Extracting Text from a PDF file with embedded font

Question

I have a PDF file containing some tabular data.

http://dl.dropbox.com/u/44235928/sample_rotate-0.pdf

I have to extract the tabular data from it. I have tried following with no success :

1. Select the text and paste it to notepad/excel-sheet. (I am getting junk characters)
2. Used save as text from Acrobat Reader. It is also giving junk characters and not the actual text.
3. Tried ApachePDFBox command line utility to extract text from PDF. It is also giving junk characters instead of real texts.
4. Finally I am trying a OCR solution. I am converting the pdf file into .tif images using ImageMagick and getting those images processed by tesseract OCR.

The OCR solution is not very accurate though( about 80% words matched ).

I tried changing density and geometry of the image created from PDF to get better results from tesseract OCR.

convert -rotate 90 -geometry 10000 -depth 8 -density 800 sample.pdf img_800_10000.tif;
tesseract img_800_10000.tif img_800_10000.tif nobatch letters;

I am not sure for what kind of image( density, geometry, monochromatic, sharpen boundary etc) would be best suited for the OCR.

Please suggest what could be the best possible parameters(density,geometry,depth etc) for generating images from a PDF file, so that the tesseract accuracy will increase.

I am open to other( non-ocr ) solutions as well.

Answer 1

Since a comment asked "How to define the geometry of an image when using Ghostscript as we do in convert?", here is an answer:

1. It does not make sense to define geometry (that is image dimensions) and resolution for a raster image created by Ghostscript at the same time.

2. Once you convert a vector based page of a given dimension (such as PDF) into a raster image (such as the TIFF G4 format) giving a desired resolution (as done in the other answer), you already indirectly and implicitly also did set the dimension:

   • The original PDF dimension of your sample file sample_rotate-0.pdf is 1008x612 points.
   • At a resolution of 72 DPI (the default Ghostscript uses if not given directly, or -r72 in the Ghostscript command if given directly) the image dimensions will be 1008x612 pixels.
   • At a resolution of 720 DPI (-r720 in the Ghostscript command) the image dimensions will be 10080x6120 pixels.
   • At a resolution of 1440 DPI (-r1440 in the Ghostscript command of my other answer) the image dimensions will be 20160x12240 pixels.
   • At a resolution of 1200 DPI (-r1200 in the Ghostscript command) the image dimensions will be 16800x10200 pixels.
   • At resolution of 1000 DPI (-r1000 in the Ghostscript command) the image dimensions will be 14000x8500 pixels.
   • At a resolution of 120 DPI (-r120 in the Ghostscript command) the image dimensions will be 1680x1020 pixels.
   • At resolution of 100 DPI (-r100 in the Ghostscript command) the image dimensions will be 1400x850 pixels.

3. If you absolutely insist to specify the dimension/geometry for the output image on the Ghostscript commandline (rather than the resolution), you can do so by adding -gNNNNxMMMM -dPDFFitPage to the commandline.

Answer 2

In this case I recommend to NOT use ImageMagick for the PDF -> TIFF conversion. Instead, use Ghostscript. Two reasons:

1. Using Ghostscript directly will give you more control over individual parameters of the conversion.

2. ImageMagick cannot do that particular conversion itself -- it will call Ghostscript as its 'delegate' anyway, but will not allow you to give all the same fine-grained control that your own Ghostscript command will give you.

Most of the text in the table of your sample PDF is extremely small (I guess, only 4 or 5 pt high). This makes it rather difficult to run a successful OCR unless you increase the resolution considerably.

Ghostscript uses -r72 by default for image format output (such as TIFF). Tesseract works best with r=300 or r=400 -- but only for a font size from 10-12 pt or higher. Therefor, to compensate for the small text size you should make Ghostscript using a resolution of at least 1200 DPI when it renders the PDF to the image.

Also, you'll have to rotate the image so the text displays in the normal reading direction (not bottom -> top).

This is the command which I would try first:

gs                              \
  -o sample.tif                 \
  -sDEVICE=tiffg4               \
  -r1200                        \
  -dAutoRotatePages=/PageByPage \
   sample_rotate-0.pdf

You may need to play with variations of the -r1200 parameter (higher or lower) for best results.

Answer 3

There you can find decoded content of your file: https://docs.google.com/open?id=0B1YEM-11PerqSHpnb1RQcnJ4cFk

A absolutely sure the OCR is the best way to read pdf file, but you can try REGEX-ing the native content. It going to be be the hard and long way.