Image super-resolution reconstruction of multi-scale deep feature distillation

Xiang LI; Ling XIONG; Daohui YE; Shufan LI

doi:10.37188/OPE.20253310.1657

Optics and Precision Engineering, Volume. 33, Issue 10, 1657(2025)

Image super-resolution reconstruction of multi-scale deep feature distillation

Xiang LI^1,2, Ling XIONG^1,2、*, Daohui YE³, and Shufan LI³

Author Affiliations

¹School of Artificial Intelligence and Automation，Wuhan University of Science and Technology，Wuhan43008， China

²Engineering Research Center for Metallurgical Automation and Measurement Technology of Ministry of Education，Wuhan University of Science and Technology，Wuhan430081， China

³Sinopec Jiang Diamond Oil Machinery Co， Ltd， Wuhan40200， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(20)

Fig. 1. Multi-scale deep feature distillation network structure

Download full size

View in Article

Fig. 2. Multi-scale deep feature distillation architecture

Download full size

View in Article

Fig. 3. ConvNeXt-3 convolutional layer

Download full size

View in Article

Fig. 4. Architecture of ESA block

Download full size

View in Article

Fig. 5. Architecture of CCA block

Download full size

View in Article

Fig. 6. Reconstruction effect comparison of “img_061” from Urban100 datasets at ×4 magnification

Download full size

View in Article

Fig. 7. Reconstruction effect comparison of “img_067” from Urban100 datasets at ×4 magnification

Download full size

View in Article

Fig. 8. Reconstruction effect comparison of “img_073” from Urban100 datasets at ×4 magnification

Download full size

View in Article

Fig. 9. Reconstruction effect comparison of “028” from PDC30 datasets at ×4 magnification

Download full size

View in Article

Fig. 10. Comparison of different scales of ConvNeXt residual block stacking sequence reconstruction results

Download full size

View in Article

Fig. 11. Comparison of parameters and reconstruction performance of different algorithms

Download full size

View in Article

Fig. 12. Semantic segmentation map and ellipse fitting results of “013.png” from PDC30 datasets before and after super-resolution reconstruction

Download full size

View in Article

Table 1. Performance of different algorithms on five benchmark datasets at ×2 magnification

View table

View in Article

Table 1. Performance of different algorithms on five benchmark datasets at ×2 magnification

Method	Scale	PSNR/SSIM
Method	Scale	Set5	Set14	BSD100	Urban100	Manga109
Bicubic	×2	33.69/0.928 4	30.34/0.867 5	29.57/0.843 8	26.88/0.843 8	30.80/0.933 9
VDSR^［9］	×2	37.53/0.958 7	33.03/0.912 4	31.90/0.896 0	30.76/0.914 0	37.22/0.972 9
DRCN^［23］	×2	37.63/0.958 8	33.04/0.911 8	31.85/0.894 2	30.75/0.913 3	37.63/0.972 3
IDN^［11］	×2	37.83/0.960 0	33.30/0.914 8	32.08/0.898 5	31.27/0.919 6	38.01/0.974 9
CARN^［24］	×2	37.76/0.959 0	33.52/0.916 6	32.09/0.897 8	31.92/0.925 6	-/-
IMDN^［12］	×2	38.00/0.960 5	33.63/0.917 7	32.19/0.899 6	32.17/0.928 3	38.88/0.977 4
RFDN^［13］	×2	38.05/0.960 6	33.68/0.918 4	32.16/0.899 4	32.12/0.927 8	38.88/0.977 3
SMSR^［25］	×2	38.00/0.960 1	33.64/0.917 9	32.17/0.899 0	32.19/0.928 4	38.76/0.977 1
ShuffleMixer^［26］	×2	38.01/0.960 6	33.63/0.918 0	32.17/0.899 5	31.89/0.925 7	38.83/0.977 4
VLESR^［27］	×2	38.01/0.960 5	33.58/0.917 7	32.16/0.899 3	32.14/0.928 0	38.75/0.977 0
SAFMN^［28］	×2	38.00/0.960 5	33.54/0.917 7	32.16/0.899 5	31.84/0.925 6	38.71/0.977 1
LBRN^［29］	×2	38.08/0.960 8	33.57/0.917 3	32.23/0.900 5	32.35/0.930 3	38.93/0.977 7
MSDFDN	×2	38.13/0.960 9	33.68/0.918 5	32.26/0.900 8	32.44/0.930 9	38.94/0.977 7
MSDFDN+	×2	38.17/0.961 1	33.82/0.920 2	32.28/0.901 0	32.58/0.932 3	39.09/0.977 8

Table 2. Performance of different algorithms on five benchmark datasets at ×3 magnification

View table

View in Article

Table 2. Performance of different algorithms on five benchmark datasets at ×3 magnification

