Low-light True Color Image Enhancement Algorithm Based on Adaptive Truncation Simulation Exposure and Unsupervised Fusion

Layer	1	2	3	4	5	6	7	8
Channel	24	24	24	24	24	24	24	1
Kernel	3×3	3×3	3×3	3×3	3×3	3×3	3×3	1×1
Dilation	1	2	5	1	2	5	1	1
Equivalent kernel	3×3	5×5	11×11	3×3	5×5	11×11	3×3	1×1
Receptive field	3×3	7×7	17×17	19×19	23×23	33×33	35×35	35×35

Table 2. Algorithm steps of depth guided filter module

View table

View in Article

Table 2. Algorithm steps of depth guided filter module

Input	Low resolution weight map $W_{k}^{l}$ Low resolution exposure sequence $X_{k}^{l}$ High resolution exposure sequence $X_{k}^{h}$
Output	High resolution weight map $W_{k}^{h}$
Step 1	$\{\begin{array}{l} \bar{X_{k}^{l}} = f_{m e a n} (X_{k}^{l}) \\ \bar{W_{k}^{l}} = f_{m e a n} (W_{k}^{l}) \\ \bar{X_{l}^{2}} = f_{m e a n} (X_{k}^{l} ⊙ X_{k}^{l}) \\ \bar{X_{l} W_{l}} = f_{m e a n} (X_{k}^{l} ⊙ W_{k}^{l}) \end{array}$	Step 3	$\{\begin{array}{l} A_{l} = \sum_{X_{l} W_{l}} / (\sum_{X_{l}} + ε) \\ B_{l} = \bar{W_{l}} - A_{l} ⊙ \bar{X_{l}} \end{array}$
Step 1		Step 4	$\{\begin{array}{l} A_{h} = f_{↑} (A_{l}) \\ B_{h} = f_{↑} (B_{l}) \end{array}$
Step 2	$\{\begin{array}{l} \sum_{X_{l}} = \bar{X_{l}^{2}} - \bar{X_{k}^{l}} ⊙ \bar{X_{k}^{l}} \\ \sum_{X_{l} W_{l}} = \bar{X_{l} W_{l}} - \bar{X_{k}^{l}} ⊙ \bar{W_{k}^{l}} \end{array}$	Step 5	$W_{k}^{h} = A_{k}^{h} ⊙ X_{k}^{h} + B_{k}^{h}$

Table 3. NIQE comparison of enhancement results under different algorithms

View table

View in Article

Table 3. NIQE comparison of enhancement results under different algorithms

	LIME	LECARM	FEM	KinD	RUAS	Zero-DCE	Proposed
DICM dataset	3.517 0	2.611 3	2.746 5	2.988 8	4.547 7	2.750 2	2.764 5
MEF dataset	4.732 9	3.509 7	3.413 2	3.375 3	3.483 1	3.380 8	3.130 1
NPE dataset	3.942 5	3.449 0	3.407 2	3.634 2	5.349 7	3.655 4	3.425 3
LIME dataset	4.411 1	4.325 8	4.404 6	4.453 0	4.284 0	4.138 8	3.952 5
Average	4.510 9	3.474 0	3.492 9	3.612 8	4.416 1	3.481 3	3.318 1

Table 4. Image quality evaluation and time spent comparison of enhancement results under different algorithms

View table

View in Article

Table 4. Image quality evaluation and time spent comparison of enhancement results under different algorithms

	LIME	LECARM	KinD	RUAS	Zero-DCE	Proposed
PSNR↑	14.766 0	15.506 5	14.990 8	15.568 1	13.742 6	16.170 6
SSIM↑	0.334 4	0.547 2	0.593 9	0.560 3	0.514 2	0.605 4
NIQE↓	6.738 2	7.132 1	4.185 5	6.831 9	7.081 9	6.481 7
Time↓	53 s	5 s	20 s	7 s	4 s	19 s

Table 5. Color difference comparison of color card image enhancement results under different illuminance
View table
View in Article
Table 5. Color difference comparison of color card image enhancement results under different illuminance
LIME LECARM KinD RUAS Zero-DCE Proposed
8.71×10^-2 lx 20.762 4 23.840 0 23.915 9 25.198 5 21.635 5 18.081 5
1.02×10^-2 lx 42.503 7 40.688 2 36.293 8 41.445 0 35.681 4 34.625 4

Tools

Get Citation

Copy Citation Text

Yongcheng HAN, Wenwen ZHANG, Weiji HE, Qian CHEN. Low-light True Color Image Enhancement Algorithm Based on Adaptive Truncation Simulation Exposure and Unsupervised Fusion[J]. Acta Photonica Sinica, 2023, 52(9): 0910002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites