End-to-end recognition of nighttime wildlife based on semi-supervised learning

No.	Label name	Source 1	Source 2	Total
1	Armadillo	3 305	0	3 305
2	Bird	3 215	0	3 215
3	Bobcat	0	3 104	3 104
4	Cat	0	2 635	2 635
5	Coyote	0	3 357	3 357
6	Dasyprocta_punctata	3 628	0	3 628
7	Deer	509	3 503	4 012
8	Dog	0	712	712
9	Fox	63	852	915
10	Giant_anteater	485	0	485
11	Leopardus	314	0	314
12	Mustelidae	59	1 058	1 117
13	Opossum	5 049	6 757	11 806
14	Paca	3 833	0	3 833
15	Peccary	7 249	0	7 249
16	Rabbit	0	3 266	3 266
17	Raccoon	1 038	3 802	4 840
18	Squirrel	53	129	182
19	Tamandua	1 120	0	1 120
20	Tapirus	1 383	0	1 383
Sum		31 303	29 175	60 478

Table 2. Detection performance for application of different enhancement modules to YOLOv8

View table

View in Article

Table 2. Detection performance for application of different enhancement modules to YOLOv8

Method	Additional modules	Extra time/ms	mAP₅₀	mAP_50∶95
YOLOv8-n	None	0	0.571	0.352
Two-stage enhancement	Zero-DCE	160.58	0.566	0.351
	Zero-DCE++	62.73	0.571	0.356
	RetinexNet	634.61	0.595	0.414
End-to-end enhancement	CPA-Enhancer	262.61	0.561	0.365
End-to-end enhancement	PENet	166.06	0.560	0.345
Attention mechanism	Channel attention （backbone）	5.77	0.583	0.367
	Channel attention （neck）	5.46	0.578	0.357
	Spatial attention （backbone）	4.06	0.591	0.370
	Spatial attention （neck）	4.10	0.586	0.361
	CBAM（backbone）	8.28	0.572	0.361
	CBAM（neck）	8.68	0.582	0.362
	FFA（backbone）	31.72	0.574	0.360
	FFA（neck）	32.16	0.596	0.379

Table 3. Performance comparison of semi-supervised and fully supervised learning
View table
View in Article
Table 3. Performance comparison of semi-supervised and fully supervised learning
Label ratio Burn-in Supervised Semi-supervised
mAP₅₀ mAP_50：95 mAP₅₀ mAP_50：95 mAP₅₀ mAP_50∶95
5% 0.484 0.278 0.596 0.379 0.697 0.509
10% 0.673 0.444 0.749 0.520 0.798 0.604

Table 4. Performance comparison of semi-supervised and fully supervised learning
View table
View in Article
Table 4. Performance comparison of semi-supervised and fully supervised learning
Attention Semi-supervision mAP₅₀ mAP_50：95
0.571 0.352
√ 0.596（↑4.38%） 0.379（↑7.67%）
√ 0.687（↑20.32%） 0.490（↑39.20%）
√ √ 0.697（↑22.07%） 0.509（↑44.60%）

Table 5. Performance comparison of semi-supervised and fully supervised learning methods

View table

View in Article

Table 5. Performance comparison of semi-supervised and fully supervised learning methods

Method	FLOPs/G	mAP₅₀	mAP_50∶95
YOLOv3-tiny	18.9	0.553	0.344
YOLOv5-n	7.2	0.554	0.368
YOLOv8-s	28.7	0.651	0.485
SSD	6.8	0.536	0.311
RT-DETR-r18	57.3	0.610	0.464
Soft-teacher	202.3	0.674	0.466
Efficient-teacher	5.72	0.662	0.450
Ours	5.90	0.697	0.509

Tools

Get Citation

Copy Citation Text

Han LU, Bolun CUI, Huayang WAN, Guofeng ZHANG, Chen SHEN, Chi WANG. End-to-end recognition of nighttime wildlife based on semi-supervised learning[J]. Optics and Precision Engineering, 2025, 33(5): 789

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Dec. 24, 2024

Accepted: --

Published Online: May. 20, 2025

The Author Email:

DOI:10.37188/OPE.20253305.0789

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Distribution of labeled data in datasets

Table 1. Distribution of labeled data in datasets

Table 2. Detection performance for application of different enhancement modules to YOLOv8

Table 2. Detection performance for application of different enhancement modules to YOLOv8

Table 3. Performance comparison of semi-supervised and fully supervised learning

Table 3. Performance comparison of semi-supervised and fully supervised learning

Table 4. Performance comparison of semi-supervised and fully supervised learning

Table 4. Performance comparison of semi-supervised and fully supervised learning

Table 5. Performance comparison of semi-supervised and fully supervised learning methods

Table 5. Performance comparison of semi-supervised and fully supervised learning methods