Optics and Precision Engineering, Volume. 33, Issue 2, 247(2025)
Enhanced edge feature extraction dual branch fusion network for real image dehazing
Figures & Tables(14)
Fig. 1. Structure of enhanced-edge-feature dual-branch fusion dehazing network
Fig. 6. Dehazing results for different methods in the Middlebury dataset
Fig. 7. Performance comparison of different methods for PSNR and SSIM evaluation
Fig. 9. Dehazing results of different methods in the BeDDE dataset
Fig. 10. Qualitative comparison of processing results for different network branches
Table 1. Quantitative comparison in the SOTS dataset and the Middlebury dataset
View tableView in ArticleTable 1. Quantitative comparison in the SOTS dataset and the Middlebury dataset
Methods RESIDE-SOTS MIDDLEBURY PSNR ↑ SSIM ↑ LPIPS ↓ PSNR ↑ SSIM ↑ LPIPS ↓ DCP 18.885 8 0.827 2 0.106 14.731 9 0.790 5 0.205 NLD 18.627 6 0.847 5 0.133 13.818 1 0.771 8 0.222 PSD 14.606 2 0.791 7 0.135 12.133 1 0.738 3 0.291 Dehamer 16.132 2 0.857 7 0.152 12.788 3 0.741 6 0.125 FFA-Net 17.009 6 0.866 4 0.091 14.035 9 0.744 7 0.231 DehazeNet 18.855 8 0.861 9 0.093 13.821 2 0.788 7 0.198 AOD-Net 19.591 1 0.860 8 0.089 13.186 3 0.777 3 0.223 MSBDN 19.236 5 0.890 5 0.074 15.386 5 0.814 3 0.191 MSCNN 19.675 1 0.848 1 0.087 13.694 8 0.791 2 0.199 EPDN 20.432 2 0.878 7 0.109 14.933 3 0.800 4 0.183 RefineDNet 20.839 4 0.885 2 0.084 14.222 1 0.792 5 0.214 Ours 22.096 1 0.916 0 0.079 14.857 9 0.806 5 0.189
Table 2. Quantitative comparisons in RTTS datasets
View tableView in ArticleTable 2. Quantitative comparisons in RTTS datasets
Methods PIQE ↓ NIQE ↓ BRISQUE ↓ FADE ↓ Entropy ↑ Hazy 36.210 3.648 39.848 3.755 7.057 DehazeNet 29.362 3.398 26.028 1.681 7.284 AOD-Net 31.875 3.616 29.185 1.259 7.199 PSD 27.189 3.190 21.326 1.371 7.591 MSCNN 24.127 3.427 23.111 1.366 7.383 FFA-Net 34.972 3.582 36.754 3.238 7.181 EPDN 23.424 3.554 25.319 1.051 7.454 RefineDNet 27.709 3.363 20.739 1.521 7.250 Ours 21.471 3.256 21.135 1.234 7.436
Table 3. Quantitative comparisons in the BeDDE dataset
View tableView in ArticleTable 3. Quantitative comparisons in the BeDDE dataset
Method RI ↑ VI ↑ DCP 0.943 7 0.884 0 PSD 0.945 2 0.877 3 DehazeNet 0.957 2 0.880 6 AOD-Net 0.949 7 0.874 0 FFA-Net 0.952 3 0.872 5 EPDN 0.941 4 0.876 0 MSCNN 0.959 2 0.881 7 RefineDNet 0.962 0 0.890 1 Ours 0.971 1 0.900 3
Table 4. Ablation study
View tableView in ArticleTable 4. Ablation study
TR Branch × √ × √ FE Branch × × √ √ PSNR ↑ 20.508 0 21.441 6 21.526 4 22.187 2 SSIM ↑ 0.900 1 0.902 2 0.913 3 0.920 2 LPIPS ↓ 0.084 0.090 0.077 0.073 PIQE ↓ 28.889 24.743 26.241 21.471 NIQE ↓ 3.313 3.028 3.215 3.256 BRISQUE ↓ 20.739 21.906 22.267 21.135 FADE ↓ 1.377 1.179 1.263 1.234 Entropy ↑ 7.101 7.339 7.258 7.436
Tools
Get Citation
Copy Citation Text
Xiongxin LI, Fengling XIA, Kaomin ZHANG, Hongliang WANG, Tao XIE. Enhanced edge feature extraction dual branch fusion network for real image dehazing[J]. Optics and Precision Engineering, 2025, 33(2): 247
Paper Information
Category:
Received: Aug. 26, 2024
Accepted: --
Published Online: Apr. 30, 2025
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20