How to compute the fourier transform of multiple images?

I am trying to compute the fourier transform of multiple images from an image dataset in order to obtain its spectrum and process it further for image recognition.

But when I try to use the cell function to apply the fft, fft2 or fourier functions to the image files of the dataset, it shows an error saying: "Invalid data type. First argument must be double, single, int8, uint8, int16, uint16, int32, uint32, or logical." Due to this, I'm also trying to get the images in matrix form by applying the imread function in the entire cell, but it's taking a lot of time to compute it. Is the size of the dataset causing this problem? the dataset contains around 28000 images, and i'm dividing it into five batches containing 5.4k images each, and further dividing each batch into 80/20 ratio of training/testing images, and applying the cellfun on these training images(approximately 4.5k in each batch, done seperately). Below is the code:

clc;clear all;close all;

%get location
location = fullfile('G:\FACEDATA_train\train');

%datastore
imds = imageDatastore(location,'IncludeSubFolders',true,...
    'LabelSource','foldernames');
countEachLabel(imds);
rng(100);
imds = shuffle(imds);
[batch1, batch2, batch3, batch4, batch5] = splitEachLabel(imds, 0.2, 0.2, 0.2, 0.2, 0.2);

[imdsTrain_b1, imdsTest_b1] = splitEachLabel(batch1,0.8);
[imdsTrain_b2, imdsTest_b2] = splitEachLabel(batch2,0.8);
[imdsTrain_b3, imdsTest_b3] = splitEachLabel(batch3,0.8);
[imdsTrain_b4, imdsTest_b4] = splitEachLabel(batch4,0.8);
[imdsTrain_b5, imdsTest_b5] = splitEachLabel(batch5,0.8);

XTrain_b1 = imdsTrain_b1.Files;
XTrain_b2 = imdsTrain_b2.Files;
XTrain_b3 = imdsTrain_b3.Files;
XTrain_b4 = imdsTrain_b4.Files;
XTrain_b5 = imdsTrain_b5.Files;

XTest_b1 = imdsTest_b1.Files;
XTest_b2 = imdsTest_b2.Files;
XTest_b3 = imdsTest_b3.Files;
XTest_b4 = imdsTest_b4.Files;
XTest_b5 = imdsTest_b5.Files;


YTrain_b1 = imdsTrain_b1.Labels;
YTrain_b2 = imdsTrain_b2.Labels;
YTrain_b3 = imdsTrain_b3.Labels;
YTrain_b4 = imdsTrain_b4.Labels;
YTrain_b5 = imdsTrain_b5.Labels;

YTest_b1 = imdsTest_b1.Labels;
YTest_b1 = imdsTest_b1.Labels;
YTest_b1 = imdsTest_b1.Labels;
YTest_b1 = imdsTest_b1.Labels;
YTest_b1 = imdsTest_b1.Labels;

XTR_b1 = cellfun(@fft2, XTrain_b1);

[Walter Roberson]

XTest_b1 = imdsTest_b1.Files;

isn't that just the file names as a cell array?

You are applying fft2 to the file names, not to the content of the images

[Utsav Pandya]

Yes, @Walter Roberson, that's what I figured. But, upon applying the cellfun(@imread, XTrain_b1), it is taking a lot of time to generate the uint8 matrices of the images (48x48). So, is there any way that time can be minimized? thanks for answering.

[Adam Danz]

Consider again @Walter Roberson's question.

splitEachLabel produces ImageDatastore objects. ImageDatastore.Files returns strings, chars, cellstr, etc. If you're sending cellstrs or any type of string into fft2 you should be getting an error.

Here's a similar example using your variable names.

imds = imageDatastore(fullfile(matlabroot, 'toolbox', 'matlab', {'demos','imagesci'}),...
    'LabelSource', 'foldernames');
[batch1, batch2] = splitEachLabel(imds,0.6);
[imdsTrain_b1, imdsTest_b1] = splitEachLabel(batch1,0.8);
XTrain_b1 = imdsTrain_b1.Files

XTrain_b1 = 6×1 cell array
    {'/MATLAB/toolbox/matlab/demos/cloudCombined.jpg'}
    {'/MATLAB/toolbox/matlab/demos/example.hdf'      }
    {'/MATLAB/toolbox/matlab/demos/example.tif'      }
    {'/MATLAB/toolbox/matlab/demos/landOcean.jpg'    }
    {'/MATLAB/toolbox/matlab/imagesci/corn.tif'      }
    {'/MATLAB/toolbox/matlab/imagesci/grid.hdf'      }

XTR_b1 = cellfun(@fft2, XTrain_b1)

Error using fftn
Invalid data type. First argument must be double, single, int8, uint8, int16, uint16, int32, uint32, or logical.

Error in fft2 (line 18)
f = fftn(x);

Have you created your own fft2 function?

[Utsav Pandya]

@Adam Danz no, the fft2 function is available in MATLAB for 2-dimensional fourier transform. I am getting it that the imageDatastore function i'm using is returning just the names etc. of the files. What I want to do now is read those files using the cellfun in which I want to apply imread to each files of the image datastore. That is where i'm getting the problem. Its not throwing any error, but taking too long. Hence what I'm asking is if there's any alternate way of reading the image content of the imageDatastore files in a shorter period of time.

[Adam Danz]

>What I want to do now is read those files using the cellfun in which I want to apply imread to each files of the image datastore.

Although cellfun could be used to read in a series of files,I recommend using a loop. For example,

images = cell(1,numberOfFiles);
for i = 1:numberOfFiles
    images{i} = imread(file{i})
end

[Image Analyst]

Sorry but I still have to go with my Answer below (scroll down). Storing 28,000 images in a cell array would take up lots (perhaps too much) memory and I don't see any need for it. You can simply read in the images one at a time and take their FFT. If you then want to have deep learning process the FFT images instead of the original images then you can save the FFT images to disk files and pass a list of their filenames to the training function.

[Adam Danz]

+1 on that advice to analyze and toss the image data.

Answer 1

[Image Analyst]

I'd not do the datastore approach and just use a regular traditional for loop like in the FAQ:

https://matlab.fandom.com/wiki/FAQ#How_can_I_process_a_sequence_of_files?