InitializeNNF.m

% Given an image "A" to rebuild in, assuming image B to rebuild from and
% given "mask", the region of A to rebuild, randomly initialize offsets and
% computes the similarity between patches centered in A and centered in B
% at the assigned offset location. "pad_A" and "pad_B" are virtually
% extended versions of A and B, "half_patch" is the size of half a patch
% side, "error" is the similarity metric to be used (0, 1 or 2) and
% "lambda" is used for Poisson metric.
function [NNF] = InitializeNNF(A, mask, pad_A, pad_B, half_patch, error, lambda)
    [m, n, ~] = size(A);

    % Random offsets, located outside of the region to fill.
    [row, col] = find(mask == 0);
    random_indices = randi([1 length(row)], m, n);
    
    offset_x = row(random_indices);
    offset_y = col(random_indices);

    NNF(:, :, 1) = offset_x;
    NNF(:, :, 2) = offset_y;
        
    % Virtually extending the offsets to ease the indices handling.
    pad_offset_x = padarray(offset_x, [half_patch half_patch]);
    pad_offset_y = padarray(offset_y, [half_patch half_patch]); 
 
    % For each pixel in A, compute the SSD between the patch centered in
    % A(i,j) and B(NNF(i,j,1), NNF(i,j,2)). Store the distance in the third
    % channel.
    for i = 1 + half_patch : m + half_patch
        for j = 1 + half_patch : n + half_patch                        
            NNF(i - half_patch, j - half_patch, 3) = Distance(pad_A, pad_B, i, j, pad_offset_x(i, j) + half_patch, pad_offset_y(i, j) + half_patch, half_patch, error, lambda);
        end
    end    
end