Self-Supervised Enhancement of Low-Light Images Based on Blind Spot Networks

Yong Chen; Jinliang Zhang; Huanlin Liu; Kaixin Shao; Shangming Chen; Hangying Xiong; Yourui Zhang

doi:10.3788/AOS240549

Acta Optica Sinica, Volume. 44, Issue 13, 1310001(2024)

Self-Supervised Enhancement of Low-Light Images Based on Blind Spot Networks

Yong Chen^1、*, Jinliang Zhang¹, Huanlin Liu², Kaixin Shao¹, Shangming Chen¹, Hangying Xiong², and Yourui Zhang¹

¹Key Laboratory of Industrial Internet of Things and Network Control, Ministry of Education, Chongqing University of Posts and Telecommunications, Chongqing 400065, China

²School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Overall framework of self-supervised network

Download full size

Fig. 2. Histogram equalization of bilateral multi-scale fusion

Download full size

Fig. 3. Structure diagram of DBSNet

Download full size

Fig. 4. DBS module

Download full size

Fig. 5. RAC module

Download full size

Fig. 6. PD and PD^-1 operations

Download full size

Fig. 7. Output effects of different stride factors and convolution kernel sizes

Download full size

Fig. 8. Visual comparison of processing results of low-light images by different algorithms. (a) Input image; (b) LIME algorithm^[27]; (c) SIRE algorithm^[28]; (d) NPE algorithm^[29]; (e) EG algorithm^[26]; (f) Zero-DCE algorithm^[25]; (g) RetinexNet algorithm^[6]; (h) DUPE algorithm^[32]; (i) LLFLow algorithm^[30]; (j) GLADNet algorithm^[7]; (k) SCI algorithm^[31]; (l) proposed algorithm

Download full size

Fig. 9. Visual comparison of processing results of real low-light images by different algorithms. (a) Input image; (b) LIME algorithm^[27]; (c) SIRE algorithm^[28]; (d) RetinexNet algorithm^[6]; (e) LLFLow algorithm^[30]; (f) EG algorithm^[26]; (i) Zero-DCE algorithm^[25]; (j) KinD++ algorithm^[8]; (k) proposed algorithm

Download full size

Fig. 10. Ablation experiments with different loss functions. (a) L_DBSNet; (b) L_spa; (c) L_col; (d) L_tv; (e) DBSNet; (f) real scene

Download full size

Table 1. Comparison of output images under different stride factors
View table
Table 1. Comparison of output images under different stride factors
$h$ PSNR↑ SSIM↑
1 15.60 0.685
2 19.55 0.723
3 20.49 0.775
4 19.21 0.801
5 22.12 0.836
6 20.56 0.831
7 20.32 0.799

Table 2. Comparison of output image indexes under different algorithms

View table

Table 2. Comparison of output image indexes under different algorithms

Index	LIME^［27］	SIRE^［28］	NPE^［29］	EG^［26］	Zero-DCE^［25］	RetinexNet^［6］
PSNR↑	16.76	11.86	16.97	17.48	14.86	15.17
SSIM↑	0.444	0.494	0.482	0.654	0.562	0.075
NIQE↓	9.779	8.073	9.788	5.238	8.811	6.261
Delta E↓	21.43	32.62	21.77	19.31	24.56	24.17
Index	DUPE^［32］	LLFlow^［30］	GLADNet^［7］	SCI^［31］	KinD++^［8］	Proposed
PSNR↑	14.77	19.19	19.72	19.81	20.12	20.81
SSIM↑	0.470	0.714	0.685	0.731	0.736	0.739
NIQE↓	9.079	4.213	7.283	4.103	4.290	4.191
Delta E↓	26.19	16.17	16.54	15.59	11.52	11.50

Table 3. Ablation experiments with different loss functions
View table
Table 3. Ablation experiments with different loss functions
$L_{D B S N e t}$ $L_{s p a}$ $L_{c o l}$ $L_{t v}$ PSNR↑ SSIM↑
× √ √ √ 10.11 0.456
√ × √ √ 18.60 0.684
√ √ × √ 20.55 0.711
√ √ √ × 17.41 0.759
√ √ √ √ 22.21 0.863

Tools

Get Citation

Copy Citation Text

Yong Chen, Jinliang Zhang, Huanlin Liu, Kaixin Shao, Shangming Chen, Hangying Xiong, Yourui Zhang. Self-Supervised Enhancement of Low-Light Images Based on Blind Spot Networks[J]. Acta Optica Sinica, 2024, 44(13): 1310001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites