This is a simple wrapper for word-segmented OCR. It provides a standardized interface to:

• PaddleOCR (local)
• EasyOCR (local)
• Azure Cognitive Skills
• Google Cloud
• AWS

For a usage example, see tryme.ipynb.

In addition to the automatically installed requirements, you also need to install the packages needed for the individual OCR solutions. These are not automatically installed, since we don't want to force dependencies which are not really needed if you only use one specific OCR solution.

Use the following dependency groups to install the required packages for the OCR solutions you want to use:

• Azure: ocr_wrapper[azure]
• Google: ocr_wrapper[google]
• GoogleAzure: ocr_wrapper[googleazure] - this also supports checkbox detection
• AWS: ocr_wrapper[aws]
• EasyOCR: ocr_wrapper[easy]
• PaddleOCR: ocr_wrapper[paddle]

Depending on your operating system you might also have to install the Noto Sans font package (e.g. noto-sans in Ubuntu)

Different OCR solutions are provided via the classes GoogleOCR, AwsOCR, AzureOCR, EasyOCR, and PaddleOCR.

All solutions have built in caching support, which can be activated by supplying a cache_file argument in the constructor. If a page has already been OCRed, the result from the cache will be used. Once an OCR class has been instantiated, an image can be OCRed with the .ocr Method which accepts an image in PIL format.

The result will be a list of BBox instances. Each BBox contains the coordinates of the four points of the bounding box and the recognized text. It can also contain an optional label, which is not used for OCR. BBox aslo supports many additional conversion functions, detecting overlap of bounding boxes, etc.

To easily visualize bounding boxes, the library also offers the method draw_bboxes.

The function autoselect_ocr_engine() can be used to automatically return the class for the needed OCR engine, using the OCR_PROVIDER environment variable. google, azure, aws, easy, and paddle are valid settings. If no provider is explicitly set, Google OCR is chosen by default. In case an invalid OCR provider is specified, an InvalidOcrProviderException will be raised.

The credentials will be obtained from one of the following:

• The environment variable GOOGLE_APPLICATION_CREDENTIALS
• A credentials file ~/.config/gcloud/credentials.json
• A credentials file /credentials.json

Credentials etc. for Azure OCR will be obtained from one of the following (in this order):

• From the endpoint and key arguments when creating AzureOCR
• From the environment variables AZURE_OCR_ENDPOINT and AZURE_OCR_KEY
• From the credentials file ~/.config/azure/ocr_credentials.json that contains the keys endpoint and key

Credentials for GoogleOCR as well as AzureOCR have to be set

All OCR wrappers support detection and decoding of QR codes as well as different forms of barcodes. Internally, the zbar library is used.

The QR/bar-codes will be added as bounding boxes, and the contained information is returned as the OCR text. The text is formatted in the form TYPE[[DATA]] (e.g. QRCODE[[Encoded Information]], CODE39[[1234567890]], ...). All valid types can be found in the pyzbar.pyzbar.ZBarSymbol enum.

This has to be activated with the argument add_qr_barcodes=True when creating the wrapper (default is False). In these cases, the pyzbar Python library, as well as the zbar system library have to be installed.