局部得分检测约束的深度特征提取方法

Deep feature extraction constrained by local score detection

针对端到端深度特征提取网络的特征点数量与定位精度难以满足运动恢复结构(SFM)几何解算的问题,该文基于深度特征图的“一图两用”思想,提出联合可变形卷积与局部得分检测的端到端特征提取方法。首先,在特征图提取阶段,利用附加可变形卷积层的轻量级网络提取影像对的多尺度特征图,并在各个尺度进行特征图加权融合生成特征检测图。其次,在关键点检测与描述阶段,不再考虑特征检测图的通道极大值约束,仅由局部得分计算特征得分图,避免描述子向量的数值分布对特征点数量和定位精度的影响。最后,基于欧式距离准则及比值测试和交叉验证策略进行初始特征匹配,并结合核线约束优化匹配结果。利用多组地面近景影像和无人机影像进行特征匹配和SFM重建试验。结果表明,该文方法能够显著增加特征匹配和重建点数量,其增加比例分别达到了22.2%~41.7%和11.4%~37.7%。同时,SFM三维重建的重投影误差优于1.3像素。

Aiming at the problem that the number and localization accuracy of feature points extracted from end-to-end feature extraction networks are not satisfactory for structure from motion(SFM),an end-to-end feature matching method that combines deformable convolution and local score detection was proposed according to the"one map,two uses"idea of feature maps in this paper.First,in feature extraction,a lightweight network with additional deformable convolution layers was used to extract multi-scale feature maps,and the feature detection maps were generated by fusing feature maps at each scale.Second,in keypoint detection and description,the channel maximum constraint was no longer considered,and the feature score map was only calculated based on local scores to avoid the influence of the numerical distribution of descriptors on feature points.Third,based on the Euclidean distance criterion and ratio test and cross-check strategies,initial feature matching was obtained,which was then optimized using the epipolar constraints.Finally,tests were conducted by using close-range and unmanned arial vehicle(UAV)images,and the results showed that our method could increase the number of feature matches and resumed 3D points with the increasing ratio within the ranges of 22.2%to 41.7%and 11.4%to 37.7%,respectively.In addition,the reprojection error of SFM was better than 1.3 pixels.