使用直方图比较c ++匹配图像

Question

I have a base image(a number or an operator) which i have to compare with 14 images(0 to 9 and * / - +) to know which one matches with the base image.

I created the base image's histogram and using a loop I create for all the 14 images' histograms and the histograms were normalized.

In the loop, for each newly created histogram i compared with the base histogram using the compareHist() function. And output the resulted double value.

Using Correlation or Chi-square or Intersection or Bhattacharyya method:

I am getting a particular set of values. And when using a different base , I am still getting that same set of values.

Why am I getting this? Do I need to alter the normalize function to get distinct values for different bases?

CODE:

void matchHistogram(){

Mat src_base, hsv_base;
Mat src_test1, hsv_test1;

/// Histograms
MatND hist_base;
MatND hist_test1;

/// Using 30 bins for hue and 32 for saturation
int h_bins = 30; int s_bins = 32;
int histSize[] = { h_bins, s_bins };

// hue varies from 0 to 255, saturation from 0 to 180
float h_ranges[] = { 0, 255 };
float s_ranges[] = { 0, 180 };

const float* ranges[] = { h_ranges, s_ranges };

// Use the o-th and 1-st channels
int channels[] = { 0, 1 };


for(int i=0;i<noOfcropped;i++){ //get base image  //noOfCropped is number of base images i'll compare to 14 images
    cout<<"     "<<i<<endl;
    stringstream croppedimage;
    croppedimage<<"CroppedImages/croppedImage"<<i;
    croppedimage<<".jpg";

    src_base = imread( croppedimage.str(), 1 );
    imshow(croppedimage.str(),src_base);

    /// Convert to HSV
    cvtColor( src_base, hsv_base, CV_BGR2HSV );


    /// Calculate the histogram for the HSV images
    calcHist( &hsv_base, 1, channels, Mat(), hist_base, 2, histSize, ranges);
    normalize( hist_base, hist_base, 0, 1,  NORM_MINMAX, -1, Mat() );


    for(int j=0;j<14;j++){//comparing 1 croppedimage with each different characters
        cout<<"  "<<j<<endl;
        stringstream test1;
        test1<<"ImagesToCompare/"<<j;
        test1<<".jpg";

        src_test1 = imread(test1.str(), 1 );

        /// Convert to HSV
        cvtColor( src_test1, hsv_test1, CV_BGR2HSV );

        /// Calculate the histogram for the HSV images
        calcHist( &hsv_test1, 1, channels, Mat(), hist_test1, 2, histSize, ranges);
        normalize( hist_test1, hist_test1, 0, 1,NORM_MINMAX, -1, Mat() );

        /// Apply the histogram comparison methods
        int compare_method = 0;
        //when 0 or 2, highest comparison values>> best match
        //when 1 or 3, lowest comparison values>> best match
        double base_test1 = compareHist( hist_base, hist_test1, compare_method );


        cout<<base_test1<<endl;
    }
}

}

Answer 1

If I understand your question correctly, you isolate and crop a character from some bitmap-like image and you then want to identify what character it is? Like an automatic character recognition?

Perhaps you could use an edge detector instead of comparing histograms?

I would try an algorithm something like:

1. find, isolate and crop the character to be identified
2. Scale it to a predetermined horizontal and vertical size to normalize it.
3. Sweep a directional edge detector over the character, something simple like a Sobel edge detector.
   • First apply a Horizontal edge detector, then horizontally "flatten" the edge-map to obtain a vector representing the edge information at each pixel row. (ie count the 1, and 0 in the edge-map for each pixel row)
   • Second apply a Vertical edge detector, and vertically flatten the edge-map giving another vector that represents the edge information at each pixel column. (ie cumulate and count the edge information for each pixel column)
4. Concatenate these two vectors [horizontal-edge information, vertical-edge information]
5. Then compare the final concatenated vector to a library of precomputed vectors of your known test samples (0-9, +/*-)

Numbers that look somewhat similar (8,6,9,3) should still have distinct peaks and valleys either in the horizontal component or the vertical component.