This isn't really a complete solution, but oh well.
Just because regionprops does binarization doesn't mean that the result needs to be binary. You can use the mask to select regions from the image. The masks don't even necessarily need to have hard edges. For example:
a=imread('majority.jpg');
c=~imbinarize(a);
cl = bwlabel(c);
S=regionprops(c,'boundingbox');
bb = vertcat(S.BoundingBox);
maxbb = max(bb(:,[4 3])); % minimal box geometry
maxbb = maxbb+5; % add some extra room
numchars = numel(S);
for j=1:numchars
tp = round(S(j).BoundingBox);
xrange = tp(1)+(-1:tp(3)+1);
yrange = tp(2)+(-1:tp(4)+1);
% start out with no masking
target = a(yrange,xrange);
% use a dilated mask
tm = double(imdilate(cl(yrange,xrange)==j,ones(3)));
% apply the mask
target = im2uint8(im2double(target).*tm + ones(size(tm)).*(1-tm));
% this pads with black to match sizes
bs = (maxbb-size(target))/2;
target = padarray(target,floor(bs),255,'pre');
target = padarray(target,ceil(bs),255,'post');
ax=subplot(3,3,j);
imshow(target);
end
Now of course, segmenting by region connectivity isn't going to solve the problem. There's no trivial way to break apart multiple letters made with the same stroke, and there's no way to associate the dots with their corresponding characters. How is that done? Fancier binary segmentation? Polygonal ROI's? I don't know, but using imcrop() isn't the answer.
