Remote Sensing Rotating Object Detection Based on Multi-Scale Feature Extraction

Luobing Wu; Yuhai Gu; Wenhao Wu; Shuaixin Fan

doi:10.3788/LOP221716

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 12 >Page 1228010 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 12, 1228010 (2023)

Remote Sensing Rotating Object Detection Based on Multi-Scale Feature Extraction

Luobing Wu¹, Yuhai Gu^1,2,*, Wenhao Wu¹, and Shuaixin Fan¹

Author Affiliations

¹Key Laboratory of Modern Measurement and Control Technology, Ministry of Education, Beijing Information Science & Technology University, Beijing 100089, China

²Mechanical Electrical Engineering School, Beijing Information Science & Technology University, Beijing 100089, China

show less

DOI: 10.3788/LOP221716 Cite this Article Set citation alerts

Luobing Wu, Yuhai Gu, Wenhao Wu, Shuaixin Fan. Remote Sensing Rotating Object Detection Based on Multi-Scale Feature Extraction[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1228010 Copy Citation Text

show less

Fig. 1. Structure diagram of CenterNet model

Download full size

Fig. 2. Structure diagram of CenterNet-RS model

Download full size

Fig. 3. Schematic of principle and structure of receptive field expansion module. (a) Schematic of principle; (b) schematic of structure

Download full size

Fig. 4. Structure of adaptive feature fusion

Download full size

Fig. 5. Presentation of DOTA dataset

Download full size

Fig. 6. Effect presentations of the CenterNet-RS on DOTA dataset

Download full size

Fig. 7. Structure of models in the ablation experiment

Download full size

Experiment	Content
CPU	Intel Core i5
GPU	Nvidia RTX 2080 Ti
RAM	16 GB
OS	Ubuntu 18.04
Plantform	CUDA 10.0
Library	PyTorch

Table 1. Hardware and software environments for experimental and model construction

Model	Backbone	Rotatable-Head	mAP /%	Running time /s
R2CNN	VGG-16	√	60.67	0.263
ICN	ResNeXt101	√	68.16	0.221
CAD-Net	ResNet101	√	69.90	0.172
RoI-Trans	ResNet101	√	69.59	0.175
BBAVectors	ResNet101	√	75.36	0.086
CenterNet	ResNet101	×	63.56	0.072
RetinaNet	DLA-34	×	62.76	0.081
CenterNet-RS	ResNet101	√	73.01	0.078

Table 2. Comparison of performance of CenterNet-RS and other models on DOTA dataset

Category	R2CNN	ICN	CAD-Net	RoI-Trans	BBA-Vectors	CenterNet-RS
mAP	60.67	68.16	69.90	69.56	75.36	73.01
PL	80.89	81.36	87.80	88.64	88.63	92.37
BD	65.75	74.30	82.40	78.52	84.06	74.86
BR	35.34	47.70	49.40	43.44	52.13	50.42
GTF	67.44	70.32	73.50	75.92	69.56	66.43
SV	59.93	64.89	71.10	68.81	78.26	64.49
LV	50.91	67.82	63.50	73.68	80.40	79.63
SH	55.81	69.98	76.70	83.59	88.06	76.74
TC	90.67	90.76	90.90	90.74	90.87	93.96
BC	66.92	79.06	79.20	77.27	87.23	75.27
ST	72.39	78.20	73.30	81.46	86.39	87.30
SBF	55.06	53.64	48.40	58.39	56.11	62.72
RA	52.23	62.90	60.90	53.54	65.62	69.16
HA	55.14	67.02	62.00	62.83	67.10	66.76
SP	53.35	64.17	67.00	58.93	72.08	59.60
HC	48.22	50.23	62.20	47.67	63.96	56.05

Table 3. AP of different algorithms detecting various objects on DOTA dataset

Method	Backbone	AP		mAP
Method	Backbone	Car	Airplane	mAP
R2CNN*	VGG-16	78.89	89.76	84.32
ICN	ResNeXt101	85.02	90.32	87.67
CAD-Net	ResNet101	88.35	89.93	89.14
RoI-Trans*	ResNet101	87.99	89.90	88.95
BBAVectors	ResNet101	90.27	91.41	90.84
RetinaNet *	DLA-34	83.64	89.51	86.57
CenterNet	ResNet101	79.90	90.61	85.25
CenterNet-RS	ResNet101	89.37	92.82	91.10

Table 4. AP of different algorithms detecting various objects on UCAS-AOD dataset

Category	CenterNet	FPN	FPN+ RFEM	AFF	Rotatable-Head	FPN+Rotatable-Head	CenterNet-RS
mAP	63.43	66.44	68.98	67.64	67.03	69.14	73.01
PL	90.20	91.24	91.94	90.81	90.99	91.13	92.37
BD	68.13	71.12	73.44	72.75	69.77	71.86	74.86
BR	44.72	47.28	46.01	45.18	48.23	49.89	50.42
GTF	57.49	59.98	64.98	62.98	58.39	59.98	66.43
SV	54.44	54.69	56.47	56.44	63.81	63.16	64.49
LV	71.37	75.72	75.29	72.35	75.67	79.12	79.63
SH	66.39	68.54	69.93	68.80	73.96	75.21	76.74
TC	89.82	90.58	91.79	93.76	92.17	92.03	93.96
BC	63.22	66.90	71.55	70.86	69.96	72.74	75.27
ST	76.59	84.25	85.98	86.77	76.45	83.21	87.3
SBF	60.05	61.16	64.56	59.87	60.61	60.82	62.72
RA	62.82	65.48	67.63	65.43	64.95	66.71	69.16
HA	57.28	60.23	62.83	63.66	63.76	65.81	66.76
SP	47.97	51.26	57.63	55.00	50.92	53.31	59.60
HC	40.93	48.17	54.68	49.99	45.72	52.06	56.05

Table 5. Comparison of AP values of different methods in ablation experiment

Luobing Wu, Yuhai Gu, Wenhao Wu, Shuaixin Fan. Remote Sensing Rotating Object Detection Based on Multi-Scale Feature Extraction[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1228010

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享