https://news.ycombinator.com/item?id=6156238 (2013)

Interestingly it wasn't the OCR that was the problem but the JBIG2 compression.