how to optimized this code

I am new to matlab and trying to optimize this matlab code through optimization

this function is get location features using superpixel image and original image

Variable Sp is superpixel image matrix (512x512) , N_Sp is number of superpixel, im is original image matrix(512,512)

function [Features, Locpixels, distances] = LocationFeatures(Sp, N_Sp, im) 

Features = zeros(3,N_Sp);
for i = 1 : N_Sp
   [x, y] = find(Sp==i);
   
   
   loc = (y-1)*size(im,1) + x;
   Locpixels{i} = loc;
   
   lengthofpix = length(loc);

   Features(1,i) = mean(x) / size(im,1);
   Features(2,i) = mean(y) / size(im,2);  
   Features(3,i) = lengthofpix;
end
temp = Features(1:2,:)';
parfor i = 1 : N_Sp
    distances(i,:) = pdist2(temp, Features(1:2,i)')';
end

and then calculate features using ColorFeatures Function in blow but it too much time consume to calculate

function [Features] = ColorFeatures(Sp, N_Sp, Locpixels, im, imlab, distances) 

Features = zeros(40, N_Sp);

im = im2double(im);
iml = mat2gray(imlab(:,:,1)); ima = mat2gray(imlab(:,:,2)); imb = mat2gray(imlab(:,:,3));
imR = im(:,:,1);  imG = im(:,:,2);   imB = im(:,:,3);  
imhsv = rgb2hsv(im);
imh = mat2gray(imhsv(:,:,1)); ims = mat2gray(imhsv(:,:,2)); imv = mat2gray(imhsv(:,:,3));
    
expdist = exp(-1/0.16 * distances);
    
histrgb = zeros(N_Sp, 48);
histlab = zeros(N_Sp, 40);
histh = zeros(N_Sp, 8);
hists = zeros(N_Sp, 8);

for i = 1 : N_Sp
	histrgb(i,:) = [imhist(imR(Locpixels{i}),16); imhist(imG(Locpixels{i}),16); imhist(imB(Locpixels{i}),16)] ./ length(Locpixels{i});
	histlab(i,:) = [imhist(iml(Locpixels{i}),8); imhist(ima(Locpixels{i}),16); imhist(imb(Locpixels{i}),16)] ./ length(Locpixels{i});
	histh(i,:) = imhist(imh(Locpixels{i}),8) ./ length(Locpixels{i});
	hists(i,:) = imhist(ims(Locpixels{i}),8) ./ length(Locpixels{i});
end
    
    
for i = 1 : N_Sp
    % The average color values
	Features(1,i) = mean(imR(Locpixels{i}));
	Features(2,i) = mean(imG(Locpixels{i}));
	Features(3,i) = mean(imB(Locpixels{i}));
	Features(4,i) = mean(iml(Locpixels{i}));
	Features(5,i) = mean(ima(Locpixels{i}));
	Features(6,i) = mean(imb(Locpixels{i}));
	Features(7,i) = mean(imh(Locpixels{i}));
	Features(8,i) = mean(ims(Locpixels{i}));
    Features(9,i) = mean(imv(Locpixels{i}));
        
    % Color Histogram Features
	Features(10,i) = sum(chi_square_statistics_fast(histrgb(i,:),histrgb));
	Features(11,i) = sum(chi_square_statistics_fast(histlab(i,:),histlab));
	Features(12,i) = sum(chi_square_statistics_fast(histh(i,:),histh));
	Features(13,i) = sum(chi_square_statistics_fast(hists(i,:),hists));
end
    
Rdist = zeros(N_Sp,N_Sp);  Gdist = zeros(N_Sp,N_Sp);  Bdist = zeros(N_Sp,N_Sp);
ldist = zeros(N_Sp,N_Sp);  adist = zeros(N_Sp,N_Sp);  bdist = zeros(N_Sp,N_Sp);
hdist = zeros(N_Sp,N_Sp);  sdist = zeros(N_Sp,N_Sp);  vdist = zeros(N_Sp,N_Sp);
for i = 1 : N_Sp
	Rdist(i,:) = pdist2(Features(1,:)', Features(1,i));
	Gdist(i,:) = pdist2(Features(2,:)', Features(2,i));
	Bdist(i,:) = pdist2(Features(3,:)', Features(3,i));
	ldist(i,:) = pdist2(Features(4,:)', Features(4,i));
	adist(i,:) = pdist2(Features(5,:)', Features(5,i));
	bdist(i,:) = pdist2(Features(6,:)', Features(6,i));
	hdist(i,:) = pdist2(Features(7,:)', Features(7,i));
	sdist(i,:) = pdist2(Features(8,:)', Features(8,i));
	vdist(i,:) = pdist2(Features(9,:)', Features(9,i));
end

