312
- PDF下载量: 0
- 被引次数: 0
邮件订阅
RSS
