I have a selection of pdfs that I want to text mine. I use tika to parse the text out of each pdf and save to a .txt with utf-8 encoding (I'm using windows)
Most of the pdfs were OCR'd before I got them but when I view the extracted text I have "pnÁnn¿¡c" instead of "Phádraig" if I view the PDF.
Is it possible for me to verify the text layer of the PDF (forgive me if thats the incorrect term) Ideally without needing the full version of Acrobat
It sounds like you are dealing with scanned books with "hidden OCR", ie. the PDF shows an image of the original document, behind which there is a layer of OCRed text.
That allows you to use the search function and to copy-paste text out of the document.
When you highlight the text, the hidden characters become visible (though this behaviour maybe depends on the viewer you use).
To be sure, you can copy-paste the highlighted text to a text editor.
This will allow you to tell if you are actually dealing with OCR quality this terrible, or if your extraction process caused mojibake.
Since OCR quality heavily depends on language resources (dictionaries, language model), I wouldn't be surprised if the output was actually that bad for a low-resource language like Gaelic (Old Irish?).
Related
I've noticed that when I use an OCR to transform a scanned PDF document into text, in this case Adobe Acrobat Pro, I'm getting very different outputs depending on how I extract the data.
In the above photo - you can see a piece of a PDF that has been OCR'ed into fairly good quality text. If I select it in Adobe and copy it to say, a word or txt doc, it paste over perfectly fine.
However, if I export it using Adobe to Rich Text Format, use Python's PDFminer, or Python Apache Tika then I get the above photo which as you can see completely jumbles it. The extraction results are very consistent between the approaches - basically all 3 jumble it in the exact same way.
Would any of you have any idea as to why an OCR'd PDF can be copied just fine to a text editor but is extracting in such a bizarre way?
Thank you!
Regards,
Mano
So what ended up working for me was running the initial parsing with Apache-Tika and then, on the few that didn't work on, pass them through PyPDF2. My theory is that PyPDF2 uses a different mechanism for parsing that doesn't rely on the root of the PDF unlike Tika and that is what seems to have become corrupted in a few of these OCR'd docs.
Not sure of the initial cause but that was my solution.
I'm curious how does PDF securing work? I can lock PDF file so system can't recognize text and manipulate with PDF file. Everything I found was about "how to lock/unlock" however nothing about "how does it work". Is there anyone who could explain it to me? Thx
The OP clarified in a comment
I mean lock on text recognition or manipulation with PDF file. There should be nothing about cryptography imho just some trick.
There are some options, among them:
You can render the text as a bitmap and include that bitmap in the PDF
-> no text information.
Or you can embed the font in question using a non-standard encoding without using standard glyph names
-> text information in an unknown encoding.
E.g. cf. the PDF analysed in this answer.
A special case: make the encoding wrong only for a few characters, maybe just one, probably a digit. This way an unalert person thinks everything was extracted ok, and only when the data is to be used, the errors start screwing things up, something which especially in case of wrong digits is hard to fix. E.g. cf. the PDF analysed in this answer.
Or you can put text in structures where text extraction software or copy&paste routines usually don't look, like creating a large pattern tile containing the text for some text area and filling the area with the matching pattern color.
-> text information present but not seen by most extractors.
E.g. cf. this answer; the technique here is used to make the text of a watermark non-extractable.
Or you can put extra text all over the page but make it invisible, e.g. under images, drawn in rendering mode 3 (invisible), located in some disabled optional content group (layer), ... Text extractors often do not check whether the text they extract actually is visible.
-> text information present but polluted by garbage text bits.
...
I've used a linux function to convert a list of PDF files to text.
Command:
pdftotext -htmlmeta
This work well for most of my files.
but for a small amount of them, this return me a blank text file.
My unsuccesssfull pdf files were not encrypted, not securised by user / password and they were not read only.
Converting PDFs to text is not a well-defined process. It can work awesome or not at all, depending on the PDF input.
Why is this? Because a PDF's task is mainly to represent the optics of a document, not the textual contents. PDFs can be everything from a pure text with positional information up to a pure graphics of the glyphs of the letters of the text. In the latter case one would need to run an OCR on the input in order to receive text information. This is not done by tools like pdftotext.