expR = exp(-1/0.16 * Rdist);
expG = exp(-1/0.16 * Gdist);
expB = exp(-1/0.16 * Bdist);
expl = exp(-1/0.16 * ldist);
expa = exp(-1/0.16 * adist);
expb = exp(-1/0.16 * bdist);
exph = exp(-1/0.16 * hdist);
exps = exp(-1/0.16 * sdist);
expv = exp(-1/0.16 * vdist);
    
for i = 1 : N_Sp 
	Features(14,i) = sum(Rdist(i,:));       %
	Features(15,i) = sum(Gdist(i,:));
	Features(16,i) = sum(Bdist(i,:));
	Features(17,i) = sum(ldist(i,:));
	Features(18,i) = sum(adist(i,:));
	Features(19,i) = sum(bdist(i,:));
	Features(20,i) = sum(hdist(i,:));
	Features(21,i) = sum(sdist(i,:));
	Features(22,i) = sum(vdist(i,:));
        
	Features(23,i) = sum(Rdist(i,:) .* expdist(i,:));   % local contrast
	Features(24,i) = sum(Gdist(i,:) .* expdist(i,:));
	Features(25,i) = sum(Bdist(i,:) .* expdist(i,:));
	Features(26,i) = sum(ldist(i,:) .* expdist(i,:));
	Features(27,i) = sum(adist(i,:) .* expdist(i,:));
	Features(28,i) = sum(bdist(i,:) .* expdist(i,:));
	Features(29,i) = sum(hdist(i,:) .* expdist(i,:));
	Features(30,i) = sum(sdist(i,:) .* expdist(i,:));
	Features(31,i) = sum(vdist(i,:) .* expdist(i,:));
        
	Features(32,i) = sum(distances(i,:) .* expR(i,:));  % element distribution
	Features(33,i) = sum(distances(i,:) .* expG(i,:));
	Features(34,i) = sum(distances(i,:) .* expB(i,:));
	Features(35,i) = sum(distances(i,:) .* expl(i,:));
	Features(36,i) = sum(distances(i,:) .* expa(i,:));
	Features(37,i) = sum(distances(i,:) .* expb(i,:));
	Features(38,i) = sum(distances(i,:) .* exph(i,:));
	Features(39,i) = sum(distances(i,:) .* exps(i,:));
	Features(40,i) = sum(distances(i,:) .* expv(i,:));
end

Can it be vectorized for lower computation time?

[Rik]

Apart from the advice to change length to numel, I don't really see obvious ways to improve this code. Vectorized functions are generally only faster if there is a directly implemented method that works on an entire array. I don't recognize what this code is calculating to suggest a replacement with something like a convolutional filter.

The only thing you could try is to rework this code to allow parfor loops instead of for loops.

[Jinto]

@Rik Thanks to reply my question.

I'm trying to optimized this code (https://github.com/jhkim89/Saliency-HDCT). It is automatically detect saliency region and generate saliency map.

I just using original code for my project but it need too much time cost. So I want to optimized this code but hard to understand documentation, and apply it.

[Dyuman Joshi]

Instead of defining every single element

iml = mat2gray(imlab(:,:,1)); ima = mat2gray(imlab(:,:,2)); imb = mat2gray(imlab(:,:,3));
imR = im(:,:,1);  imG = im(:,:,2);   imB = im(:,:,3);  
imhsv = rgb2hsv(im);
imh = mat2gray(imhsv(:,:,1)); ims = mat2gray(imhsv(:,:,2)); imv = mat2gray(imhsv(:,:,3));

You can keep them in a cell array in the order required -

%perform operations on the whole data, rather than individually
imLAB=mat2gray(imlab);
imhsv = mat2gray(rgb2hsv(im));
%Order - R G B l a b h s v
IM={im(:,:,1); im(:,:,2); im(:,:,3); imLAB(:,:,1); imLAB(:,:,2); imLAB(:,:,3); ...
    imhsv(:,:,1); imhsv(:,:,2); imhsv(:,:,3)};

and then do operations accordingly, using indices, get the results and then assign the final values.

[Alvaro]

@Jinto A couple of questions:

1. What are your computer specs and what version of MATLAB are you running this on?
2. How big is your image dataset? Does it still take too long if you run this for 1 or 2 images?

I would also consider sending a short email to the corresponding author of that paper and see if they can give you some clues as to why the code is running so inefficiently. It is likely they've run into the same issue.