Intensity scaling in RGB image

I'm going to compress the intensity scale of RGB image (unit8) to eliminate the extreme bright and extreme dark pixels through remapping the the unscaled intensity I, given by I=R+G+B, of range 0-765 to a narrower range 300-490. So, I think the following actions can be used:

1. RGB--->Grayscale (calculating the Intensity)
2. intensity adjustment (through imadjust)
3. Gray---->RGB

as expalined above, I want to calculate the Intensity given by I=R+G+B; but I think the (rgb2gray) can not be used for the first action since it calculates the intensity by I=gray = 0.2989 * rgb(:,:,1) + 0.5870 * rgb(:,:,2) + 0.1140 * rgb(:,:,3); So, I really appreciate that guide me that my approach is correct or not? If it's wrong, please let me know the solution in Matlab.

Answer 1

[Image Analyst]

Why are you doing all that? Why can't you just do

RGBNew = double(rgbImage) / 765 * 190 + 300;

[hamidreza mohafez]

Hi and good day, Many thanks for your prompt reply. I did not get your point, that is how you get 190 in the formula: RGBNew = double(rgbImage) / 765 * 190 + 300; Thanks

[Image Analyst]

190 is the range = 490-300. That's just the standard normalization formula to scale any data to any new range. Here's a more general form:

originalRange = originalMaxValue - originalMinValue;
desiredRange = desiredMax - desiredMin;
normalized = desiredRange * (double(InputImage) - originalMinValue) / originalRange + desiredMin;

[hamidreza mohafez]

Actually, I am going to remap the unscaled intensity I=R+G+B as follow: 1. if 300<I<490 then do linear mapping same as imadjust with gamma=1) otherwise: if I<300 then make the pixel brighter by adding white such that I_scaled=300 and if I>490 then make the pixel darker such that I_scaled=490. But, I could not get this from your proposed formula. I am looking forward hearing from you soon Sir.

[hamidreza mohafez]

Actually,I applied the above formula to RGB image (unit8) which its min=6 and max=255 and the newimage range was 301-363. Should I change the newrgbimage to unit8?

[Image Analyst]

Casting to uint8 will clip the values to the range 0-255 - you won't have any values like 300 or 700 or 301-363 because they'll get clipped to 255. You can though if the output image is double.

You should be able to do what you want like this

I_scaled = rgbImage; % Initialize
% Identify out of range pixels.
darkPixels = rgbImage < 300;
brightPixels = rgbImage > 490;
% Set dark pixels to 300;
I_scaled(darkPixels) = 300;
% Set bright pixels equal to 490.
I_scaled(brightPixels) = 490;

I haven't tested but it looks like it should do what you asked for.

You won't be able to display that RGB image though you can do mathematical operations on it. To display you need to get it into a range of 0-255. You can use uint8(255 * mat2gray(I_scaled)) to create an image for display.

imshow(uint8(255 * mat2gray(I_scaled)));

[hamidreza mohafez]

I applied the formula (RGBNew = double(rgbImage) / 765 * 190 + 300) to 'football.jpg' and the result were as attached figure. I really appreciate if you guide me.

[Image Analyst]

It's probably all white because it's still a double image and you didn't scale to 0-255 and cast to uint8 for display like I said. Why do you want to do this unusual scaling anyway? And the standard football demo image does not have gray levels in the range or more than 255 - did you alter it?

[hamidreza mohafez]

Dear Sir, Many thanks for your fruitful information. Actually, I have series of wound images taken from diabetic foot patients and I am going to implement one the proposed algorithm by Berris which created the 3D RGB histogram of region of interest and then did the bin clustering to segment the comprising tissues. So, he suggested scaling the RGB intensity as a pre-processing action and as a result I tried to find the best way for compressing the intensity scale and tested it on ‘football.jpg’.

