基于Resinv-Unet的图像特征点检测方法  

Image feature point detection based on Resinv-Unet

作　　者：高梓皓 张巧芬[1] 王桂棠[1,2] 温腾腾 庞亮雨 贾林锋 吴铭扬 李云飞 Gao Zihao;Zhang Qiaofen;Wang Guitang;Wen Tengteng;Pang Liangyu;Jia Linfeng;Wu Mingyang;Li Yunfei(School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China;Cangke Intelligent Technology Co.,Ltd.,Foshan 528225,China;School of Medical Information Engineering,Guangdong Pharmaceutical University,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广州510006 [2]佛山沧科智能科技有限公司,佛山528225 [3]广东药科大学医药信息工程学院,广州510006

出　　处：《国外电子测量技术》2023年第4期1-7,共7页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金(61705045);佛山市2021年高校教师特色创新研究项目(2021DZXX15)资助。

摘　　要：针对传统特征点检测算法需人为制定检测机制和基于深度学习的特征点检测网络泛化能力不强的问题,引入灰度不变量和残差结构,设计并实现具备像素级特征点检测能力的残差不变量神经网络(residual-invariant neural network,Resinv-Unet)。采用自标注的方式,在真实场景图像数据集的基础上构建用于训练神经网络的数据集。实验结果表明,Resinv-Unet相较于现有的特征点检测算法和特征点检测网络,在真实场景图像上具有更强的泛化能力和鲁棒性,在平均精确度、精确度和召回率上均取得更好的性能指标,其中,平均精确度达到0.7155、精确度达到0.7762、召回率达到0.7137。Aiming at the problem that the traditional feature point detection algorithm needs to develop the detection mechanism manually and the weak generalization ability of feature point detection network based on deep learning.The Resinv-Unet neural network with pixel-level feature point detection capability is designed and implemented by introducing gray-value invariant and residual.The data source for training neural networks is constructed by self-labeling on the basis of the real scene image data set.The experimental results show that compared with the existing feature point detection algorithms and feature point detection networks,Resinv-Unet has stronger generalization ability and robustness in real scene images,and has better performance in terms of average accuracy,accuracy and recall,with average accuracy of 0.7155,accuracy of 0.7762,and recall 0.7137.

关 键 词：图像处理 特征点提取 三维重建 机器视觉 神经网络 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]