How to normalize/Rescale Image Datastore

Hi, I want to Normilze/Rescale the Dataset between 0 and 1

In Keras the Following Function is used to Normalize the Data between 0 and 1

train_image_generator = ImageDataGenerator(
    rescale=1./255,)
    
 %% in MATLAB i am using ImageDatastore 
 imageFolder = fullfile(downloadFolder,'Classdata');
 imds = imageDatastore(imageFolder, 'LabelSource', 'foldernames', 'IncludeSubfolders',true);
 

How can i Normalize the ImageDatastore in MATLAB?

[Image Analyst]

Do you mean you just want to read in images one at a time and rescale them to a range of 0-1 with mat2gray() or rescale()?

[hammad younas]

@Image Analyst mat2gray() convert image into grayscale. I want to rescale the whole imagedatastore between 0-1

Answer 1

[Image Analyst]

@hammad younas an imageDatastore is essentially just a structure that is a listing of file names. It is not image data itself. So you'd loop over the images and call rescale(thisImage, 0, 1) on each image, and then do something with it, like call image processing functions or something.

% Demo by Image Analyst
clc;    % Clear the command window.
close all;  % Close all figures (except those of imtool.)
clear;  % Erase all existing variables. Or clearvars if you want.
workspace;  % Make sure the workspace panel is showing.
format long g;
format compact;
fontSize = 20;
markerSize = 40;
folder = pwd;
imds = imageDatastore(folder, "FileExtensions", [".jpg",".tif", ".png"])
numImages = length(imds.Files)
%numImages = 4; % For debugging.
plotRows = ceil(sqrt(numImages))
for k = 1 : numImages %length(imds.Files)
    thisFile = imds.Files{k}
    [f, baseFileName, ext] = fileparts(thisFile);
    % Read in the image.
    thisImage = imread(thisFile);
    % Scale the min to 0 and the max to 1.
    thisImage = rescale(thisImage, 0, 1);
    % Now do something with the rescaled image.
    subplot(plotRows, plotRows, k);
    imshow(thisImage);
    title(baseFileName, 'FontSize', fontSize);
    text(10, 10, baseFileName, 'Color', 'r', ...
        'BackgroundColor', 'y', 'FontSize', fontSize, 'VerticalAlignment', 'top')
end
g = gcf;
g.WindowState = 'maximized'

[hammad younas]

@Image Analyst Thanks for you answer but How can i give this Loop Image for prediction of the trained network?

[Image Analyst]

I imagine you'd pass thisImage into a function like predict() or whatever it's called. Didn't your deep learning example show you both how to train the network, and how to apply it once it's been trained?

[hammad younas]

basically i am using keras trained model to MATLAB but my accuracy is different from keras network. The keras code apply Rescalling on Image Dataset and then predict on it.

Thats why i am doing rescalling in matlab to obtain the same result using the following code

imageFolder = fullfile(downloadFolder,'Classdata');
 imds = imageDatastore(imageFolder, 'LabelSource', 'foldernames', 'IncludeSubfolders',true);
augimdsTrain = augmentedImageDatastore(inputSize(1:2),imds);
[YPred,scores] = classify(netTransfer,augimdsTrain);

[Image Analyst]

I don't think that normalizes the image. I think if they need to be normalized you will have to normalize the images individually and save them. Then they can be passed in to classify().

[hammad younas]

@Image Analyst thanks for your reply i will check it. Can you please update the above code to save the rescale image with same name in the folder

[Image Analyst]

Just add

imwrite(thisImage, thisFile);

[hammad younas]

@Image Analyst It will save image in same folder?

[hammad younas]

@Image Analyst In keras we normalize image using

img = cv2.resize(cv2.imread(path),(150,150)) 
img_normalized = img/255

How can i do this in above code?

[Image Analyst]

% Read in the image.
thisImage = imread(thisFile);
% Resize to 150,150.
thisImage = imresize(thisImage, [150, 150]);
% Scale the max to 1, and everything else linearly scaled.
thisImage = thisImage / 255;
% Now do something with the resized and rescaled image.

Of course you could string that all together into one really long line of code (like your other code) if you want, though it might make it harder to read and follow.