摘要
为提高特征点检测的质量和性能,实现间接提高三维重建的模型效果,建立一个多尺度卷积的特征点检测模型。参考多尺度模型思想,设计一个通过有监督的学习,使用卷积运算检测点特征点的方法;在融合多尺度与视觉注意力机制的基础上,通过恒等映射改善图像特征丢失的缺点,提高特征点检测的重复率。通过在不同数据集上进行测试,有效检测到大量准确且重复的特征点,相比于同类方法,有效降低了检测的时间成本。
To improve the quality and performance of feature point detection and to indirectly improve the model effect of 3D reconstruction,a multi-scale convolution feature point detection model was established.Using the idea of multi-scale model,a method was designed to detect point feature points through supervised learning and convolution operation.Based on the fusion of multi-scale and visual attention mechanisms,identity mapping was used to improve the shortcomings of image feature loss and increase the repetition rate of feature point detection.By testing on different data sets,a large number of accurate and repeated feature points are effectively detected,which effectively reduces the time cost of detection compared with similar methods.
作者
赵亚帅
李宏伟
张进鹏
张彭昱
张鹏飞
ZHAO Ya-shuai;LI Hong-wei;ZHANG Jin-peng;ZHANG Peng-yu;ZHANG Peng-fei(School of Information Engineering,Zhengzhou University,Zhengzhou 450052,China;School of the Geo-Science and Technology,Zhengzhou University,Zhengzhou 450052,China)
出处
《计算机工程与设计》
北大核心
2023年第4期1159-1165,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(41571394)
郑州大学高层次人才科研启动基金项目(237-32310319)。
关键词
三维重建
特征点
卷积神经网络
多尺度
注意力模型
特征丢失
重复率
3D reconstruction
feature point
convolutional neural network
multiscale
attention module
feature loss
repetition rate
