Train OCRNet to read text in cropped images.
Recent revisions are kept. Restoring one creates a new revision.
