基于改进KeyPointNet网络的特征点检测和描述

Feature Point Detection and Description Based on an Improved KeyPointNet Network

传统手工设计的特征提取方法如SIFT、ORB等,在光照或视角变化等挑战性场景中特征提取鲁棒性、精度都不如基于深度学习的特征点检测网络。启发于KeyPointNet网络在图像特征提取任务中表现的鲁棒性,文章利用轻量化网络设计KeyPointNet改进模型,旨在使其满足一定精度的情况下,在资源受限的平台上实时运行。实验结果表明,改进后的KeyPointNet在HPatches数据集上,重复性与单应性精度都优于原KeyPointNet模型,并且改进后的网络模型参数量大约压缩了88.83%,浮点运算次数减少了约86.62%,更适合部署在实际场景中。

Traditional hand-designed feature extraction methods such as SIFT and ORB are not as robust and accurate as deep learning-based feature point detcction nctworks for feature cextraction in challenging scnarios such as lighting or viewpoint changes.Inspired by the robustness of KeyPointNet network performance in image feature extraction tasks,this paper uses lightweight networks to design an improved model of KeyPointNet,aiming to make it satisfy a certain level of accuracy and run in real-time on resource-constrained platforms.The experimental results show that the improved KeyPointNet outperforms the original KeyPointNet model in terms of repeatability and single response accuracy on the HPatches dataset,and the improved network model compresses the number of parameters by approximately 88.83% and reduces the number of floating-point operations by about 86.62%,which is more suitable to be deployed in realworld sccnarios.