Trying OCR with GCloud Document AI
- 22 Dec 2023
- Manuel Capel
- Tags: deep learning cloud
OCR stands for “Optical Character Recognition”, and is a powerful technique for extracting texts (and possibly also their position, fonts etc.) out of images. This task is far from being trivial, given all the possible fonts, colors, image qualitiesi out there. The text may also not lay on a horizontal straight line… Well you guess it, everything is possible in the wild, and the first step to make sense out of it is to extract the characters.
Google has been very strong in this area over the last years, especially through its open-source tool Tesseract OCR. This tool comes also in wrappers for the Python or Go etc. programming language. It also scores high in detection accuracy and support 100+ languages.
That’s why I got very curious to try out the OCR capabilities served on the Document AI service of Google Cloud.
Setup
First you have to log in the Google Cloud console, select one of your projects (or create a new one for that purpose) and activate Document AI for that project. There you have to create a processor. Either you can create your own processor, or chose one from the Processor gallery. I would recommend the latter to start. At the moment you can create it in only in two regions, namely us and eu. Once created, you the ID of your processor will be displayed. You will need it.
Personally I create a general-purpose OCR processor in the eu region.
For a local PDF file
If you want to try it out yourself, you can do it with this
Python command-line tool.
Don’t forget to follow the instruction at the beginning regarding pip install
and creation of a .env file.
I try it with a 30 MB / 153 pages PDF file in English.
Then I get an error related to quota limit, which makes sense. According to the doc, online processing (what we have just done through this script) accepts files of max 20 MB in size.
But batch processing accepts documents up to 1 GB, so let’s try it out:
For a batch of documents on GCloud Storage
For batch processing, your documents (PDF files) have to be stored in a bucket on GCloud storage. So I uploaded 16 documents, between 3.5 and 30 MB. This should keep was below the quota limits for batch processing. (Code for a command-line tool in Python for Document AI OCR batch processing available here) There I got an unclear error message back, related to some tokens in one of the documents
Conclusion
On the pro side:
- Large gallery of ready-to-use OCR processors
- Possibility to create / fine-tune your owm OCR processor, also with Human-in-the-loop
- You can also define the fields you want back (Box positions etc.), allowing fine-granular post-processing (see fieldMask in the output config)
Cons:
- Processor has to be created on the console. So no automated “Terraformed” deployment
- Quota limits not suited for industrial use
In brief, it’s a top-notch engine, but imho. still lacking packaging and integration for serious stuffs in production. It may come in a near future.