Histogram comparation for Content Based Image Retrieval. Different color spaces (RGB,HSV,etc..)

Question

I'm doing a final degree proyect in Content Based Image Retrieval using OpenCv. I have started comparing histograms. The thing is that I have seen a lot of post saying that RGB is the worst color space to operate, and it's better to use HSV or YCrCb . However, when I compare images with RGB, the results are always better than when I use other color spaces.

This is the code for the YCrCb color:

void Histogram::calculateYCCHist(const cv::Mat3b& img_base, const cv::Mat1b& mask)
{
cv::Mat3f ycbcr;
cvtColor( Mat3f(img_base), ycbcr, CV_BGR2YCrCb);
int hist_size[] = {100, 100, 100};
float y_range[] = { 0, 1 }; // luma (Y) value have a nominal range from 0 to 1
float chr_range[] = { -0.5, 0.5 }; // chroma (CB and CR) values will have a nominal range from -0.5 to +0.5
const float* ranges_Y[] = {y_range};
const float* ranges_Cb[] = {chr_range};
const float* ranges_Cr[] = {chr_range};
int channel_y[] = {0};
int channel_cb[] = {1};
int channel_cr[] = {2};

// Compute histogram
calcHist(&ycbcr, 1, channel_y, mask, m_histogram_b, 1, hist_size, ranges_Y, true, false);
normalize( m_histogram_b, m_histogram_b, 0, m_histogram_b.rows, NORM_MINMAX, -1, Mat() );
calcHist(&ycbcr, 1, channel_cb, mask, m_histogram_g, 1, hist_size, ranges_Cb, true, false);
normalize( m_histogram_g, m_histogram_g, 0, m_histogram_g.rows, NORM_MINMAX, -1, Mat() );
calcHist(&ycbcr, 1, channel_cr, mask, m_histogram_r, 1, hist_size, ranges_Cr, true, false);
normalize( m_histogram_r, m_histogram_r, 0, m_histogram_r.rows, NORM_MINMAX, -1, Mat() );
 }

Are the ranges right?

I normalize the image but it doesn't change anything at all.

Do you think that I should try to use other methods?

I have also noticed that the number of bins is of great importance and I get very different results if I change this value, is there any way to control this?

Regards

Answer 1

I realise that this may not answer your specific question but rather your more general one.... To be honest each colour space is very different from each other. In my experience using a different colour space in your algorithm usually requires changes to how the algorithm works in order to get useful results. An example of this would be how in the HSV space the H channel is circular i.e. a value of 100 = 0, which isnt the case for RGB. For example is that in YCbCr the Y channel might have more emphasis/importance than the Cr adn Cb channels if you are dealing with one type of images where illumination strength is of importance or the opposite way round if you are dealing with colours regardless of illumination strength.

Secondly when you say that one colour space is worse than another you are being unfair to each. They each have their uses and their limitations. The article that you link to on which space is better is subjective to "stability". There are many many reasons why you might choose a colour space over another.

On your specific question on YCbCr ranges...Im sorry I dont know... Im a bit rusty on this. :)