Lightweight Remote Sensing Image Super-Resolution Reconstruction Based on Saliency Analysis and Information Distillation

Xueli Shen; Xiaoming Zhu; Haibo Jin

doi:10.3788/LOP241966

Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1028001(2025)

Lightweight Remote Sensing Image Super-Resolution Reconstruction Based on Saliency Analysis and Information Distillation

Xueli Shen, Xiaoming Zhu^*, and Haibo Jin

Author Affiliations

School of Software, Liaoning Technical University, Huludao 125105, Liaoning , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(20)

Fig. 1. Overall network architecture diagram

Download full size

Fig. 2. Structure diagram of the saliency detector. (a) Saliency detection module; (b) residual block

Download full size

Fig. 3. Different feature extraction paths for DRM

Download full size

Fig. 4. FRU structure diagram

Download full size

Fig. 5. Diagrams of the multiscale void separable convolution module. (a) MDAM; (b) MSCM

Download full size

Fig. 6. Improved CPAM

Download full size

Fig. 7. Dual-path upsampling module

Download full size

Fig. 8. Images in the GeoEye-1 dataset

Download full size

Fig. 9. Comparison of the visual effects when different algorithms are magnified 4 times. (a) Google Earth; (b) GeoEye-1

Download full size

Fig. 10. Comparison of reconstruction effect of different algorithms when enlarging 4 times on GeoEye-1 dataset

Download full size

Fig. 11. Influence of the number of MDAM on PSNR

Download full size

Table 1. Comparison of different algorithms in parameters, PSNR, and SSIM with magnification of 2 in the two test sets

View table

Table 1. Comparison of different algorithms in parameters, PSNR, and SSIM with magnification of 2 in the two test sets

Method	Parameter /10³	GeoEye-1		Google Earth
Method	Parameter /10³	PSNR /dB	SSIM	PSNR /dB	SSIM
Bicubie	‒	24.84	0.7900	28.03	0.8243
DRRN^［27］	301	26.68	0.8571	29.82	0.8821
IMDN^［13］	694	26.70	0.8573	29.98	0.8816
CTN^［28］	412	26.67	0.8583	29.80	0.8782
CARN^［11］	1592	26.69	0.8583	30.00	0.8818
CFSRCNN^［12］	1310	26.72	0.8594	30.00	0.8823
FeNet^［29］	618	26.72	0.8594	29.95	0.8810
DSRNet^［30］	751	26.78	0.8611	29.99	0.8819
SCNet-B^［31］	561	26.76	0.8606	29.98	0.8817
SalDRN^［26］	571	26.80	0.8617	29.97	0.8811
Proposed	542	26.85	0.8612	30.02	0.8820

Table 2. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 3 in the two test sets

View table

Table 2. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 3 in the two test sets

Method	Parameter /10³	GeoEye-1		Google Earth
Method	Parameter /10³	PSNR /dB	SSIM	PSNR /dB	SSIM
Bicubie	‒	22.52	0.6414	25.25	0.6841
DRRN^［27］	301	23.88	0.7434	26.96	0.7724
IMDN^［13］	703	23.97	0.7422	26.99	0.7719
CTN^［28］	412	23.83	0.7365	26.83	0.7679
CARN^［11］	1592	23.83	0.7365	26.83	0.7679
CFSRCNN^［12］	1495	23.98	0.7429	27.01	0.7729
FeNet^［29］	627	23.91	0.7407	26.95	0.7705
DSRNet^［30］	759	23.99	0.7435	27.01	0.7726
SCNet-B^［31］	593	23.97	0.7429	26.96	0.7728
SalDRN^［26］	571	23.98	0.7437	26.97	0.7722
Proposed	542	24.04	0.7440	27.05	0.7735

Table 3. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 4 in the two test sets

View table

Table 3. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 4 in the two test sets

Method	Parameter /10³	GeoEye-1		Google Earth
Method	Parameter /10³	PSNR /dB	SSIM	PSNR /dB	SSIM
Bicubie	‒	21.25	0.5307	23.75	0.5800
DRRN^［27］	301	22.36	0.6439	25.14	0.6764
IMDN^［13］	715	22.43	0.6408	25.19	0.6780
CTN^［28］	412	22.30	0.6324	25.04	0.6698
CARN^［11］	1580	22.39	0.6383	25.15	0.6761
CFSRCNN^［12］	1458	22.49	0.6453	25.22	0.6799
FeNet^［29］	639	22.39	0.6371	25.16	0.6753
DSRNet^［30］	766	22.45	0.6427	25.21	0.6800
SCNet-B^［31］	582	22.44	0.6422	25.18	0.6786
SalDRN^［26］	571	22.42	0.6408	25.17	0.6768
Proposed	547	22.54	0.6441	25.28	0.6801

