OCRTools

Tools for scraping pdfs with OCR

Installation

Clone the repo and install with:

python3 -m pip install .

On MacOS you may need to run

python3 -m pip install --no-binary tesserocr .

Basic Usage

import ocrtools as ocrt
import re

ocr = ocrt.TesseractReader()

# Find dates within long strings like paragraphs
date_tagger = ocrt.DateTagger()

extractor = ocrt.Extractor([
    ocrt.FieldExtractor("date", ocrt.get_date(date_tagger), ocrt.BBox.from_xyxy(0.12, 0.12, 0.2, 0.2)),
    ocrt.FieldExtractor("address", ocrt.match(re.compile("address: {.*}")), ocrt.BBox.from_xyxy(0.12, 0.12, 0.2, 0.2))
])

doc = ocrt.get_doc("/path/to/doc.pdf")
for page in doc.pages():
    if res := extractor.extract(ocr, page):
        print(f"address: {res['address']}")
        print(f"date: {res['date']}")

This searches the PDF and extracts dates and addresses from the specified regions with the fewest possible calls to the OCR engine with only a subset of each page.

Logging

Log the OCR results:

ocrt.logging.configure_loggers("/tmp/ocr-logs")
logger = ocrt.logging.get_logger("test")
doc = ocrt.get_doc("/path/to/doc.pdf")
for page in doc.pages():
    if res := extractor.extract(ocr, page, logger=logger):
        print(f"address: {res['address']}")
        print(f"date: {res['date']}")

This writes images annotated with OCR and extraction result to /tmp/ocr-logs/test.

Available OCR Engines

OCRTools comes with Tesseract (via Tesserocr) and Textract interfaces. To use your own simply provide a function of the following type to extractor.extract.

IOCREngine = Callable[[OCRResource], OCRResult]

Taggers

OCRTools uses spacy and huggingface tranformers to identify entities such as dates and names within text. To use your own, simply provide a function of the following type:

INERTagger = Callable[[str], List[TokenSpan]]

Requirements

• fitz (pymupdf)
• dateparse
• pillow
• pandas
• tesserocr
• spacy
• transformers

For AWS Textract

• boto3
• textractcaller
• trp