Added module for Ni-Black thresholding

[samyak-268]

This is in response to the feature request: 4024 and is same as PR.

[StevenPuttemans]

You probably forgot to add the adapted header file to your PR and therefore it crashes. Prototypes of your functions need to be in the public part of the module namespace.

[samyak-268]

@StevenPuttemans : Working on that.

[StevenPuttemans]

👍

[samyak-268]

@StevenPuttemans : Added function prototype. Could you review this?
Thanks!

[StevenPuttemans]

I do not have review powers, but I can take a look at it tomorrow!

[samyak-268]

Sure! Thanks a lot 👍

[vpisarev]

@samyak-268, thanks a lot! Can you please add some doxygen comments with the function and parameter description?

[samyak-268]

@vpisarev : Sure, will be doing that very soon.

[samyak-268]

@vpisarev I added some doxygen comments along with the code.

[samyak-268]

@vpisarev @StevenPuttemans : I was thinking of adding some sample/tutorial code for the Niblack module. Should I send a fresh PR for that or updating this one would be fine?

[StevenPuttemans]

O simply include the sample here!

[samyak-268]

@StevenPuttemans @vpisarev Updated PR with the following changes:

• Changed Niblack threshold from (mean - (k*std_dev)) to (mean + (k*std_dev)) which is the correct implementation as per the original text.
• Added sample code that demonstrates the use of Niblack thresholding. The value of k in the sample can be modified using track-bar.

Adding the sample made me realize that I was missing a CV_EXPORTS_W in the header!

[vpisarev]

@samyak-268, thanks! May I ask you to do 2 more things:

1. as far as I can see, using double for stddev image is not mandatory in this case; can you change it do float? (which is ~5x faster than double on ARM)
2. can you add some simple regression test?

[samyak-268]

@vpisarev : Sure. I am a little busy for the next few days. Will look into this as soon as I find time. Hope that's all right.

[vpisarev]

ok, I have yet another request. This seems to be just one relatively small function. Can you please put it to ximgproc (extended/advanced image processing) instead of creating a whole new module?

[samyak-268]

Coming back to this PR after quite a long time!

@vpisarev : I have changed the type of Mat stddev from double to float. Regarding your other suggestion, what sort of checks is a regression test supposed to perform? It would be helpful if you could guide me to some existing regression tests within the repo. Is the modules/ximgproc/test/ directory the right place to look for the same?

[samyak-268]

@StevenPuttemans : Included everything within ximgproc. The code no longer resides in it's own module. Is this what you were looking for?
Thanks!

[StevenPuttemans]

Seems good to me, but you will need acceptance of @vpisarev

[samyak-268]

Hey! Just a gentle reminder. It's been quite a while, anything that needs to be done from my end in regards to this PR? Thanks!

[StevenPuttemans]

@samyak-268 I am afraid not. Seeing that I know of PR's waiting since June ... I want to ask you to wait at least a little bit more :)

Ow and you could merge the 3 commits into 1, to facilitate the merging faster.

[mshabunin]

I'd suggest to add image conversion to grayscale (IMREAD_GRAYSCALE flag), otherwise the exception will be thrown here.

[mshabunin]

I get some strange results using your example:

It seems something goes wrong.

[samyak-268]

@mshabunin : A little busy for a few days. I'll look into your comments soon. Thanks!

[mschoeneck]

I have written some statistical threshold functions some time ago from this paper http://www.math-info.univ-paris5.fr/~vincent/articles/DRR_nick_binarization_09.pdf .

I have never really used this, but the first impression looked good. Perhabs you can use this for your implementation.

void statisticalThreshold(InputArray inputImage, OutputArray outputImage, int maxValue, int blockSize, double k, int algorithmType) {
    CV_Assert(!inputImage.empty() && (inputImage.kind() == _InputArray::MAT || inputImage.kind() == _InputArray::UMAT));
    CV_Assert(outputImage.kind() == _OutputArray::MAT || outputImage.kind() == _OutputArray::UMAT);

    UMat _inputImage = inputImage.getUMat();
    outputImage.create(inputImage.size(), CV_8U);
    UMat _outputImage = outputImage.getUMat();
    int nCompareType = CMP_GT;
    //int nNumberOfPixels = _inputImage.total();

