Infrared and Laser Engineering, Volume. 54, Issue 1, 20240376(2025)
FDLIE-YOLO: Frequency domain enhanced end-to-end low-light image target detection method
Figures & Tables(14)
Fig. 7. Detection effect of each end-to-end target detection method in Exdark dataset
Fig. 8. Test results of mAP metrics in Exdark dataset with different detection modulators
Table 1. Metrics averages for different network models on the LOL-Real dataset
View tableView in ArticleTable 1. Metrics averages for different network models on the LOL-Real dataset
Model MBLLEN[ 11 ]Zero-DCE[ 13 ]KinD[ 12 ]FECNet[ 21 ]SNR-Aware[ 22 ]PENet[ 15 ]FDLIENet(Ours) PSNR↑/dB 17.90 14.83 20.01 20.67 23.48 21.48 23.18 SSIM↑ 0.702 0.531 0.841 0.795 0.847 0.838 0.858 Parameter/×1012 0.45 0.07 8.02 0.15 39.12 0.21 0.14
Table 2. Comparison of Exdark metrics for each augmented network on Exdark data
View tableView in ArticleTable 2. Comparison of Exdark metrics for each augmented network on Exdark data
Model MBLLEN[ 11 ]Zero-DCE[ 13 ]KinD[ 12 ]FECNet[ 21 ]SNR-Aware[ 22 ]PENet[ 15 ]FDLIENet (Ours) NIQE↓ 4.38 5.05 4.19 4.33 3.88 4.28 3.98 Parameter/×1012 0.45 0.07 8.02 0.15 39.12 0.21 0.16
Table 3. Comparison of the detection performance of FDLIE-YOLO with the combined method of preprocessing network and target detection network
View tableView in ArticleTable 3. Comparison of the detection performance of FDLIE-YOLO with the combined method of preprocessing network and target detection network
Metrics Bicycle Boat Bottle Bus Car Cat Chair Cup Dog Mbike People Table mAP YOLOv8n 81.8 75.3 79.2 91.2 83.2 70.0 76.1 79.2 78.2 77.3 81.8 56.5 77.4 MBLLEN+YOLOv8n[ 11 ]83.0 78.1 80.5 93.3 84.1 71.6 77.3 81.1 80.4 78.6 82.1 58.6 79.0 Zero-DCE+YOLOv8n[ 13 ]82.1 76.9 80.3 92.8 83.7 72.3 77.8 80.6 79.4 77.9 81.9 56.8 78.5 KinD+YOLOv8n[ 12 ]83.1 78.7 82.1 93.1 83.9 72.9 78.1 81.4 80.3 78.3 82.0 59.8 79.4 FDLIE-YOLO(YOLOv8n ) 83.9 79.4 81.9 94.2 85.8 71.8 77.9 82.5 81.5 78.6 83.1 61.8 80.2
Table 4. Comparison of detection performance of each end-to-end low-light target detection method
View tableView in ArticleTable 4. Comparison of detection performance of each end-to-end low-light target detection method
Metrics Bicycle Boat Bottle Bus Car Cat Chair Cup Dog Mbike People Table mAP YOLOv8 81.8 75.3 79.2 91.2 83.2 70.0 76.1 79.2 78.2 77.3 81.8 56.5 77.4 IA-YOLO[ 22 ]86.2 79.8 81.2 92.3 84.1 71.9 71.1 82.2 80.8 77.9 82.2 58.8 79.0 PE-YOLO 85.8 79.3 80.7 93.8 83.9 70.5 77.5 81.4 80.2 79.3 83.8 58.5 79.5 FDLIE-YOLO(Ours) 85.2 79.4 82.1 94.2 85.8 72.6 78.9 82.5 81.5 78.6 83.1 64.1 80.6
Table 5. Comparison of the detection performance of each detection method
View tableView in ArticleTable 5. Comparison of the detection performance of each detection method
Metrics mAP FPS Param/×1012 Dynamic R-CNN[ 8 ]60.5% 38.65 41.3 RT-DETR[ 10 ]70.7% 84.89 42.0 YOLOv5n 66.4% 81.15 1.77 YOLOv7 75.8% 85.33 36.5 YOLOv8n 77.4% 89.89 3.01 FDLIE-YOLO(Ours) 80.6% 88.82 3.16
Table 6. Performance analysis of each component
View tableView in ArticleTable 6. Performance analysis of each component
Metrics Exdark mAP FPS FDLIE-YOLO (Ours) 80.6% 88.82 Ours w/o FDPB 79.8% 89.28 Ours w/o Lamp 80.1% 89.68
Tools
Get Citation
Copy Citation Text
Yang LI, Xianguo LI, Lian CHEN, Qingyong YANG, Changyu XU, Sheng XU. FDLIE-YOLO: Frequency domain enhanced end-to-end low-light image target detection method[J]. Infrared and Laser Engineering, 2025, 54(1): 20240376
Paper Information
Category: 图像处理
Received: Oct. 3, 2024
Accepted: --
Published Online: Feb. 12, 2025
The Author Email: CHEN Lian (chenlhit@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20