Objects of interest are small and of varying size

•Upper row: Iq =50,1 = 70, dl = 5,10,20,30
•Lower row: Iq =80,1 = 70, dl = 5,10,20,30


• The methods which use constant contextual region will fail if
- objects of interest are small and of varying size.

• 
  Solution: adaptive ( growing ) contextual region.
  
• 
  also time saving.
  
• 
  Require library of ‘curves’ for different backgrounds luminosity levels.
  

• 
  The function is calculated in respect to the image background luminosity level.
  

• 
  Time consuming: calculating the contextual region
  
• 
  Noise enhancement: as small object contrast enhanced so does the noise
  
• 
  lots of methods to improve those problems
  

• Contrast Discrimination

• 
  usual methods don’t consider the brightness perception by the human eye.
  
• 
  experiments in contrast sensitivity with different background luminosity
  

An important class of functions are: monotonic functions

• 
  preserve order relationship between pixel.
  
• 
  affecting only the relative differences.
  

Local Contrast Lnhancemcnt
Yehuda Gan-El

• 
  It is then possible to look at the contrast enhancement as a function.
  
• 
  New gray level = F( old gray level)
  

•Many function had been proposed
•Each pixel (x,y) interpolated from 4 sample points


•Pixels at the edges and corners are treated specialy

• 
  Good contrast enhancement is good noise enhancement
  
• 
  Noise is mostly disturbing in flat regions.
  
• 
  Use of local statistics to limit the noise effect in those regions
  

- Contextual Region Size

• 
  each pixel is effected by 4 contextual regions forming equivalent contextual region (ECR)
  
• 
  it has been found empirically that different sample rates with the same ECR produce approximately the same results.
  
• 
  For wide range of clinical images the optimum is: 1/64
  
• 
  as the ECR < 1/64 the contrast becomes too sensitive, which cause artifacts.
  

.liter;
Method for speedup the process:

• 
  

  BB

  Calculate the mapping only for sample points
  

• 
  interpolate the mapping for the other pixels
  

• 
  Mapping the pixels with new gray levels
  
• 
  Try to make the histogram as equalized as possible
  
• 
  Small object are faded into the background
  

Blurring Level: control the flatness of the histogram
- Applied by smoothing the histogram with Gaussian { of width S )
Region Size: control the contextual region size.


— The region size and pixel weights are determined by the width of Gaussian ( R)

• 
  for each pixel (x,y) we define two neighborhoods:
  
• 
  foreground, defined by 8-connected pixels (i,j) which have the property
  

I p(x,y) - p(i,j) 1 <= t

• 
  background, 8-connected pixels which are grown around the foreground up to width of S layers.
  

Example for
•for each pixel we calculate a neighborhood.
•after the neighborhood has been set, regular histogram equalization is preformed with the neighborhood as contextual region.
neighborhood with
S=4 •Neighbor pixels with the same gray level
will grow the same neighborhood

• 
  local statistics (L): entropy, edge entropy, histogram spread function.
  
• 
  New C= f(C,L, a ).
  
• 
  a = user parameters (min and max of a)
  
• 
  for example: f(C)=0‘^^’
  
• 
  Each local statistic has different function
  

• 
  One family of functions which satisfy this are: f(C) = C“
  
• 
  where a> 0
  
• 
  As a varies from 0 to 1 the contrast enhancement decreases
  
• 
  a =1 mean no change
  
• 
  a > 1 will result contrast de-enhancement
  

Im a!

• 
  what is the optimal contextual region size?
  

Have we lost details during the enhancement?

• 
  In order to view all the possibilities we want to span the image contrast space.
  

• 
  for each pixel we define a surrounding area - Contextual Region
  
• 
  each pixel is mapped by his contextual region
  
• 
  there are many way to define the contextual region (rectangle, circle, ...)
  
• 
  pixels can have different weights
  

Sample Rate

• 
  finer sampling produce better quality but consume more time
  

• no significant difference was found between mosaic sampling and full sampling ( with the same CR ) in clinical medical images.



during the histogram enhancement process the histogram is made flat.
This flat histogram determines the contrast of the image.
If we take histograms of the same image in different flatness level and calculate the contrast out of them we will the desired space.

• 
  [1] S.M. Pizer, E.P. Ambum, J.D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B.H. Romeny, J.B. Zimmerman, and K. Zuiderveld, Adaptive histogram equalization and its variations computer vision, graphics, and image processing, vol. 39, pp. 355-368, 1987.
  
• 
  [2] D. Laxmikant and R. Cromartie, Adaptive Contrast Enhancement and De-Enhancement Pattern Regocnition,vo]. 24, pp. 289-302, 1991.
  
• 
  [3] R.B. Pai'anjape, W.M. Morrow, and R.M. Rangayyan, Adaptive- Neighborhood Histogram Equalization for Image Enhancement CVGIP: Graphial Models and image Processing, vol. 54, pp. 259-267, 1992.
  
• 
  [4] A. Mokrane, A New Image Contrast Enhancement Technique Based on a Contarst Discrimination Model CVGIP: Graphial Models and Image Processing, vol. 54, pp. 171-180, 1992.
  

• 
  [5] J.M. Gauch, Investigations of Image Contrast Space Defined by Variations on Histogram Equalization CVGIP: Graphial Models and Image Processing, vol. 54, pp. 269-280, 1992.