Lightweight and robust ship detection method driven by self-attention mechanism

Feng MA; Zihui SHI; Jie SUN; Chen CHEN; Xianbin MAO; Xinping YAN

doi:10.19693/j.issn.1673-3185.03389

Journals >Chinese Journal of Ship Research >Volume 19 >Issue 5 >Page 188 > Article

Chinese Journal of Ship Research
Vol. 19, Issue 5, 188 (2024)

Lightweight and robust ship detection method driven by self-attention mechanism

Feng MA^1,2, Zihui SHI^1,2, Jie SUN³, Chen CHEN^3,4..., Xianbin MAO⁵ and Xinping YAN²|Show fewer author(s)

Author Affiliations

¹School of Computer Science and Artificial Intelligence, Wuhan University of Technology, Wuhan 430063, China

²Intelligent Transportation Systems Research Center, Wuhan University of Technology, Wuhan 430063, China

³Nanjing Smart Water Transportation Technology Co., Ltd, Nanjing 210028, China

⁴School of Computer Science and Technology, Wuhan Institute of Technology, Wuhan 430205, China

⁵Zhoushan Haihua Passenger Transport Co., Ltd, Zhoushan 316111, China

show less

DOI: 10.19693/j.issn.1673-3185.03389 Cite this Article

Feng MA, Zihui SHI, Jie SUN, Chen CHEN, Xianbin MAO, Xinping YAN. Lightweight and robust ship detection method driven by self-attention mechanism[J]. Chinese Journal of Ship Research, 2024, 19(5): 188 Copy Citation Text

show less

Fig. 1. [in Chinese]

Download full size | View in the Article

Fig. 2. [in Chinese]

Download full size | View in the Article

Fig. 3. [in Chinese]

Download full size | View in the Article

Fig. 4. [in Chinese]

Download full size | View in the Article

Fig. 5. [in Chinese]

Download full size | View in the Article

Fig. 6. [in Chinese]

Download full size | View in the Article

Fig. 7. [in Chinese]

Download full size | View in the Article

Fig. 8. [in Chinese]

Download full size | View in the Article

Fig. 9. [in Chinese]

Download full size | View in the Article

参数	数值
学习率	0.01
余弦退火超参数	0.1
动量	0.937
权重衰退	0.000 5
批量大小	16
输入尺寸	640
优化器	SGD

Table 1. Initial parameters of the model

View in the Article

算法模型	平均精度	平均召回率	mAP@0.5	mAP@0.5:0.95	参数量
YOLOv3	0.877	0.901	0.915	0.615	61 513 575
YOLOv5	0.919	0.868	0.916	0.618	46 124 433
YOLOv7	0.905	0.869	0.926	0.620	36 497 954
Faster R-CNN	0.538	0.331	0.533	0.268	138 357 544
SSD	0.643	0.558	0.521	0.281	48 306 395
ShipDet	0.921	0.903	0.929	0.629	35 366 310

Table 2. Comparison results of evaluation indicators for algorithm models

View in the Article

算法模型	mAP@0.5:0.95
算法模型	普通船	集装箱船	拖轮	客船	全类
YOLOv3	0.523	0.602	0.715	0.62	0.615
YOLOv5	0.527	0.600	0.718	0.627	0.618
YOLOv7	0.508	0.631	0.719	0.621	0.620
Faster R-CNN	0.129	0.403	0.255	0.285	0.268
SSD	0.143	0.399	0.240	0.342	0.281
ShipDet	0.546	0.618	0.719	0.623	0.629

Table 3. Experimental results of different types of ships

View in the Article

算法模型	平均精度(mAP@0.5:0.95)			平均召回率( mAP@0.5:0.95)
算法模型	小尺寸船舶	中尺寸船舶	大尺寸船舶	小尺寸船舶	中尺寸船舶	大尺寸船舶
YOLOv3	0.389	0.664	0.763	0.474	0.729	0.806
YOLOv5	0.410	0.671	0.785	0.483	0.729	0.827
YOLOv7	0.421	0.683	0.770	0.501	0.754	0.813
Faster R-CNN	0.004	0.252	0.551	0.009	0.378	0.600
SSD	0.142	0.355	0.593	0.102	0.411	0.616
ShipDet	0.457	0.677	0.767	0.503	0.739	0.808

Table 4. Comparison results of different models on different scales

View in the Article

模型	SIoU损失函数	STR模块	平均精度	平均召回率	mAP@0.5	mAP@0.5:0.95
YOLOv5L	×	×	0.903	0.868	0.916	0.618
SD-NET	√	×	0.898	0.902	0.919	0.618
	×	√	0.911	0.893	0.918	0.621
	√	√	0.921	0.903	0.929	0.629

Table 5. Results of ablation experiments

View in the Article

算法模型	平均精度	平均召回率	mAP@0.5	mAP@0.5:0.95
YOLOv3	0.980	0.991	0.993	0.829
YOLOv5	0.985	0.976	0.923	0.832
YOLOv7	0.976	0.977	0.991	0.836
Faster R-CNN	0.937	0.959	0.900	0.786
SSD	0.887	0.890	0.892	0.564
ShipDet	0.990	0.991	0.993	0.899