Citation: Luo, S.; Yang, K.; Yang, L.;
Wang, Y.; Gao, X.; Jiang, T.; Li, C.
Laser Curve Extraction of Wheelset
Based on Deep Learning Skeleton
Extraction Network. Sensors 2022, 22,
859. https://doi.org/10.3390/
s22030859
Academic Editor: Anastasios
Doulamis
Received: 29 November 2021
Accepted: 12 January 2022
Published: 23 January 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Laser Curve Extraction of Wheelset Based on Deep Learning
Skeleton Extraction Network
Shuai Luo
1
, Kai Yang
1,
*, Lijuan Yang
2
, Yong Wang
1
, Xiaorong Gao
1
, Tianci Jiang
3
and Chunjiang Li
4
1
School of Physical Science and Technology, Southwest Jiaotong University, Chengdu 610031, China;
yeluo@my.swjtu.edu.cn (S.L.); wangyonga@swjtu.edu.cn (Y.W.); gxrr@home.swjtu.edu.cn (X.G.)
2
School of Mathematics, Sichuan Normal University, Chengdu 610066, China; 20200892014@stu.sicnu.edu.cn
3
School of Mechanical Engineering, Waseda University, Kitakyushu 8080135, Japan; jiangtianci@ruri.waseda.jp
4
School of Mechanics, Zhejiang University, Hangzhou 310058, China; lcj@zju.edu.cn
* Correspondence: yangkai@swjtu.edu.cn; Tel.: +86-17716474578
Abstract:
In this paper, a new algorithm for extracting the laser fringe center is proposed. Based
on a deep learning skeleton extraction network, the laser stripe center can be extracted quickly
and accurately. Skeleton extraction is the process of reducing the shape image to its approximate
central axis representation while maintaining the image’s topological and geometric shape. Skeleton
extraction is an important step in topological and geometric shape analysis. According to the
characteristics of the wheelset laser curve dataset, a new skeleton extraction network, a hierarchical
skeleton network (LuoNet), is proposed. The proposed architecture has three levels of the encoder–
decoder network, and YE Module interconnection is designed between each level of the encoder
and decoder network. In the wheelset laser curve dataset, the F1_score can reach 0.714. Compared
with the traditional laser curve center extraction algorithm, the proposed LuoNet algorithm has the
advantages of short running time, high accuracy, and stable extraction results.
Keywords: deep learning; semantic segmentation; laser curve extraction; image processing
1. Introduction
In recent years, deep learning has made remarkable progress in the three main fields of
computer vision image recognition, target detection, and image segmentation. While deep
learning approaches are comparable to human vision in many areas, some areas require
designing different models for different tasks. In this paper, a new skeleton extraction
method based on deep learning is proposed to extract the laser curve of a subway wheelset.
Skeletonization is the process of extracting or generating an approximate geometric
representation of a shape (skeleton) with the aim of reducing it to clean skeleton pixels to
preserve the range and connectivity of the original shape. Skeleton extraction combines
local and global knowledge of shapes. The skeleton is a compact and intuitive central
axis representation of the shape, which preserves the topology and geometry of the shape.
Skeleton representations of shapes can be used for a variety of purposes, such as modeling,
manipulation, composition, matching, registration, compression, and analysis.
At present, the laser stripe center extraction of a subway wheelset mainly relies on the
traditional algorithm, which has low efficiency and low accuracy. The gray center of gravity
method [
1
], curve fitting method [
2
], and extreme value method [
3
] often require that the
gray value of the laser fringe be distributed into an ideal Gaussian distribution, but the
obtained fringe is not an ideal Gaussian distribution and is easily disturbed by noise. The
template matching method is limited by the limited template direction, and the accuracy of
fringe center extraction is affected by the surface roughness of the object [
4
]. The Steger
method has a large amount of computation, low efficiency, and the improper selection of
a Gaussian kernel will lead to image information distortion [
5
]. Due to the influence of
light, the stripe center line extracted by Zhang-Suen method has burrs, which increases
Sensors 2022, 22, 859. https://doi.org/10.3390/s22030859 https://www.mdpi.com/journal/sensors
