Strong Interference Target Detection on the Sea Surface Based on Feature Augmentation

Xiang Long; Huajie Chen; Haoyu Wu; Di Yu

doi:10.3788/LOP241156

Journals >Laser & Optoelectronics Progress >Volume 62 >Issue 2 >Page 0212008 > Article

Laser & Optoelectronics Progress
Vol. 62, Issue 2, 0212008 (2025)

Strong Interference Target Detection on the Sea Surface Based on Feature Augmentation

Xiang Long^*, Huajie Chen, Haoyu Wu, and Di Yu

Author Affiliations

School of Automation, Hangzhou Dianzi University, Hangzhou 310018, Zhejiang , China

show less

DOI: 10.3788/LOP241156 Cite this Article Set citation alerts

Xiang Long, Huajie Chen, Haoyu Wu, Di Yu. Strong Interference Target Detection on the Sea Surface Based on Feature Augmentation[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0212008 Copy Citation Text

show less

Fig. 1. Structure of YOLOv8 network

Download full size

Fig. 2. Overall structure of TFF module

Download full size

Fig. 3. Structure of TS-YOLOv8 network

Download full size

Fig. 4. Partial images in datasets

Download full size

Fig. 5. Variation trend of mAP₅₀. (a) On the AFO dataset; (b) on the SeaDronesSee dataset

Download full size

Fig. 6. Comparison of detection results and heatmaps on the AFO test set by baseline model and proposed model. (a) Input images; (b)‒(c) detection results; (d)‒(e) heatmaps

Download full size

Fig. 7. Comparison of detection results and heatmaps on the SeaDronesSee test set by baseline model and proposed model. (a) Input images; (b)‒(c) detection results; (d)‒(e) heatmaps

Download full size

Parameter	Configuration
Python version	3.11
PyTorch version	2.0.0
CUDA version	11.7
Learning rate	0.01
Optimizer	Stochastic gradient descent （SGD）
Batch size	16
Epoch	100
Momentum	0.937
Weight decay	5×10^-4

Table 1. Configuration and parameters setting for experiment

Algorithm	Parameters /10⁶	GFLOPs	P /%	R /%	mAP₅₀/%	mAP₉₅ /%	FPS /（frame·s^-1）
Faster RCNN	136.79	369.84	42.96	47.97	41.01	17.39	17
SSD	24.28	61.25	72.26	19.96	44.10	22.26	39
YOLOv3	61.55	65.63	83.70	46.12	53.99	20.69	35
YOLOv4	63.97	59.99	68.28	20.82	35.79	13.80	62
YOLOv5	46.17	108.31	94.01	51.01	57.83	31.69	73
YOLOX	54.15	155.69	94.31	75.26	76.26	46.10	69
YOLOv7	37.22	105.20	87.40	76.76	84.88	44.80	81
RetinaNet	36.43	146.91	77.69	28.06	35.24	19.36	71
CenterNet	32.67	70.22	90.89	52.02	61.71	28.89	82
RT-DETR	42.31	136.74	89.45	83.79	88.09	47.68	108
YOLOv8	11.64	28.43	90.1	84.69	89.54	53.67	114
TS-YOLOv8	12.17	29.06	94.54	91.56	95.14	61.05	110

Table 2. Overall performance of different algorithms on the AFO dataset

Algorithm	AP₅₀						mAP₅₀
Algorithm	Human	Board	Boat	Buoy	Sailboat	Kayak	mAP₅₀
Faster RCNN	19.26	69.32	45.78	—	34.01	77.69	41.01
SSD	22.12	87.83	31.83	10.06	34.13	78.63	44.10
YOLOv3	62.11	87.92	30.34	30.08	31.11	82.40	53.99
YOLOv4	32.49	78.54	12.90	—	28.70	62.13	35.79
YOLOv5	82.04	97.90	46.53	16.65	15.80	88.05	57.83
YOLOX	91.43	98.88	58.10	86.20	34.69	88.23	76.26
YOLOv7	82.72	98.72	68.40	79.34	80.88	99.24	84.88
RetinaNet	4.29	74.36	37.72	—	33.79	61.26	35.24
CenterNet	65.54	82.38	55.02	45.85	33.56	87.93	61.71
RT-DETR	84.78	98.26	77.51	84.18	84.92	98.79	88.07
YOLOv8	82.75	98.19	80.64	88.18	87.92	99.54	89.54
TS-YOLOv8	89.82	99.14	91.54	94.48	96.28	99.60	95.14

