Does image augmentation increase the number of training data?

My data availability is very less i.e only 745 images (340, 225,130,50 images in each category). I used augmentedImageDatastore function but the result is still poor. I would like to know whether image augmentation actually increases the number of data during training or it just transforms the existing images but the number of data remains the same. if its the latter case, how can I increase the number of data?

Answer 1

[Arun Mathamkode]

augmentedImageDatastore increases the training data by performing resizing, rotation, and reflection etc. Technically it is transforming the existing images but for the network it will be a different data and it prevents the network from overfitting and memorizing the exact details of the training images. But in your case, I can see the available images are very less and hence even with data argumentation I belive you are still running short of data.

In case you are not tried, you can also try transfer learning .

[adam R]

 The actual number of training images at each epoch does not change.

https://www.mathworks.com/help/nnet/ref/augmentedimagedatastore.html#mw_53d6e643-585d-4781-bcc6-1e8b98758e78