/img2table

img2table is a table identification and extraction Python Library for PDF and images, based on OpenCV image processing

Primary LanguagePythonMIT LicenseMIT

img2table

img2table is a simple, easy to use, table identification and extraction Python Library based on OpenCV image processing that supports most common image file formats as well as PDF files.

Thanks to its design, it provides a practical and lighter alternative to Neural Networks based solutions, especially for usage on CPU.

Table of contents

Installation

The library can be installed via pip:

pip install img2table: Standard installation, supporting Tesseract
pip install img2table[paddle]: For usage with Paddle OCR
pip install img2table[easyocr]: For usage with EasyOCR
pip install img2table[gcp]: For usage with Google Vision OCR
pip install img2table[aws]: For usage with AWS Textract OCR
pip install img2table[azure]: For usage with Azure Cognitive Services OCR

Features

  • Table identification for images and PDF files, including bounding boxes at the table cell level
  • Handling of complex table structures such as merged cells
  • Handling of implicit rows - see example
  • Table content extraction by providing support for OCR services / tools
  • Extracted tables are returned as a simple object, including a Pandas DataFrame representation
  • Export extracted tables to an Excel file, preserving their original structure

Supported file formats

Images

Images are loaded using the opencv-python library, supported formats are listed below.

Supported image formats
  • Windows bitmaps - .bmp, .dib
  • JPEG files - .jpeg, .jpg, *.jpe
  • JPEG 2000 files - *.jp2
  • Portable Network Graphics - *.png
  • WebP - *.webp
  • Portable image format - .pbm, .pgm, .ppm .pxm, *.pnm
  • PFM files - *.pfm
  • Sun rasters - .sr, .ras
  • TIFF files - .tiff, .tif
  • OpenEXR Image files - *.exr
  • Radiance HDR - .hdr, .pic
  • Raster and Vector geospatial data supported by GDAL
    OpenCV: Image file reading and writing
Multi-page images are not supported.

PDF

Both native and scanned PDF files are supported.

Usage

Documents

Images

Images are instantiated as follows :

from img2table.document import Image

image = Image(src, 
              detect_rotation=False)

Parameters

src : str, pathlib.Path, bytes or io.BytesIO, required
Image source
detect_rotation : bool, optional, default False
Detect and correct skew/rotation of the image

The implemented method to handle skewed/rotated images supports skew angles up to 45° and is based on the publication by Huang, 2020.
Setting the detect_rotation parameter to True, image coordinates and bounding boxes returned by other methods might not correspond to the original image.

PDF

PDF files are instantiated as follows :

from img2table.document import PDF

pdf = PDF(src, 
          pages=[0, 2],
          detect_rotation=False,
          pdf_text_extraction=True)

Parameters

src : str, pathlib.Path, bytes or io.BytesIO, required
PDF source
pages : list, optional, default None
List of PDF page indexes to be processed. If None, all pages are processed
detect_rotation : bool, optional, default False
Detect and correct skew/rotation of extracted images from the PDF
pdf_text_extraction : bool, optional, default True
Extract text from the PDF file for native PDFs

PDF pages are converted to images with a 200 DPI for table identification.

OCR

img2table provides an interface for several OCR services and tools in order to parse table content.
If possible (i.e for native PDF), PDF text will be extracted directly from the file and the OCR service/tool will not be called.

Tesseract 
from img2table.ocr import TesseractOCR

ocr = TesseractOCR(n_threads=1, 
                   lang="eng", 
                   psm=11,
                   tessdata_dir="...")

Parameters

n_threads : int, optional, default 1
Number of concurrent threads used to call Tesseract
lang : str, optional, default "eng"
Lang parameter used in Tesseract for text extraction
psm : int, optional, default 11
PSM parameter used in Tesseract, run tesseract --help-psm for details
tessdata_dir : str, optional, default None
Directory containing Tesseract traineddata files. If None, the TESSDATA_PREFIX env variable is used.

Usage of Tesseract-OCR requires prior installation. Check documentation for instructions.
For Windows users getting environment variable errors, you can check this tutorial

PaddleOCR

PaddleOCR is an open-source OCR based on Deep Learning models.
At first use, relevant languages models will be downloaded.

from img2table.ocr import PaddleOCR

ocr = PaddleOCR(lang="en",
                kw={"kwarg": kw_value, ...})

Parameters

lang : str, optional, default "en"
Lang parameter used in Paddle for text extraction, check documentation for available languages
kw : dict, optional, default None
Dictionary containing additional keyword arguments passed to the PaddleOCR constructor.

NB: For usage of PaddleOCR with GPU, the CUDA specific version of paddlepaddle-gpu has to be installed by the user manually as stated in this issue. 
# Example of installation with CUDA 11.8
pip install paddlepaddle-gpu==2.5.0rc1.post118 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
pip install paddleocr img2table

If you get an error trying to run PaddleOCR on Ubuntu, please check this issue for a working solution.


EasyOCR

EasyOCR is an open-source OCR based on Deep Learning models.
At first use, relevant languages models will be downloaded.

from img2table.ocr import EasyOCR

ocr = EasyOCR(lang=["en"],
              kw={"kwarg": kw_value, ...})

Parameters

lang : list, optional, default ["en"]
Lang parameter used in EasyOCR for text extraction, check documentation for available languages
kw : dict, optional, default None
Dictionary containing additional keyword arguments passed to the EasyOCR Reader constructor.

docTR

docTR is an open-source OCR based on Deep Learning models.
In order to be used, docTR has to be installed by the user beforehand. Installation procedures are detailed in the package documentation

from img2table.ocr import DocTR

ocr = DocTR(detect_language=False,
            kw={"kwarg": kw_value, ...})

Parameters

detect_language : bool, optional, default False
Parameter indicating if language prediction is run on the document
kw : dict, optional, default None
Dictionary containing additional keyword arguments passed to the docTR ocr_predictor method.

