当前位置: X-MOL 学术J. Vis. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Shape evaluation of highly overlapped powder grains using U-Net-based deep learning segmentation network
Journal of Visualization ( IF 1.7 ) Pub Date : 2021-04-12 , DOI: 10.1007/s12650-021-00748-0
Daehee Kwon , Eunseop Yeom

Abstract

With the advancement of electron microscopy, industrial microscale objects are analyzed through image-based characterization. However, the automated and objective assessment of a vast number of images required for quality control is limited by the incomplete segmentation of individual objects in the image. In this study, the scanning electron microscope images of powder grains are selected as target images representing industrial microscale objects. A deep neural network based on the U-Net is developed and trained by manually labeled ground truth. Although the U-Net is a basic network originally devised for biomaterials, the network in this study achieves approximately 90% accuracy and outperforms conventional thresholding methods. However, the boundaries distinguishing individual are not completely classified. The inference results are further processed with morphological operations and watershed algorithms to quantitatively measure grain shapes. Discrepancies in shape parameters between ground truth and network prediction are also discussed.

Graphic Abstract



中文翻译:

使用基于U-Net的深度学习分割网络评估高度重叠的粉末颗粒的形状

摘要

随着电子显微镜的发展,通过基于图像的表征分析了工业微尺度的物体。但是,质量控制所需的大量图像的自动和客观评估受到图像中单个对象的不完整分割的限制。在这项研究中,选择粉末颗粒的扫描电子显微镜图像作为代表工业微米级物体的目标图像。基于U-Net的深度神经网络是通过手动标记的地面事实开发和训练的。尽管U-Net是最初为生物材料设计的基本网络,但该研究中的网络可达到约90%的准确度,并且优于传统的阈值处理方法。但是,区分个人的边界并没有完全分类。推断结果将通过形态学运算和分水岭算法进行进一步处理,以定量地测量晶粒形状。还讨论了地面真实性和网络预测之间形状参数的差异。

图形摘要

更新日期:2021-04-12
down
wechat
bug