Contents

Niblack Image Thresholding

Sample to compare Niblack thresholding against other algorithms (global thresholding and adaptive thresholding) for an image with varying illumination.

Sources:

function niblack_thresholding_demo()
    % Input 8-bit grayscale image + Parameters
    % - BS: block size (local neighborhood) [niblack, adaptive]
    % - K : constant multiplied by std dev next subtracted from mean [niblack]
    % - C : constant subtracted from mean [adaptive]
    if ~mexopencv.isOctave() && mexopencv.require('images')
        % image with dark pixels being foreground
        im = which('printedtext.png');
        K = -0.7;
        C = 7;
    elseif true
        % image with dark pixels being foreground
        im = fullfile(mexopencv.root(),'test','sudoku.jpg');
        K = -0.7;
        C = 7;
    elseif ~mexopencv.isOctave() && mexopencv.require('images')
        % image with white pixels being foreground
        im = which('rice.png');
        K = 0.7;
        C = -17;
    end
    assert(~isempty(im) && exist(im, 'file') == 2);
    img = cv.imread(im, 'Grayscale',true);
    BS = min(floor(size(img)/16) * 2 + 1);
    assert(~isempty(img), 'Failed to load image');

    % Preprocess image
    if true
        % no processing
        src = img;
    elseif false
        src = cv.medianBlur(img, 'KSize',3);
    elseif true
        % really effective for global thresholding [otsu]
        src = localNormalization(img, 11, 33);
    else
        % rice image, estimate and subtract non-uniform illumination background
        % (see NonuniformIlluminationExample.mlx example)
        if mexopencv.require('images')
            src = imtophat(img, strel('disk',15));
            src = imadjust(src);
        else
            el = cv.getStructuringElement('Shape','Ellipse', 'KSize',[15 15]*2-1);
            src = cv.morphologyEx(img, 'Tophat', 'Element',el);
            obj = cv.SimpleWB();
            src = obj.balanceWhite(src);
        end
    end

    % Threshold
    opts = {'Type','Binary', 'MaxValue',255};
    bw1 = cv.threshold(src, 'Otsu', opts{:});
    bw2 = cv.adaptiveThreshold(src, 'Method','Mean', ...
        'C',C, 'BlockSize',BS, opts{:});
    bw3 = cv.adaptiveThreshold(src, 'Method','Gaussian', ...
        'C',C, 'BlockSize',BS, opts{:});
    bw4 = cv.niBlackThreshold(src, K, 'Method','Niblack', ...
        'BlockSize',BS, opts{:});
    bw5 = cv.niBlackThreshold(src, -K, 'Method','Sauvola', ...
        'BlockSize',BS, opts{:});
    bw6 = cv.niBlackThreshold(src, -K, 'Method','Wolf', ...
        'BlockSize',BS, opts{:});
    bw7 = cv.niBlackThreshold(src, K, 'Method','Nick', ...
        'BlockSize',BS, opts{:});
    %bw8 = my_niblack(src, K, BS);

    % Results
    subplot(331), imshow(img), title('Source')
    subplot(332), imshow(src), title('Processed')
    subplot(333), imshow(bw1), title('Otsu')
    subplot(334), imshow(bw2), title('Adaptive Mean')
    subplot(335), imshow(bw3), title('Adaptive Gaussian')
    subplot(336), imshow(bw4), title('Niblack')
    subplot(337), imshow(bw5), title('Sauvola')
    subplot(338), imshow(bw6), title('Wolf')
    subplot(339), imshow(bw7), title('Nick')
end

Helper function

function out = localNormalization(img, s1, s2)
    %LOCALNORMALIZATION  local normalization to get uniform local mean and variance
    %
    %     out = localNormalization(img)
    %     out = localNormalization(img, s1, s2)
    %
    % The local normalization tends to uniformize the mean and variance of an
    % image around a local neighborhood. This is especially useful for correct
    % non-uniform illumination or shading artifacts.
    %
    % ## Input
    % * __img__ 8-bit input image
    %
    % ## Output
    % * __out__ output image of same size and type.
    %
    % ## Options
    % * __s1__ sigma to estimate the local mean. default 5
    % * __s2__ sigma to estimate the local variance. Often `s2` should be
    %   larger than `s1`. default 15
    %
    % ## References
    % > http://bigwww.epfl.ch/sage/soft/localnormalization/
    %

    % check arguments
    if nargin < 2, s1 = 5; end
    if nargin < 3, s2 = 15; end
    validateattributes(img, {'uint8'}, {});

    % convert to grayscale
    if size(img,3) == 3
        gray = cv.cvtColor(img, 'RGB2GRAY');
    else
        gray = img;
    end

    % convert to floating-point image
    gray = cv.convertTo(gray, 'RType','single', 'Alpha',1.0/255.0);

    % numerator = img - gauss_blur(img)
    blur = cv.GaussianBlur(gray, 'KSize',[0 0], 'SigmaX',s1, 'SigmaY',s1);
    num = gray - blur;

    % denominator = sqrt(gauss_blur(img^2))
    den = sqrt(cv.GaussianBlur(num.^2, 'KSize',[0 0], 'SigmaX',s2, 'SigmaY',s2));

    % output = numerator / denominator
    out = num ./ den;

    % normalize output into [0,1]
    out = cv.normalize(out, 'Alpha',0.0, 'Beta',1.0, 'NormType','MinMax');

    % convert to 8-bit
    out = cv.convertTo(out, 'RType','uint8', 'Alpha',255.0);
end

function bw = my_niblack(img, K, BS)
    %MY_NIBLACK  Manual implementation of Niblack thresholding

    img = im2double(img);
    mu = imboxfilt(img, [BS BS]);
    sd = sqrt(imboxfilt(img.^2, [BS BS]) - mu.^2);
    bw = img > (mu + K*sd);
end