OCR Captchas

This repository demonstrates a practice of building an Optical Character Recognition (OCR) model to read 4-digit CAPTCHA images using TensorFlow and Keras.

Overview

The project focuses on developing an OCR system capable of recognizing 4-digit CAPTCHAs. The approach involves:

• Data Preparation: Utilizing a dataset of CAPTCHA images where each image's filename corresponds to its label.
• Model Architecture: Implementing a Convolutional Recurrent Neural Network (CRNN) combined with Connectionist Temporal Classification (CTC) loss to effectively recognize sequences in images.
• Training and Evaluation: Training the model on the prepared dataset and evaluating its performance on unseen data.

Dataset

The dataset comprises CAPTCHA images stored in the captcha_images directory. Each image filename represents the correct 4-digit label, facilitating supervised learning.

Dependencies

To set up the environment, install the necessary dependencies using:

pip install -r requirements.txt

Usage

Training the Model

To train the OCR model:

1. Ensure your dataset is in the captcha_images directory, with each image named according to its 4-digit label.

2. Run the training script:

   python ocr_captcha.py

   This script will preprocess the data, define the CRNN model, and commence training. Upon completion, the trained model will be saved for future use.

Predicting with the Model

To make predictions on new CAPTCHA images:

1. Place the image you wish to predict in the root directory and name it test.png.

2. Execute the prediction script:

   python predict_example.py

   The script will load the trained model and output the predicted 4-digit code for test.png.

Results

The model's performance can be visualized through training metrics and sample predictions. Below is an example of the learning curve:

And a visualization of the model's architecture:

References

This project is inspired by various OCR implementations for CAPTCHA recognition, including:

• EVOL-ution/Captcha-Recognition-using-CRNN
• nbswords/ocr-captchas
• Keras Example: OCR model for reading Captchas

License

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

---

Note: This README is based on the structure and information from similar OCR CAPTCHA projects and may need adjustments to align perfectly with the specific implementations and results of this repository.