基于联合特征块学习的特征直线描述方法  

Feature Line Description Method Based on Joint Feature Block Learning

作　　者：付苗苗 FU Miaomiao(College of Information Technology,Luoyang Normal University,Luoyang Henan 471000,China)

机构地区：[1]洛阳师范学院信息技术学院,河南洛阳471000

出　　处：《信息与电脑》2022年第19期157-159,199,共4页Information & Computer

基　　金：河南省高等学校重点科研项目计划(项目编号:22A120009);洛阳师范学院2021年度校级教改(项目编号:2021xjgj024、2021xjgj022)。

摘　　要：特征直线描述是特征直线匹配过程的关键和基础。为了改善手工设计方法在复杂场景下存在区分性弱、鲁棒性差的缺陷,提出了一种基于联合特征块学习的特征直线描述方法。首先,在原有的小型直线数据集上重新构造直线块,即通过联合每幅图像中直线支撑区域内像素的亮度与梯度获得固定大小的直线块。其次,将获得的直线块输入预先训练的L2-Net,使用微调策略和三元组损失函数训练网络。最后,输出紧凑且强区分性的特征直线描述子。特征匹配任务的实验结果证明,所提出的基于联合特征块学习的特征直线描述方法与最先进的手工直线描述符相比,更具有优越性和有效性。Feature line description is the key and foundation of feature line matching process. Compared with feature point descriptors, its development speed has been slow in the past few decades, and the methods based on manual design are still mainly used to solve this problem. In order to improve the shortcomings of weak discrimination and poor robustness in complex scenes, this paper proposes a feature line description method based on joint feature block learning. First, reconstruct the line block on the original small line dataset, that is, obtain a line block of fixed size by combining the intensity and gradient of the pixels in the line support area in each image. It is then fed into a pretrained L2-Net to train the network using a fine-tuning strategy and triplet loss function.Finally, a compact and strongly discriminative feature line descriptor is output. Compared with the state-of-the-art handcrafted line descriptors, the experimental results on the feature matching task demonstrate the superiority and effectiveness of the proposed feature line description method based on joint feature block learning.

关 键 词：特征直线描述 深度学习 特征直线匹配 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]