Improved Algorithm for Super-Resolution Reconstruction of Remote-Sensing Images Based on Generative Adversarial Networks

Qiang Li; Xiyuan Wang; Jiawei He

doi:10.3788/LOP213046

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 10 >Page 1028010 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 10, 1028010 (2023)

Improved Algorithm for Super-Resolution Reconstruction of Remote-Sensing Images Based on Generative Adversarial Networks

Qiang Li, Xiyuan Wang^*, and Jiawei He

Author Affiliations

College of Physics and Electronic Engineering, Ningxia University, Yinchuan 750021, Ningxia, China

show less

DOI: 10.3788/LOP213046 Cite this Article Set citation alerts

Qiang Li, Xiyuan Wang, Jiawei He. Improved Algorithm for Super-Resolution Reconstruction of Remote-Sensing Images Based on Generative Adversarial Networks[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1028010 Copy Citation Text

show less

Fig. 1. Improved SR reconstruction network. (a) Structure of generating network; (b) structure of adversarial network

Download full size

Fig. 2. Structure of RRDB

Download full size

Fig. 3. Structure of RRFDB

Download full size

Fig. 4. Structure of RFB

Download full size

Fig. 5. PSNR values of different convolution combinations

Download full size

Fig. 6. Image reconstruction of different algorithms on Kaggle test dataset. (a) HR; (b) Bicubic; (c) SRGAN; (d) EDSR; (e) ESRGAN; (f) proposed algorithm

Download full size

Fig. 7. Image reconstruction of different algorithms on WHU-RS19 dataset. (a) HR; (b) Bicubic; (c) SRGAN; (d) EDSR; (e) ESRGAN; (f) proposed algorithm

Download full size

Fig. 8. Image reconstruction of different algorithms on AID dataset. (a) HR; (b) Bicubic; (c) SRGAN; (d) EDSR; (e) ESRGAN; (f) proposed algorithm

Download full size

Module	PSNR /dB	SSIM	FSIM
GAN+RFDB（16）	29.83	0.862	0.975
GAN+RRDB（16）	30.49	0.884	0.990
GAN+RRDB（16）+RFDB（4）	30.92	0.892	0.992
GAN+RRDB（16）+RFDB（6）	31.67	0.897	0.993
GAN+RRDB（16）+RFDB（8）	31.68	0.895	0.993

Table 1. Performance of algorithm on Kaggle test dataset under different module settings

Module	PSNR /dB	SSIM	FSIM
SRGAN original loss	29.71	0.844	0.930
$L_{c o n t}$	30.63	0.874	0.931
$L_{c o n t}$ + $L_{p e r c e p}$	31.21	0.883	0.931
$L_{c o n t}$ + $L_{p e r c e p}$ + $L_{a d v}$	31.42	0.886	0.964
$L_{c o n t}$ + $L_{p e r c e p}$ + $L_{a d v}$ + $L_{T V}$	31.67	0.897	0.993

Table 2. Performance of algorithm on Kaggle test datasets with different loss function settings

Dataset	Scale	Bicubic	EDSR	SRGAN	ESRGAN	Proposed algorithm
Kaggle	2	29.01	37.51	36.91	37.76	37.99
	3	26.03	33.31	32.55	33.76	34.10
	4	24.34	30.71	30.16	31.23	31.67
WHU-RS19	2	25.59	27.86	27.15	28.75	29.06
	3	24.55	26.83	25.84	27.80	28.08
	4	22.96	24.74	23.94	25.71	26.08
AID	2	25.43	29.18	28.37	29.44	29.55
	3	22.82	25.91	25.02	26.33	26.52
	4	21.34	23.89	23.18	24.35	24.63

Table 3. Average PSNR of different algorithms on Kaggle, WHU-RS19, and AID

Dataset	Scale	Bicubic	EDSR	SRGAN	ESRGAN	Proposed algorithm
Kaggle	2	0.856	0.970	0.960	0.962	0.972
	3	0.794	0.927	0.906	0.918	0.935
	4	0.737	0.874	0.848	0.881	0.897
WHU-RS19	2	0.800	0.942	0.844	0.834	0.854
	3	0.742	0.900	0.797	0.796	0.822
	4	0.689	0.848	0.746	0.764	0.788
AID	2	0.712	0.978	0.802	0.790	0.798
	3	0.660	0.935	0.757	0.754	0.767
	4	0.613	0.881	0.708	0.724	0.736

Table 4. Average SSIM of different algorithms on Kaggle, WHU-RS19, and AID

Dataset	Scale	Bicubic	EDSR	SRGAN	ESRGAN	Proposed algorithm
Kaggle	2	0.861	0.993	0.994	0.998	0.999
	3	0.850	0.990	0.990	0.993	0.997
	4	0.834	0.983	0.981	0.986	0.993
WHU-RS19	2	0.832	0.910	0.910	0.914	0.915
	3	0.822	0.907	0.906	0.909	0.912
	4	0.806	0.901	0.898	0.903	0.908
AID	2	0.824	0.903	0.904	0.908	0.906
	3	0.814	0.900	0.900	0.903	0.903
	4	0.798	0.894	0.892	0.897	0.899