Alternative OCR Libraries

[pgaviganHC]

It may be worth trying some alternative OCR libraries, as discussed in this article: https://www.statcan.gc.ca/en/data-science/network/character-recognition

Might be a good idea to have these alternatives be available options in this library.

[benjaminLuoHC]

Based on my understanding, Tesseract (the current OCR library) excels for documents while CRAFT (the model mentioned in the link) is for text-detection in more complicated images. They work well in conjunction

Another option is EasyOCR which does not require any external dependencies (Issue #4). It uses CRAFT for text-detection and then its own OCR engine so it's essentially Tesseract+CRAFT in a single library, but potentially less powerful due to it being more light-weight

[pgaviganHC]

Based on my understanding, Tesseract (the current OCR library) excels for documents while CRAFT (the model mentioned in the link) is for text-detection in more complicated images. They work well in conjunction

Another option is EasyOCR which does not require any external dependencies (Issue #4). It uses CRAFT for text-detection and then its own OCR engine so it's essentially Tesseract+CRAFT in a single library, but potentially less powerful due to it being more light-weight

Good insight here, thanks. I wonder if we could quantify the performance difference between these options with a simple test of some sort?

[BrennanHCSC]

• This may become more relevant in the near future as multi-model LLMs become more effective and available for OCR purposes.
• For example, personal experimentation with GPT-4V shows promising results for table extraction.

[benjaminLuoHC]

EasyOCR has been recommended to us by Microsoft for use in MS Fabric (the current Tesseract implementation cannot be installed in the MS Fabric environment)

Need to do a security check on EasyOCR