    UMat normImage, meanImage, meanImageSquare, tempImage1, tempImage2, varianceImage, stdDevImage;
    normalize(_inputImage, normImage, 0, 255, NORM_MINMAX, CV_32F);
    boxFilter(normImage, meanImage, CV_32F, Size(blockSize, blockSize), Point(-1, -1), true, BORDER_REFLECT101);
    sqrBoxFilter(normImage, tempImage1, CV_32F, Size(blockSize, blockSize), Point(-1, -1), true, BORDER_REFLECT101);
    multiply(meanImage, meanImage, meanImageSquare);
    subtract(tempImage1, meanImageSquare, varianceImage);

    // Mean and squared mean image with integral image
    //int windowSizeHalf = blockSize / 2;
    //copyMakeBorder(_inputImage, borderImage, windowSizeHalf, windowSizeHalf, windowSizeHalf, windowSizeHalf, BORDER_REFLECT101);
    //integral(borderImage, integralImage, integralSquareImage, CV_32F, CV_32F);
    //UMat addImage1, addImage2, stdDevImage;
    //add(integralImage(Rect(0, 0, _inputImage.cols, _inputImage.rows)), integralImage(Rect(blockSize-1, blockSize-1, _inputImage.cols, _inputImage.rows)), addImage1);
    //add(integralImage(Rect(blockSize-1, 0, _inputImage.cols, _inputImage.rows)), integralImage(Rect(0, blockSize-1, _inputImage.cols, _inputImage.rows)), addImage2);
    //addWeighted(addImage1, 1. / (blockSize*blockSize), addImage2, -1. / (blockSize*blockSize), 0, meanImage, CV_32F);
    //multiply(meanImage, meanImage, meanSquareImage, 1.0, CV_32F);
    //add(integralSquareImage(Rect(0, 0, _inputImage.cols, _inputImage.rows)), integralSquareImage(Rect(blockSize, blockSize, _inputImage.cols, _inputImage.rows)), addImage1);
    //add(integralSquareImage(Rect(blockSize, 0, _inputImage.cols, _inputImage.rows)), integralSquareImage(Rect(0, blockSize, _inputImage.cols, _inputImage.rows)), addImage2);
    //addWeighted(addImage1, 1. / (blockSize*blockSize), addImage2, -1. / (blockSize*blockSize), 0, squareMeanImage, CV_32F);
    //subtract(tempImage1, tempImage2, subImage, noArray(), CV_32F);
    //cv::sqrt(subImage, stdDevImage);

    switch (algorithmType) {
    case CV_THRESH_NIBLACK:
        nCompareType = CMP_LE;
    case CV_THRESH_NIBLACK_INV:
        sqrt(varianceImage, stdDevImage);
        scaleAdd(stdDevImage, k, meanImage, tempImage1);
        break;
    case CV_THRESH_SAUVOLA:
        nCompareType = CMP_LE;
    case CV_THRESH_SAUVOLA_INV:
        sqrt(varianceImage, stdDevImage);
        stdDevImage.convertTo(tempImage1, CV_32F, (1. / 128), -1.);
        tempImage1.convertTo(tempImage2, CV_32F, k, 1.);
        multiply(meanImage, tempImage2, tempImage1, 1.);
        break;
    case CV_THRESH_WOLF:
        nCompareType = CMP_LE;
    case CV_THRESH_WOLF_INV:
        sqrt(varianceImage, stdDevImage);
        multiply(stdDevImage, meanImage, tempImage2, (k / 128.));
        scaleAdd(meanImage, (1. - k), tempImage2, tempImage1);
        break;
    case CV_THRESH_NICK:
        nCompareType = CMP_LE;
    case CV_THRESH_NICK_INV:
        add(meanImageSquare, varianceImage, tempImage1);
        //scaleAdd(meanImageSquare, nNumberOfPixels / (nNumberOfPixels - 1.), varianceImage, tempImage2);
        sqrt(tempImage1, tempImage2);
        scaleAdd(tempImage2, k, meanImage, tempImage1);
        break;
    default:
        CV_Error(Error::StsBadArg, "Threshold type unknown.");
    }
    compare(normImage, tempImage1, meanImage, nCompareType);
    meanImage.convertTo(_outputImage, CV_8U, maxValue, 0);

    return;
}

[vpisarev]

thanks! let's put it in 👍