Satellite Image Translation Method Based on Attention Residual Network

Jinyu Wang; Changgong Zhang; Haitao Yang; Bodi Feng; Gaoyuan Li; Yuge Gao

doi:10.3788/LOP202259.0210017

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 2 >Page 0210017 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 2, 0210017 (2022)

Satellite Image Translation Method Based on Attention Residual Network

Jinyu Wang, Changgong Zhang, Haitao Yang^*, Bodi Feng, Gaoyuan Li, and Yuge Gao

Author Affiliations

School of Space Information, Space Engineering University, Beijing 101416, China

show less

DOI: 10.3788/LOP202259.0210017 Cite this Article Set citation alerts

Jinyu Wang, Changgong Zhang, Haitao Yang, Bodi Feng, Gaoyuan Li, Yuge Gao. Satellite Image Translation Method Based on Attention Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0210017 Copy Citation Text

show less

Fig. 1. Generating adversarial network architecture

Download full size

Fig. 2. Separating attention residual module

Download full size

Fig. 3. Consistent principle of circulation

Download full size

Fig. 4. Network architecture

Download full size

Fig. 5. Generator

Download full size

Fig. 6. Image data of SAR

Download full size

Fig. 7. Dataset visualization

Download full size

Fig. 8. Results of semantic translation

Download full size

Fig. 9. FID change during generation phase

Download full size

Fig. 10. Loss function curves. (a) CycleGAN; (b) proposed method

Download full size

Fig. 11. Subjective evaluation of ablation experiments

Download full size

Fig. 12. Translation results for other datasets. (a) Dataset 2; (b) dataset 3; (c) dataset 4

Download full size

Type	Procedure	Name	Number	Input size	Output size	Normalization	Activation function	Padding
Encoder	1	Conv	1	3 $\times$ 256 $\times$ 256	64 $\times$ 256 $\times$ 256	In	ReLU	Yes
	2	Conv	1	64 $\times$ 256 $\times$ 256	128 $\times$ 128 $\times$ 128	In	ReLU	No
	3	Conv	1	128 $\times$ 128 $\times$ 128	256 $\times$ 64 $\times$ 64	In	ReLU	No
Converter	4	SeparConv	9	256 $\times$ 64 $\times$ 64	256 $\times$ 64 $\times$ 64	In	ReLU	Yes
Converter	5	SERestnet	9	256 $\times$ 64 $\times$ 64	256 $\times$ 64 $\times$ 64	In	ReLU	No
Decoder	6	DeConv	1	256 $\times$ 64 $\times$ 64	128 $\times$ 128 $\times$ 128	In	ReLU	No
	7	DeConv	1	128 $\times$ 128 $\times$ 128	64 $\times$ 256 $\times$ 256	In	ReLU	No
	8	DeConv	1	64 $\times$ 256 $\times$ 256	3 $\times$ 256 $\times$ 256	In	ReLU	Yes

Table 1. Generator network configuration

Type of translation	Dataset	Terrestrial	DataA	DataB	Data source	Size	Data volume（A）	Data volume（B）
SemanticTranslation	1	City	Maps	Images	CycleGAN^［12］	256	1098	1096
Style Translation	2	Factory	Images	Images	RSI-CB	256	2511	2445
	3	Vegetation	Images	SAR	RSI-CB、GF-3	256	2358	848
	4	Farmland	Images	SAR	RSI-CB、GF-3	256	2491	656

Table 2. Content of dataset

method	Parameter /MB	Model size /MB	PSNR		SSIM		FID
method	Parameter /MB	Model size /MB	Map→image	Image→map	Map→imag	Image→map	Map→imag	Image→map
CycleGAN	13.7	43.4	14.02	24.98	0.18	0.6	190.36	257.59
Pixpix	54.4	207	14.96	17.09	0.21	0.51	233.69	322.38
Proposed method	4.8	25.8	14.67	26.68	0.22	0.67	178.44	197.56

Table 3. Comparison of indicators

Algorithm	Parameter /MB	Model size /MB	FID（map→image）	FID（image→map）
CycleGAN	13.7	43.4	190.36	257.59
CycleGAN+A	13.9↑	43.7↑	181.68↓	208.89↓
CycleGAN+B	4.8↓	25.5↓	180.96↓	200.61↓
CycleGAN+C	13.7	43.4	181.74↓	199.74↓
SA-CycleGAN	4.8↓	25.8↓	178.44↓	197.56↓

Table 4. Objective evaluation of ablation experiments

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information