metadata

license: mit
language:
  - he
tags:
  - languages
  - manuscripts
  - hebrew
  - ocr
  - letters
  - manuscript
  - digital-humanities
datasets:
  - bsesic/HebrewManuscripts

Hebrew Letter Recognition Model

Model Description

This is a Convolutional Neural Network (CNN) model trained to recognize Hebrew letters and a stop symbols in images. The model can identify individual letters from a provided image, outputting their respective class along with probabilities.

Model Details:

Model Type: Convolutional Neural Network (CNN)
Framework: TensorFlow 2.x / Keras
Input Size: 64x64 grayscale images of isolated letters.
Output Classes: 28 Hebrew letters + 1 stop symbol (.)
Use Case: Recognizing handwritten or printed Hebrew letters and punctuation in scanned images or photos of documents.

Intended Use

This model is designed for the automatic recognition of Hebrew letters from images. The model can be used in applications such as:

Optical character recognition (OCR) systems for Hebrew text.
Educational tools to help learners read Hebrew text.
Historical document digitization of Hebrew manuscripts.

How to Use:

from tensorflow.keras.models import load_model
import numpy as np
import cv2

# Load the model
model = load_model('path_to_model.hebrew_letter_model.keras')

# Preprocess an input image (example for one letter)
img = cv2.imread('path_to_image.jpg', cv2.IMREAD_GRAYSCALE)
img_resized = cv2.resize(img, (64, 64)) / 255.0
img_array = np.expand_dims(img_resized, axis=0)

# Predict
predictions = model.predict(img_array)
predicted_class = np.argmax(predictions, axis=1)[0]

# Class names for Hebrew letters
class_names = ['stop', 'א', 'ב', 'ג', 'ד', 'ה', 'ו', 'ז', 'ח', 'ט', 'י', 'ך', 'כ', 'ל', 'ם', 'מ', 'ן', 'נ', 'ס', 'ע', 'ף', 'פ', 'ץ', 'צ', 'ק', 'ר', 'ש', 'ת']

print("Predicted letter:", class_names[predicted_class])

Example:

If given an image with the Hebrew word "אברם" (Abram), the model can detect and classify the letters and stop symbols with probabilities.

Limitations:

Font Variations: The model performs best on specific fonts (e.g., square Hebrew letters). Performance may degrade with highly stylized or cursive fonts.
Noise Sensitivity: Images with a lot of noise, artifacts, or low resolution may lead to incorrect predictions.
Stop Symbol: The stop symbol is particularly recognized by detecting three vertical dots. However, false positives can occur if letters with similar shapes are present.

Training Data:

The model was trained on a dataset containing Hebrew letters and stop symbols. The training dataset includes:

28 Hebrew letters.
1 stop symbol representing three vertical dots (.).

Training Procedure:

Optimizer: Adam
Loss function: Categorical Crossentropy
Batch size: 32
Epochs: 10

Data augmentation was applied to reduce overfitting and increase the model's generalizability to unseen data. This includes random rotations, zooms, and horizontal flips.

Model Performance

Metrics:

Accuracy: 95% on the validation dataset.
Precision: 94%
Recall: 93%
Performance may vary depending on the quality of the input images, noise levels, and whether the letters are handwritten or printed.

Known Issues:

False Positives for Stop Symbols: The model sometimes incorrectly identifies letters that resemble three vertical dots as stop symbols.
Overfitting to Specific Fonts: Performance can degrade on handwritten texts or cursive fonts not represented well in the training set.

Ethical Considerations

Bias: The model was trained on a specific set of Hebrew fonts and may not perform equally well across all types of Hebrew texts, particularly historical or handwritten documents. Fairness: The model may produce varying results depending on font style, quality of input images, and preprocessing applied.

Future Work:

Improving Generalization: Future work will focus on improving the model's robustness to different fonts, handwriting styles, and noisy inputs. Multilingual Expansion: Adding support for other Semitic scripts or expanding the model for multilingual OCR tasks. Citation:

If you use this model in your work, please cite it as follows:

@misc{hebrew-letter-recognition,
  title={Hebrew Manuscripts Letter Recognition Model},
  author={Benjamin Schnabel},
  year={2024},
  howpublished={\url{https://huggingface.co/bsesic/HebrewManuscriptsMNIST}},
}

License:

This model is licensed under MIT License.