Optics and Precision Engineering, Volume. 33, Issue 13, 2108(2025)
Three-stage low-light image enhancement based on illumination guidance
Figures & Tables(13)
Fig. 2. Network structure of pixel-wise enhancement and Fourier reconstruction stages
Fig. 4. Comparison of image enhancement results in outdoor low-light environments
Fig. 5. Comparison of image enhancement results in indoor low-light environments
Fig. 6. Detail comparison of enhancement results by various algorithms on LOLv1 dataset
Fig. 7. Detail comparison of enhancement results by various algorithms on LOLv2-Real dataset
Fig. 8. Detail comparison of enhancement results by various algorithms on LOLv2-Syn dataset
Fig. 9. Comparison of enhancement results on three unpaired datasets
Table 1. Comparative experimental results of PSNR/SSIM for different algorithms on LOL datasets （v1 and v2）
View tableView in ArticleTable 1. Comparative experimental results of PSNR/SSIM for different algorithms on LOL datasets （v1 and v2）
Method Complexity LOLv1 LOLv2-Real LOLv2-Syn Normal GT Mean Normal GT Mean Normal GT Mean Params/M FLOPs/G PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM RetinexNet［ 9 ］0.84 584.47 16.774 0.419 18.915 0.427 16.097 0.401 18.323 0.447 17.137 0.762 19.099 0.774 KinD［ 10 ］8.02 34.99 17.650 0.775 20.860 0.802 14.740 0.641 17.544 0.669 13.290 0.578 16.259 0.591 ZeroDCE［ 11 ］0.075 4.83 14.861 0.559 21.880 0.640 16.059 0.580 19.771 0.671 17.712 0.815 21.463 0.848 3DLUT［ 40 ］0.59 7.67 14.350 0.445 21.350 0.585 17.590 0.721 20.190 0.745 18.040 0.800 22.173 0.854 DRBN［ 41 ］5.47 48.61 16.290 0.617 19.550 0.746 20.290 0.831 - - 23.220 0.927 - - MIRNet［ 35 ］31.76 785.10 24.100 0.845 26.519 0.856 20.020 0.820 27.173 0.865 21.940 0.876 25.955 0.898 RUAS［ 36 ］0.003 0.83 16.405 0.500 18.654 0.518 15.326 0.488 19.061 0.510 13.765 0.638 16.584 0.719 LLFlow［ 37 ］17.42 358.40 21.149 0.854 24.998 0.871 17.433 0.831 25.421 0.877 24.807 0.919 27.961 0.930 EnlightenGAN［ 12 ］114.35 61.01 17.480 0.651 20.003 0.691 18.230 0.617 - - 16.570 0.734 - - Restormer［ 21 ］26.13 144.25 22.365 0.816 26.682 0.853 18.693 0.834 26.116 0.853 21.413 0.830 25.428 0.859 LEDNet［ 39 ］7.07 35.92 20.627 0.823 25.470 0.846 19.938 0.827 27.814 0.870 23.709 0.914 27.367 0.928 SNR-Aware［ 13 ］4.01 26.35 24.610 0.842 26.716 0.851 21.480 0.849 27.209 0.871 24.140 0.927 27.787 0.941 PairLIE［ 8 ］0.33 20.81 19.510 0.736 23.526 0.755 19.885 0.778 24.025 0.803 21.351 0.857 25.326 0.873 LLFormer［ 38 ］24.55 22.52 23.649 0.816 25.758 0.823 20.056 0.792 26.197 0.819 24.038 0.909 28.006 0.927 Retinexformer［ 14 ］1.53 15.85 25.153 0.845 27.140 0.850 22.794 0.840 27.694 0.856 25.670 0.930 28.992 0.939 FourLLIE［ 17 ］0.12 2.57 22.526 0.831 26.037 0.849 22.127 0.843 27.591 0.871 24.763 0.924 28.121 0.926 PRNet［ 42 ］1.93 8.08 23.452 0.837 26.860 0.856 21.314 0.848 27.661 0.883 25.274 0.931 28.657 0.943 Ours 1.81 14.49 23.202 0.849 26.968 0.867 22.260 0.856 27.880 0.882 25.225 0.936 28.939 0.947
