目的航空遥感图像中多为尺寸小、方向错乱和背景复杂的目标。传统的目标检测算法由于模型的特征提取网络对输入图像进行多次下采样,分辨率大幅降低,容易造成目标特征信息丢失,而且不同尺度的特征图未能有效融合,检测目标之间存在的相似...目的航空遥感图像中多为尺寸小、方向错乱和背景复杂的目标。传统的目标检测算法由于模型的特征提取网络对输入图像进行多次下采样,分辨率大幅降低,容易造成目标特征信息丢失,而且不同尺度的特征图未能有效融合,检测目标之间存在的相似特征不能有效关联,不仅时间复杂度高,而且提取的特征信息不足,导致目标漏检率和误检率偏高。为了提升算法对航空遥感图像目标的检测准确率,本文提出一种基于并行高分辨率结构结合长短期记忆网络(long short-term memory,LSTM)的目标检测算法。方法首先,构建并行高分辨率网络结构,由高分辨率子网络作为第1阶段,分辨率从高到低逐步增加子网络,将多个子网并行连接,构建子网时对不同分辨率的特征图反复融合,以增强目标特征表达;其次,对各个子网提取的特征图进行双线性插值上采样,并拼接通道特征;最后,使用双向LSTM整合通道特征信息,完成多尺度检测。结果将本文提出的检测算法在COCO(common objects in context)2017数据集、KITTI(Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago)车辆检测和UCAS-AOD(University of Chinese Academy of Sciences-Aerial Object Detection)航空遥感数据集上进行实验验证,平均检测准确率(mean average precision,m AP)分别为41.6%、69.4%和69.3%。在COCO 2017、KITTI和VCAS-AOD数据集上,本文算法与SSD513算法相比,平均检测准确率分别提升10.46%、7.3%、8.8%。结论本文方法有效提高了航空遥感图像中目标的平均检测准确率。展开更多
Currently,the manual contact rail measurement that was basically adopted in China has low detection efficiency,poor accuracy and poor stability.In order to improve the function of the system,we propose a non-contact m...Currently,the manual contact rail measurement that was basically adopted in China has low detection efficiency,poor accuracy and poor stability.In order to improve the function of the system,we propose a non-contact measurement method based on the flatness and verticality ruler model.The flatness measurement model was built by employing the string measurement method.In addition,the verticality measurement model was built by the dihedral method to measure the rail comprehensively.By extracting curvature information of feature points,in this system,each laser sensor is used to collect rail profile curves.A large number of three-dimensional point clouds data are generated by the unit quaternion method of coordinate transformation,and the contour curves of the characteristic points of the four laser sensors are matched with the corresponding point sets one to one,and the rail contour splicing is finally completed.The experimental results show that this method has better measurement effect compared with the traditional manual measurement method.展开更多
文摘目的航空遥感图像中多为尺寸小、方向错乱和背景复杂的目标。传统的目标检测算法由于模型的特征提取网络对输入图像进行多次下采样,分辨率大幅降低,容易造成目标特征信息丢失,而且不同尺度的特征图未能有效融合,检测目标之间存在的相似特征不能有效关联,不仅时间复杂度高,而且提取的特征信息不足,导致目标漏检率和误检率偏高。为了提升算法对航空遥感图像目标的检测准确率,本文提出一种基于并行高分辨率结构结合长短期记忆网络(long short-term memory,LSTM)的目标检测算法。方法首先,构建并行高分辨率网络结构,由高分辨率子网络作为第1阶段,分辨率从高到低逐步增加子网络,将多个子网并行连接,构建子网时对不同分辨率的特征图反复融合,以增强目标特征表达;其次,对各个子网提取的特征图进行双线性插值上采样,并拼接通道特征;最后,使用双向LSTM整合通道特征信息,完成多尺度检测。结果将本文提出的检测算法在COCO(common objects in context)2017数据集、KITTI(Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago)车辆检测和UCAS-AOD(University of Chinese Academy of Sciences-Aerial Object Detection)航空遥感数据集上进行实验验证,平均检测准确率(mean average precision,m AP)分别为41.6%、69.4%和69.3%。在COCO 2017、KITTI和VCAS-AOD数据集上,本文算法与SSD513算法相比,平均检测准确率分别提升10.46%、7.3%、8.8%。结论本文方法有效提高了航空遥感图像中目标的平均检测准确率。
基金Supported by the National Natural Science Foundation of China(U1831133)Shanghai Natural Science Foundation(17ZR1443500)Baoshan Science and Technology Innovation Special Fund(17-C-21)。
文摘Currently,the manual contact rail measurement that was basically adopted in China has low detection efficiency,poor accuracy and poor stability.In order to improve the function of the system,we propose a non-contact measurement method based on the flatness and verticality ruler model.The flatness measurement model was built by employing the string measurement method.In addition,the verticality measurement model was built by the dihedral method to measure the rail comprehensively.By extracting curvature information of feature points,in this system,each laser sensor is used to collect rail profile curves.A large number of three-dimensional point clouds data are generated by the unit quaternion method of coordinate transformation,and the contour curves of the characteristic points of the four laser sensors are matched with the corresponding point sets one to one,and the rail contour splicing is finally completed.The experimental results show that this method has better measurement effect compared with the traditional manual measurement method.