Method	Scale	PSNR/SSIM
Method	Scale	Set5	Set14	BSD100	Urban100	Manga109
Bicubic	×3	30.39/0.868 2	27.55/0.774 2	27.21/0.738 5	24.46/0.734 9	26.95/0.855 6
VDSR^［9］	×3	33.66/0.921 3	29.77/0.831 4	28.82/0.797 6	27.14/0.827 9	32.01/0.934 0
DRCN^［23］	×3	33.82/0.922 6	29.76/0.831 1	28.80/0.796 3	27.15/0.827 6	32.24/0.934 3
IDN^［11］	×3	34.11/0.925 3	29.99/0.835 4	28.95/0.801 3	27.42/0.835 9	32.71/0.938 1
CARN^［24］	×3	34.29/0.925 5	30.29/0.840 7	29.06/0.803 4	28.06/0.849 3	-/-
IMDN^［12］	×3	34.36/0.927 0	30.32/0.841 7	29.09/0.804 6	28.17/0.851 9	33.61/0.944 5
RFDN^［13］	×3	34.41/0.927 3	30.34/0.842 0	29.09/0.804 2	28.21/0.852 5	33.67/0.944 9
SMSR^［25］	×3	34.40/0.927 0	30.33/0.841 2	29.10/0.805 0	28.25/0.853 6	33.68/0.944 5
ShuffleMixer^［26］	×3	34.40/0.927 2	30.37/0.842 3	29.12/0.805 1	28.08/0.849 8	33.69/0.944 8
VLESR^［27］	×3	34.40/0.927 2	30.34/0.841 5	29.08/0.804 3	28.16/0.851 9	33.61/0.944 5
SAFMN^［28］	×3	34.34/0.926 7	30.33/0.841 8	29.08/0.804 8	27.95/0.847 4	33.52/0.943 7
LBRN^［29］	×3	34.43/0.927 6	30.39/0.842 9	29.13/0.805 9	28.29/0.854 5	33.72/0.945 5
MSDFDN	×3	34.54/0.928 2	30.45/0.844 8	29.18/0.806 9	28.47/0.858 6	33.96/0.946 8
MSDFDN+	×3	34.61/0.928 9	30.52/0.845 9	29.23/0.808 3	28.64/0.861 7	34.12/0.947 9

Table 3. Performance of different algorithms on five benchmark datasets at ×4 magnification

View table

View in Article

Table 3. Performance of different algorithms on five benchmark datasets at ×4 magnification

Method	Scale	PSNR/SSIM
Method	Scale	Set5	Set14	BSD100	Urban100	Manga109
Bicubic	×4	28.42/0.810 4	26.00/0.702 7	25.96/0.667 5	23.14/0.657 7	24.89/0.786 6
VDSR^［9］	×4	31.35/0.883 8	28.01/0.767 4	27.29/0.725 1	25.18/0.752 4	28.83/0.887 0
DRCN^［23］	×4	31.53/0.885 4	28.02/0.767 0	27.23/0.723 3	25.14/0.751 0	28.93/0.885 4
IDN^［11］	×4	31.82/0.890 3	28.25/0.773 0	27.41/0.729 7	25.41/0.763 2	29.41/0.894 2
CARN^［24］	×4	32.13/0.893 7	28.60/0.780 6	27.58/0.734 9	26.07/0.783 7	-/-
IMDN^［12］	×4	32.21/0.894 8	28.58/0.781 1	27.56/0.735 3	26.04/0.783 8	30.45/0.907 5
RFDN^［13］	×4	32.24/0.895 2	28.61/0.781 9	27.57/0.736 0	26.11/0.785 8	30.58/0.908 9
SMSR^［25］	×4	32.12/0.893 2	28.55/0.780 8	27.55/0.735 1	26.11/0.786 8	30.54/0.908 5
ShuffleMixer^［26］	×4	32.21/0.895 3	28.66/0.782 7	27.61/0.736 6	26.08/0.783 5	30.65/0.909 3
VLESR^［27］	×4	32.17/0.894 5	28.55/0.780 2	27.55/0.734 5	26.03/0.783 0	30.48/0.907 3
SAFMN^［28］	×4	32.18/0.894 8	28.60/0.781 3	27.58/0.735 9	25.97/0.780 9	30.43/0.906 3
LBRN^［29］	×4	32.33/0.896 4	28.62/0.782 6	27.60/0.737 7	26.17/0.788 2	30.60/0.910 2
MSDFDN	×4	32.28/0.896 5	28.67/0.783 9	27.63/0.738 2	26.27/0.791 7	30.73/0.911 4
MSDFDN+	×4	32.46/0.898 2	28.78/0.785 9	27.68/0.740 1	26.49/0.797 6	30.96/0.914 4

