Lightweight Multi‑Scale Synthetic Aperture Radar Ship Detection Algorithm

Changzhen Xiong; Xiyu Li; Heyi Zhao; Songming Xie

doi:10.3788/AOS241848

Journals >Acta Optica Sinica >Volume 45 >Issue 5 >Page 0528003 > Article

Acta Optica Sinica
Vol. 45, Issue 5, 0528003 (2025)

Lightweight Multi‑Scale Synthetic Aperture Radar Ship Detection Algorithm

Changzhen Xiong¹, Xiyu Li^1,*, Heyi Zhao², and Songming Xie¹

Author Affiliations

¹School of Electrical and Control Engineering, North China University of Technology, Beijing 100144, China

²School of Mechanical and Materials Engineering, North China University of Technology, Beijing 100144, China

show less

DOI: 10.3788/AOS241848 Cite this Article Set citation alerts

Changzhen Xiong, Xiyu Li, Heyi Zhao, Songming Xie. Lightweight Multi‑Scale Synthetic Aperture Radar Ship Detection Algorithm[J]. Acta Optica Sinica, 2025, 45(5): 0528003 Copy Citation Text

show less

Fig. 1. Network structure of proposed algorithm

Download full size

Fig. 2. Structures of Conv, bottleneck, and C2f

Download full size

Fig. 3. Structure of MSPPF

Download full size

Fig. 4. Structure of EMA

Download full size

Fig. 5. Structure of LSCH

Download full size

Fig. 6. Comparisons of detection results between YOLOv8 and proposed algorithm on HRSID dataset

Download full size

Dataset	LSCH	MSPPF	Focaler-ShapeIoU	Params /M	FLOPs /G	P /%	R /%	AP_0.5 /%	AP_0.5∶0.95 /%
HRSID				3.0	8.2	91.0	82.0	90.5	65.8
	√			2.2	6.1	90.9	82.3	91.0	66.6
	√	√		2.3	6.3	92.3	83.8	91.7	67.2
	√	√	√	2.3	6.3	92.5	84.0	92.0	67.6
SSDD				3.0	8.2	94.9	91.4	97.0	69.1
	√			2.2	6.1	94.8	91.9	98.2	70.1
	√	√		2.3	6.3	95.8	94.4	98.4	72.0
	√	√	√	2.3	6.3	96.1	94.1	98.6	72.2

Table 1. Results of ablation experiments on HRSID and SSDD datasets

Dataset	Method	Focaler	AP_0.5 /%	AP_0.5∶0.95 /%
HRSID	CIoU	×	91.7	67.2
	CIoU	√	91.9	67.4
	SIoU	×	91.6	67.1
	SIoU	√	91.7	67.2
	ShapeIoU	×	91.8	67.3
	ShapeIoU	√	92.1	67.6
SSDD	CIoU	×	98.4	72.0
	CIoU	√	98.5	72.1
	SIoU	×	98.3	72.0
	SIoU	√	98.4	72.2
	ShapeIoU	×	98.5	72.1
	ShapeIoU	√	98.6	72.2

Table 2. Results of contrast experiments on loss functions

Dataset	Algorithm	Params /M	FLOPs /G	P /%	R /%	AP_0.5 /%	AP_0.5∶0.95 /%
HRSID	Faster R-CNN	41.4	134.4	82.4	90.5	88.0	—
	Libra R-CNN	42.8	141.3	83.6	89.7	88.8	—
	YOLOv3	103.7	283.0	90.3	84.0	91.0	66.3
	YOLOv8n	3.0	8.2	91.0	82.0	90.5	65.8
	FCOS	32.1	126.0	62.9	86.1	84.5	—
	FENDet	—	120.6	79.0	88.7	89.9	—
	GCBANet	—	948.0	—	—	89.8	—
	FL-CSEvROIE	107.6	—	—	—	90.2	—
	YOLO-CLF	—	—	91.8	83.7	90.4	62.7
	MCMA-Net	—	—	80.4	88.9	87.4	—
	Ours	2.3	6.3	92.5	84.0	92.0	67.6
SSDD	Faster R-CNN	41.4	134.4	67.2	94.5	89.5	—
	Libra R-CNN	42.8	141.3	66.8	94.7	88.9	—
	YOLOv3	103.7	283.0	90.7	93.7	95.0	70.2
	YOLOv8n	3.0	8.2	94.9	91.4	97.0	69.1
	FCOS	32.1	126.0	84.7	93.8	95.8	—
	FENDet	—	120.6	95.8	94.4	95.9	—
	GCBANet	—	948.0	—	—	95.4	—
	FL-CSE-ROIE	107.6	—	—	—	95.9	—
	YOLO-CLF	—	—	95.0	93.7	98.2	63.1
	MCMA-Net	—	—	84.7	97.9	96.6	—
	Ours	2.3	6.3	96.1	94.1	98.6	72.2

Table 3. Contrast experimental results on HRSID and SSDD datasets

Dataset	Input size	Model	AP_0.5 /%	AP_0.5∶0.95 /%
HRSID	320	YOLOv8	77.6	52.6
	320	Ours	79.0	53.3
	480	YOLOv8	86.2	60.1
	480	Ours	88.4	63.3
	640	YOLOv8	90.5	65.8
	640	Ours	92.0	67.6
SAR-Ship-Dataset	320	YOLOv8	92.4	54.6
	320	Ours	92.6	55.0
	480	YOLOv8	92.2	54.0
	480	Ours	92.5	55.3
	640	YOLOv8	91.1	53.5
	640	Ours	91.6	54.8
SSDD	320	YOLOv8	94.0	66.2
	320	Ours	94.8	65.1
	480	YOLOv8	97.2	70.0
	480	Ours	97.4	70.2
	640	YOLOv8	97.0	69.1
	640	Ours	98.6	72.2

Table 4. Contrast experimental results at different resolutions

Changzhen Xiong, Xiyu Li, Heyi Zhao, Songming Xie. Lightweight Multi‑Scale Synthetic Aperture Radar Ship Detection Algorithm[J]. Acta Optica Sinica, 2025, 45(5): 0528003

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