Integrating global information and dual-domain attention mechanism for optical remote sensing aircraft target detection

Shanling LIN; Xue ZHANG; Yan CHEN; Jianpu LIN; Shanhong LÜ; Zhixian LIN; Tailiang GUO

doi:10.37188/OPE.20243220.3085

Journals >Optics and Precision Engineering >Volume 32 >Issue 20 >Page 3085 > Article

Optics and Precision Engineering
Vol. 32, Issue 20, 3085 (2024)

Integrating global information and dual-domain attention mechanism for optical remote sensing aircraft target detection

Shanling LIN^1,2, Xue ZHANG^1,2, Yan CHEN^1,2, Jianpu LIN^1,2,*..., Shanhong LÜ^1,2, Zhixian LIN^1,2,3 and Tailiang GUO^2,3|Show fewer author(s)

Author Affiliations

¹School of Advanced Manufacturing， Fuzhou University， Quanzhou36225， China

²Fujian Science and Technology Innovation Laboratory for Photoelectric Information， Fuzhou350116， China

³College of Physics and Information Engineering， Fuzhou University， Fuzhou50116， China

show less

DOI: 10.37188/OPE.20243220.3085 Cite this Article

Shanling LIN, Xue ZHANG, Yan CHEN, Jianpu LIN, Shanhong LÜ, Zhixian LIN, Tailiang GUO. Integrating global information and dual-domain attention mechanism for optical remote sensing aircraft target detection[J]. Optics and Precision Engineering, 2024, 32(20): 3085 Copy Citation Text

show less

Fig. 1. Overall structure of the network model

Download full size | View in the Article

Fig. 2. Structure of SPPF_Global module

Download full size | View in the Article

Fig. 3. Schematic of the structure of DDA module.

Download full size | View in the Article

Fig. 4. Structure of PDS module

Download full size | View in the Article

Fig. 5. Schematic of CIoU and PIoU Parameters

Download full size | View in the Article

Fig. 6. Training curve

Download full size | View in the Article

Fig. 7. Visualization results of different models in complex environments

Download full size | View in the Article

Fig. 8. Visualization results of different models with small targets

Download full size | View in the Article

Module	Precision	Recall	mAP50	mAP50-95	Parameters/M	FLOPs/B	Latency/ms
SPP	78.3%	77.2%	83.2%	62.2%	3.01	8.1	0.89
SPPF	81.0%	78.7%	84.6%	63.2%	3.01	8.1	0.88
ASPP	81.5%	77.4%	84.5%	63.2%	3.83	8.7	0.91
RFB	81.1%	77.0%	84.0%	62.8%	3.18	8.2	0.89
SPPCSPC	81.3%	78.1%	84.2%	63.1%	4.62	9.4	0.97
SimCSPSPPF	81.0%	78.1%	84.9%	63.4%	3.36	8.4	0.90
SPPF_Global	82.2%	78.5%	85.6%	64.0%	3.06	8.0	0.88

Table 1. Comparison of performance of different spatial pyramid pooling modules

View in the Article

Model	Precision	Recall	mAP50	mAP50-95	Parameters/M	FLOPs/B	Latency/ms
YOLOv8n	81.0%	78.7%	84.6%	63.2%	3.01	8.1	0.88
+CBAM	80.7%	79.0%	84.6%	63.1%	3.08	8.2	0.89
+CA	81.0%	78.4%	84.8%	63.6%	3.02	8.1	0.95
+CPCA	81.5%	78.8%	85.0%	63.6%	3.14	8.3	1.21
+DLKA	82.4%	78.9%	86.0%	64.2%	4.64	9.4	0.90
+EMA	81.9%	76.5%	84.6%	63.3%	3.02	8.2	0.90
+LSKA	80.6%	78.0%	84.8%	63.5%	3.08	8.2	0.89
+SA	80.6%	77.7%	84.8%	63.4%	3.01	8.1	0.89
+SimAM	82.7%	77.3%	85.1%	63.6%	3.01	8.1	0.89
+TA	80.4%	79.1%	85.1%	63.7%	3.01	8.1	0.90
+DDA	81.7%	79.5%	85.9%	64.1%	3.01	8.0	0.88

Table 2. Comparison of performance of different attention mechanisms

View in the Article

SPPF_ Global	DDA	PDS	PIoU	Precision	Recall	mAP 50	mAP 50-95	Parameters /M	FLOPs /B	Latency /ms
×	×	×	×	81.0%	78.7%	84.6%	63.2%	3.01	8.1	0.88
√	×	×	×	82.2%	78.5%	85.6%	64.0%	3.06	8.0	0.88
×	√	×	×	81.7%	79.5%	85.9%	64.1%	3.01	8.1	0.88
×	×	√	×	83.1%	78.5%	86.0%	64.6%	2.74	7.8	1.04
√	√	×	×	82.6%	79.7%	86.5%	64.6%	3.06	8.0	0.89
√	√	√	×	82.7%	80.8%	87.0%	65.6%	2.81	7.8	1.06
√	√	√	√	84.3%	81.3%	87.8%	65.8%	2.81	7.8	1.06

Table 3. Ablation experiments

View in the Article

Model	Precision	Recall	mAP50	mAP50-95	Parameters/M	FLOPs/B	Latency/ms
Faster R-CNN	66.1%	61.1%	80.9%	58.7%	41.75	91.23	8.74
RT-DETR	87.7%	83.4%	83.0%	61.9%	19.9	57.0	4.34
Gold-YOLO-n	78.2%	75.7%	81.8%	59.8%	5.61	12.07	1.35
YOLOv5n	80.6%	75.4%	81.9%	59.6%	1.79	4.2	0.72
YOLOv5s	82.6%	79.5%	85.9%	63.5%	7.06	15.9	1.53
YOLOv6n	79.7%	74.3%	81.9%	59.4%	4.63	11.35	0.99
YOLOv7-tiny	79.9%	75.5%	82.5%	59.5%	6.06	13.2	0.85
YOLOv8n	81.0%	78.7%	84.6%	63.2%	3.01	8.1	0.88
Ours	84.3%	81.3%	87.8%	65.8%	2.81	7.8	1.06

Table 4. Comparative experiments on MAR20 dataset

View in the Article

Model	Precision	Recall	mAP50	mAP50-95	Parameters/M	FLOPs/B	Latency/ms
RT-DETR	87.6%	86.3%	88.9%	56.1%	19.88	57.0	4.34
Gold-YOLO	89.9%	79.0%	86.0%	48.2%	5.61	12.06	2.07
YOLOv5n	90.2%	78.6%	86.1%	46.8%	1.77	4.2	1.20
YOLOv5s	92.5%	85.3%	89.6%	51.6%	7.04	15.8	1.79
YOLOv6n	85.2%	80.9%	85.6%	47.7%	4.63	11.34	1.55
YOLOv7-tiny	87.2%	87.0%	89.4%	49.5%	6.03	13.1	2.29
YOLOv8n	86.1%	82.1%	87.7%	54.1%	3.01	8.1	0.87
Ours	91.1%	87.2%	90.2%	54.4%	2.80	7.8	1.05

Table 5. Comparative experiments on NWPU VHR-10 dataset

Download Citation

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享