Opto-Electronic Engineering, Volume. 50, Issue 12, 230225-1(2024)
Low-light image enhancement based on dual-frequency domain feature aggregation
Figures & Tables(15)
Fig. 6. Comparison of enhancement results of mit-adobe fivek dataset
Fig. 7. Comparison of experimental effects of modular attention structure
Fig. 9. Comparison of effect diagrams of modular ablation experiments
Fig. 10. Test results of monitoring images of low-light vehicles at night
Table 1. LOL real-world dataset results
View tableView in ArticleTable 1. LOL real-world dataset results
Method PSNR SSIM LPIPS RetinexNet[ 26 ]16.7740 0.4250 0.4739 Zero-DCE[ 27 ]14.8607 0.5624 0.3352 DSLR[ 28 ]14.9822 0.5964 0.3757 KinD[ 29 ]17.6476 0.7715 0.1750 EnGAN[ 30 ]17.4829 0.6515 0.3223 GLAD[ 32 ]19.7182 0.6820 0.3994 RUAS[ 31 ]16.4047 0.5034 0.2078 R2RNet[ 10 ]20.2070 0.8160 - UHDFour[ 8 ]23.0926 0.8720 - URetinexNet[ 6 ]21.3282 0.8348 - Ours 24.3714 0.8937 0.1525
Table 2. MIT-Adobe FiveK dataset results
View tableView in ArticleTable 2. MIT-Adobe FiveK dataset results
Method PSNR SSIM LPIPS Exposure[ 33 ]18.7412 0.8159 0.1674 CycleGAN[ 34 ]19.3823 0.7852 0.1636 RetinexNet[ 26 ]12.5146 0.6708 0.2535 DSLR[ 28 ]20.2435 0.8289 0.1526 KinD[ 29 ]16.2032 0.7841 0.1498 EnGAN[ 30 ]17.9050 0.8361 0.1425 Zero-DCE[ 27 ]15.9312 0.7668 0.1647 Zero-DCE++[ 35 ]14.6111 0.4055 0.2309 RUAS[ 31 ]15.9953 0.7863 0.1397 Ours 22.7214 0.8726 0.1153
Table 3. Comparison results of module attention structure testing
View tableView in ArticleTable 3. Comparison results of module attention structure testing
Method PSNR SSIM LPIPS Baseline 22.7052 0.8147 0.2078 With serial of CA & SA 23.6042 0.8283 0.1837 With parallel of CA & SA 24.3714 0.8937 0.1525
Table 4. Experimental results of network module ablation
View tableView in ArticleTable 4. Experimental results of network module ablation
Model FDIEM MSAM DDFAM PSNR SSIM Baseline × × × 20.8620 0.8515 Model-1 × √ √ 21.3582 0.8653 Model-2 √ × √ 22.0401 0.8878 Model-3 √ √ × 21.9068 0.8919 Ours √ √ √ 24.3714 0.8937
Table 5. Comparison of different network average processing time, model size and floating-point operations
View tableView in ArticleTable 5. Comparison of different network average processing time, model size and floating-point operations
Model Time/ms Params/M FLOPs/G PSNR SSIM RetinexNet[ 26 ]20 9.2 136.0151 16.7740 0.4250 Zero-DCE[ 27 ]2 0.97 5.2112 14.8671 0.5624 KinD[ 29 ]10 35 29.1303 20.3792 0.7715 EnGAN[ 30 ]20 33 61.0102 17.4828 0.6515 GLAD[ 32 ]25 11 252.1410 19.7182 0.6820 MBLLEN[ 36 ]80 1.95 19.9560 17.8583 0.7247 LPNet[ 37 ]18 0.15 0.7700 21.4612 0.8020 URetinexNet[ 6 ]2.93 0.34 1801.4110 21.3282 0.8348 Ours 48 1.61 288.3776 24.3714 0.8937
Tools
Get Citation
Copy Citation Text
Shengjun Xu, Hua Yang, Minghai Li, Guanghui Liu, Yuebo Meng, Jiuqiang Han. Low-light image enhancement based on dual-frequency domain feature aggregation[J]. Opto-Electronic Engineering, 2024, 50(12): 230225-1
Paper Information
Category: Research Articles
Received: Sep. 8, 2023
Accepted: Dec. 3, 2023
Published Online: Mar. 26, 2024
The Author Email: Yang Hua (杨华)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20