摘要： 为提高直线特征匹配的可靠性，提出一种基于卷积神经网络（CNN）学习的直线特征描述方法。构建用于网络学习的大规模直线数据集，该数据集包含约20.8万对匹配直线对，每条直线用其周围的局部图像块表征。将图像块输入CNN，利用HardNet网络结构提取特征，使用三元组损失函数进行训练，输出强鲁棒性的直线特征描述子。实验结果表明，与手工设计的描述子MSLD和IOCD相比，该描述子在视角、模糊、尺度和旋转变化下均具有较好的区分性，在图像拼接应用中同样表现出良好的描述性能。

关键词: 直线匹配, 直线特征描述子, 深度学习, 大规模直线数据集, 卷积神经网络

Abstract: To improve the reliability of line feature matching,this paper proposes a line feature description method based on Convolutional Neural Network(CNN).For the learning of neural network,a large-scale line dataset is constructed.The dataset contains about 208,000 pairs of matched lines,and each line is characterized by the local image block around it.The image blocks are input into a CNN,and the HardNet structure is used for feature extraction.At the same time,the triple loss function is used for training to output the required line feature descriptor with high robustness.The experimental results show that compared with the existing manual design descriptors such as MSLD and IOCD,the proposed descriptor has better discriminability for the changes of image perspective,blur,scale and rotation,and also has excellent description performance in image mosaic applications.

Key words: line matching, line feature descriptor, deep learning, large-scale line dataset, Convolutional Neural Network(CNN)

中图分类号: 

• TP391