Infrared Ship Detection Using Attention Mechanism and Multiscale Fusion

Table 1. Performance comparison of light weight network
View table
Table 1. Performance comparison of light weight network
Network FLOPs /10⁶ Memory /MB Latency /ms
MobileNetv3 68.62 18.99 27
ShuffleNetv2 149.58 20.84 41
GhostNetv1 148.17 40.04 90

Table 2. Comparison experiment of target detection networks

View table

Table 2. Comparison experiment of target detection networks

Method	Precision /%	Recall /%	F1 score	mAP@ 0.5 /%	mAP@ 0.5∶0.95 /%	Speed /（frame·s^-1）
CenterNet	88.16	91.21	0.90	91.71	58.80	33
SSD	82.03	80.50	0.81	86.86	42.80	39
EfficientDet	83.66	88.98	0.86	91.85	57.40	26
RetinaNet	85.00	86.17	0.85	90.88	53.30	13
YOLOv3	87.35	75.66	0.81	84.87	39.75	20
YOLOv4s	89.26	78.04	0.83	86.71	41.30	22
YOLOv5s	91.50	88.70	0.90	92.78	60.88	62
YOLOv7	91.80	89.64	0.91	93.11	61.45	60
YOLOv8s	92.20	87.33	0.90	92.21	64.72	105

Table 3. Results of ablation experiments

View table

Table 3. Results of ablation experiments

Alias	Improvement strategy				Precision	Recall	mAP@0.5	mAP@0.5∶0.95
Alias	MobileNetv3	BiFPN	SENet	Wise IoU	Precision	Recall	mAP@0.5	mAP@0.5∶0.95
A	×	×	×	×	0.914	0.884	0.923	0.609
B	√	×	×	×	0.894	0.847	0.899	0.582
C	√	√	×	×	0.911	0.884	0.919	0.604
D	√	√	√	×	0.921	0.894	0.928	0.613
E	√	√	√	√	0.923	0.909	0.935	0.621

Table 4. Comparative experiments of attention mechanisms
View table
Table 4. Comparative experiments of attention mechanisms
Model Precision Recall mAP@0.5 mAP@0.5∶0.95
YOLOv7-C 0.911 0.884 0.919 0.604
YOLOv7-C+SENet 0.921 0.894 0.928 0.613
YOLOv7-C+CA 0.925 0.890 0.927 0.612
YOLOv7-C+CBAM 0.928 0.886 0.925 0.615

Table 5. Lightweighting results
View table
Table 5. Lightweighting results
Model Inference time for GPU /ms Inference time for CPU /ms FLOPs /10⁹ Parameter /10⁶
YOLOv7 12.1 554.6 105.2 37.2
Proposed algorithm 5.1 150.5 36.3 22.9

Tools

Get Citation

Copy Citation Text

Shen Zhang, Lin Hu, Xiang'e Sun, Meihua Liu. Infrared Ship Detection Using Attention Mechanism and Multiscale Fusion[J]. Laser & Optoelectronics Progress, 2023, 60(22): 2212008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites