摘要
为了解决大尺寸航拍图像下的多尺度车辆检测问题,在YOLO v2检测框架的基础上提出了一种多尺度目标检测算法。首先,将大尺寸航拍图像切分成若干有重叠区域的小图像块;然后,将各图像切片依次输入检测网络,主干网络针对输入图像提取不同尺度的特征,并对3种尺度的特征进行融合以获取不同的感受野,同时解决了浅层特征语义信息不足的问题;最后,各图像块的检测结果通过非极大值抑制的方法进行合并。在实际的航拍车辆数据集上,所提方法在不增加额外预测框的情况下,相比原YOLO v2检测算法的平均精度提高了约8个百分点。
In order to address the problem of multiple scale vehicle detection in large size satellite imagery,a multi-scale object detection method based on YOLO v2 framework was proposed.Firstly,the satellite imagery was cropped into several small slices with overlaps.Then every image slice was fed into the detection network successively.The backbone network extracted multi-scale features of the input images,and fused features at three different scales for diverse receptive fields,which also complements the semantic information of shallow layers.Lastly,non-maximal suppression was used to combine the detection results from every image slice.On the target dataset for realistic satellite imagery,the proposed method obtained 8 mAP gain without extra predication boxes compared with YOLO v2.
作者
赵爽
黄怀玉
胡一鸣
娄小平
王欣刚
ZHAO Shuang;HUANG Huaiyu;HU Yiming;LOU Xiaoping;WANG Xingang(College of Instrument Science and Optoelectronic Engineering,Beijing Information Science and Technology University,Beijing 100192,China;Center for Precision Perception and Control,Institue of Automation,Chinese Academy of Sciences,Beijing 100190,China)
出处
《计算机应用》
CSCD
北大核心
2019年第S02期91-96,共6页
journal of Computer Applications
基金
国家自然科学基金资助项目(61573349)
北京信息科技大学2019年度实培计划项目
