function CV2011_1HW4_ben_delillo()

imdir = 'C:\Users\Ben\Downloads\'; % Change this as appropriate
I = imread(strcat(imdir,'TextImage.tif'));

% Use inverted binary image. Inversion changes the letters and other ink
% from holes into connected components.
Iinv = im2bw(I) == 0;

% After many variatons on refining edge detection I stumbled upon
% morphological operations and that using a structuring element made up
% of one element at the origin and one element three pixels above,
% approximating the separation between the base of an i and its tittle,
% was enough to change an i into a single connected component without
% compromising the integrity of the rest of the image as a blur would have.
Iprocessed = imdilate(Iinv, strel('pair', [-3 0]));

% Manualy extract example instance of what we want to find
target = imcrop(Iprocessed, [682 467 16 30]);

% Split into connected comonents
[components count] = bwlabel(Iprocessed);

% Using regionprops FilledImage option lets us easily analyze each
% candidate letter
stats = regionprops(components, 'BoundingBox', 'FilledImage');

% Making all letter candidates the same size as the target makes
% comparison easier (also allows for finding off-scale matches)
% It also makes rejecting small (< 10 pixels) clumps as they stretch
% out as to almost make a solid white box.
sizeToTemplate = @(img) imresize(img{1},size(target));

% Embed each image in a cell because arrayfun needs uniformly sized input
letters = arrayfun(sizeToTemplate, {stats.FilledImage},...
    'UniformOutput', false); % non-scalar output requires manual override

% This algorithm relies on two assumptions. The first is that shape
% matching the target shape will have a similar 'density' of white pixels
% per row as the target shape. The second is that the pool of candidates
% will not have non-matching shapes with a similar density 'fingerprint'.
% That second assumption actually fails for this text when using columns to
% calculate that fingerprint because of how letters are taller than they
% are wide and how similar l is to i. However, using the sum of rows
% instead is high enough resolution that the algorithm can successfully
% differentiate between l and i.
sumOfDiffsOfRowPixelSums = @(ltr) sum(abs(sum(ltr, 2) - sum(target,2)));
colSumDiffs = cellfun(sumOfDiffsOfRowPixelSums, letters);

% Is there an equivalent of the functional 'filter' idiom?
results = []; bbs = [];
for n=1:count
    if colSumDiffs(n) < 50
        results = cat(4, results, letters{n});
        bbs = cat(2, bbs, stats(n).BoundingBox.');
    end
end

foundCount = length(results);
titleStr = ['Found ', num2str(foundCount), ' instances of the letter i'];

figure('Name', titleStr), imshow(I);
plotBoundingBoxes(bbs);

function plotBoundingBoxes(boxes)
hold on;
for cnt = 1:length(boxes)
    rectangle('position', boxes(:,cnt), 'edgecolor', 'r');
end

function imSxS(im1, im2)
figure;
subplot(1,2,1), subimage(im1);
subplot(1,2,2), subimage(im2);