A lightweight detection of multi-spectral infrared ship targe

Spectral Band	Backbone	mAP₅₀/%	mAP_50～95/%	Param/M	Flops/G	MemFootprint /M
B3	Darknet	88.4	58.3	3.0	8.2	6.0
B3	FasterNet	87.2	57.3	4.1	10.7	8.6
B3	EfficientViT	87.7	58.2	4.0	9.5	8.7
B3	HGnetv2	88.6	59.7	2.4	7.1	5.0
B3	GSHGnetv2	88.5	59.5	2.3	6.9	4.7
Fusion	DarkNet	95.2	66.9	7.6	18.6	13.4
Fusion	FasterNet	94.3	65.1	14.1	22.3	27.9
Fusion	EfficientViT	94.6	65.6	13.7	19.7	27.1
Fusion	HGnetv2	95.3	67.6	5.0	14.9	11.2
Fusion	GSHGnetv2	95.3	67.5	4.8	14.6	10.7

Table 2. Comparison of different pruning algorithms

View table

View in Article

Table 2. Comparison of different pruning algorithms

Model	Pruning	mAP₅₀/%	mAP_50～95/%	Param/M	Flops/G	MemFootprint /M
YOLOv5	L1-norm	80.1	50.9	1.47	1.3	3.3
	L2-norm	81.8	51.4	1.39	1.3	3.1
	Group-Slim	84.2	53.2	1.26	1.3	2.8
	DepGraph	86.1	55.3	0.39	1.2	0.9
	La-Depgraph	86.3	55.7	0.37	1.2	0.9
YOLOv8	L1-norm	80.3	51.2	1.25	1.7	3.1
	L2-norm	82.7	52.1	1.16	1.7	2.9
	Group-Slim	85.4	54.3	1.08	1.7	2.6
	DepGraph	88.2	58.3	0.52	1.6	1.2
	La-Depgraph	88.7	59.1	0.48	1.6	1.2
Ours	L1-norm	89.7	57.8	2.48	3.5	6.6
	L2-norm	91.1	59.2	2.35	3.5	6.1
	Group-Slim	92.2	61.3	2.23	3.6	5.6
	DepGraph	94.2	66.4	0.96	3.5	2.4
	La-Depgraph	94.9	67.4	0.90	3.5	2.3

Table 3. Comparison of different training set ratios

View table

View in Article

Table 3. Comparison of different training set ratios

Ratio	Model	mAP₅₀/%	mAP_50～95/%
7∶3	Teacher	95.6	70.1
	Student	93.8	66.2
	Distill-Student	95.1	69.6
8∶2	Teacher	96.2	70.5
	Student	94.5	66.9
	Distill-Student	95.9	69.8
9∶1	Teacher	96.6	71.0
	Student	94.9	67.4
	Distill-Student	96.4	70.6

Table 4. Ablation experiments
View table
View in Article
Table 4. Ablation experiments
LW-Net Pruning KD mAP₅₀/% mAP_50～95/%
Param
/M
Flops
/G
- - - 95.2 66.9 7.6 18.6
√ - - 95.5 67.8 4.7 14.3
√ √ - 94.9 67.4 0.9 3.5
√ √ √ 96.4 70.6 0.9 3.5

Table 5. Performance of different models

View table

View in Article

Table 5. Performance of different models

Spectral Band	Model	mAP₅₀/%	mAP_50～95/%	Param/M	Flops/G	MemFootprint /M
B3	SSD512	84.3	53.8	24.0	87.8	92.1
B3	Faster RCNN	86.6	56.5	136.8	200.9	521.9
B3	YOLOv5n	86.9	56.8	1.9	4.5	5.0
B3	YOLOv7tiny	87.4	57.2	6.0	13.0	12.3
B3	YOLOv8n	88.4	58.3	3.0	8.2	6.0
B3	YOLOv8s	90.0	60.5	11.1	28.6	22.5
B3	YOLOv10n	87.9	57.6	2.7	8.4	5.7
B1＆B3	YOLOv5-ACDF	91.5	63.4	8.0	12.7	16.1
B1＆B3	YOLOv7-CMF	92.2	64.3	16.6	50.9	33.4
B1＆B3	YOLOv8-CFT	91.7	63.8	11.5	28.8	23.1
B1＆B3	Ours	93.5	65.9	0.7	2.5	1.6
Fusion	YOLOv8n-MF	95.2	66.9	7.6	18.6	13.4
Fusion	Ours	96.4	70.6	0.9	3.5	2.3

Tools

Get Citation

Copy Citation Text

Jiale ZHAO, Shuli LOU, Chao LIN. A lightweight detection of multi-spectral infrared ship targe[J]. Optics and Precision Engineering, 2025, 33(8): 1328

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites