Image Augmentation

For an image classification problem, we are only concerned if the object is present in the image. We do not care if it's rotated (rotation invariance), flipped, to the right or left (translation invariance), We basically want our algorithm to learn an invariant representation of the image. We also do not want our model changing its prediction based on the size of the object (scale invariance).

So how do we make our algorithm good with invariant representation of images?

We just need to feed the training network with rotated/flipped/translated/scaled images.