how to calculate Contrast Improvment Ratio - CIR

Hey everyone,

Im trying to calculate to Contrast Improvment Ratio (CIR) between enhanced and unenhanced images within region of interest R. This is the formula Im using:

where C(x, y) and C¯(x,y)C¯(x,y) are the local contrast values at (x, y) of the unenhanced and enhanced images, respectively.

The Local contrast value C(x, y) computed as

where p and a are the mean values within the center region (3 × 3) pixels and the neighborhood, or surrounding region, (7 × 7) pixels, respectively.

I have a problem to actually write this into a matlab code.

I would appreciate your help, thanks!

[Simon Chan]

Does the surrounding region includes or excludes the center region?

[Itzhak Mamistvalov]

Actually, Im not sure about it. I guess it is, I wrote here the defenition and the explanation of the formula (this is an academic task).

Answer 1

[Image Analyst]

Try this:

windowSize7 = 7;
windowSize3 = 3;
kernel3 = ones(windowSize3) / windowSize3^2;
% Counting center 3x3 region
kernel7 = ones(windowSize7) / windowSize7^2; 
% If you don't want to count the center 3x3 region, Make the kernel this way.
kernel7Ring = ones(windowSize7) / windowSize7^2; 
kernel7Ring(3:5, 3:5) = 0;

% Here you choose.  If you want the center expluded, add in the line below.
% kernel7 = kernel7Ring;

% Compute C1
p3 = conv2(double(grayImage1), kernel3, 'same');
countImage = conv2(ones(size(grayImage1)), kernel7, 'same');
glSumImage = conv2(double(grayImage1), kernel7, 'same');
p7 = conv2(double(grayImage1), kernel7, 'same');
C1 = imabsdiff(p3,p7) / (p3 + p7);
subplot(2, 2, 3);
imshow(C1, []);
impixelinfo;
title('C1', 'FontSize', fontSize);

% Compute Cbar
p3 = conv2(double(grayImage2), kernel3, 'same');
countImage = conv2(ones(size(grayImage2)), kernel7, 'same');
glSumImage = conv2(double(grayImage2), kernel7, 'same');
p7 = glSumImage ./ countImage;
Cbar = imabsdiff(p3,p7) / (p3 + p7);
subplot(2, 2, 4);
imshow(Cbar, []);
impixelinfo;
title('C2', 'FontSize', fontSize);

numerator = (C1 - Cbar).^2;
denominator = C1 .^ 2;;
CIR = sum(numerator(:)) / sum(denominator(:))

[Itzhak Mamistvalov]

First of all thanks for your detailed answer.

Secondly, while computing p3 there is no matter for that the kernel3 supposed to be centered in the kernel7? I mean if there is a need to cut the borders (2 from each top bottom left and right) of the original image?

[Image Analyst]

Often it is assumed that the image is large enough that you don't need to worry about edge effects. If you really care about those, I'd use imfilter() which has many more options for handling edge effect than conv2() which has only 3. If you want you can use conv2() with 'valid' instead of 'same' but you'll have images with 2 different sizes after that and you'll have to crop the one with 7 to make it the same size as the one with 3. My guess is it won't make any significant difference, as long as your image is more than a few dozen pixels wide, but you're welcome to try (by yourself)

[Itzhak Mamistvalov]

Ok. Thanks for your help.

Just last qusetoin, may you explain me please the glSumImage part?

[Image Analyst]

If you're going to exclude the central 3x3 region, then there are not 7x7=49 pixels to be averaged over, you need to average over 49 - 9 = 40 because you're not going to consider the central 3x3 = 9 pixels. So while a convolution gets you the average over the whole 49 pixels if you do

kernel7 = ones(windowSize7) / windowSize7^2;

if you use a kernel that is basically a ring shape, you need to get the average by summing the 40 gray levels and then dividing by 40. So to add up the gray levels of the 40, I apply that ring-shaped kernel to the image. That gets me the sum of the 40 pixels around that point. Now to find out how many pixels are in there (40 except near the edges) I apply the ring-shaped kernel to an image of all 1's. That will count the pixels we are considering at each point. Then we divide them to get the average of the (mostly) 40 pixels.

[Itzhak Mamistvalov]

Thanks again!