Image Normalization before Fine-Tuning a pretrained CNN for image classification

Hello,

Is it possible to directly add an image normalization step, to this training code below, to normalize all the dataset images before training the CNN pretrained model ? I need to train my model with pixel values ranging between 0 and 1 instead of 0 and 255.

imds = imageDatastore(dataset, 'IncludeSubfolders',true,'LabelSource','foldernames')
tbl = countEachLabel(imds);
numClasses = height(tbl); 
[trainingSet, testSet] = splitEachLabel(imds, 0.7,'randomize');

I tried to modify the image input layer (Normalization 'rescale-zero-one') of the model but it did not work because this option does not exist effectively ( previous question asked related: https://fr.mathworks.com/matlabcentral/answers/1441834-imageinputlayer-normalization-data-normalization-options?s_tid=srchtitle )

Is there any way to normalize directly images in augmentedImageDatastore ?

    augmentedTrainingSet = augmentedImageDatastore(imageSize, ...
        trainingSet, 'ColorPreprocessing', 'gray2rgb'); 
    augmentedTestSet = augmentedImageDatastore(imageSize, ...
        testSet, 'ColorPreprocessing', 'gray2rgb');

Thank you in advance !! Appreciate any kind of help !

Answer 1

[yanqi liu]

sir, may be you shoud use function handle to define your read image style, pleaes read the follow code

clc; clear all; close all;

dataset = fullfile(matlabroot,'toolbox','matlab');

imds = imageDatastore(dataset,'IncludeSubfolders',true,...

'FileExtensions','.tif',...

'LabelSource','foldernames',....

'ReadFcn',@data_preporcess);

tbl = countEachLabel(imds);

numClasses = height(tbl);

[trainingSet, testSet] = splitEachLabel(imds, 0.7,'randomize');

function data = data_preporcess(file)

data = imread(file);

% ranging between 0 and 1 instead of 0 and 255

data = mat2gray(data);

end

you can read the book 《计算机视觉与深度学习实战》

[Andrea Daou]

Thank you for your help.

Answer 2

[Image Analyst]

Sure. Use mat2gray() or rescale() or im2double().

[Andrea Daou]

Thank you for your help.