Table 4. Performance of different algorithms on PDC bit composite dataset

View table

View in Article

Table 4. Performance of different algorithms on PDC bit composite dataset

Dataset	Scale	PSNR/SSIM
Dataset	Scale	Bicubic	RFDN^［13］	ShuffleMixer^［26］	SAFMN^［28］	MSDFDN
PDC30	×2	35.40/0.954 8	41.87/0.978 6	42.02/0.978 5	41.81/0.977 4	42.13/0.978 9
	×3	32.49/0.917 2	37.17/0.946 3	37.24/0.946 5	37.18/0.946 1	37.26/0.946 7
	×4	30.49/0.879 5	34.37/0.915 1	34.32/0.915 1	34.32/0.914 7	34.38/0.915 2

Table 5. Influence of the number of MSDFDB modules on the performance of the proposed algorithm

View table

View in Article

Table 5. Influence of the number of MSDFDB modules on the performance of the proposed algorithm

模块数量	Params/K	PSNR/SSIM
模块数量	Params/K	Set5	Set14	BSD100	Urban100	Manga109
n=5	481	38.10/0.960 8	33.67/0.918 8	32.24/0.900 5	32.31/0.930 2	38.86/0.977 3
n=6	572	38.13/0.961 0	33.67/0.918 5	32.26/0.900 8	32.44/0.930 9	38.91/0.977 6
n=7	664	38.17/0.961 1	33.79/0.920 2	32.26/0.900 7	32.52/0.931 8	39.01/0.977 6
n=8	755	38.13/0.961 0	33.76/0.919 1	32.25/0.900 6	32.50/0.931 9	38.97/0.977 2

Table 6. Experimental results of attention mechanism module ablation

View table

View in Article

Table 6. Experimental results of attention mechanism module ablation

Method	ESA	CCA	PSNR/SSIM
Method	ESA	CCA	Set5	Set14	BSD100	Urban100	Manga109
Model1	×	×	38.07/0.960 7	33.71/0.918 5	32.22/0.900 4	32.23/0.929 3	38.59/0.977 0
Model2	×	√	38.10/0.960 8	33.70/0.918 4	32.22/0.900 3	32.36/0.930 3	38.90/0.977 4
Model3	√	×	38.11/0.960 9	33.69/0.918 5	32.25/0.900 7	32.41/0.930 8	38.87/0.977 3
Model4	√	√	38.13/0.961 0	33.67/0.918 5	32.26/0.900 8	32.44/0.930 9	38.91/0.977 6

Table 7. Performance of different algorithms on five benchmark datasets

View table

View in Article

Table 7. Performance of different algorithms on five benchmark datasets

Method	PSNR/SSIM
Method	Set5	Set14	BSD100	Urban100	Manga109
MSDFDB-333	38.07/0.960 8	33.65/0.918 6	32.23/0.900 6	32.32/0.930 0	38.90/0.977 6
MSDFDB-753	38.09/0.960 9	33.68/0.918 6	32.24/0.900 5	32.38/0.930 5	38.87/0.977 5
MSDFDB-357	38.13/0.961 0	33.67/0.918 5	32.26/0.900 8	32.44/0.930 9	38.91/0.977 6

Table 8. Comparison of ellipse fitting errors before and after super-resolution reconstruction
View table
View in Article
Table 8. Comparison of ellipse fitting errors before and after super-resolution reconstruction
Group 1 2
MAE/px 0.071 0.043
MSE/px 0.013 0.005
Max Error/px 0.152 0.110

Tools

Get Citation

Copy Citation Text

Xiang LI, Ling XIONG, Daohui YE, Shufan LI. Image super-resolution reconstruction of multi-scale deep feature distillation[J]. Optics and Precision Engineering, 2025, 33(10): 1657

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Nov. 12, 2024

Accepted: --

Published Online: Jul. 23, 2025

The Author Email: Ling XIONG (xiongling@wust.edu.cn)

DOI:10.37188/OPE.20253310.1657

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Performance of different algorithms on five benchmark datasets at ×2 magnification

Table 1. Performance of different algorithms on five benchmark datasets at ×2 magnification

Table 2. Performance of different algorithms on five benchmark datasets at ×3 magnification

Table 2. Performance of different algorithms on five benchmark datasets at ×3 magnification

Table 3. Performance of different algorithms on five benchmark datasets at ×4 magnification

Table 3. Performance of different algorithms on five benchmark datasets at ×4 magnification

Table 4. Performance of different algorithms on PDC bit composite dataset

Table 4. Performance of different algorithms on PDC bit composite dataset

Table 5. Influence of the number of MSDFDB modules on the performance of the proposed algorithm

Table 5. Influence of the number of MSDFDB modules on the performance of the proposed algorithm

Table 6. Experimental results of attention mechanism module ablation

Table 6. Experimental results of attention mechanism module ablation

Table 7. Performance of different algorithms on five benchmark datasets

Table 7. Performance of different algorithms on five benchmark datasets

Table 8. Comparison of ellipse fitting errors before and after super-resolution reconstruction

Table 8. Comparison of ellipse fitting errors before and after super-resolution reconstruction