Is it possible to loop through and set image pixels with a parfor loop?

This code transforms a 360 degree image, but it is it slow with images at high resolutions. I was wondering if it was possible to convert one of the loops to a parfor loop to cut down on the time it takes to run the transformations.

Below is the code:

% loop through every pixel of am image
for row = 1 : width
    % find pixel theta (elevation)
    theta = (90/width) * (row + (row-1));

    % transform theta into z cartesian coordinates
    z = cosd(theta);

    for col = 1 : 1: length

        % find pixel psi (azimuth)
        psi = (180/length) * (col + (col-1)) - 180;
        
        % transform theta and psi into x and y cartesian coordinates
        x = sind(theta) * cosd(psi);
        y = sind(theta) * sind(psi);

        % roll transformation
        yNew = y * cosd(roll) - z * sind(roll);
        zNew = y * sind(roll) + z * cosd(roll);

        % pitch transformation
        xNew = x * cosd(-pitch) + zNew * sind(-pitch);
        zNew = zNew * cosd(-pitch) - x * sind(-pitch);
        
        % yaw transformation
        xYaw = xNew;
        xNew = xYaw * cosd(yaw) - yNew * sind(yaw);
        yNew = xYaw * sind(yaw) + yNew * cosd(yaw);
        
        % transform x, y, and z back to spherical coordinates
        psiNew = atan2d(yNew, xNew); 
        thetaNew = acosd(zNew);
        
        % find new pixel row from theta
        rowNew = round((thetaNew*width)/180 + 0.5);
        % find new pixel column from psi
        colNew = round((psiNew*length)/360 + length/2 + 0.5);
        
        % map pixels from the original image to a new one
        try
            transformedImage(row, col, :) = app.newImage(rowNew, colNew, :); % THIS WONT WORK W/ A PARFOR LOOP
        catch
            disp("Psi/theta out of bounds:");
            disp(psiNew + ", " + thetaNew);
            disp(rowNew + ", " + colNew);
        end
    end
end

[Constantino Carlos Reyes-Aldasoro]

In general it is best not to use loops through matrices, that is the whole power of Matlab that considers everything a matrix. I.e., you can run through every pixel of an image and compare to a certain value:

myImage = rand(32,32);
for k1=1:32
    for k2=1:32
        myImage2(k1,k2) = myImage(k1,k2)>0.5;
    end
end
imagesc(myImage2)

Or you can simply work as matrices:

myImage3 = (myImage>0.5);
imagesc(myImage3)

Thus, try to avoid loops and work with your data as a matrix.

Answer 1

[Walter Roberson]

            transformedImage(row, col, :) = app.newImage(rowNew, colNew, :); % THIS WONT WORK W/ A PARFOR LOOP

Change that to something like

            transformedrow(col, :) = newImage(rowNew, colNew, :);

and after the end of the col loop

            transformedImage(row, :, :) = transformedrow;

And before the parfor

newImage = app.newImage;

When you have a variable indexed by the parfor loop variable, all other indices of the variable should be constants or arithmetic expressions that evaluate to things that are constant for the entire parfor.

[Bao]

I added your changes, and it seems to make the code run faster without a parfor loop. When I make the inner loop a parfor loop though, the code runs very slowly, which makes a 2 minute operation seem like an infinite one.

[Walter Roberson]

transformedImage(row, :, :) = transformedrow;

If you can rewrite the code to work down columns instead of across rows, then the performance would be improved.

Answer 2

[Matt J]

If you have the Image Processing Toolbox, you should use imwarp to perform an image transformation.