Image patch Data Augmentation during training

In some research paper, authors are using image patches as augmentation during training and use whole image in testing phase( Input and Output to CNN is image).

  1. How image patch helps instead resize image as CNN input?
  2. Can we use this technique to all types of CNN architecture?
  3. Should it be fail in some specific cases?

I have recently read this paper: Patch-based CNN evaluation for bark classification, and the authors did a great comparison about classification using patches, with and without rescaling the original image. In short, the obtained results shows that using patches for classification (without rescaling) helps improve the accuracy, since we will not lose details of the object. But when classifying patches from a rescaled image we will lose in term of patch accuracy. I really suggest you to go through the paper, and you will understand more from the authors.

In general, this paper doesn’t reply to the exact question that you are asking but it could give you an idea about patch classification.