Methods and systems for training a neural network to distinguish between text documents and image documents are described. A corpus of text and image documents is obtained. A page of a text document is scanned by shifting a text window to a plurality of locations. In accordance with a determination that the text in the window at a respective location meets text line criteria, the text in the window is stored as a respective text snippet. A plurality of image windows are superimposed over at least one page of an image document. In accordance with a determination that the content of a respective image window meets image criteria, content of the image window is stored as a respective image snippet. The respective text snippet and the respective image snippet are provided to a classifier.