化工进展 ›› 2021, Vol. 40 ›› Issue (12): 6499-6504.DOI: 10.16085/j.issn.1000-6613.2021-1387

• 专栏:多相流测试 • 上一篇    下一篇

基于深度学习的显微离焦图像法颗粒深度测量

徐日辛1,2(), 周骛1,2(), 张翔云1,2   

  1. 1.上海理工大学能源与动力工程学院,上海 200093
    2.上海市动力工程多相流动与传热重点实验室,上海 200093
  • 收稿日期:2021-07-01 修回日期:2021-08-08 出版日期:2021-12-05 发布日期:2021-12-21
  • 通讯作者: 周骛
  • 作者简介:徐日辛(1995—),男,硕士研究生,研究方向为图像法颗粒测量。E-mail:gd_xurixin@163.com
  • 基金资助:
    国家科技重大专项(2017-V-0016-0069);国家自然科学基金(51576130)

Particle depth position measurement using microscopic defocused imaging method based on deep learning

XU Rixin1,2(), ZHOU Wu1,2(), ZHANG Xiangyun1,2   

  1. 1.School of Energy and Power Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
    2.Shanghai Key Laboratory of Multiphase Flow and Heat Transfer in Power Engineering, Shanghai 200093, China
  • Received:2021-07-01 Revised:2021-08-08 Online:2021-12-05 Published:2021-12-21
  • Contact: ZHOU Wu

摘要:

显微成像条件下的三维流场测量是微通道流动等研究的基础,其难点在于颗粒深度位置的测量。由于显微镜头景深极小,成像时通道内大部分颗粒处于离焦状态。本文首先基于几何光学原理分析了显微成像前后离焦 不对称的特点,随后基于Inception V3卷积神经网络搭建了颗粒深度预测模型,并通过光线追踪方法生成粒径 1~10μm的10种颗粒在深度范围-50~50μm内的仿真显微图像,应用深度预测模型对其进行训练及预测,颗粒深度预测结果显示1~3μm颗粒的相对误差在±13%以内,4~10μm颗粒的相对误差小于±5%。最后在微通道中拍摄粒径分别为2.6μm和5μm的聚苯乙烯微球在深度范围-50~50μm内的显微图像,应用同一深度预测模型进行训练及预测,两种颗粒深度预测结果的相对误差分别为小于±15%和±5%。结果表明,所提出的基于深度学习的显微离焦图像法能够有效测量微通道内颗粒深度位置,为图像法流场测量技术增加了新的思路。

关键词: 卷积神经网络, 显微离焦, 深度位置测量, 微通道

Abstract:

Three-dimensional flow field measurement using microscopic imaging is the foundation of research in microchannel flow, and one difficulty lies in the measurement of particle depth position. Due to the limited depth of field of the microscope, most particles in the microchannel will be out of focus during imaging. In this paper, based on the principle of geometrical optics, the characteristics of asymmetric defocus was analyzed for microscopic imaging. A particle depth prediction model was built based on the Inception V3 convolutional neural network, and the simulated microscopic images of ten sizes of particles with diameters 1—10μm in the depth range of -50—50μm were generated by the optical ray tracing method. The particle depth prediction model was trained and used for prediction with these synthetic images. The results showed that the relative error of prediction for 1—3μm particles was within ±13%, and less than ±5% for 4—10μm particles. Finally, the microscopic images of polystyrene microspheres with sizes of 2.6μm and 5μm in the depth range of -50—50μm were captured in the microchannel, and the same depth prediction model was used for training and prediction. The relative errors of depth prediction for the two sizes of particles were less than ±15% and ±5%, respectively. The microscopic defocused imaging method based on deep learning can measure the depth position of particles in microchannel effectively, adding new ideas to the imaging method for flow field measurement technology.

Key words: convolutional neural networks (CNN), microscopic defocus, depth measurement, microchannel

中图分类号: 

京ICP备12046843号-2;京公网安备 11010102001994号
版权所有 © 《化工进展》编辑部
地址:北京市东城区青年湖南街13号 邮编:100011
电子信箱:hgjz@cip.com.cn
本系统由北京玛格泰克科技发展有限公司设计开发 技术支持:support@magtech.com.cn