Crnn Base Fa V2

🚀 Hezar Persian OCR CRNN Model

A CRNN model for Persian OCR, offering enhanced performance and features.

🚀 Quick Start

This is a CRNN model designed for Persian OCR. It's based on a simple CNN + LSTM architecture, inspired by this paper. It's a successor to our previous model hezarai/crnn-base-fa-v1, with several key improvements.

✨ Features

• Larger Dataset: Utilizes a dataset 5X larger than the previous model.
• Input Image Size Change: Alters the input image size from 64x256 to 32x384.
• Increased Output Length: Raises the max output length from 64 to 96 (the max length of samples in the dataset is 48 to handle CTC loss issues).
• Expanded Character Support: Supports numbers and special characters (check id2label in model_config.yaml).
• Auto-Handling of LTR Characters: Automatically manages LTR characters like digits within the text.

Note that this model is optimized for printed/scanned documents and performs best on texts up to around 50 characters in length. For an end-to-end OCR pipeline, first use a text detector model such as hezarai/CRAFT to extract text boxes, preferably at the word level, and then use this model. However, it can also be fine - tuned for other domains like license plates or handwritten texts.

📦 Installation

pip install hezar

💻 Usage Examples

Basic Usage

from hezar.models import Model

crnn = Model.load("hezarai/crnn-base-fa-v2")
texts = crnn.predict(["sample_image.jpg"])
print(texts)

📄 License

This project is licensed under the Apache-2.0 license.