Trans-YOLO: Improved YOLOv8 with RT-DETR Decoder &amp; Head for Infrared Small Target Detection

Jiannan Liu; Shuxian Liu; Hankiz Yilahun; Askar Hamdulla

doi:10.3788/LOP241915

Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1037001(2025)

Trans-YOLO: Improved YOLOv8 with RT-DETR Decoder & Head for Infrared Small Target Detection

Jiannan Liu... Shuxian Liu, Hankiz Yilahun and Askar Hamdulla* |Show fewer author(s)

Author Affiliations

Department of Computer Science and Technology, Xinjiang University, Urumqi 830046, Xinjiang , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Network architecture schematic of RT-DETR

Download full size

Fig. 2. Network structure diagram of proposed algorithm

Download full size

Fig. 3. Structure diagrams of the three modules. (a) Network structure diagram of C2f; (b) structure diagram of RGCSPELAN module; (c) network structure diagram of RepConv

Download full size

Fig. 4. Structure diagram of CAFM

Download full size

Fig. 5. CAFMFusion mechanism

Download full size

Fig. 6. Detection performance of the BiFPN and CAFMFusion across different datasets. (a)‒(c) BiFPN; (d)‍‒‍(f) CAFMFusion

Download full size

Fig. 7. Evaluation metrics comparing Trans-YOLO with RT-DETR-r18 and YOLOv8 on IRSTD-1K and NUDT-SIRST datasets. (a)‒(d) Precision, recall, mAP@50, and mAP@50∶95 on dataset NUDT-SIRST; (e)‒(h) precision, recall, mAP@50, and mAP@50∶95 on dataset IRSTD-1K

Download full size

Fig. 8. Comparison of different target detection algorithms on infrared images with Trans-YOLO detection results in multiple scenes

Download full size

Fig. 9. Impact of each improved module on detection results

Download full size

Table 1. Comparative experiments of RGCSPELAN module with common blocks
View table
Table 1. Comparative experiments of RGCSPELAN module with common blocks
Block Mean time /s FPS /frame/s FLOPs /Gbit Parameter /10³
C3 0.00114 875.5 4.83 295.6
ELAN 0.00143 698.6 8.05 492.2
C2f 0.00128 780.1 7.51 459.5
RepNCSPELAN 0.00224 446.4 3.69 226.1
RGCSPELAN 0.00112 891.4 3.56 217.9

Table 2. Introduction of publicly available infrared small target datasets used in the experiments

View table

Table 2. Introduction of publicly available infrared small target datasets used in the experiments

Class	Dataset	Number of sample		Label form	Dataset introduction
Class	Dataset	Training set	Test set	Label form	Dataset introduction
Single frame	NUDT-SIRST	1061	265	Anchor， mask	Dataset contains multiple target types， rich target sizes， and different clutter backgrounds
	IRSTD-1K	800	201	Anchor， mask	Dataset contains objects of different shapes and sizes， rich backgrounds， and accurate pixel-level annotations
	SIRST	341	86	Anchor， mask	First real infrared small target dataset with high-quality images and labels

Table 3. Hyperparameter settings for experiments
View table
Table 3. Hyperparameter settings for experiments
Hyperparameter Detail
Epoch 450
Image size /（pixel×pixel） 640×640
Batch size 16
Optimizer AdamW
Initial LR 0.01
Patience 80

Table 4. Comparison of performance metrics of the BiFPN and CAFMFusion on the public datasets
View table
Table 4. Comparison of performance metrics of the BiFPN and CAFMFusion on the public datasets
Algorithm NUDT-SIRST SIRST IRSTD-1K
mAP@50 F1-score mAP@50 F1-score mAP@50 F1-score
Detr+BiFPN 0.953 0.940 0.834 0.870 0.811 0.850
Detr+CAFMFusion 0.969 0.960 0.853 0.844 0.824 0.840

Table 5. Comparison of experimental results of different algorithms on the datasets

View table

Table 5. Comparison of experimental results of different algorithms on the datasets

