Author Affiliations
1School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China2School of Physics and Electronic Information Engineering, Qinghai Nationalities University, Xining, Qinghai 810007, Chinashow less
Fig. 1. Framework of StarGAN
Fig. 2. RRDB network structure. (a) RRDB module; (b) dense block structure
Fig. 3. Framework of the proposed method
Fig. 4. Network structure of generative model
Fig. 5. Network structure of discriminative model
Fig. 6. Four categories of label images. (a)Indoor image; (b) bluish underwater image; (c) greenish underwater image; (d) yellowish underwater image
Fig. 7. Sharpness results of synthesized underwater images. (a) Underwater images; (b) results of CLAHE algorithm; (c) results of RED algorithm; (d) results of CycleGAN algorithm; (e) results of UWGAN algorithm; (f) results of StarGAN algorithm; (g) results of proposed method; (h) original images
Fig. 8. Sharpness results of the bluish underwater images. (a) Underwater images; (b) results of CLAHE algorithm; (c) results of RED algorithm; (d) results of CycleGAN algorithm; (e) results of UWGAN algorithm; (f) results of StarGAN algorithm; (g) results of proposed method
Fig. 9. Sharpness results of the greenish underwater images. (a) Underwater images; (b) results of CLAHE algorithm; (c) results of RED algorithm; (d) results of CycleGAN algorithm; (e) results of UWGAN algorithm; (f) results of StarGAN algorithm; (g) results of proposed method
Fig. 10. Sharpness results of the yellowish underwater images. (a) Underwater images; (b) results of CLAHE algorithm; (c) results of RED algorithm; (d) results of CycleGAN algorithm; (e) results of UWGAN algorithm; (f) results of StarGAN algorithm; (g) results of proposed method
Fig. 11. Comparison of experimental results with different modules. (a) Underwater images; (b) results with residual block; (c) results of proposed method
Image style | Number of images |
---|
Train | Test for realunderwater image | Test for synthesizedunderwater image |
---|
Indoor image | 3154 | 280 | 0 | Bluish underwater image | 2921 | 449 | 280 | Greenish underwater image | 1265 | 319 | 280 | Yellowish underwater image | 968 | 129 | 280 |
|
Table 1. Number of images used for training and testing
Method | Bluish underwater image | Greenish underwater image | Yellowish underwater image |
---|
UCIQE | Informationentropy | UCIQE | Informationentropy | UCIQE | Informationentropy |
---|
CLAHE | 0.4705 | 7.0841 | 0.4873 | 7.0974 | 0.4630 | 7.0848 | RED | 0.5460 | 6.9678 | 0.5541 | 7.1506 | 0.5459 | 6.6468 | CycleGAN | 0.5647 | 7.2450 | 0.5614 | 7.1677 | 0.5582 | 7.4000 | UWGAN | 0.5707 | 7.2809 | 0.5614 | 7.3234 | 0.5817 | 7.3333 | StarGAN | 0.5651 | 6.9598 | 0.5583 | 7.2112 | 0.5589 | 7.0824 | Ours | 0.5776 | 7.2996 | 0.5750 | 7.3177 | 0.5868 | 7.6801 |
|
Table 2. Comparison of UCIQE and information entropy using different methods in the synthesized image
Method | Bluish underwater image | Greenish underwater image | Yellowish underwater image |
---|
SSIM | MSE | SSIM | MSE | SSIM | MSE |
---|
CLAHE | 0.8408 | 1823.5 | 0.8713 | 815.3 | 0.8445 | 1686.6 | RED | 0.8752 | 1274.9 | 0.9212 | 332.7 | 0.7977 | 2353.5 | CycleGAN | 0.4231 | 4367.0 | 0.4173 | 4280.1 | 0.3371 | 5105.6 | UWGAN | 0.8587 | 514.4 | 0.8848 | 492.9 | 0.8650 | 561.3 | StarGAN | 0.7490 | 836.4 | 0.8388 | 313.7 | 0.7996 | 540.9 | Ours | 0.9052 | 241.8 | 0.9223 | 203.2 | 0.9019 | 545.1 |
|
Table 3. Comparison of SSIM and MSE using different methods in the synthesized image
Method | Bluish underwater image | Greenish underwater image | Yellowish underwater image |
---|
UCIQE | Informationentropy | UCIQE | Informationentropy | UCIQE | Informationentropy |
---|
Original image | 0.4937 | 7.3015 | 0.4951 | 7.4309 | 0.4741 | 6.9084 | CLAHE | 0.4872 | 7.2702 | 0.5108 | 7.4257 | 0.4987 | 7.1475 | RED | 0.5401 | 7.1654 | 0.5568 | 7.3826 | 0.5261 | 6.7265 | CycleGAN | 0.5686 | 7.4598 | 0.5897 | 7.5928 | 0.5763 | 7.3997 | UWGAN | 0.5681 | 7.2693 | 0.5871 | 7.4656 | 0.5990 | 7.1673 | StarGAN | 0.5918 | 7.4753 | 0.5876 | 7.5821 | 0.5872 | 7.4947 | Ours | 0.6221 | 7.6430 | 0.6092 | 7.6902 | 0.6012 | 7.5778 |
|
Table 4. Comparison of UCIQE and information entropy using different methods in the real image
Network | Bluish underwater image | Greenish underwater image | Yellowish underwater image |
---|
UCIQE | Informationentropy | UCIQE | Informationentropy | UCIQE | Informationentropy |
---|
Original image | 0.4937 | 7.3015 | 0.4951 | 7.4309 | 0.4741 | 6.9084 | Residual block | 0.5744 | 7.3729 | 0.5819 | 7.6010 | 0.5779 | 7.3933 | Ours | 0.6221 | 7.6430 | 0.6092 | 7.6902 | 0.6012 | 7.5778 |
|
Table 5. Comparison results of UCIQE and information entropy of different networks