Neural networks used for image recognition not only are difficult to set up and train, they also require a lot of data to train. Also, they tend to overfit on the images used during training. For example, when you only use photos of faces in an upright position, your model will have a hard time recognizing faces that are rotated in another direction.
To help overcome problems with rotation and shifts in certain directions, you can use image augmentation. CNTK supports specific transforms when creating a minibatch source for images.
We've included an additional notebook for this chapter that demonstrates how to use the transformations. You can find the sample code for this section in the Recognizing hand-written digits with augmented data.ipynb file in the samples for this chapter.
There are several transformations that you...
