Low-Light Image Enhancement Based on Deep Convolutional Neural Network

Fig. 4. Subjective visual comparison of different methods for synthetic low-light images. (a) Image “caps”; (b) image “carnivaldolls”; (c) image “cemetry”; (d) image “building 2”

Download full size

Fig. 5. Convergence performance of HSI and RGB enhancement methods with BN and without BN. (a) Average SSIM within 50 epochs; (b) average PSNR within 50 epochs

Download full size

Fig. 6. Subjective visual comparison of different methods for real low-light images. (a) Image from DICM dataset; (b) image from VV dataset; (c)-(d) image from NASA dataset; (e) enlarged result of part shown in blue box of Fig. 6(d)

Download full size

Table 1. Tested PSNR under different network layers and convolution kernel numbers
View table
Table 1. Tested PSNR under different network layers and convolution kernel numbers
Number of layers Number of convolution kernels PSNR /dB
5 n₁=64, n_p_-1=32 21.74
5 n₁=64, n_p_-1=64 21.87
7 n₁=64, n_p_-1=32 22.23
7 n₁=64, n_p_-1=64 22.31
9 n₁=64, n_p_-1=32 22.04
9 n₁=64, n_p_-1=64 22.17

Table 2. Objective evaluation index of different methods for synthetic low-light images
View table
Table 2. Objective evaluation index of different methods for synthetic low-light images
Method PSNR /dB SSIM MSE LOE
HE^[25] 16.19 0.7985 1928.3 505
Dong^[10] 16.29 0.7947 1699.6 2040
SRIE^[8] 21.08 0.9579 686.4 776
LIME^[9] 13.47 0.8097 3230.7 1277
Proposed method 22.23 0.9204 389.0 402

Table 3. Objective evaluation index of differentmethods for real low-light images

View table

Table 3. Objective evaluation index of differentmethods for real low-light images

Method	Entropy of information	Degree of chromaticity change	LOE	VIF
HE^[25]	7.1412	0.2474	571	0.4782
Dong^[10]	6.9541	0.0313	1484	0.4262
SRIE^[8]	6.9542	0.0045	972	0.6153
LIME^[9]	7.2612	0.0124	1390	0.3444
Proposed method	7.1425	0.0032	378	0.7356

Tools

Get Citation

Copy Citation Text

Hongqiang Ma, Shiping Ma, Yuelei Xu, Mingming Zhu. Low-Light Image Enhancement Based on Deep Convolutional Neural Network[J]. Acta Optica Sinica, 2019, 39(2): 0210004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: Jul. 25, 2018

Accepted: Sep. 25, 2018

Published Online: May. 10, 2019

The Author Email:

DOI:10.3788/AOS201939.0210004

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Tested PSNR under different network layers and convolution kernel numbers

Table 1. Tested PSNR under different network layers and convolution kernel numbers

Table 2. Objective evaluation index of different methods for synthetic low-light images

Table 2. Objective evaluation index of different methods for synthetic low-light images

Table 3. Objective evaluation index of differentmethods for real low-light images

Table 3. Objective evaluation index of differentmethods for real low-light images