Remote Sensing Target Detection Based on Multilevel Self-Attention Enhancement

Xiegen Wei; Lin Cao; Shu Tian; Kangning Du; Peiran Song; Yanan Guo

doi:10.3788/LOP223048

Laser & Optoelectronics Progress, Volume. 60, Issue 20, 2028004(2023)

Remote Sensing Target Detection Based on Multilevel Self-Attention Enhancement

Xiegen Wei^1,2, Lin Cao^2,3, Shu Tian^3、*, Kangning Du³, Peiran Song³, and Yanan Guo³

¹School of Instrument Science and Opto-Electronics Engineering, Beijing Information Science & Technology University, Beijing 100101, China

²Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument, Beijing Information Science & Technology University, Beijing 100101, China

³Key Laboratory of Information and Communication Systems, Ministry of Information Industry, Beijing Information Science & Technology University, Beijing 100101, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Schematic of midpoint offset representation method

Download full size

Fig. 2. Schematic of Rotated RoIAlign

Download full size

Fig. 3. Overall framework of the proposed algorithm

Download full size

Fig. 4. Swin Transformer Block structure

Download full size

Fig. 5. Comparison between MSA principle and W-MSA principle

Download full size

Fig. 6. Comparison betwwen W-MSA principle and SW-MSA principle

Download full size

Fig. 7. Some samples in DOTA dataset

Download full size

Fig. 8. Some samples in HRSC2016 dataset

Download full size

Fig. 9. Detection results of different algorithms on DOTA dataset

Download full size

Fig. 10. Detection results of different algorithms on HRSC2016 dataset

Download full size

Fig. 11. Detection speed and accuracy of different algorithms on DOTA dataset

Download full size

Table 1. AP of different algorithms on DOTA dataset

View table

Table 1. AP of different algorithms on DOTA dataset

Category	ICN	Faster R-CNN-O	Gliding Vertex	RoI Transformer	Oriented R-CNN	Proposed algorithm
mAP	68.16	73.37	75.02	74.61	75.66	77.20
PL	81.36	89.12	89.64	88.65	89.19	89.24
BD	74.30	83.06	85.00	82.60	82.53	82.88
BR	47.70	50.26	52.26	52.53	51.86	52.95
GTF	70.32	67.49	77.34	70.87	72.21	75.50
SV	64.89	78.64	73.01	77.93	78.86	78.85
LV	67.82	73.44	73.14	76.67	81.87	84.26
SH	69.98	85.97	86.82	86.87	87.91	88.24
TC	90.76	90.89	90.74	90.71	90.90	90.91
BC	79.06	84.58	79.02	83.83	86.70	86.91
ST	78.20	82.92	86.81	82.51	85.13	86.08
SBF	53.64	54.34	59.55	53.95	63.85	64.57
RA	62.90	66.09	70.91	67.61	65.85	67.89
HA	67.02	66.22	72.94	74.67	73.24	75.33
SP	64.17	68.99	70.86	68.75	68.76	70.97
HC	50.23	58.52	59.32	61.03	56.07	63.43

Table 2. Detection results of different algorithms on HRSC2016 dataset
View table
Table 2. Detection results of different algorithms on HRSC2016 dataset
Method BackBone AP₅₀ /% AP₇₅ /%
RetinaNet-O ResNet-50 84.8 59.9
RoI Transformer ResNet-101 86.1 65.3
S²A-Net ResNet-50 89.7 74.6
Oriented R-CNN ResNet-50 90.1 76.9
Proposed method Swin Transformer 90.6 79.8

Table 3. Speed and accuracy of different algorithms on DOTA dataset

View table

Table 3. Speed and accuracy of different algorithms on DOTA dataset

Method	Framework	s /（frame·s^-1）	mAP /%
RetinaNet-O	One-stage	16.8	69.79
S²A-Net	One-stage	15.5	73.85
RoI Transformer	Two-stage	14.3	74.61
Gliding Vertex	Two-stage	15.3	75.02
Oriented R-CNN	Two-stage	15.0	75.86
Proposed method	Two-stage	15.2	77.20

Table 4. Comparison of Parameters and FLOPs of main algorithms on DOTA dataset

View table

Table 4. Comparison of Parameters and FLOPs of main algorithms on DOTA dataset

Model	Framework	mAP /%	Parameters /10⁶	FLOPs /10⁹
RetinaNet-O	One-stage	69.79	43.56	213.87
S²A-Net	One-stage	73.85	44.66	215.36
RoI Transformer	Two-stage	74.61	59.76	231.69
Gliding Vertex	Two-stage	75.02	64.98	240.24
Oriented R-CNN	Two-stage	75.86	42.14	213.43
Proposed method	Two-stage	77.20	42.10	213.58

Table 5. Results of ablation experiment
View table
Table 5. Results of ablation experiment
Baseline Swin Transformer KLD mAP /% Parameters /10⁶ FLOPs /10⁹
√ 75.86 42.14 213.43
√ √ 76.73 44.75 215.66
√ √ 76.21 41.15 213.44
√ √ √ 77.20 42.10 213.58

Tools

Get Citation

Copy Citation Text

Xiegen Wei, Lin Cao, Shu Tian, Kangning Du, Peiran Song, Yanan Guo. Remote Sensing Target Detection Based on Multilevel Self-Attention Enhancement[J]. Laser & Optoelectronics Progress, 2023, 60(20): 2028004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites