Latent Space Diffusion Model Digital Radiography Image Super-Resolution Enhancement Algorithm

Zhihui Liang; Xinyi Wu; Wei Wu

doi:10.3788/LOP250503

Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1634001(2025)

Latent Space Diffusion Model Digital Radiography Image Super-Resolution Enhancement Algorithm

Zhihui Liang, Xinyi Wu^*, and Wei Wu

Key Laboratory of Nondestructive Testing, Ministry of Education, Nanchang Hangkong University, Nanchang 330063, Jiangxi , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Overall structure of the algorithm. (a) Autoencoder; (b) latent space diffusion

Download full size

Fig. 2. Forward diffusion process and reverse diffusion process

Download full size

Fig. 3. Diffusion model denoising network structure

Download full size

Fig. 4. Cross attention structure guided by frequency domain information

Download full size

Fig. 5. Partial original images in the dataset

Download full size

Fig. 6. Dataset display after pruning

Download full size

Fig. 7. Comparison of the super-resolution enhancement results of DR images. (a) Workpiece 1; (b) workpiece 2

Download full size

Fig. 8. Comparison of the super-resolution enhancement results of CT reconstruction images. (a) Local comparison; (b) global comparison

Download full size

Fig. 9. Spatial resolution comparison experiments. (a) Original image; (b) grayscale distribution of the original image; (c) quantitative identification of spatial resolution; (d) grayscale distribution of super-resolution reconstruction image

Download full size

Fig. 10. Comparison of reasoning performance

Download full size

Table 1. Pseudo code for training and inference algorithms

View table

Table 1. Pseudo code for training and inference algorithms

Algorithm 2 diffusion model sampling

Input： $x_{T} ~ N (0,1)$ ，condition $y$

sampling steps $N$ ，previous time step $k$ ，next time step $s$

Output：reconstructed feature $x_{0}$

1：for $k$ in linespace（ $T$ ，1， $N$ ）

2： $ϵ^{'} = ϵ_{θ} (x_{k}, y, t)$

3： $x_{s} = \sqrt[]{{\bar{α}}_{s}} (\frac{x_{k} - ϵ^{'} \sqrt[]{1 - {\bar{α}}_{k}}}{\sqrt[]{{\bar{α}}_{k}}}) + \sqrt[]{1 - {\bar{α}}_{s} - σ^{2}} ϵ^{'}$

4：end

5： $x_{0} = x_{s}$

Table 2. Performance of each method

View table

Table 2. Performance of each method

Method	PSNR/dB $↑$	SSIM $↑$
Bilinear	35.300	0.930
Bicubic	36.733	0.942
VDSR	38.345	0.943
SRCNN	39.451	0.939
SRGAN	39.579	0.963
EDSR	40.562	0.954
FSRCNN	40.654	0.932
ESRGAN	41.886	0.956
RDN	42.754	0.966
Real-ESRGAN	43.559	0.962
SwinIR	44.889	0.969
DLGSANet	46.073	0.974
SRFormer	46.153	0.970
Proposed	47.106	0.983

Table 3. Performance at different sampling time steps

View table

Table 3. Performance at different sampling time steps

Sampling step	Inference time /s	PSNR /dB	SSIM
1	0.094	23.5225	0.2394
5	0.266	46.9629	0.9837
10	0.524	47.1052	0.9835
20	1.028	47.1061	0.9831
50	2.470	46.9000	0.9824
100	5.001	46.6152	0.9814
200	9.960	46.4195	0.9802
500	24.905	45.8416	0.9760
1000	50.005	45.0766	0.9714
2000	99.249	45.3313	0.9717

Table 4. Inter module ablation experiment
View table
Table 4. Inter module ablation experiment
Module Evaluation metric
PSNR /dB SSIM Inference time /s
Cross attention 46.865 0.974 26.67
Autoencoder 45.214 0.953 0.51
Autoencoder+cross
attention
47.106 0.983 0.53

Tools

Get Citation

Copy Citation Text

Zhihui Liang, Xinyi Wu, Wei Wu. Latent Space Diffusion Model Digital Radiography Image Super-Resolution Enhancement Algorithm[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1634001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites