Take a PDF, OCR it, and add OCR Text as background layer to original PDF to make it searchable

ocrpdf.sh

#!/bin/bash

# NOTE: I recommend pdfsandwich instead of this script, partly because imagemagick (and pdftoppm) fail on large detailed images.
# While that technique does not preserve the original graphics, it can come close.
# To preserve color:
# pdfsandwich -rgb input.pdf
# To preserve grey tones:
# pdfsandwich -gray input.pdf
# To disable all preprocessing:
# pdfsandwich -nopreproc input.pdf

set -m # turn on job control for parallel processes

# Source:
# https://gist.github.com/wcaleb/7337097
# https://gist.github.com/jburon/d31e0132dfb291dc804bac019f9d9023
#
# Changes:
# - Don't delete files with wildcards. Always use a (random) prefix.
# - Fix extensions of generated files.
# - Don't use greyscale because it's not compatible with some versions of tesseract.
# - Clean up all generated files afterwards.
# - Keep the hocr2pdf command from another fork, but comment it out because it failed in my tests.
# - Process pages in parallel instead of using multithreading in tesseract (which is less efficient).
#   Override the job parallelism by setting THREAD_COUNT.

# Take a PDF, OCR it, and add OCR Text as background layer to original PDF to make it searchable.
# Hacked together using tips from these websites:
#      http://www.jlaundry.com/2012/ocr-a-scanned-pdf-with-tesseract/
#      http://askubuntu.com/questions/27097/how-to-print-a-regular-file-to-pdf-from-command-line
# Dependencies: pdftk, tesseract, imagemagick, enscript or hocr2pdf/pdfsandwich


function process_page() {
	local FILE=$1
	echo "Processing $FILE"
	local PAGE=$(basename "$FILE" .pdf)

	# Convert the PDF page into a TIFF file
	local IMG=$PAGE.tif
	convert -density 600 "$FILE" "$IMG"

	# OCR the TIFF file and save text to output.txt or output.hocr

	OMP_THREAD_LIMIT=1 tesseract "$IMG" "${PAGE}_output"
	# Turn text file outputed by tesseract into a PDF, then put it in background of original page
	#enscript output.txt -B -o - | ps2pdf - output.pdf && pdftk $FILE background output.pdf output new-"$FILE"
	enscript "${PAGE}_output.txt" -B -o - | ps2pdf - "${PAGE}_output.pdf" && pdftk "$FILE" background "${PAGE}_output.pdf" output "new-$FILE"

	#tesseract "$IMG" "${PAGE}_output" hocr
	## Turn html outputed by tesseract into a PDF, combined with the original image as foreground
	#hocr2pdf -i "$IMG" -o "new-${FILE}" < "${PAGE}_output.hocr"

	# Clean up
	rm "$PAGE"*
}

function wait_jobs() {
	while [[ $(jobs -r | wc -l) -gt $((${THREAD_COUNT:-$(nproc)} - 1)) ]]; do
		sleep 0.25
	done
}

if [[ $# -eq 0 || ! -e $1 ]]
then
	echo "Adds an OCR text layer to a PDF file to make searching easier."
	echo "Usage: $0 <pdf file>"
	exit
fi

TEMPNAME=$(mktemp -p . -u)
TEMPNAME=${TEMPNAME/.\//} # remove "./"
[[ -e $TEMPNAME ]] && echo "Could not create temp filenames" && exit

cp $1 $1.bak
pdftk $1 burst output "${TEMPNAME}_tesspage_%05d.pdf"
for FILE in ${TEMPNAME}_tesspage*
do
	process_page "$FILE" &
	wait_jobs
done
wait

pdftk "new-${TEMPNAME}"* cat output $1

# Clean up
rm doc_data.txt "new-${TEMPNAME}"*

Don't delete files with wildcards. Always use a (random) prefix.
Fix extensions of generated files.
Don't use greyscale because it's not compatible with some versions of tesseract.
Clean up all generated files afterwards.
Keep the hocr2pdf command from another fork, but comment it out because it failed in my tests.

Process pages in parallel instead of using multithreading in tesseract (which is less efficient). Override the job parallelism by setting THREAD_COUNT.