Complex image processing for loop vectorization

Hello everyone, I am new to MatLab. I would like to reuse this code but I was wondering if there is a way to vectorize the double for loop ?

link to paper : Guided Image Filtering

function q = guidedfilter_color(I, p, r, eps)
%   GUIDEDFILTER_COLOR   O(1) time implementation of guided filter using a color image as the guidance.
%
%   - guidance image: I (should be a color (RGB) image)
%   - filtering input image: p (should be a gray-scale/single channel image)
%   - local window radius: r
%   - regularization parameter: eps

if ~(size(I,3) == 3) 
	error('The guidance image input should have 3 channels');
end
[hei, wid] = size(p);
if r<2*min(hei, wid), r = round(min(hei, wid)/4); end;
N = boxfilter(ones(hei, wid), r); % the size of each local patch; N=(2r+1)^2 except for boundary pixels.

mean_I = zeros(size(I));
for ii =1:size(I,3)
    mean_I(:,:,ii) = boxfilter(I(:, :, ii), r) ./ N;
end

mean_p = boxfilter(p, r) ./ N;

mean_Ip = zeros(size(I));
for ii =1:size(I,3)
    mean_Ip(:,:,ii) = boxfilter(I(:, :, ii).*p, r) ./ N;
end

% covariance of (I, p) in each local patch.
cov_Ip = zeros(size(I));
for ii =1:size(I,3)
    cov_Ip(:,:,ii) = mean_Ip(:,:,ii) - mean_I(:,:,ii) .* mean_p;
end

% variance of I in each local patch: the matrix Sigma in Eqn (14).
% Note the variance in each local patch is a 3x3 symmetric matrix:
%           rr, rg, rb
%   Sigma = rg, gg, gb
%           rb, gb, bb
var_I_rr = boxfilter(I(:, :, 1).*I(:, :, 1), r) ./ N - mean_I(:,:,1) .*  mean_I(:,:,1); 
var_I_rg = boxfilter(I(:, :, 1).*I(:, :, 2), r) ./ N - mean_I(:,:,1) .*  mean_I(:,:,2); 
var_I_gg = boxfilter(I(:, :, 2).*I(:, :, 2), r) ./ N - mean_I(:,:,2) .*  mean_I(:,:,2); 
var_I_rb = boxfilter(I(:, :, 1).*I(:, :, 3), r) ./ N - mean_I(:,:,1) .*  mean_I(:,:,3); 

var_I_gb = boxfilter(I(:, :, 2).*I(:, :, 3), r) ./ N - mean_I(:,:,2) .*  mean_I(:,:,3); 
var_I_bb = boxfilter(I(:, :, 3).*I(:, :, 3), r) ./ N - mean_I(:,:,3) .*  mean_I(:,:,3); 

a = zeros(hei, wid, 3);
for y=1:hei
    for x=1:wid        
        Sigma = [var_I_rr(y, x), var_I_rg(y, x), var_I_rb(y, x);
            var_I_rg(y, x), var_I_gg(y, x), var_I_gb(y, x);
            var_I_rb(y, x), var_I_gb(y, x), var_I_bb(y, x)];
        %Sigma = Sigma + eps * eye(3);
        
        cov_Ip1 = [cov_Ip(y, x,1), cov_Ip(y, x,2), cov_Ip(y, x,3)];        
        
        a(y, x, :) = cov_Ip1 * inv(Sigma + eps * eye(3)); % Eqn. (14) in the paper;
    end
end

b = mean_p - a(:, :, 1) .* mean_I(:,:,1) - a(:, :, 2) .* mean_I(:,:,2) - a(:, :, 3) .* mean_I(:,:,3); % Eqn. (15) in the paper;

q = (boxfilter(a(:, :, 1), r).* I(:, :, 1)...
+ boxfilter(a(:, :, 2), r).* I(:, :, 2)...
+ boxfilter(a(:, :, 3), r).* I(:, :, 3)...
+ boxfilter(b, r)) ./ N;  % Eqn. (16) in the paper;
end

[Rik]

Vectorization of code will only speed up your code if there is a way to process the data as an array. A good example of this is a convolution, which you could implement by looping over every dimension, but for which there are mathematical shortcuts that Matlab implements.

Sometimes array processing will cause a slowdown because of memory requirements.

If you're looking for a speed increase: a change in algorithm has a much larger potential impact.

[Artur MKRTCHYAN]

hi, tanks for answer. maybe you some propositions of solution code ?

[Rik]

I don't have a ready-made solution for you. You could start yourself by running the profiler to look where your code spends the most time.

[Artur MKRTCHYAN]

Of course I've already used the profiler, and that's why I want to vectorize the double loop because that's exactly what takes more time. So I'm not asking to rewrite the code, I'm looking to see if it's possible to optimize only the double loop part.

Answer 1

[Image Analyst]

I would not reuse that code. I'd use the built-in imguidedfilter() function.

[Artur MKRTCHYAN]

This is a good idea, but it is part of a project and the input parameters of the function do not match the parameters of the gidedfilter() function.

[Rik]

Would it be possible to write an interface function that converts the input parameters to what imguidedfilter needs?

[Artur MKRTCHYAN]

It seems to me that this function has been created because the basic function is not adapted for RGB images. I'm not sure that's possible to make it this way.