roadsigns.m

% code adapted from Laurel's Night 7 (14.5) submission
% make sure to clear variables before running again (otherwise vector
% lengths might not match up for error plotting)
load signs.mat

% other items to use later
k_list = 1:5:num_signs; % number of eigenvectors to use
% k_list = [100]; % simplifying for testing single images
plot_error = true; % decides whether or not to show error graph
rec = 1; % index for recording error later
tic;  % start timing

% set up training data facespace
train_r = train_data.'; % reshape to signs x pixels (signs are row vectors)
train_m = train_r - mean(train_r); % mean center training data
test_r = test_data.'; % reshape to signs x pixels (signs are row vectors)
test_m = test_r - mean(test_r); % mean center test data
train_nums = 1:(num_signs*2); % placeholder until the text file can be read in
test_nums = 1:num_signs; % placeholder until the text file can be read in

% compute eigendecomposition
R = train_m.' * train_m; % find covariance matrix of pixels of training data
[V, D] = eig(R); % find eigenvectors/values of R
D_r = diag(D); % put eigenvalues in one vector
for k = k_list
    t_in_loop_start = tic; % start timing one loop
    [values, indices] = maxk(D_r, k); % choose k largest eigenvalues
    vectors = V(:,indices); % choose associated eigenvectors
    train_c = train_m * vectors; % represent training signs with chosen vectors

    % recognize signs
    test_c = test_m * vectors; % represent test signs with chosen vectors
    [close_index, distance] = knnsearch(train_c, test_c); % find closest match between test sign and training

    % compute accuracy
    matches = 0; % start from 0 matches
    for i = 1:num_signs
        if train_nums(close_index(i)) == test_nums(i) % if the indices match
            matches = matches + 1; % count this as a match
        elseif train_nums(close_index(i)) - num_signs == test_nums(i) % to account for more training data
            matches = matches + 1;
        end
    end
    t_in_loop(rec) = toc(t_in_loop_start); % record how long each loop takes
    per_matches(rec) = matches/num_signs * 100; % find percent accurate matches
    rec = rec+1;
end
% print final error

% use this section if k_list is long
% disp("Number of iterations");
% disp(numel(k_list));
% disp("Percent accurate (max)");
% disp(max(per_matches));
% disp("Time per loop (avg)");
% disp(mean(t_in_loop));

% use this section if k_list is manageable
disp("Eigenvectors");
disp(k_list);
disp("Percent accurate");
disp(per_matches);
disp("Time per loop");
disp(t_in_loop);

if plot_error == true
    % plot error
    figure(1);
    plot(k_list, per_matches);
    hold on;
    plot(k_list, t_in_loop * 1000);
    legend('Percent Accurate', 'Computation Time (ms)', 'Location', 'northwest');
    xlabel('Number of eigenvectors');
    title('Error Analysis');
end

save results.mat train_m test_m train_c test_c vectors close_index n num_signs