Masayuki Tanaka
Takashi Shibata
Masatoshi Okutomi
Gradient-Based
Low-Light Image
Enhancement
Background
1
Most common mistake is
blurring of moving objects or
by camera-shake.
Blurred image
Dark image
Increase shutter speed
Sensitivity of the sensor has been
improved.
But, we still have dark images
with fast shutter speed.
Low-Light Image Enhancement
2
Take photo
under
low-lightscene
Proposed Image EnhancementLow-Light Image Enhancement is
highly demanded.
We propose a gradient-
based low-light image
enhancement.
Gradient-Based Image Processing
3
Gradient
information
Intensity
information>
Shibata, Gradient-Domain Image Reconstruction Framework with Intensity-Range and Base-Structure
Constraints, CVPR2016
Important
for human visual system
Many gradient-
based image
processing
applications have
been proposed.
Gradient-based image processing
4
Gradient
extraction
Gradient
manipulation
Gradient
integration
Post
processing
Processing pipeline
Input
image
Output
image
Drawback of Gradient-based image processing
The intensity range of output image is unknown unless the integration is
performed. The intensity range of the output image often exceeds the
typical intensity range of [0,255].
We’ve gotten
saturation in the
output images.
Image
enhancement
Example of post-processing
5
Input image Processed image
×1.5
Intensity range of
processed image
exceeds [0,255]
Rescaling Intensity clipping Proposed
Proposed Integration
6
Shibata, Gradient-Domain Image Reconstruction Framework with Intensity-Range and Base-Structure
Constraints, CVPR2016
Gradient
extraction
Gradient
manipulation
Proposed
integration
Input
image
Output
image
Intensity range
information
𝐹𝐹 𝑢𝑢 𝑥𝑥 = � �
𝑑𝑑=ℎ,𝑣𝑣
𝜕𝜕𝑑𝑑 𝑢𝑢 𝑥𝑥 − 𝑞𝑞𝑑𝑑 𝑥𝑥 2
𝑑𝑑𝑑𝑑 + G𝑅𝑅[𝑢𝑢(𝑥𝑥)]
𝑢𝑢 𝐱𝐱
𝑔𝑔𝑅𝑅(𝑢𝑢 𝐱𝐱 )
𝑅𝑅 𝑚𝑚𝑚𝑚 𝑚𝑚 𝑅𝑅 𝑚𝑚𝑚𝑚𝑚𝑚
Intensity range
We penalize if the
intensity exceeds the
intensity range.This cost function means to
preserving the gradient
information while keeping
the intensity range within the
given intensity range.
Gradient Enhancement Strategy
7
Gradient
extraction
Gradient
manipulation
Proposed
integration
Input
image
Output
image
Intensity range
information
Intensity
Gradient manipulation: Higher gain for darker region
Input
gradient
Gain
generation
Enhanced
gradient
Intensity
Gain
β:15
1
τ:50
Experimental Comparisons
8
Input image Histogram equalization
Automatic tone correction
by photoshop
HDR tone mapping
by matlab
LIME (Guo2016)
Proposed
Experimental Comparisons
9
Input image Histogram equalization
Automatic tone correction
by photoshop
HDR tone mapping
by matlab
LIME (Guo2016)
Proposed
Experimental Results
10 Input images
Experimental Results
11 Proposed results
Real Time Demonstration
12
Conclusions
13
We propose have a gradient-based low-light
image enhancement.
• Higher gain for darker region
• Image integration with
intensity range constraint
Input image
Proposed result
matlab code is available on our project page:
http://www.ok.sc.e.titech.ac.jp/res/IC/LowLight/

Gradient-Based Low-Light Image Enhancement