Memory Unit-Based Multistage Self-Supervised Image Denoising Method

Xiaodong Zhang; Linghan Zhu; Shaoshu Gao; Xinrui Wang; Shuo Wang

doi:10.3788/LOP241376

Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0437003(2025)

Memory Unit-Based Multistage Self-Supervised Image Denoising Method

Xiaodong Zhang^*, Linghan Zhu, Shaoshu Gao, Xinrui Wang, and Shuo Wang

Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China), Qingdao 257061, Shandong , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(10)

Fig. 1. Comparison of loss function calculation methods between existing method and MU-based method. (a) Existing method; (b) MU-based method

Download full size

Fig. 2. Strcture of multistage self-supervised denoising method based on MU

Download full size

Fig. 3. Comparison of visualization results while introducing MU in different training stages. (a) without MU; (b) with MU

Download full size

Fig. 4. Visual comparison results of different methods on SIDD dataset

Download full size

Fig. 5. Visual comparison results of different methods on DND dataset

Download full size

Table 1. Results of MU ablation experiments
View table
Table 1. Results of MU ablation experiments
Stage 1 Stage 2 Stage 3 PSNR /dB
BNN — — 36.02
BNN+MU — — 36.23
BNN+MU DSN — 34.91
BNN+MU DSN+MU — 34.60
BNN+MU DSN UNet 36.90
BNN+MU DSN UNet+MU 37.30

Table 2. Impact of different training stages on denoising performance
View table
Table 2. Impact of different training stages on denoising performance
Stage 1 Stage 2 Stage 3 PSNR /dB
√ √ 36.62
√ √ 35.21
√ √ √ 37.30

Table 3. Comparison of performance for different training strategies
View table
Table 3. Comparison of performance for different training strategies
Training strategy PSNR /dB SSIM Training time /h
Joint training 37.05 0.925 265
Multistage training 37.30 0.930 121

Table 4. Comparison of denoising performances of different algorithms on SIDD and DND datasets

View table

Table 4. Comparison of denoising performances of different algorithms on SIDD and DND datasets

Type	Algorithm	PSNR /dB		SSIM
Type	Algorithm	SIDD dateset	DND dateset	SIDD dateset	DND dateset
Non-learning based	BM3D^［7］	25.65	34.51	0.685	0.851
Non-learning based	WNNM^［24］	25.78	34.67	0.809	0.865
Self-supervised	Noise2Void^［18］	27.68	—	0.668	—
	Noise2Self^［25］	29.56	—	0.808	—
	NAC^［26］	—	36.20	—	0.925
	R2R^［27］	35.45	—	0.898	—
	CVF-SID^［28］	34.71	36.50	0.917	0.924
	AP-BSN^［21］	36.91	38.09	0.931	0.937
	SASSID^［23］	37.08	38.18	0.928	0.938
	MM-SSID （proposed）	37.30	38.52	0.930	0.941

Table 5. Comparison of model efficiency for different denoising algorithms

View table

Table 5. Comparison of model efficiency for different denoising algorithms

Algorithm	Parameters /10⁶	FLOPs /10⁹	Time /ms
Noise2Void^［18］	7.18	73.1	17.2
Noise2Self^［25］	0.56	73.1	15.6
NAC^［26］	0.78	51.2	24.9
R2R^［27］	0.56	73.1	15.6
CVF-SID^［28］	1.19	155.7	38.5
AP-BSN^［21］	3.66	3788.1	634.1
MM-SSID	1.08	35.0	13.0

Tools

Get Citation

Copy Citation Text

Xiaodong Zhang, Linghan Zhu, Shaoshu Gao, Xinrui Wang, Shuo Wang. Memory Unit-Based Multistage Self-Supervised Image Denoising Method[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0437003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites