Revision history [back]

Subtract operation creates result with noise

Hi all, I'm trying to subtract two Mat instances to obtain a result that represents a person's signature. Unfortunately, my resulting Mat seems to show noise along the top edge. Here are the source images and output:

The templateToMatch: C:\fakepath\template.png

The original source from which the ROI will be extracted, referred to as "formImage" in the code:C:\fakepath\sep-ira-completed-handwritten-p5.png

The ROI extracted via matchTemplate and Rect:C:\fakepath\signed.png

The result of running subtract on the template and extracted ROI: C:\fakepath\difference.png

As you can see, difference.png shows noise creeping down over the actual signature.

Can someone have a look at my implementation and tell me what I may be doing wrong? Is there a way I can mitigate this, if not avoid it altogether?

Here's where the subtraction takes place:

cv::Mat extractSignatureFromRegion(cv::Mat & templateToMatch, cv::Mat & formImage, bool* sigFound) {
    Mat signature;
    Mat result;
    bool hasSignature = false;
    Mat sigRegion = getSignatureRegion(templateToMatch, formImage);

    subtract(templateToMatch, sigRegion, signature);
    cvtColor(signature, signature, CV_BGR2GRAY);
    if (countNonZero(signature) > 0) {
        hasSignature = true;
    }
    *sigFound = hasSignature;
    return signature;
}

Here's the getSignatureRegion that uses template matching to extract the potentially signed region in the source doc:

    cv::Mat getSignatureRegion(cv::Mat & templateToMatch, cv::Mat & formImage) {
    //Create result matrix from template and original source
    Mat resultImg;
    Mat resultMatrix;
    Mat regionMatchingTemplate;
    Mat result;
    double minVal;
    double maxVal;
    Point minLoc;
    Point maxLoc;
    Point matchLoc;

    int resultColumns = formImage.cols - templateToMatch.cols + 1;
    int resultRows = formImage.rows - templateToMatch.rows + 1;
    resultMatrix.create(resultColumns, resultRows, CV_32FC1);


    // Match template and normalize
    matchTemplate(formImage, templateToMatch, resultMatrix, CV_TM_SQDIFF);
    normalize(resultMatrix, resultMatrix, 0, 1, NORM_MINMAX, -1, Mat());
    minMaxLoc(resultMatrix, &minVal, &maxVal, &minLoc, &maxLoc, Mat());
    matchLoc = minLoc;
    regionMatchingTemplate = formImage(Rect(matchLoc.x, matchLoc.y, templateToMatch.cols, templateToMatch.rows));
    regionMatchingTemplate.copyTo(result);
    return result;
}

Subtract operation creates result with noise

The templateToMatch: C:\fakepath\template.png

The original source from which the ROI will be extracted, referred to as "formImage" in the code:C:\fakepath\sep-ira-completed-handwritten-p5.png

The ROI extracted via matchTemplate and Rect:C:\fakepath\signed.png

The result of running subtract on the template and extracted ROI: C:\fakepath\difference.png

As you can see, difference.png shows noise creeping down over the actual signature.

Can someone have a look at my implementation and tell me what I may be doing wrong? Is there a way I can mitigate this, if not avoid it altogether?

Here's where the subtraction takes place:

cv::Mat extractSignatureFromRegion(cv::Mat & templateToMatch, cv::Mat & formImage, bool* sigFound) {
    Mat signature;
    Mat result;
    bool hasSignature = false;
    Mat sigRegion = getSignatureRegion(templateToMatch, formImage);

    subtract(templateToMatch, sigRegion, signature);
    cvtColor(signature, signature, CV_BGR2GRAY);
    if (countNonZero(signature) > 0) {
        hasSignature = true;
    }
    *sigFound = hasSignature;
    return signature;
}

Here's the getSignatureRegion that uses template matching to extract the potentially signed region in the source doc:

    cv::Mat getSignatureRegion(cv::Mat & templateToMatch, cv::Mat & formImage) {
    //Create result matrix from template and original source
    Mat resultImg;
    Mat resultMatrix;
    Mat regionMatchingTemplate;
    Mat result;
    double minVal;
    double maxVal;
    Point minLoc;
    Point maxLoc;
    Point matchLoc;

    int resultColumns = formImage.cols - templateToMatch.cols + 1;
    int resultRows = formImage.rows - templateToMatch.rows + 1;
    resultMatrix.create(resultColumns, resultRows, CV_32FC1);


    // Match template and normalize
    matchTemplate(formImage, templateToMatch, resultMatrix, CV_TM_SQDIFF);
    normalize(resultMatrix, resultMatrix, 0, 1, NORM_MINMAX, -1, Mat());
    minMaxLoc(resultMatrix, &minVal, &maxVal, &minLoc, &maxLoc, Mat());
    matchLoc = minLoc;
    regionMatchingTemplate = formImage(Rect(matchLoc.x, matchLoc.y, templateToMatch.cols, templateToMatch.rows));
    regionMatchingTemplate.copyTo(result);
    return result;
}

Here's another example where this time, I used a template to extract a region from an empty form where no signature is present in the ROI. I would expect the subtract operation to result in a nearly empty Mat, but again there is "noise" along the bottom edge of the result:

Unsigned form: C:\fakepath\rollover-signature-page-unsigned.jpg

Template to find ROI of empty signature box: C:\fakepath\template.png

ROI that was extracted from unsigned form: C:\fakepath\signed.png

And finally, the subtract result with a weird black rectangled along bottom edge: C:\fakepath\difference.png