Model	Parameter /10³	IRSTD-1K				NUDT-SIRST
Model	Parameter /10³	F1-score	mAP@50	Precision	Recall	F1-score	mAP@50	Precision	Recall
Faster RCNN	37.60	‒	0.637	‒	‒	‒	0.734	‒	‒
MBFormer-YOLO	7.10	0.836	0.894	0.826	0.847	0.940	0.979	0.982	0.961
FCDN	4.10	0.827	0.768	0.844	0.831	0.895	0.917	0.923	0.907
SFM-YOLOv8	2.83	0.499	0.462	0.532	0.471	0.604	0.579	0.621	0.589
YOLOv5	2.00	0.823	0.820	0.839	0.817	0.934	0.956	0.950	0.920
YOLOv7	36.40	0.844	0.811	0.862	0.809	0.920	0.929	0.911	0.932
YOLOv8	3.00	0.765	0.784	0.834	0.687	0.930	0.965	0.920	0.928
YOLOv9	20.00	0.783	0.774	0.799	0.771	0.930	0.960	0.926	0.936
YOLOv10	2.60	0.735	0.730	0.717	0.755	0.921	0.958	0.906	0.940
RT-DETR-r18	19.00	0.807	0.783	0.813	0.802	0.990	0.992	0.990	0.991
Trans-YOLO	5.00	0.871	0.861	0.865	0.878	0.992	0.995	0.994	0.991

Table 6. Experimental results of improved algorithm on the SIRST dataset and its comparison with other algorithms

View table

Table 6. Experimental results of improved algorithm on the SIRST dataset and its comparison with other algorithms

Algorithm	Precision	Recall	mAP@50	mAP@95	Parameter /10³	F1-score
SSD	0.717	0.660	0.732	0.291	7.7	0.689
YOLOHG-LAD	0.872	0.814	0.815	0.299	1.7	0.840
MBFormer-YOLO	0.839	0.828	0.887	0.392	7.1	0.836
FCDN	0.841	0.819	0.801	0.303	4.1	0.827
SFM-YOLOv8	0.644	0.675	0.571	0.220	2.8	0.499
YOLOv5	0.852	0.834	0.847	0.324	2.0	0.823
YOLOv7	0.827	0.817	0.813	0.311	36.4	0.844
YOLOv8	0.688	0.743	0.729	0.284	3.0	0.720
YOLOv9	0.820	0.845	0.831	0.305	20.0	0.870
YOLOv10	0.730	0.746	0.721	0.295	2.6	0.737
Trans-YOLO	0.828	0.817	0.866	0.394	5.9	0.860

Table 7. Ablation experiments on dataset IRSTD-1K and dataset NUDT-SIRST

View table

Table 7. Ablation experiments on dataset IRSTD-1K and dataset NUDT-SIRST

Algorithm	IRSTD-1K			NUDT-SIRST
Algorithm	mAP@50	Precision	Recall	mAP@50	Precision	Recall
YOLOv8	0.784	0.834	0.687	0.965	0.920	0.928
YOLOv8+Decoder &Head	0.826	0.832	0.831	0.991	0.978	0.975
YOLOv8+RGCSPELAN	0.799	0.784	0.802	0.985	0.966	0.948
YOLOv8+CAFMFusion	0.804	0.812	0.824	0.994	0.994	0.995
YOLOv8+RGCSPELAN+CAFMFusion	0.833	0.828	0.882	0.981	0.968	0.942
YOLOv8+Decoder &Head+RGCSPELAN+CAFMFusion	0.861	0.865	0.878	0.995	0.994	0.991

Tools

Get Citation

Copy Citation Text

Jiannan Liu, Shuxian Liu, Hankiz Yilahun, Askar Hamdulla. Trans-YOLO: Improved YOLOv8 with RT-DETR Decoder & Head for Infrared Small Target Detection[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1037001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Aug. 27, 2024

Accepted: Oct. 28, 2024

Published Online: May. 13, 2025

The Author Email: Askar Hamdulla (askar@xju.edu.cn)

DOI:10.3788/LOP241915

CSTR:32186.14.LOP241915

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparative experiments of RGCSPELAN module with common blocks

Table 1. Comparative experiments of RGCSPELAN module with common blocks

Table 2. Introduction of publicly available infrared small target datasets used in the experiments

Table 2. Introduction of publicly available infrared small target datasets used in the experiments

Table 3. Hyperparameter settings for experiments

Table 3. Hyperparameter settings for experiments

Table 4. Comparison of performance metrics of the BiFPN and CAFMFusion on the public datasets

Table 4. Comparison of performance metrics of the BiFPN and CAFMFusion on the public datasets

Table 5. Comparison of experimental results of different algorithms on the datasets

Table 5. Comparison of experimental results of different algorithms on the datasets

Table 6. Experimental results of improved algorithm on the SIRST dataset and its comparison with other algorithms

Table 6. Experimental results of improved algorithm on the SIRST dataset and its comparison with other algorithms

Table 7. Ablation experiments on dataset IRSTD-1K and dataset NUDT-SIRST

Table 7. Ablation experiments on dataset IRSTD-1K and dataset NUDT-SIRST

微信扫一扫：分享