Object Detection in Remote Sensing Images Using Multiscale Convolutional Neural Networks

Qunli Yao; Xian Hu; Hong Lei

doi:10.3788/AOS201939.1128002

Journals >Acta Optica Sinica >Volume 39 >Issue 11 >Page 1128002 > Article

Acta Optica Sinica
Vol. 39, Issue 11, 1128002 (2019)

Object Detection in Remote Sensing Images Using Multiscale Convolutional Neural Networks

Qunli Yao^1、2、*, Xian Hu^1、2, and Hong Lei¹

Author Affiliations

¹Department of Space Microwave Remote Sensing Systems, Institute of Electronics, Chinese Academy of Sciences, Beijing 100190, China

²School of Electronics, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China

show less

DOI: 10.3788/AOS201939.1128002 Cite this Article Set citation alerts

Qunli Yao, Xian Hu, Hong Lei. Object Detection in Remote Sensing Images Using Multiscale Convolutional Neural Networks[J]. Acta Optica Sinica, 2019, 39(11): 1128002 Copy Citation Text

show less

Fig. 1. Target detection framework of MSCNN

Download full size

Structures of EFPN and dilated bottleneck. (a) Structure of EFPN module; (b) dilated bottleneck structure; (c) dilated bottleneck structure with 1×1 Conv

Fig. 2. Structures of EFPN and dilated bottleneck. (a) Structure of EFPN module; (b) dilated bottleneck structure; (c) dilated bottleneck structure with 1×1 Conv

Download full size

Fig. 3. Network structure of EFPN-NoProj

Download full size

Fig. 4. Visual detection results of MSCNN

Download full size

Small	Medium	Large
(0,60]	(60,120]	(120,+¥)

Table 1. Definition of bounding box areas based on distribution of instance scales

Method	RICNN^[25]	FRCN-VGG-16^[16]	YOLO^[15]	SSD^[13]	R-FCN^[26]	FRCN-Deform^[27]	FPN^[18]	MSDN^[28]	MSCNN
Airplane	0.884	0.830	0.874	0.956	0.961	0.983	0.964	0.998	0.994
Ship	0.773	0.776	0.847	0.937	0.983	0.892	0.931	0.972	0.953
Storage tank	0.853	0.525	0.427	0.617	0.725	0.817	0.914	0.838	0.918
Baseball diamond	0.881	0.963	0.931	0.995	0.994	0.984	0.947	0.991	0.963
Tennis court	0.408	0.629	0.658	0.860	0.907	0.859	0.944	0.973	0.954
Basketball court	0.585	0.688	0.870	0.944	0.978	0.927	0.959	0.999	0.967
Ground track field	0.867	0.984	0.975	0.987	0.981	0.988	0.990	0.986	0.993
Harbor	0.686	0.819	0.800	0.950	0.924	0.946	0.921	0.972	0.955
Bridge	0.615	0.793	0.903	0.966	0.934	0.947	0.838	0.927	0.972
Vehicle	0.711	0.639	0.704	0.745	0.884	0.816	0.900	0.901	0.933
mAP	0.726	0.764	0.799	0.894	0.928	0.917	0.931	0.956	0.960

Table 2. Comparison of the detection precision of different algorithms on NWPU VHR-10 dataset

Backbone	Average precision
Backbone	I_OU=0.5:0.95	I_OU=0.5	I_OU=0.75	Small	Medium	Large
RetinaNet	0.690	0.945	0.809	0.532	0.559	0.682
EFPN	0.706	0.960	0.824	0.547	0.578	0.701
EFPN-NoProj	0.700	0.950	0.819	0.544	0.562	0.698

Table 3. Ablation experimental parameters of MSCNN

Backbone	Average precision			Average precision(I_OU =0.5:0.95)			Average recall(I_OU=0.5:0.95)
Backbone	I_OU=0.5:0.95	I_OU=0.5	I_OU=0.75	Small	Medium	Large	Small	Medium	Large
RetinaNet	0.690	0.945	0.809	0.532	0.559	0.682	0.573	0.586	0.754
MSCNN	0.706	0.960	0.824	0.547	0.578	0.701	0.600	0.605	0.755
EFPN-NoProj	0.700	0.950	0.819	0.544	0.562	0.698	0.597	0.600	0.753

Table 4. Average precision and average recall under different IOU thresholds and different bounding box areas

Qunli Yao, Xian Hu, Hong Lei. Object Detection in Remote Sensing Images Using Multiscale Convolutional Neural Networks[J]. Acta Optica Sinica, 2019, 39(11): 1128002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information