Question about function augmentedImageDatastore

I would like to ask a question about Matlab's function augmentedImageDatastore. If I understood correctly, it doesn't increase the (train) dataset by creating (even if it doesn't store them) additional (transformed) dataset samples, it just performs a bunch of mathematical operations at the samples of mini batch (with some predefined probability).Does that lead to the conclusion that at each epoch, the CNN is being trained at a slightly different dataset?

Answer 1

[Jayanti]

Hi Panos,

The “augmentedImageDatastore” in MATLAB does not create or store additional transformed samples as part of the dataset. It applies transformations such as resizing, rotation and reflection for each epoch, so that it uses a slightly different data set.

Because of this augmentation, the CNN sees a slightly different version of the dataset. This can help improve the model's generalization ability by exposing to a wider variety of scenarios, reducing overfitting, and making the model more robust.

You can also refer to the below official MathWorks documentation link on “augmentedImageDatastore” for more details:

https://www.mathworks.com/help/deeplearning/ref/augmentedimagedatastore.html