I've been experimenting with using Tesseract to OCR my PDFs, and it has been mostly successful, particularly with German Fraktur texts (the old style gothic print), which tools like Adobe Acrobat can't recognize properly.
The problem is that the output files from Tesseract are rather large, and I want to compress them after OCRing. However, when I use Ghostscript to compress the files, the embedded OCR text he's messed up. Similarly, if I use ImageMagick, the embedded text is removed. Is there a way around this? Theoretically I could compress before OCRing but that would make the OCR accuracy worse.
Generally speaking, my goal is to have high-quality OCR embedded text in my output PDF files, and have the embedded images be highly compressed so that the files don't take up nearly as much space. I have found that the Adobe Acrobat Pro feature "Save as Other > Reduced Size PDF" highly compresses the images but screws up any OCR'd text. This is true whether the files were OCR'd in Acrobat, or using a tool like Tesseract.
Here's my current workflow, using a sample pdf.
Split PDF into TIFF files
pdftk infile.pdf burst output "temp/page_%03d.pdf"
dpi=130 #this is the dpi of the particular file
parallel convert -verbose -density $dpi "{}" -depth 8 -background white -compress zip "{}.tiff" ::: temp/*.pdf
Run Tesseract on each of the TIFF files (see sample file's output)
language=deu_frak
parallel tesseract {} {} -l $language pdf ::: temp/*.tiff
- When I combine the output PDF files with Ghostscript, I get a file like this one, which screws up the embedded text
- When I combine them with PDFtk (e.g. pdftk temp/*.pdf cat output outfile.pdf`), I get a file like this one, which maintains the embedded text but somehow makes the file larger
- And then when I try to compress that file using ImageMagic (e.g.
convert -density 130x130 -quality 5 -compress jpeg outfile-pdftk.pdf outfile-pdftk-imagemagick.pdf) it removes the embedded OCR text (output)
It seems that Tesseract doesn't compress the images in the output PDF, which is to be expected - its job is to OCR the files, not compress the output.
For instance, on the initial Tesseract OCR'd files, pdfimages -list temp/page_001.pdf.tiff.pdf produces:
page num type width height color comp bpc enc interp object ID x-ppi y-ppi size ratio
--------------------------------------------------------------------------------------------
1 0 image 1067 1508 rgb 3 8 jpeg no 11 0 129 130 318K 6.7%
... which indicates that the image object in the PDF isn't exactly stored optimally. It is still in RGB, not black & white. Upon compressing with ImageMagick, by contrast, pdfimages -list gives:
pdfimages -list outfile-pdftk-imagemagick.pdf
page num type width height color comp bpc enc interp object ID x-ppi y-ppi size ratio
--------------------------------------------------------------------------------------------
1 0 image 1075 1520 gray 1 8 jpeg no 8 0 130 131 54.0K 3.4%
1 1 smask 1075 1520 gray 1 8 image no 8 0 130 131 25.1K 1.6%
2 2 image 1075 1520 gray 1 8 jpeg no 22 0 130 131 59.9K 3.8%
2 3 smask 1075 1520 gray 1 8 image no 22 0 130 131 25.1K 1.6%
3 4 image 1075 1520 gray 1 8 jpeg no 36 0 130 131 45.2K 2.8%
3 5 smask 1075 1520 gray 1 8 image no 36 0 130 131 25.1K 1.6%
4 6 image 1075 1520 gray 1 8 jpeg no 50 0 130 131 62.8K 3.9%
4 7 smask 1075 1520 gray 1 8 image no 50 0 130 131 25.1K 1.6%
5 8 image 1075 1520 gray 1 8 jpeg no 64 0 130 131 61.1K 3.8%
5 9 smask 1075 1520 gray 1 8 image no 64 0 130 131 25.1K 1.6%
6 10 image 1075 1520 gray 1 8 jpeg no 78 0 130 131 63.4K 4.0%
6 11 smask 1075 1520 gray 1 8 image no 78 0 130 131 25.1K 1.6%
7 12 image 1075 1520 gray 1 8 jpeg no 92 0 130 131 65.1K 4.1%
7 13 smask 1075 1520 gray 1 8 image no 92 0 130 131 25.1K 1.6%
8 14 image 1075 1520 gray 1 8 jpeg no 106 0 130 131 61.0K 3.8%
8 15 smask 1075 1520 gray 1 8 image no 106 0 130 131 25.1K 1.6%
9 16 image 1075 1520 gray 1 8 jpeg no 120 0 130 131 66.8K 4.2%
9 17 smask 1075 1520 gray 1 8 image no 120 0 130 131 25.1K 1.6%
10 18 image 1075 1520 gray 1 8 jpeg no 134 0 130 131 65.6K 4.1%
10 19 smask 1075 1520 gray 1 8 image no 134 0 130 131 25.1K 1.6%
As we can see the images take up less space, however the OCR-embedded text was removed and, somehow, the file is less. By comparison, if I take the original file (without OCR-embedded text) and compress it using Adobe Acrobat's "Save As Other > Reduced Size PDF", I get:
pdfimages -list infile-adobe.pdf
page num type width height color comp bpc enc interp object ID x-ppi y-ppi size ratio
--------------------------------------------------------------------------------------------
1 0 image 1000 1499 gray 1 8 jpx no 38 0 129 129 78.1K 5.3%
2 1 image 1000 1499 gray 1 8 jpx no 3 0 129 129 89.1K 6.1%
3 2 image 1000 1499 gray 1 8 jpx no 6 0 129 129 65.6K 4.5%
4 3 image 1000 1499 gray 1 8 jpx no 9 0 129 129 97.7K 6.7%
5 4 image 1000 1499 gray 1 8 jpx no 12 0 129 129 95.4K 6.5%
6 5 image 1000 1499 gray 1 8 jpx no 15 0 129 129 98.7K 6.7%
7 6 image 1000 1499 gray 1 8 jpx no 18 0 129 129 102K 6.9%
8 7 image 1000 1499 gray 1 8 jpx no 21 0 129 129 94.6K 6.5%
9 8 image 1000 1499 gray 1 8 jpx no 24 0 129 129 105K 7.2%
10 9 image 1000 1499 gray 1 8 jpx no 27 0 129 129 103K 7.1%
... As we can see, Adobe Acrobat seems to compress images using JPEG2000 (JPX) which isn't available to Ghostscript or ImageMagick due to patent issues.
On the whole, any suggestions on how to compress Tesseract-OCR'd PDF files?
