CharNet OCR

💡 Being the very first version, CharNet's performance is currently a baseline, and it only accepts images having solid #FFF background and solid #000 text.

❓ It is being planned to improve the model's performance and make it more robust in the upcoming versions.

---

An Optical Character Recognition library for Python. It recognizes character-by-character and produces text of words present in the image. An ensemble of 3 YOLOv8-nano models that determines the character-level bounding boxes and labels.

About the models

Model	Description
CharNetLC	Character-level detection model for lower-case alphabets
CharNetUC	Character-level detection model for upper-case alphabets
CharNetDC	Character-level detection model for digits and special characters

The models were trained on custom dataset of of ~2,000 fonts collected from Dafont's "Script > Handwritten" section (under "100% free" filter).

Getting Started

Installation

pip install CharNetOCR

Usage

from CharNetOCR import recognize 

my_img = './test.png'

data = recognize(
    img_path=my_img,

    # Additional parameters (optional):
    verbose=False, # True by default
    word_configs={
        'kernel': (5, 5), # kernel size for morphological operations
        'iterations': 2, # number of iterations for morphological operations
    },
    char_weights={
        'lowercase': 1.0,
        'uppercase': 0.95, 
        'digits_and_special_chars': 0.85
    }
)

# data = (word_str, bbox_coords, confidence)[]

For the recognize function, you can pass the following parameters:

Parameter	Default	Description
img_path	None	Path to the image file
verbose	True	Whether to print the progress or not
word_configs	{ 'kernel': (5, 5), 'iterations': 2 }	Morphological operations configurations
char_weights	{ 'lowercase': 1.0, 'uppercase': 0.95, 'digits_and_special_chars': 0.85 }	Weights for the CharNet models

How it works?

1. Image segmentation

1. First, all words are segmented in the image through word-level bounding boxes determination using OpenCV.

2. Then, we iterate through each "bbox" in the list of word's bounding boxes. On each iteration,

   • The segment of the image is cropped using the bounding box coordinates.
   • The segmented image is padded to remove dramatic enlargement of the word.
   • The processed image is saved temporarily and passed into the ensemble of CharNet models.
   • We receive a combined character-level prediction. We delete the temporary image
   • And append the predicted text to the final list of words.

2. CharNet Ensembling

1. The cropped image is passed through the ensemble of CharNet models.
2. Each model predicts the character-level bounding boxes and labels.
3. The highest-confidence bounding box is selected from the 3 models for each character.

🎯 And the best part? You can control the weights of the models to get the best results!

Contributing

1. Fork the repository
2. Create a new branch (git checkout -b feat/awesome-feature)
3. Commit the changes (git commit -am 'Add awesome feature')
4. Push to the branch (git push origin feat/awesome-feature)
5. Create a new Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Credits

This project was developed by Aakash and Vikas Jha