Table 4. Comparison of runtime of each algorithm
View table
Table 4. Comparison of runtime of each algorithm
Method Runtime /ms
CARN^［11］ 4.6
IMDN^［13］ 5.3
SCNet-B^［31］ 6.2
CFSRCNN^［12］
FeNet^［29］
11.4
12.8
Proposed 13.6
DSRNet ^［30］ 14.7
SalDRN^［26］ 15.4
CTN^［28］ 18.4

Table 5. Comparative experiments
View table
Table 5. Comparative experiments
Methods Parameter /10³ PSNR/dB
IMDN（IMDB） 715 25.19
IMDN（MDAM） 723 25.21
SDM+DRM（IMDB） 570 25.23
SDM+DRM（MDAM） 579 25.26
SDM+DRM（MDAM）+DSM 547 25.28

Table 6. Comparison of ablation experiments of each module
View table
Table 6. Comparison of ablation experiments of each module
Module MSCM CPAM Parameter /10³ PSNR /dB
3×3+CCA × × 538 25.23
MSCM+CCA √ × 543 25.26
3×3+CPAM × √ 541 25.25
Proposed √ √ 547 25.28

Table 7. Comparison of results using different activation functions
View table
Table 7. Comparison of results using different activation functions
Activation function Parameter /10³ PSNR /dB
ReLU 531 26.82
LeakyReLU 533 26.83
Meta-Acon 542 26.85

Table 8. Effect of image block size on model reconstruction quality

View table

Table 8. Effect of image block size on model reconstruction quality

Patch size /（pixel×pixel）	GeoEye-1		Google Earth
Patch size /（pixel×pixel）	PSNR /dB	SSIM	PSNR /dB	SSIM
32×32	22.52	0.6440	25.27	0.6800
40×40	22.49	0.6438	25.26	0.6801
48×48	22.54	0.6441	25.28	0.6801
56×56	22.50	0.6435	25.24	0.6796
64×64	22.46	0.6432	25.23	0.6795

Table 9. Influence of path selection switch threshold setting on model in DRM

View table

Table 9. Influence of path selection switch threshold setting on model in DRM

Threshold settings	Passing ratios /%	FLOPs /10⁶	PSNR /dB
［0，0，0］	［100， 100， 100］	3520	26.87
［0，0.25，0.50］	［100， 75， 46］	2624	26.85
［0，0.50，0.75］	［100， 46， 25］	2126	26.80
［0，0.75，1.00］	［100， 25， 0］	1501	26.75

Tools

Get Citation

Copy Citation Text

Xueli Shen, Xiaoming Zhu, Haibo Jin. Lightweight Remote Sensing Image Super-Resolution Reconstruction Based on Saliency Analysis and Information Distillation[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1028001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Remote Sensing and Sensors

Received: Sep. 6, 2024

Accepted: Oct. 28, 2024

Published Online: May. 9, 2025

The Author Email: Xiaoming Zhu (1945595956@qq.com)

DOI:10.3788/LOP241966

CSTR:32186.14.LOP241966

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison of different algorithms in parameters, PSNR, and SSIM with magnification of 2 in the two test sets

Table 1. Comparison of different algorithms in parameters, PSNR, and SSIM with magnification of 2 in the two test sets

Table 2. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 3 in the two test sets

Table 2. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 3 in the two test sets

Table 3. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 4 in the two test sets

Table 3. Comparison of different algorithms in parameter, PSNR, and SSIM with magnification of 4 in the two test sets

Table 4. Comparison of runtime of each algorithm

Table 4. Comparison of runtime of each algorithm

Table 5. Comparative experiments

Table 5. Comparative experiments

Table 6. Comparison of ablation experiments of each module

Table 6. Comparison of ablation experiments of each module

Table 7. Comparison of results using different activation functions

Table 7. Comparison of results using different activation functions

Table 8. Effect of image block size on model reconstruction quality

Table 8. Effect of image block size on model reconstruction quality

Table 9. Influence of path selection switch threshold setting on model in DRM

Table 9. Influence of path selection switch threshold setting on model in DRM