Learning OpenCV: Computer Vision with the OpenCV Library

This is a good book about computer vision. You can find how to implement not only color histogram computation and comparison but also other computer vision algorithms, by using OpenCV and C/C++. It also provide the principle of color histogram computation and comparison.

You can do this easily in matlab. Read a color image in matlab and then run the following code. The color image has three different channels (red, green, blue). You can then plot each channel separately as follows:

im = imread('color_img.jpeg');

red_hist= imhist(im(:,:,1));

plot(red_hist,'red'); title('RED IMAGE HISTOGRAM')

green_hist=imhist(im(:,:,2));

plot(green_hist,'green'); title('GREEN IMAGE HISTOGRAM')

blue_hist=imhist(im(:,:,3));

plot(blue_hist,'blue'); title('BLUE IMAGE HISTOGRAM');

As for histogram comparison, usual choices include L1 distance, histogram intersection, chi-square, hellinger distance, etc. For building the histogram, both OpenCV and Matlab are excellent tools. If you have to use C or C++, OpenCV is a better choice. Otherwise, Matlab is more efficient for coding.

You can also use Scilab5.3.3 which is freely downloadable .. and use very similar code

given by Christos..

Cheers

Dear Hadjer Lacheheb,

if you want to extract a color histogram from an RGB-Image you should perform first a conversion to the HSV color space ( http://en.wikipedia.org/wiki/HSL_and_HSV ) to isolate color information in one channel (H). In addition, you might want to weigh the contribution of a pixel to your color histogram by its saturation (S channel) and its value (V channel) as pixels with either low saturation or low value will have more or less random color (hue) values.

To compare color histograms you could use a simple norm like L1, but a cross-bin distance like the earth movers distance (EMD) will probably work much better. A nice paper in this regard is:

Approximate earth mover's distance in linear time.

Sameer Shirdhonkar, David W. Jacobs

In proceeding of: 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2008), 24-26 June 2008, Anchorage, Alaska, USA

https://www.researchgate.net/publication/221363234_Approximate_earth_mover's_distance_in_linear_time

Cheers,

Jochen

Conference Paper Approximate earth mover's distance in linear time

thanks a lot for your answer Sir Jochen Kerdels

Jochen Kerdels..it is a nice answer. It would be interesting if the following can be commented on:

1. Does it matter if the image is computer generated or captured with a camera?

2. Is it better to use RIFF (raw inf format) and then use color/model or space?

3. Can the jpeg output actually invert back to real frequency of captured image?

4. How would the specular and dispersive/ texture properties affect the histogram ?

5. Is it possible to obtain the complete wavelength spectrum from the RIFF as

only red/ green/ blue filters are used .( open to correction).

6. would the lens structure affect the capture of different wavelength in a

practical camera?

Thanks ...

Dear Narasim Ramesh,

in general you cannot determine the exact wavelength that caused a color in an image - independent of the format. You cannot even determine if it was just one specific wavelength or if it was a combination of wavelengths. The latter may be the case in most situations.

The situation is a bit better, if you know the exact type of sensor your camera uses. Let's say you would use a ICX424AL from Sony. In the corresponding datasheet you will find this diagramm:

http://html.alldatasheet.com/html-pdf/95421/SONY/ICX424AL/1213/14/ICX424AL.html

It describes the sensitivity of the sensor to specific wavelengths of light. But if you get a high pixel value, you can't tell if it was, e.g., a weak signal at 500nm or a very strong signal at 900nm. Both signals could lead to the same pixel value.

The situation gets still better, if you place color filters in front of the sensor. These color filters act as a bandpass-filter for a specific subrange of wavelengths. But the typical RGB-Filter used in a Bayer-Pattern have still a pretty wide range wavelengths that they cover. If you know the exact filter properties of your color filters in addition to the sensitivity of your sensor, you may come close to recovering the original wavelengths of the signal.

If you have a specific application that needs information about a certain, well defined wavelength, you may get color filters with a very narrow band that just pass the wavelength you want to detect. Here is a nice example of such a filter:

http://www.iridian.ca/english/irdProducts.asp?x=915&a=mpp&pcid=27&pscid=15&pid=85&mpid=242&d=1

Cheers,

Jochen

Dear Jochen Kerdels Thanks for a very informative answer. My apologies to the questioner as I am straying from the main question . . I was interested

in feasible and challenging assignments for my students in DSP on the following and hence

the questions

. 1 To be able to find spectral components from images of stars / electrical arcs (Plasma) using ordinary cameras on mobile phones

2 To be able to predict ripeness/tastiness/properties of fruits and vegetables using ordinary cameras on mobile phones

Thanks once again ..it was helpful and useful..

Cheers

Dear Narasim Ramesh,

for a students project this may be of interest:

http://hackaday.com/2012/08/27/turning-a-webcam-into-a-spectrometer/

http://publiclab.org/wiki/video-spectrometer-construction

It describes how you can make a makeshift spectrometer based on diffraction.

Cheers,

Jochen

Dear Jochen Kerdels Thanks for pointing out a .very useful site . It was amazing

how nicely it was described. i hope to pass this on to our students..On their and my own behalf thank you once again .

Cheers

Narasim

First RGB converted into ycbcr. Seperate each channel. Modify y with imhist command. Append cbcr color information

Some metrics for colour histogram comparisions are :

Now, opencv provides interfaces to calculate these distances for histogram distributions.

Color histogram shows distribution for each channels and split into RGB Channels:

Red = image(:,:,1);

Green = image(:,:,2);

Blue = image(:,:,3);

%Get histValues for each channel

[yRed, x] = imhist(Red);

[yGreen, x] = imhist(Green);

[yBlue, x] = imhist(Blue);

%Plot them together in one plot

plot(x, yRed, 'Red', x, yGreen, 'Green', x, yBlue, 'Blue');

OR

image=imread(' .jpg')

Red=image(:,:,1);

Green=image(:,:,2);

Red=image(:,:,1);

hFig = figure;

subplot(2,1,1);

imshow(image);

title(' image');

subplot(2,1,2);

h1=histogram(Red);

hold on

h2=histogram(Green);

hold on

h3=histogram(Blue);

title('RGB histogram');