基于多敏感度最优传输的深度图匹配方法
Image Matching Based on Multi-Sensitivity Optimal Transport
杜千1
作者信息
- 1. 江苏华博在线传媒有限责任公司,南京,211135
- 折叠
摘要
本文针对不同图像中关键点之间的关系匹配问题,提出一种基于多敏感度最优传输的深度图像匹配网络框架.该框架利用视觉几何群网络(VGG16)从图像中提取一阶表观特征,进而使用德洛内三角测量建立关键点二阶邻接关系图.在此基础上,采用图卷积学习关键点的特征嵌入,并引入多敏感度的最优传输进行关键点之间的结构化关系度量,以增强关键点对齐能力.该框架在计算机视觉挑战赛(Pascal VOC Keypoints)数据集取得了良好的性能.
Abstract
This paper proposes a deep image matching network framework based on multi-sensitivity opti-mal transport to address the problem of matching relationships between key points in different images.It utili-zes VGG16 to extract first-order appearance features from images,and then uses Delaunay triangulation to es-tablish a second-order adjacency relationship graph of key points.On this basis,graph convolution is used to learn the feature embedding of key points,and a multi-sensitivity optimal transport is introduced to measure the structured relationship between key points,in order to enhance the key point alignment ability of the image matching framework.This framework achieves good performance on the Pascal VOC Keypoints dataset.
关键词
图像匹配/深度图神经网络/多敏感度最优传输Key words
image matching/deep map neural network/multi-sensitivity optimal transport引用本文复制引用
出版年
2024
NETL
NSTL
万方数据