Creating searchable PDFs

Source: vignettes/articles/searchable_pdfs.Rmd

Somewhat surprisingly, the Google Document AI tool suite does not include a straightforward way to embed OCR output in PDFs. This functionality exists in many other OCR products – in Tesseract, for example, it’s as simple as running(tesseract input.jpg output pdf) – so I suspect Google just hasn’t prioritized it. Interestingly, the Python client library “Document AI Toolbox” includes a hOCR conversion script) but stops short of providing a method for creating searchable PDFs.

This said, writing code to reliably create searchable PDFs from custom OCR output is not straightforward, especially outside of Python. The main challenge is embedding text in PDFs in the right places, handling font sizes and encoding issues along the way. As far as I am aware, the only open-source tool that can do it well is the Reportlab PDF Toolkit (a Python library). But the Reportlab toolkit itself does not include a direct solution for converting OCR output formats such as HOCR and ALTO to searchable PDFs, so it is left to small community efforts such as hocr-tools to implement it (by adapting Reportlab’s code). Meanwhile, in R, there are no packages that can embed text in PDFs, let alone create searchable PDFs from custom OCR output.

This is why daiR does not provide a native R solution for embedding OCR-extracted text in PDFs. However, as of version 0.9.9, daiR includes the function make_hocr, which creates HOCR files from Document AI output, and this takes us some of the distance. This article will describe a simple pipeline which combines dair::make_hocr and the command-line tool hocr-pdf from the aforementioned hocr-tools project.

Before we begin, you may want to install hocr-tools on your computer. Instructions are here. Note that the package version on PyPI currently lags behind the Github codebase, and I had to clone the repository and build from source to get the hocr-pdf function to work. This situation will hopefully change in due course.

Note that hocr-pdf only handles individual .jpg images, and it works best if the hocr file is generated from an OCR process involving the same .jpg file. This means you cannot process a multi-page pdf in Document AI and take the output to hocr-pdf; you need to convert the pdf to a set of .jpg files first and then feed them individually to Document AI, so that you get hocr files that match the .jpg files. Other than this quirk, the procedure works well.

In this illustration, we will use a single, “ready made” .jpg. Conversion from PDF to .jpg can be done quite easily with packages like pdftools or magick.

Let us start by downloading the image into a temporary subdirectory of our working directory.

dir.create("temp")
download.file("https://dair.info/articles/sample.jpg", "temp/sample.jpg")

Then we process the image synchronously with Document AI.

library(daiR)
resp <- dai_sync("temp/sample.jpg")

Then we generate the HOCR file:

make_hocr(type = "sync", output = resp, dir = "temp", outfile_name = "sample.hocr")

Note that the two files sample.jpg and sample.hocr carry the same name (sample) and reside in the same directory (temp). Both conditions are necessary for hocr-pdf to work.

If hocr-tools is installed, we can now run the hocr-pdf utility, either in the terminal or from within R using the system() command.

In the terminal:

hocr-pdf temp > temp/sample.pdf

In R:

system("hocr-pdf temp > temp/sample.pdf")

And you should now have a searchable PDF in the temp directory.

Note that you can also open the .hocr file in a web browser and read the extracted text there. Just right-click on it in the file explorer, click “open with”, and select your browser. This works for .hocr files generated from documents in all the languages that DAI covers.

Note on encoding

make_hocr handles non-Latin scripts and right-to-left languages just fine, but unfortunately hocr-pdf cannot as yet embed non-Latin script in pdfs. This means there is currently no way of creating searchable pdfs from documents with such scripts OCRed in Document AI – at least to my knowledge. To generate searchable pdfs from such documents, you can use Tesseract, but then you get the output from a different OCR engine.

Languages tested:

Cleanup

To remove the directory with the example files:

unlink("temp", recursive = TRUE)