torchfree-ocr 1.0.1

End-to-End English Optical Character Recognition

Torchless EasyOCR

This package is EasyOCR-based English optical character recognition. Unlike EasyOCR, the package uses a pre-saved with onnx English language model, so it doesn't need a 1-2 Gb pytorch dependency. This is particularly useful for developing light-weight applications that utilize text recognition.

Comparison for virtual env size

• 365 MB torchfree_ocr with dependencies
• 1.52 GB easyocr with dependencies
• 4.79 GB easyocr with dependencies + GPU enabled

Limitations

Each language requires generating a corresponding .onnx file, therefore, for the purpose of reducing package size only English is supported. Additionally, there is no GPU CUDA support.

Performance of torchfree_ocr vs easyocr

In terms of recognition, there is no visible recognition quality difference between torchfree_ocr and easyocr. In terms of speed torchfree_ocr works slightly faster than easyocr in CPU mode (~20% faster). Obviously, easyocr generally runs much faster in GPU mode, which torchfree_ocr doesn't support.

Examples

Installation

Install using pip

For the latest release:

pip install torchfree_ocr

Usage

import torchfree_ocr
reader = torchfree_ocr.Reader() # English recognition is default, other languages are not supported
result = reader.readtext('english.png')

The output will be in a list format, each item represents a bounding box, the text detected and confident level, respectively.

[([[231, 32], [672, 32], [672, 64], [231, 64]], 'Reduce your risk of coronavirus infection:', 0.8413621448628567), 
 ([[326, 98], [598, 98], [598, 124], [326, 124]], 'Clean hands with soap and water', 0.9633979603853523), 
 ([[328, 124], [540, 124], [540, 148], [328, 148]], 'or alcohol-based hand rub', 0.802668636048309), 
 ([[248, 170], [595, 170], [595, 196], [248, 196]], 'Cover nose and mouth when coughing and', 0.9529594602295661), 
 ([[248, 196], [546, 196], [546, 222], [248, 222]], 'sneezing with tissue or flexed elbow', 0.8406205896147358), 
 ([[320, 240], [624, 240], [624, 266], [320, 266]], 'Avoid close contact with anyone with', 0.8602271367787114), 
 ([[318, 265], [528, 265], [528, 293], [318, 293]], 'cold or flu-like symptoms', 0.9378307488433589), 
 ([[248, 322], [510, 322], [510, 348], [248, 348]], 'Thoroughly cook meat and eggs', 0.7159722535422908), 
 ([[332, 370], [640, 370], [640, 396], [332, 396]], 'No unprotected contact with live wild', 0.8346977728209518), 
 ([[334, 396], [464, 396], [464, 420], [334, 420]], 'or farm animals', 0.7179850171130348), 
 ([[595, 427], [683, 427], [683, 447], [595, 447]], 'World Health', 0.9979501800152029), 
 ([[597, 445], [685, 445], [685, 463], [597, 463]], 'Organization', 0.9977550970521537)]

Note 1: Instead of the filepath english.png, you can also pass an OpenCV image object (numpy array) or an image file as bytes. A URL to a raw image is also acceptable.

Note 2: The line reader = easyocr.Reader() is for loading a model into memory. It takes some time but it needs to be run only once.

You can also set detail=0 for simpler output.

reader.readtext('english.png', detail = 0)

Result:

['Reduce your risk of coronavirus infection:', 'Clean hands with soap and water', 'or alcohol-based hand rub', 'Cover nose and mouth when coughing and', 'sneezing with tissue or flexed elbow', 'Avoid close contact with anyone with', 'cold or flu-like symptoms', 'Thoroughly cook meat and eggs', 'No unprotected contact with live wild', 'or farm animals', 'World Health', 'Organization']

Averall, usage is the same as with EasyOCR, except Reader in this package only has recognizer= parameter and there is no readtextlang() method.

Usage for EasyOCR can be found in their tutorial and API Documentation.

Run on command line

$ torchfree_ocr -f english.png --detail=1