Table 3. AP of each category in the AFO dataset detected by different algorithms

Model	Parameters /10⁶	GFLOPs	P /%	R /%	mAP₅₀ /%	mAP₉₅ /%
YOLOv8	11.64	28.43	90.10	84.69	89.54	53.67
YOLOv8+EMA	11.64	28.40	91.07	85.79	89.75	53.69
YOLOv8+SimAM	11.64	28.43	91.29	85.30	90.06	55.01
YOLOv8+CA	11.75	28.91	90.95	85.90	90.18	54.92
YOLOv8+GAM	11.71	28.77	90.10	84.97	89.66	53.70
YOLOv8+CBAM	11.64	28.53	92.10	86.36	91.09	55.15
YOLOv8+SA	11.64	28.43	91.91	87.02	91.45	55.74

Table 4. Comparison of different attention mechanisms

YOLOv8	TFF	SA	Parameters /10⁶	GFLOPs	P /%	R /%	mAP₅₀ /%	mAP₉₅ /%	FPS /（frame·s^-1）
√			11.64	28.43	90.10	84.69	89.54	53.67	114
√	√		12.17	29.06	92.35	88.61	93.46	59.79	112
√		√	11.64	28.43	91.91	87.02	91.45	55.74	112
√	√	√	12.17	29.06	94.54	91.56	95.14	61.05	110

Table 5. Overall performance of each ablation algorithm on the AFO dataset

YOLOv8	TFF	SA	AP₅₀						mAP₅₀
YOLOv8	TFF	SA	Human	Board	Boat	Buoy	Sailboat	Kayak	mAP₅₀
√			82.75	99.19	80.64	78.18	90.92	99.54	89.54
√	√		88.27	99.29	92.52	86.60	94.49	99.56	93.46
√		√	87.09	98.48	87.27	84.35	91.98	99.53	91.45
√	√	√	89.82	99.14	91.54	94.48	96.28	99.60	95.14

Table 6. AP of each category in AFO dataset detected by each ablation algorithm

YOLOv8	TFF	SA	Parameters /10⁶	GFLOPs	P /%	R /%	mAP₅₀/%	mAP₉₅/%	FPS /（frame·s^-1）
√			11.64	28.43	88.42	82.82	86.87	57.61	110
√	√		12.17	29.06	90.37	86.14	90.09	61.91	107
√		√	11.64	28.43	88.93	85.75	88.71	59.47	108
√	√	√	12.17	29.06	91.66	87.59	91.34	63.53	106

Table 7. Overall performance of each ablation algorithm on the SeaDronesSee dataset

YOLOv8	TFF	SA	AP₅₀					mAP₅₀
YOLOv8	TFF	SA	Swimmer	Boat	Jetski	Life jacket	Buoy	mAP₅₀
√			80.15	92.63	90.76	81.01	89.80	86.87
√	√		85.75	94.13	92.93	85.08	92.56	90.09
√		√	83.99	93.49	91.86	83.05	91.16	88.71
√	√	√	87.04	95.41	94.60	86.38	93.27	91.34

Table 8. AP of each category in SeaDronesSee dataset detected by each ablation algorithm

Xiang Long, Huajie Chen, Haoyu Wu, Di Yu. Strong Interference Target Detection on the Sea Surface Based on Feature Augmentation[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0212008

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