% author: M.M.Haji
% email : mehdi.haji@gmail.com
%
function y = remove_nontext4(x)
% x: input image
% y: text mask
% remove non-text components using decision tree and dct-18 features.
% example:
%
% x = imread('image031.jpg');
% y = remove_nontext4(x);
% z = rgb2gray(x);
% z(find(y==false)) = 255;
% imshow(z);

load text_dtree
dct18_coefs = [4 5 6 12 13 14 20 21 22 44 45 46 52 53 54 60 61 62];

if ndims(x) == 3
    x = rgb2gray(x);
end

[nr nc] = size(x);

NR = ceil(nr/8) * 8;
NC = ceil(nc/8) * 8;
x = padarray(x,[NR-nr NC-nc],0,'post');
y = logical(zeros(size(x)/8));

for i1 = 1:8:NR
    for i2 = 1:8:NC
        J = dct2(x(i1:i1+7, i2:i2+7));
        J = J';
        J = J(:);
        tree_output = (treeval(text_dtree,J(dct18_coefs(1:18))'));
        if tree_output == 1
            y(ceil(i1/8),ceil(i2/8)) = 1;
        else
            y(ceil(i1/8),ceil(i2/8)) = 0;
        end
    end
end

% apply rule-based smoothing:
imwrite(y,'tmp.bmp');
dos('rbsmooth tmp.bmp tmp.bmp 1 10');
dos('rbsmooth tmp.bmp tmp.bmp 3 10');

y = imread('tmp.bmp');
y = imopen(y,strel('rect',[1 3]));
imwrite(y,'tmp.bmp');

dos('rbsmooth tmp.bmp tmp.bmp 2 10');
y = imread('tmp.bmp');

y = imclose(y,strel('rect',[1 3]));
y = imopen(y,strel('rect',[1 3]));

y = imresize(y,8);
y = y(1:nr,1:nc);