Google Vision

Authentication to GCP can be done by setting the standard GOOGLE_APPLICATION_CREDENTIALS environment variable.
If this variable is missing, an API key should be provided via the api_key parameter.

from img2table.ocr import VisionOCR

ocr = VisionOCR(api_key="api_key", timeout=15)

Parameters

api_key : str, optional, default None
Google Vision API key
timeout : int, optional, default 15
API requests timeout, in seconds

AWS Textract

When using AWS Textract, the DetectDocumentText API is exclusively called.

Authentication to AWS can be done by passing credentials to the TextractOCR class.
If credentials are not provided, authentication is done using environment variables or configuration files. Check boto3 documentation for more details.

from img2table.ocr import TextractOCR

ocr = TextractOCR(aws_access_key_id="***",
                  aws_secret_access_key="***",
                  aws_session_token="***",
                  region="eu-west-1")

Parameters

aws_access_key_id : str, optional, default None
AWS access key id
aws_secret_access_key : str, optional, default None
AWS secret access key
aws_session_token : str, optional, default None
AWS temporary session token
region : str, optional, default None
AWS server region

Azure Cognitive Services 
from img2table.ocr import AzureOCR

ocr = AzureOCR(endpoint="abc.azure.com",
               subscription_key="***")

Parameters

endpoint : str, optional, default None
Azure Cognitive Services endpoint. If None, inferred from the COMPUTER_VISION_ENDPOINT environment variable.
subscription_key : str, optional, default None
Azure Cognitive Services subscription key. If None, inferred from the COMPUTER_VISION_SUBSCRIPTION_KEY environment variable.

Table extraction

Multiple tables can be extracted at once from a PDF page/ an image using the extract_tables method of a document.

from img2table.ocr import TesseractOCR
from img2table.document import Image

# Instantiation of OCR
ocr = TesseractOCR(n_threads=1, lang="eng")

# Instantiation of document, either an image or a PDF
doc = Image(src)

# Table extraction
extracted_tables = doc.extract_tables(ocr=ocr,
                                      implicit_rows=False,
                                      borderless_tables=False,
                                      min_confidence=50)

Parameters

ocr : OCRInstance, optional, default None
OCR instance used to parse document text. If None, cells content will not be extracted
implicit_rows : bool, optional, default False
Boolean indicating if implicit rows should be identified - check related example
borderless_tables : bool, optional, default False
Boolean indicating if borderless tables are extracted on top of bordered tables.
min_confidence : int, optional, default 50
Minimum confidence level from OCR in order to process text, from 0 (worst) to 99 (best)

NB: Borderless table extraction can, by design, only extract tables with 3 or more columns.

Method return

The ExtractedTable class is used to model extracted tables from documents.

Attributes

bbox : BBox
Table bounding box
title : str
Extracted title of the table
content : OrderedDict
Dict with row indexes as keys and list of TableCell objects as values
df : pd.DataFrame
Pandas DataFrame representation of the table
html : str
HTML representation of the table
Images

extract_tables method from the Image class returns a list of ExtractedTable objects.

output = [ExtractedTable(...), ExtractedTable(...), ...]
PDF

extract_tables method from the PDF class returns an OrderedDict object with page indexes as keys and lists of ExtractedTable objects.

output = {
    0: [ExtractedTable(...), ...],
    1: [],
    ...
    last_page: [ExtractedTable(...), ...]
}

Excel export

Tables extracted from a document can be exported to a xlsx file. The resulting file is composed of one worksheet per extracted table.
Method arguments are mostly common with the extract_tables method.

from img2table.ocr import TesseractOCR
from img2table.document import Image

# Instantiation of OCR
ocr = TesseractOCR(n_threads=1, lang="eng")

# Instantiation of document, either an image or a PDF
doc = Image(src)

# Extraction of tables and creation of a xlsx file containing tables
doc.to_xlsx(dest=dest,
            ocr=ocr,
            implicit_rows=False,
            borderless_tables=False,
            min_confidence=50)

Parameters

dest : str, pathlib.Path or io.BytesIO, required
Destination for xlsx file
ocr : OCRInstance, optional, default None
OCR instance used to parse document text. If None, cells content will not be extracted
implicit_rows : bool, optional, default False
Boolean indicating if implicit rows should be identified - check related example
borderless_tables : bool, optional, default False
Boolean indicating if borderless tables are extracted. It requires to provide an OCR to the method in order to be performed - feature in alpha version
min_confidence : int, optional, default 50
Minimum confidence level from OCR in order to process text, from 0 (worst) to 99 (best)

Returns

If a io.BytesIO buffer is passed as dest arg, it is returned containing xlsx data

Examples

Several Jupyter notebooks with examples are available :

  • Basic usage: generic library usage, including examples with images, PDF and OCRs
  • Borderless tables: specific examples dedicated to the extraction of borderless tables
  • Implicit rows: illustrated effect of the parameter implicit_rows of the extract_tables method

Caveats / FYI

  • For table extraction, results are highly dependent on OCR quality. By design, tables where no OCR data can be found are not returned.
  • The library is tailored for usage on documents with white/light background. Effectiveness can not be guaranteed on other type of documents.
  • Table detection using only OpenCV processing can have some limitations. If the library fails to detect tables, you may check CNN based solutions like CascadeTabNet or the PaddleOCR implementation.