[hamidreza mohafez]

Dear Sir, I tried the following commands on my attached file (dz1.jpg) as follows: rgbImage=imread('dz1.jpg'); RGBNew = double(rgbImage) / 765 * 190 + 300; imshow(uint8(255 * mat2gray(RGBNew))); but I think it does not work, that is e.g. the pixel(1,1) of the original image (dz1.jpg) has the value of (255,255,255) and since I=R+G+B=765>490; as a result it sould be scaled to I=490; however the RGBNew(1,1) rsturned the value [363.33 363.33 363.33] or even if you scale and cat it to unit8 it returned the value [255 255 255] that is not correct. I mean the returned value for pixel(1,1) should be in such a way that I=r+g+b=490. This is because of the fact that the command ( RGBNew = double(rgbImage) / 765 * 190 + 300;) affect all 3 color component (R,G,B) of the original image in the same way. I want that each pixel of original image which its intensity I=R+G+B>490 scale to 490 (I'=R+G+B=490) and if I=R+G+B<300,then scaled to 300. But if 300<I=R+G+B<490 the pixel intensity should be mapped linearly to range of 300-490 same as imadjust does. The problem of imadjust and the above commands is that both works on individual color componenet of RGB image. Please let me know if you have any idea. bests

[Image Analyst]

What I gave you didn't work because I thought that those were the gray values of the individual color channels, not the sum of all 3 color channels. Rather than give me some values that might only work for some specific image, why don't you give the formula that they recommend you use for images in general? For example some people use normalized RGB where they extract the individual color channels and then divide by the sum of (R+G+B).

redChannel = double(rgbImage(:, :, 1));
greenChannel = double(rgbImage(:, :, 2));
blueChannel = double(rgbImage(:, :, 3));
% Display the individual color channels.
subplot(2, 4, 2);
imshow(redChannel, []);
title('Red Channel of Original Image', 'FontSize', fontSize);
subplot(2, 4, 3);
imshow(greenChannel, []);
title('Green Channel of Original Image', 'FontSize', fontSize);
subplot(2, 4, 4);
imshow(blueChannel, []);
title('Blue Channel of Original Image', 'FontSize', fontSize);

magnitudeImage = sqrt(redChannel .^2 + greenChannel .^2 + blueChannel.^2);
subplot(2, 4, 5);
imshow(magnitudeImage, []);
title('Magnitude Image');

normRed = redChannel ./ magnitudeImage;
normGreen = greenChannel ./ magnitudeImage;
normBlue = blueChannel ./ magnitudeImage;
subplot(2, 4, 6);
imshow(normRed, []);
title('Normalized Red Image');
subplot(2, 4, 7);
imshow(normGreen, []);
title('Normalized Green Image');
subplot(2, 4, 8);
imshow(normBlue, []);
title('Normalized Blue Image');

[hamidreza mohafez]

Dear Sir, As I mentioned, I am going to implement a tested method for image segmentation though bin clustering of 3D RGB histogram which is proposed by Berris (attached pdf file). In the article (as I highlighted) they did not mention that how they come up with the values (300,490) and just said that "to eliminate the extreme white and extreme dark pixels" the intensity scale was compressed by remapping the unscaled intensity I=R+G+B which ranges 0-765 to a narrower range 300-490. it is just mentioned that the chosen values (300,490) were found empirically without more details. Anyway, as you mentioned the values will be different for different database( images) and as a result I should find the values for my images. Regarding your last proposed commands, would you please help me to find the answer of following questions: 1. you mentioned that one way is to divide normalized RGB by sum of(R+G+B);however why did you divide the normalized RGB by magnitude which is (magnitudeImage = sqrt(redChannel .^2 + greenChannel .^2 + blueChannel.^2);) and not by sum of (R+G+B)?

2.Again the color component of RGB (red,green, and blue plane) were normalized individually and not as whole. So, how we can make sure that the I=R+G+B was scaled? 3. at the end the of above commnands, should we use 'rgb_scaled=cat(3,normRed,normGreen,normBlue) to form the scaled RGB image? 4. If you have any idea to choose the values (same as done by Berris 300-490) for my images (e.g. dz1.jpg)systematically? 5. finally, is there any way to compress the intensity scale of the RGB image (as I asked) as a whole and not through individual normalization of color components? I am looking forward hearing from you soon Sir. Bests