Table 2. Comparative experimental results of LPIPS for different algorithms on LOL datasets （v1 and v2）
View tableView in ArticleTable 2. Comparative experimental results of LPIPS for different algorithms on LOL datasets （v1 and v2）
Method LOLv1 LOLv2-Real LOLv2-Syn Complexity Normal GT Mean Normal GT Mean Normal GT Mean FLOPs/G RetinexNet 0.474 0.470 0.543 0.519 0.255 0.247 587.47 KinD 0.207 0.201 0.375 0.368 0.460 0.435 34.99 MIRNet 0.131 0.128 0.144 0.142 0.058 0.057 758.10 RUAS 0.270 0.257 0.309 0.296 0.364 0.341 0.83 LLFlow 0.119 0.117 0.176 0.158 0.067 0.063 358.40 EnlightenGAN 0.322 0.317 0.309 0.301 0.220 0.213 61.01 Restormer 0.141 0.138 0.231 0.214 0.144 0.139 144.25 SNR-Aware 0.151 0.150 0.156 0.150 0.057 0.056 26.35 LLFormer 0.175 0.167 0.211 0.209 0.066 0.061 22.52 LEDNet 0.118 0.113 0.120 0.114 0.061 0.056 35.92 Retinexformer 0.131 0.129 0.171 0.166 0.059 0.056 15.85 Ours 0.105 0.099 0.154 0.141 0.051 0.047 14.49
Table 3. Comparative experimental results of BRISQUE/NIQE for different algorithms on unpaired datasets （DICM/MEF/NPE）
View tableView in ArticleTable 3. Comparative experimental results of BRISQUE/NIQE for different algorithms on unpaired datasets （DICM/MEF/NPE）
Method DICM MEF NPE BRIS NIQE BRIS NIQE BRIS NIQE KinD 48.72 5.15 49.94 5.47 36.85 4.98 ZeroDCE 27.56 4.58 17.32 4.93 20.72 4.53 MIRNet 28.86 3.98 28.63 3.98 25.56 4.33 RUAS 38.75 5.21 23.68 3.83 47.85 5.53 LLFlow 26.36 4.06 30.27 4.70 28.86 4.67 SNR-Aware 37.35 4.71 31.28 4.18 26.65 4.32 PairLIE 33.31 4.03 27.53 4.06 28.27 4.18 Retinexformer 14.10 3.83 11.78 3.67 18.18 3.94 Ours 25.67 3.79 20.51 4.09 18.80 4.02
Table 4. Quantitative analysis result of ablation experiments
View tableView in ArticleTable 4. Quantitative analysis result of ablation experiments
Group Model LOL-v1 LOL-v2-Real LOL-v2-Syn 一阶段 二阶段 三阶段 PSNR SSIM LPIPS PSNR SSIM LPIPS PSNR SSIM LPIPS 1 √ 18.105 0.736 0.267 18.389 0.772 0.249 19.852 0.873 0.129 2 √ √ 18.656 0.774 0.252 21.895 0.837 0.183 21.486 0.894 0.164 3 √ √ 22.370 0.826 0.145 21.430 0.840 0.180 23.710 0.917 0.107 4 √ √ 22.269 0.824 0.140 21.330 0.842 0.146 23.249 0.915 0.090 5 √ √ √ 23.202 0.849 0.105 22.260 0.856 0.154 25.225 0.936 0.051
Tools
Get Citation
Copy Citation Text
Yihang ZHANG, Han ZHONG. Three-stage low-light image enhancement based on illumination guidance[J]. Optics and Precision Engineering, 2025, 33(13): 2108
Paper Information
Category:
Received: Mar. 24, 2025
Accepted: --
Published Online: Aug. 28, 2025
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20