摘要
检测图像中的显著关键点并提取特征描述子是视觉里程计和同步定位与建图系统等计算机视觉任务中的重要环节。特征点提取算法的主要目标是检测准确的关键点位置并提取可靠的特征描述子。可靠的特征描述子应对旋转、尺度缩放、光照变化、视角变化、噪声等保持一定程度的稳定性。目前基于深度学习的方法由于描述子特征在下采样过程中存在图像信息丢失,导致描述子可靠性和特征匹配准确度降低。针对这一问题,提出了一种面向细节保持的特征描述子提取网络。该网络融合浅层细节特征和深层语义特征,将描述子特征上采样到更高的空间分辨率,并结合注意力机制,使用局部特征(角点、线段、纹理等)、语义特征和全局特征来改进特征点检测,提高特征描述子可靠性。在Hpatches数据集上的实验结果表明,所提方法的匹配准确度为55.5%。输入图像分辨率为480×640时,所提方法的单应性估计准确度比现有方法高5.9个百分点。实验结果表明了所提方法的有效性。
Detecting salient key points in images and extracting feature descriptors are important components of computer vision tasks such as visual odometry and simultaneous localization and mapping systems.The main goal of the feature point extraction algorithm is to detect accurate key point positions and extract reliable feature descriptors.Reliable feature descriptors should maintain stability against rotation,scale scaling,illumination changes,viewing angle changes,noise,etc.Due to the loss of image information during the downsampling process in recent deep learning-based feature point extraction algorithms,the reliability of the descriptor and accuracy of feature matching are reduced.This study proposes a network structure to detect detail-preserving oriented feature descriptors to solve this problem.The proposed network fuses shallow detail and deep semantic features to sample the descriptors to a higher resolution.Combined with the attention mechanism,local(corners,lines,textures,etc.),semantic,and global features are used to improve the detection of feature points and the reliability of feature descriptors.Experiments on the Hpatches dataset show that the matching accuracy of the proposed method is 55.5%.Additionally,when the input image resolution is 480×640,the homography estimation accuracy of the proposed method is 5.9percentage points higher than that of the existing method.These results demonstrate the effectiveness of the proposed method.
作者
龙涛
苏畅
王建
Long Tao;Su Chang;Wang Jian(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2022年第22期212-219,共8页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61632018)。
关键词
机器视觉
特征点检测
深度学习
卷积神经网络
machine vision
feature point detection
deep learning
convolutional neural network
