How to implement a datastore for labeled image sequences?

I am fairly new to the datastore format, having only used imagedatastore before. However, now my dataset is a large set of 16x8x800 image sequences each with a single categorical label.

I am using this for deep learning, so I would like to have a datastore that is setup so that each {16x8x800} sequence is paired with a single {[label]}, both of which are formatted so that they can be imported directly by the trainNetwork function.

I was thinking I should just use a filedatastore and save a cell array containing the many image series in the first column and their corresponding label in the second column, like below.

Data{1} = {16x8x800 double, 1x1 categorical}

Data{2} = {16x8x800 double, 1x1 categorical}

.

.

.

Data{n} = {16x8x800 double, 1x1 categorical}

Would this be usable by the trainNetwork function? Is there a better way to perform this?

Answer 1

[Prateek Rai]

To my understanding, you want to create a datastore from dataset which is a large set of 16x8x800 image sequences each with a single categorical label.

You can keep all your files in one folder and use imageDatastore to create a datastore. Set 'LabelSource' Name-value pair to "none" to keep the "Label" property empty initially.

% location specifies location of the folder containing all images
imds = imageDatastore(location,'LabelSource','none'); 

Now, create a categorical array of size (n,1) where ith row indicates label of ith image.

Set this categorical array to "Labels" property of imds.

imds.Labels = x; % here, x is the categorical array of label

You can refer to imageDatastore MathWorks Documentation page to learn more on datastore for image data.

[Thomas Hyatt]

Thanks for this response, but what save format does one use on a 16x8x800x1 image series? (I forgot to include these images were grayscale) Or rather, a 3D image input. Can I save each as just a .mat file?