
http://www.paper.edu.cn
- 1 -
中国科技论文在线
一种新的基于神经网络的直线匹配算法
陈忠尧,李永
**
(北京邮电大学电子工程学院,北京 100876)
5
作者简介:陈忠尧(1991-),男,硕士研究生,主要研究方向:直线匹配,深度学习
通信联系人:李永(1978-),男,副教授,主要研究方向:深度学习,计算机视觉. E-mail: yli@bupt.edu.cn
摘要:直线匹配是计算机视觉领域的重要技术之一,是图像拼接、三维重建以及图像配准的
一个关键环节,具有非常广泛的应用场景。相比于特征点匹配,特征直线包含物体的结构信
息,并且对环境噪声具有较强的抗干扰能力而备受国内外学者关注。然而,直线端点的不精
确定位给直线匹配的研究带来了巨大的挑战,为了解决这个问题,本文提出了一种利用深度
神经网络提取直线特征并进行直线匹配的新方法。首先,检测图像对中的特征直线,然后根10
据直线裁剪固定大小的图像块。其次,将裁剪后的图像块输入到本文设计的适合直线匹配的
网络(LMNet,Line Matching Network)中,判断直线是否匹配。最后,利用后处理机制剔
除重复匹配和错误匹配的直线对。实验证明,本文所提出的方法显著提高了直线匹配的性能,
并且有效地解决了直线端点的不精确定位问题。
关键词:直线匹配;计算机视觉;卷积神经网络 15
中图分类号:TP399
A Novel Line Matching Algorithm Based On Neural
Networks
CHEN Zhongyao, LI Yong 20
(Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876)
Abstract: Line matching is one of the most important technologies in the field of computer vision. It is
a key link in image stitching, 3D reconstruction and image registration and widely used in many
applications. Compared with point matching, line contains more structure information of object and has
strong anti-jamming ability to environmental noise, which attracts much attention of scholars. However,
25
the inaccurate location of line endpoints brings great challenges to the research of line matching. In
order to solve this problem, this paper proposed a novel line matching algorithm based on
convolutional neural networks to extract line features. Firstly, the feature lines in the images are
detected, and then the fixed size image blocks are cropped according to the lines. Secondly, the
cropped image block is input to the Line Matching Network(LMNet), which is designed to extract line
30
features, to judge whether the feature lines matched. Finally, the post-processing mechanism is used to
eliminate the duplicate matching and the wrong matching. Experimental results show that the proposed
algorithm can significantly improve the performance of line matching and effectively solve the
problem of inaccurate location of line endpoints.
Keywords: line matching; computer vision; convolutional neural network
35
0 引言
直线匹配是图像处理领域的一个重要课题,具有重大的研究意义,同时,也是图像处理
领域的一个研究难点。近年来,学者对直线匹配进行了大量的研究并取得了较大进展。如40
Zhang 等人
[1]
提出了一种基于直线的局部外观及其几何属性的直线匹配算法。该方法通过提
取尺度空间中的直线以解决子线断裂和几何变化问题,并设计了一种描述直线外观的直线描
述符(LBD,Line Band Descriptor)。该算法对低纹理图像匹配精度很高,但是,当存在直线
评论0