去雾代码和图片

function L = matLap(I) % L = matLap(I) % Computes the "Matting Laplacian" matrix for image I with a window size % of 3x3 (default). To adjust window size, change parameter 'win_size' % Run time is approximately 70 seconds for 400x600 image on % AMD Turion Neo x2 processor with 4GB RAM win_size = 3; % Set window size to 3x3 win_els = win_size.^2; % calculate number of window elements l_els = win_els.^2; %# of elements calculated in each forloop iteration % Calculate the size of the border of pixels not in matlap calculation % - For example, if win_size = 3x3, there is a 1 pixel border to avoid % around the whole image. win_bord = floor(win_size./2); % epsilon in Matting Laplacian equation. Levin et al. report using this % value, but can be adjusted depending on image. e = 0.000001; [m,n,c] = size(I); % find number of rows, columns, and colors of I numpix = m*n; % calculate number of rows or columns for L k = reshape(1:numpix, m, n); % create matrix of indices for each pixel U = eye(win_size); % create a an identity matrix of win_size D = eye(win_els); % create an identity matrix of win_els % calculate number of terms that will be created in for loop num_els = l_els*(m-2*win_bord)*(n-2*win_bord); % create empty index vectors for L ind_i = ones(1,num_els); ind_j = ind_i; % create empty vector to store calculated L terms els = zeros(1,num_els); % initialize counter for calculating where to deposit new index and L terms count = 0; % Begin Matting Laplacian term calculations! for x = (1 + win_bord):(n - win_bord) % start and end within border for y = (1 + win_bord):(m - win_bord) %start and end within border % Extract (win_size x win_size) window from image and reshape to % (win_els x c) matrix wk = reshape(I(y-win_bord:y+win_bord,x-win_bord:x+win_bord,:),... win_els, c); % Extract pixel indices for window, reshape to (win_els x 1) w_ind = reshape(k(y-win_bord:y+win_bord,x-win_bord:x+win_bord),... 1, win_els); % Create grid of L locations [i j] = meshgrid(w_ind, w_ind); % Reshape to vectors i = reshape(i,1,l_els); j = reshape(j,1,l_els); % Add locations to L index vectors ind_i((count*l_els + 1):(count*l_els+l_els)) = i; ind_j((count*l_els + 1):(count*l_els+l_els)) = j; win_mu = mean(wk)'; % Calculate mean of colors in window % Calculate normalized covariance matrix % - direct calculation is slightly faster than using cov function win_cov = wk'*wk/win_els-win_mu*win_mu'; % Calculate difference between window values and mean dif = wk' - repmat(win_mu,1,win_els); % Compute L terms % Note that both '\' and inv have been tried and don't result in % a noticeable time difference, likely due to the fact that the % matrix being inverted is very small. So inv was used for clarity. elements = D - (1 + dif(:,1:win_els)'*inv(... win_cov + e./win_els.*U)*dif(:,1:win_els))... ./win_els; % Add terms to L elements vector els((count*l_els + 1):(count*l_els+l_els)) = ... reshape(elements,1,l_els); % increment count index count = count + 1; end end % Compute final L matrix. Note that sparse automatically adds together % terms sharing the same location. L = sparse(ind_i, ind_j, els, numpix, numpix);