Sometimes the text in the PDF is scattered throughout the file, e. g. because first all standard-font letters are mentioned in the PDF, then, later in the file, all the italics-font letters are mentioned (of course with positional information, so a reader of the optical representation won't notice this, even if standard and italics are mixed throughout the text on the page). To rearrange this mess to a fluent text is a major task not very many converters are capable of.
So I guess all you can do is try some more converters for PDF to text (some are better than others, and some are better just for some specific input) or see that you can get the text from another source than the PDF files.
My problem:
I have a pdf with lots of roman characters with complex diacritical marks (e.g., ṣ, ś, ṝ, ǎ, etc.). To make it easier to search within the pdf, I would like to add an additional layer, much as one does with hocr, where the same text is present without the diacritics.
When using full-text search engines I can index multiple terms at the same position (vector) - I would like to achieve the same effect here.
I have read lots about adding a hocr layer to scanned images, but I really just want to duplicate the text layer, pass it through a script that strips the diacritics (straightforward enough) and then adds it back in as a hidden but searchable layer.
Anyone have any suggestions? (Solutions involving any platform, language, library or toolchain will be useful!)
Thanks :)
Edit: please let me know if the question is unclear.
Well I have a (slightly ugly and hackish) solution, so I thought I'd share it.
I'm using PDFMiner to extract the text, along with the co-ordinates. Then I'm using ReportLab to write the normalized versions of the text to a new pdf, in exactly the same position, as hidden text. To make the positions line up properly, I found I had to use exactly the same font, so I've used a combination of FontForge and MuPDF to extract the required font(s) from the original pdf.
Finally, having created the new pdf, I'm using pdftk to merge it with the original.
It works pretty well, but has the downside that copying text out of the pdf results in the normalized text being copied too. But this is acceptable for my present purposes, and I can't see any way around it. The pdf spec. doesn't really support my objective, and so I don't imagine I can do better than this hackish solution.
I have written something similar to add searchable text by OCR'ing images and converting it to PDF in C#. I used QuickPDF from www.quickpdf.com to create hidden white text objects on top of the image and this worked reasonably well.
In your case QuickPDF would allow you to extract the text strings along with bounding boxes and font details. You could then normalize your text and create the invisible text objects using the existing font and position information and then save it out to a new file.
This would basically give you the same PDF as you have now and also give you both the original and normalised text as you are getting now.
QuickPDF is a commercial library. If your solution works well for you then there is no used buying a commercial engine though. The nice thing though is that it only requires 1 SDK and you would look at it if you had a more than a few PDF's to convert.
I'm trying to convert this delimited PDF to an excel (or some other delimited format). Using Adobe Acrobat 9, I attempt to save it and copy it) as Excel but it gives the error message "BAD PDF; error in processing fonts. [348]".
I'm open to any solution that will create a delimited file, ranging from using Adobe Acrobat, to programming to using other apps. The only limitation is that I don't have a budget to buy anything (such as Able2Extract).
The way I was able to export my images and fonts without buying any extra software to do the conversion was this way. go to Advanced, PDG Optimizer, select all of the options you want on the LEFT COLUMN and where it says MAKE COMPATIBLE WITH select Acrobat 8.0 and later, OK....you are in your road to success
Note: not really an answer, but some suggestions.
Sounds to me that Crystal Reports is not following the PDF spec close enough.
I'd make sure CR is fully updated/patched and try genning another file making sure that "tagging" is enabled - tagging defines the layout structure. I don't have a copy of CR handy, but you may have to define a distiller template to use so when you print to PDF you can select that job option.
You can also tell its a bad PDF by using Preflight in Acrobat, it says there is no tag structure and you can do it manually (draw boxes around each item...). Also that there is no language set, and it is somehow compatible with Acrobat 1.3? which isn't supported anymore and should be 4 at the lowest?
Once you have a "good" pdf can export to xml/word and import that to excel. Also, with Acrobat 8+ you can highlight using the select tool, right-click and choose Open As SpreadSheet. You might be able to get away with just highlighting the whole document -- though I'd hope the xml approach would be best.
Able2Extract does some OCRing and tricky fuzzy logic not only to define tags/layout so it is exportable, but also avoids any font, encoding, etc issue - at least to my knowledge.
In the rare case that you can't get a new file, then exporting to plain text/accessible seems to generate a nice flat text file. You could write a vbscript to parse it (adding your delimiter) and import that into excel.