Backscattered Light Repairing Method for Underwater Laser Image Based on Improved Generative Adversarial Network

Qingbo Zhang; Xiaohui Zhang; Hongwei Han

doi:10.3788/LOP56.041004

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 4 >Page 041004 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 4, 041004 (2019)

Backscattered Light Repairing Method for Underwater Laser Image Based on Improved Generative Adversarial Network

Qingbo Zhang^*, Xiaohui Zhang, and Hongwei Han

Author Affiliations

College of Weaponry Engineering, Naval University of Engineering, Wuhan, Hubei 430033, China

show less

DOI: 10.3788/LOP56.041004 Cite this Article Set citation alerts

Qingbo Zhang, Xiaohui Zhang, Hongwei Han. Backscattered Light Repairing Method for Underwater Laser Image Based on Improved Generative Adversarial Network[J]. Laser & Optoelectronics Progress, 2019, 56(4): 041004 Copy Citation Text

show less

Fig. 1. Network structure of the GAN

Download full size

Fig. 2. Schematic of training set. (a) Sample image; (b) image with backscattered light; (c) image with mixed noise

Download full size

Fig. 3. Analysis of dilated convolution

Download full size

Fig. 4. Analysis of jumping network

Download full size

Fig. 5. Curve of model training

Download full size

Fig. 6. Processing results of noise parameter (0, 20 dB, 0.01). (a) Target image; (b) image to be repaired; (c) Denoise+DCP; (d) Denoise+HEMSRCR; (e) proposed method

Download full size

Fig. 7. Processing results of noise parameter (0, 25 dB, 0.015). (a) Target image; (b) image to be repaired; (c) Denoise+DCP; (d) Denoise+HEMSRCR; (e) proposed method

Download full size

Fig. 8. Processing results of noise parameter (0, 30 dB, 0.02). (a) Target image; (b) image to be repaired; (c) Denoise+DCP; (d) Denoise+HEMSRCR; (e) proposed method

Download full size

Name	Kernel size	Stride	Dilation rate	Output size	BN	Dropout
input	-	-	-	256×256×1	-	-
conv1…conv2conv3	64×3×3128×3×3	12	11	256×256×64128×128×128	PP	--
conv4	128×3×3	1	1	128×128×128	P	-
conv5	256×3×3	2	1	64×64×256	P	-
conv6…conv7dilaconv8	256×3×3256×3×3	11	12	64×64×25664×64×256	PP	--
dilaconv9	256×3×3	1	4	64×64×256	P	-
dilaconv10	256×3×3	1	8	64×64×256	P	-
dilaconv11	256×3×3	1	16	64×64×256	P	-
conv12…conv13	256×3×3	1	1	64×64×256	P	-
transconv14	128×4×4	2	-	128×128×128	P	-
merge1(conv4 + transconv14)	-	-	-	128×128×256	-	P
conv15	128×3×3	1	1	128×128×128	P	-
transconv16	64×4×4	2	1	256×256×64	P	-
merge2(conv2 + transconv16)	-	-	-	256×256×128	-	P
conv17	32×3×3	1	1	256×256×32	P	-
output	1×3×3	1	1	256×256×1	-	-

Table 1. Detailed configuration information of the generator networkpixel

Name	Kernel size	Stride	Dilation rate	Output size	BN	Dropout
Input	-	-	-	256×256×1	-	-
Conv1Conv2	64×5×5128×5×5	22	11	128×128×6464×64×128	PP	--
Conv3	256×5×5	2	1	32×32×256	P	-
Conv4	512×5×5	2	1	16×16×512	P	-
Conv5	512×5×5	2	1	8×8×512	P	-
Conv6	512×5×5	2	1	4×4×512	P	-
FC	-	-	-	1	-	-

Table 2. Detailed configuration information of the discriminator networkpixel

Item	Train set	Validation set		Test set
Number ofimages	5850	650	40	40	40
Size /(pixel×pixel)	256×256	256×256	256×256	512×512	960×960

Table 3. Structure of data set

Image	Denoise			Denoise+DCP			Denoise+ HEMSRCR			Proposed method
Image	Lv1	Lv2	Lv3	Lv1	Lv2	Lv3	Lv1	Lv2	Lv3	Lv1	Lv2	Lv3
1	16.94	16.96	16.90	10.67	10.75	10.74	6.05	6.07	6.06	22.61	22.31	21.50
2	14.78	14.80	14.76	8.60	8.67	8.69	5.93	5.96	5.98	21.97	19.14	21.15
3	11.86	11.93	11.98	4.55	4.57	4.61	5.51	5.52	5.51	17.38	20.65	18.24
4	13.40	13.45	13.47	6.92	7.06	7.09	5.67	5.70	5.71	18.52	18.64	17.81
5	12.33	12.36	12.39	5.61	5.74	5.77	5.56	5.57	5.59	10.40	10.27	9.50
6	15.11	15.14	15.18	8.65	8.76	8.84	5.84	5.88	5.89	20.72	17.24	20.49

Table 4. The PSNR of different test imagesdB

Image	Denoise			Denoise+DCP			Denoise+ HEMSRCR			Proposed method
Image	Lv1	Lv2	Lv3	Lv1	Lv2	Lv3	Lv1	Lv2	Lv3	Lv1	Lv2	Lv3
1	0.91	0.91	0.91	0.90	0.91	0.91	0.59	0.58	0.53	0.93	0.92	0.92
2	0.88	0.88	0.89	0.88	0.88	0.89	0.61	0.60	0.57	.90	0.91	0.90
3	0.92	0.91	0.90	0.90	0.91	0.92	0.65	0.62	0.55	0.92	0.91	0.91
4	0.87	0.87	0.88	0.87	0.87	0.88	0.61	0.59	0.56	0.92	0.89	0.89
5	0.89	0.89	0.90	0.89	0.89	0.90	0.68	0.66	0.60	0.91	0.89	0.88
6	0.89	0.89	0.90	0.89	0.89	0.90	0.60	0.59	0.55	0.93	0.91	0.91