Vous êtes sur la page 1sur 37

Efficient Image/Video Dehazing through

Haze Density Analysis Based on Pixel-based


Dark Channel Prior

1
Outline
 Introduction
 Overview of image dehazing
 Motivation
 Background review
 Single image haze removal using dark channel prior [1]
 Proposed method
 Hazing imaging model
 Atmospheric light estimation
 Transmission map estimation and refinement
 Radiance recovery
 Experimental results
 Conclusions and discussions

2
Outline
 Introduction
 Overview of image dehazing
 Motivation
 Background review
 Single image haze removal using dark channel Prior
 Proposed method
 Hazing imaging model
 Atmospheric light estimation
 Transmission map estimation and refinement
 Radiance recovery
 Experimental results
 Conclusions and discussions

3
Overview of Image Dehazing
 What is image dehazing?
 Image dehazing is a technique of image processing that can
increase the image quality from an input with haze

Haze
images

Dehazed
images

4
Motivation
 Why is image dehazing?
 Dehazing could increase the visibility for driving assistance
 Provide the clear input image for applications

5
Pedestrian Detector Comparison

Pedestrian Detections [threshold = 0.00] Pedestrian Detections [threshold = 0.00]

0.74
0.11 0.98
0.70 0.69 0.01
0.17
0.61
0.25
0.99

Before dehazing After dehazing

6 S. Maji, A. C. Berg, and J. Malik, “Classification using intersection kernel support vector machines is efficient,” in
Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Anchorage, Alaska, USA, June 2008, pp. 1-8.
Blob Detector

Before dehazing After dehazing

7 M. Jahangiri and M. Petrou, “An attention model for extracting components that merit identification,” in Proc.
IEEE Int. Conf. Image Processing, Cairo, Egypt, Nov. 2009, pp. 965-968.
Outline
 Introduction
 Overview of image dehazing
 Motivation
 Background review
 Single image haze removal using dark channel Prior [1]
 Proposed method
 Hazing imaging model
 Atmospheric light estimation
 Transmission map estimation and refinement
 Radiance recovery
 Experimental results
 Conclusions and discussions

8
Haze Imaging Model
A widely used haze imaging model is present as:

I( x )  J( x )t( x )  A(1 - t( x ))
J(x):haze free image
A:atmospheric light
t(x):transmission
Haze image Clear image

Haze

Camera Atmosphere light

Scattering
Transmission

9
Single Image Haze Removal using
Dark Channel Prior
Haze image
Define patch-based dark channel prior:

Atmospheric light Jdark ( x )  min ( min (Jc ( y )))


estimation c{ r ,g ,b } y( x )

Transmission map Jc : a color channel of J


estimation (x ) : a local patch center at x

Refine
transmission map

dark channel
Radiance recovering
prior

Dehazed image

10
Single Image Haze Removal using
Dark Channel Prior
Haze image

Atmospheric light
estimation

Transmission map
estimation

Refine
transmission map

Radiance recovering

Find the 0.1% brightest point in dark channel prior


Dehazed image

11
Single Image Haze Removal using
Dark Channel Prior
Haze image

Atmospheric light
estimation

Transmission map
estimation

Refine
transmission map

Radiance recovering

Mapping to the original image, and choose the pixel value


Dehazed image as atmospheric light color Ac

12
Single Image Haze Removal using
Dark Channel Prior
Haze image
The haze imaging model reduced by dark channel
prior is described as:
Atmospheric light
estimation
min ( min (Ic ( y ))  t( x ) min ( min (Jc ( y ))  (1  t ( x ))A c
c{ r ,g ,b } y( x ) c{ r ,g ,b } y( x )

Transmission map
estimation According to [1], the haze image model modified by:

min ( min (Jc ( y ))  0


Refine c{ r ,g ,b } y ( x )

transmission map
The estimated transmission function is:

Radiance recovering Ic ( y )
t( x )  1  w min ( min ( c ))
c{ r ,g ,b } y ω( x ) A

Where w fixed to 0.95 in paper [1]


Dehazed image

13
Single Image Haze Removal using
Dark Channel Prior
Haze image

Atmospheric light
estimation

Transmission map
estimation

Refine
transmission map

Radiance recovering

Transmission map before refine


Dehazed image

14
Single Image Haze Removal using
Dark Channel Prior
Haze image

Atmospheric light
estimation

Transmission map
estimation

Refine
transmission map

Radiance recovering

Transmission map after refinement by Soft matting


Dehazed image

15
Single Image Haze Removal using
Dark Channel Prior
Haze image

Finally, the haze free image is recovered by:


Atmospheric light
estimation I( x )  A c
J( x )   Ac
max(t( x ), t 0 )
Transmission map
estimation
J(x) : the dehazed pixels

Refine I(x) : the input pixels


transmission map
t(x) : the transmission value

t0 : the lower bound of transmission, paper [1] set as 0.1


Radiance recovering
Ac : the air light color

Dehazed image

16
Outline
 Introduction
 Overview of image dehazing
 Motivation
 Background review
 Single image haze removal using dark channel Prior [1]
 Proposed method
 Hazing imaging model
 Atmospheric light estimation
 Transmission map estimation and refinement
 Radiance recovery
 Experimental results
 Conclusions and discussions

