Laser & Optoelectronics Progress, Volume. 55, Issue 3, 031012(2018)
Convolution Neural Network Image Defogging Based on Multi-Feature Fusion
Figures & Tables(13)
Fig. 2. Fog image database. (a) Original images; (b) transmission images; (c) generated foggy images
Fig. 3. Comparison of defogging results of cones. (a) Original image; (b) foggy image; images defogged by (c) Fattal algorithm, (d) He algorithm, (e) Meng algorithm, (f) Galdran algorithm, and (g) proposed method
Fig. 4. Comparison of defogging results of reindeer. (a) Original image; (b) foggy image; images defogged by (c) Fattal algorithm, (d) He algorithm, (e) Meng algorithm, (f) Galdran algorithm, and (g) proposed method
Fig. 5. Comparison of defogging reconstruction results of teddy. (a) Original image; (b) foggy image; images defogged by (c) Fattal algorithm, (d) He algorithm, (e) Meng algorithm, (f) Galdran algorithm, and (g) proposed method
Fig. 6. Contrastive experimental networks. (a) Compared network 1; (b) compared network 2
Fig. 7. Comparison of different defogging algorithms in natural scene. (a) Original image; images defogged by (b) Fattal algorithm, (c) He algorithm, (d) Meng algorithm, (e) Galdran algorithm, and (f) proposed method
Table 1. Multi-scale feature fusion parameters
View tableTable 1. Multi-scale feature fusion parameters
Layer type Configurations Convolution Fm-16, Kernel-1x1, pad-0, stride-1, PReLU Convolution Fm-16, Kernel-3x3, pad-1, stride-1, PReLU Convolution Fm-16, Kernel-5x5, pad-2, stride-1, PReLU Convolution Fm-16, Kernel-7x7, pad-3, stride-1, PReLU
Table 2. Quantitative comparison of different algorithms on the dataset
View tableTable 2. Quantitative comparison of different algorithms on the dataset
Image Fattal algorithm He algorithm Meng algorithm Galdran algorithm Proposed algorithm RMSE PSNR /dB RMSE PSNR /dB RMSE PSNR /dB RMSE PSNR /dB RMSE PSNR /dB Art 8.571 9.271 6.938 13.396 4.794 18.895 5.712 16.396 1.701 27.531 Bowling 8.091 9.526 5.354 16.529 4.221 19.245 6.010 15.674 1.691 28.018 Dolls 9.072 8.628 5.246 17.816 4.332 19.638 6.779 15.458 1.785 27.808 Reindeer 8.106 9.572 6.887 13.036 4.899 17.869 7.248 11.901 1.891 27.628 Cones 8.853 9.026 5.772 16.041 4.505 18.769 6.046 15.399 1.789 27.762 Teddy 8.669 9.472 6.211 15.309 4.565 18.743 5.650 16.529 1.816 27.719
Table 3. Comparison of PSNR results of different networksdB
View tableTable 3. Comparison of PSNR results of different networksdB
Image Multi-feature fusion network Compared network 1 Compared network 2 Art 27.531 27.013 26.556 Bowling 28.018 27.223 27.021 Dolls 27.808 27.396 26.528 Reindeer 27.628 27.118 26.829 Cones 27.762 27.369 27.010 Teddy 27.719 27.282 26.903
Table 4. Comparison of experimental results with different networks
View tableTable 4. Comparison of experimental results with different networks
Network PSNR /dB RMSE SSIM Dehazenet 70.9767 0.0787 0.9993 Multi-scale 71.0122 0.0774 0.9993
Table 5. Comparison of algorithm evaluation index in natural scene
View tableTable 5. Comparison of algorithm evaluation index in natural scene
Method Figure Average Standard deviation Information entropy Color entropy Mean gradient Contrast Fattal Pumpkin 68.4971 61.3326 4.8398 14.2389 10.3731 7.6360 Rice field 45.9848 64.0856 5.0569 14.9456 11.6958 14.4347 Forest 81.1157 66.5824 7.0815 20.6306 13.8976 13.9009 He Pumpkin 99.3012 64.9204 7.6468 22.3456 11.1668 8.9083 Rice field 77.0238 59.2287 7.2348 21.0209 12.2187 13.3602 Forest 87.1907 61.9286 7.3366 21.0584 12.1417 13.6283 Meng Pumpkin 84.5235 60.2172 7.3457 21.3919 11.6948 7.5397 Rice field 74.0616 63.7294 6.9783 20.0772 11.9825 13.5021 Forest 85.5998 60.5597 7.2569 20.5494 11.4450 13.3889 Galdran Pumpkin 94.2411 50.1966 7.5738 22.5294 11.3859 7.6224 Rice field 85.7172 47.1278 7.5238 21.4681 11.1032 10.4738 Forest 87.6054 43.5790 7.3710 20.3288 10.4662 11.5028 Proposed Pumpkin 84.5925 66.6132 7.6645 20.8777 10.1556 10.4233 Rice field 79.6453 67.8168 7.1421 21.7686 11.2558 15.6302 Forest 87.5713 58.2448 7.3712 20.5125 12.3077 14.0203
Tools
Get Citation
Copy Citation Text
Yan Xu, Meishuang Sun. Convolution Neural Network Image Defogging Based on Multi-Feature Fusion[J]. Laser & Optoelectronics Progress, 2018, 55(3): 031012
Paper Information
Category: Image processing
Received: Sep. 26, 2017
Accepted: --
Published Online: Sep. 10, 2018
The Author Email: Xu Yan (xuyan@tju.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20