Fig. 1. Framework of our algorithm
Fig. 2. Images processed by different stretching algorithms. (a) Original image; (b) image processed by traditional histogram stretching algorithm; (c) image processed by improved histogram stretching algorithm improved histogram stretching algorithm
Fig. 3. Structure of the GAN. (a) Generator; (b) convolution block; (c) discriminator
Fig. 4. Enhancement results of different methods on underwater images. (a) Original image; (b) HE; (c) RGHS; (d) UDCP; (e) IBLA; (f) Fusion; (g) Water-Net; (h) Uresnet; (i) FunieGAN; (j) UGAN; (k) ours
Fig. 5. Test results of different methods on the synthetic data set. (a) Orginal image; (b) Water-Net; (c) Uresnet; (d) FunieGAN; (e) UGAN; (f) ours
Fig. 6. Detail retention capabilities of different methods. (a) Original image; (b) details of the solid line frame; (c) details of the dotted line frame; (d) details of the dotted frame
Fig. 7. Detection results of underwater targets. (a) Original image; (b) enhanced image
Type | Method | PSNR /dB | SSIM |
---|
Model-based method | HE | 16.28 | 0.781 | RGHS | 19.49 | 0.832 | Fusion | 22.12 | 0.846 | UDCP | 12.64 | 0.578 | IBLA | 12.42 | 0.463 | Deeplearning-based method | Water-Net | 19.12 | 0.813 | Uresnet | 17.66 | 0.740 | FUnieGAN | 17.63 | 0.761 | UGAN | 19.60 | 0.840 | Ours | 20.59 | 0.844 |
|
Table 1. PSNR and SSIM of different methods on the UIEBD test set
Type | Method | UIQM | UICM | UISM | UIConM |
---|
Model-based method | HE | 2.906 | 11.127 | 7.554 | 0.101 | RGHS | 2.169 | 7.483 | 5.796 | 0.069 | Fusion | 2.922 | 8.177 | 8.004 | 0.091 | UDCP | 1.978 | 7.711 | 5.609 | 0.029 | IBLA | 1.718 | 8.115 | 4.474 | 0.046 | Deep learning-based method | Water-Net | 3.030 | 5.014 | 7.545 | 0.184 | Uresnet | 2.884 | 5.193 | 6.822 | 0.202 | FUnieGAN | 2.782 | 6.121 | 7.021 | 0.149 | UGAN | 3.423 | 5.988 | 8.307 | 0.235 | Ours | 3.471 | 5.483 | 8.413 | 0.232 |
|
Table 2. UIQM of different methods in EUVP data set
Dataset | Method | PSNR /dB | SSIM |
---|
EUVP | without penalty | 19.71 | 0.833 | with penalty | 20.59 | 0.844 | Synthesis | without penalty | 15.25 | 0.709 | with penalty | 15.60 | 0.716 |
|
Table 3. Influence of penalty items on the image enhancement effect
Data set | Method | PSNR /dB | SSIM |
---|
EUVP | UDCP | 18.65 | 0.832 | CycleGAN | 19.93 | 0.843 | Ours | 20.59 | 0.844 | Synthesis | UDCP | 14.19 | 0.687 | CycleGAN | 15.37 | 0.711 | Ours | 15.60 | 0.716 |
|
Table 4. Influence of different preprocessing methods on the image enhancement effect