17
The Proposed Dehazing Method:
Observation

Different haze density

18
The Proposed Dehazing Method:
Haze Imaging Model
Proposed pixel-based haze image model is present as:

I( x )  J( x )t( x )  A(x)(1 - t( x ))
J(x) : haze free image
A(x) : pixel-based atmospheric light
t(x) : transmission
Haze image Clear image

Haze

Camera Atmosphere light

Scattering
Transmission

19
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image

Atmospheric light estimation


Atmospheric light Pixel-based
estimation dark channel prior

Transmission map
estimation
Pixel-based Atmospheric light
bright channel prior mapping function
Refine
transmission map

HSV distance
Radiance recovery

Dehazed image

20
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image
Define pixel-based dark channel prior:

Atmospheric light Jdark_Pix ( x )  min (Jc ( x ))


c{ r ,g ,b }
estimation

Jc : color channel of J
Transmission map
estimation x  pixels index

Refine
transmission map pixel-based dark
channel
prior
Radiance recovery

Dehazed image

21
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image
Define pixel-based bright channel prior:

Atmospheric light Jbright_Pix( x )  max (Jc ( x ))


estimation c{ r ,g ,b }

Jc : color channel of J
Transmission map
estimation x  pixels index

Refine
transmission map pixel-based
bright channel
prior
Radiance recovery

Dehazed image

22
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image
Define HSV distance:
Atmospheric light
estimation
d( x )  || IHSV ( x )  W ||

W: the brightest pixel in HSV


Transmission map
estimation

Refine
transmission map

Radiance recovery

Dehazed image

23
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image
Atmospheric light estimating function present as:
Atmospheric light
estimation d( x )
s(x )  1 
max(d( x ))
Transmission map
estimation
c
A highest  A lowest
c

Ac (x)  | | s(x )
max
HSV
(1  1/ d( y ))  min
HSV
(1  1/ d( x ))
y I y I
Refine
transmission map
c
A highest
is the brightest value of Jdark_Pix mapping to original
Radiance recovery c image
A lowest
is the darkest value of Jbright_Pix mapping to original
image
Dehazed image

24
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image

Atmospheric light
estimation

Transmission map
estimation
Ac
c
Refine
A highest
transmission map

Radiance recovery c
A lowest

s(x)
Dehazed image

25
The Proposed Dehazing Method:
Atmospheric Light Estimation
Haze image

Atmospheric light
estimation

Transmission map
estimation
Ac
c
Refine
A highest
transmission map

Radiance recovery c
A lowest

s(x)
Dehazed image

26
The Proposed Dehazing Method:
Transmission Map Estimation
Haze image

The haze image model reduced by pixel-based dark


Atmospheric light
estimation
channel prior is described as:
min Ic ( x )  t( x ) min Jc ( x )  (1  t( x ))A c ( x )
c{ r ,g ,b } c{ r ,g ,b }
Transmission map
estimation According to [1], the haze image model modified by:
min Jc ( x )  0
Refine c{ r ,g ,b }
transmission map
The estimated transmission function is:

Radiance recovery Ic ( x )
t( x )  1  w min ( c )
c{ r ,g ,b } A ( x )

Where w fixed to 0.95 in paper [1]


Dehazed image

27
The Proposed Dehazing Method:
Transmission Map Estimation
Haze image

Atmospheric light
estimation

Transmission map
estimation

Refine
transmission map

Radiance recovery
Transmission map before refinement

Dehazed image

28
The Proposed Dehazing Method:
Refine Transmission Map
Haze image

Atmospheric light
estimation

Transmission map
estimation

Refine
transmission map

Radiance recovery
Transmission map after Bilateral filter

Dehazed image

29
The Proposed Dehazing Method:
Radiance Recovery
Haze image

Finally, the haze free image J(x) is recovered by:


Atmospheric light
estimation
I( x )  A c ( x )
J( x )   Ac (x)
Transmission map max(t( x ), t 0 )
estimation
J(x) : the dehazed pixels
Refine
transmission map I(x) : the input pixels

t(x) : the transmission value


Radiance recovery t0 : the lower bound of transmission, we fix it as 0.1
Ac(x) : the air light color
Dehazed image

30
Outline
 Introduction
 Overview of image dehazing
 Motivation
 Background review
 Single image haze removal using dark channel Prior [1]
 Proposed method
 Hazing imaging model
 Atmospheric light estimation
 Transmission map estimation and refinement
 Radiance recovery
 Experimental results
 Conclusions and discussions

31
Experimental Results

Original image Fattal’s dehazing results [2]


2008

He’s dehazing results [1] Proposed


2011 Dehazing method

32
33
34
Video Dehazing

Before dehazing After dehazing

35
Outline
 Introduction
 Overview of image dehazing
 Motivation
 Background review
 Single image haze removal using dark channel Prior [1]
 Proposed method
 Hazing imaging model
 Atmospheric light estimation
 Transmission map estimation and refinement
 Radiance recovery
 Experimental results
 Conclusions and discussions

36
Conclusions and Discussions
 The pixel-based dark/bright channel prior and fog density
estimate method have been proposed in this thesis
 Our proposed method can accurately estimate the
atmospheric light via haze density analysis
 The adaptive atmospheric light provide the high accuracy of
transmission map

37

Vous aimerez peut-être aussi