More efficient way of analyzing the Fourier transformation of an image?

Hello,

I'm trying to automatically analyze the direction of carbon fibres in a plastic part via thermography.

Currently, the process is as follows: Take Fourier transformation of the image, then get a line through the centre of the (shifted) Fourier transformed image, and sum up the amplitude values along that line.

The problem is, I've got hundreds of images, and they're split up into dozens of sub-images each. The Fourier analysis part takes hours. Is there any way to do this faster than my current code:

    phi1=double(rgb2gray(imread('peppers.png')));
    xsize=min(size(phi1));
    fftsize=2.^(nextpow2(xsize));
    phi1m=phi1-mean2(phi1);
    imfft=fftshift(fft2(phi1m,fftsize,fftsize)/fftsize);
    imfftnorm=sqrt(imfft.*conj(imfft))/max(max(abs(imfft)));
    xmp=size(imfft,1)/2;
    ymp=size(imfft,2)/2;
    radius=round(fftsize/3);
    imrot=imrotate(imfftnorm,-90);
    
    ampl=zeros(1800,1);
    for ang=1:1800
        xend=xmp-round(cosd(-ang/10)*radius);
        yend=ymp-round(sind(-ang/10)*radius);
        l1=improfile(imrot,[xmp xend],[ymp yend]);
        xend2=xmp+round(cosd(-ang/10)*radius);
        yend2=ymp+round(sind(-ang/10)*radius);
        l2=improfile(imrot,[xmp xend2],[ymp yend2]);
        ampl(ang)=nansum(l1(:))+nansum(l2(:));
    end

The Fourier transformation is rather fast, it's only the for loop and especially the improfile stuff that takes a comparatively long time.

[Image Analyst]

What do you want to know? The Computer Vision Toolbox lets you compute the HOG (histogram of oriented gradients).

[Korbi]

Thanks for your response.

I've got pictures like these (left):

and want to automatically identify the directions of the depicted fibres.

The picture on the right shows the corresponding Fourier transformation.

With the code above, I'd get a graph like this (not the same picture, but a similar one with slightly different orientations):

and could then use the peaks function to identify the directions of fibres.

From what I've seen, the HOG is more useful if you want to identify regions of the image, for which the lines aren't clear enough.

Answer 1

[Raunak Gupta]

Hi,

From the code I see that ang for loop doesn’t seems to have dependency between the iterations on the data values, so maybe you can use parfor to replace the for loop. Also make sure you are not plotting the output of improfile on a figure otherwise the code will become serial as it will require to plot each figure first and then move forward.

For better understanding about parfor you may refer the following.

Decide When to Use parfor

[Korbi]

Thanks a lot, that cut the loop time in half!