CAPTCHA Recognition Project

Introduction

This project automates CAPTCHA recognition using machine learning. It handles file management, image generation, preprocessing, and deep learning model training to achieve accurate predictions.

• Font Families: Blooming of Lilian, King of Kerry, and Eamon Wide
• Dataset Size: 330,000 images
• Tools: Python, TensorFlow/Keras, Gaussian Object Detection

Technical Workflow

1. Data Retrieval

The project uses ThreadPoolExecutor for parallelized file downloads. Missing files are logged and re-attempted to ensure dataset consistency.

2. Preprocessing

Images are normalized and resized to a standard dimension of 192x96. Augmentation techniques, such as rotation and noise addition, are applied for robustness.

3. Model Architecture

The model uses a Convolutional Neural Network (CNN) with the following layers:

• Input Layer: Accepts preprocessed CAPTCHA images.
• Convolutional Layers: Extract features using filters and ReLU activations.
• Pooling Layers: Downsamples the feature maps to reduce dimensionality.
• Dense Layers: Fully connected layers for character classification.
• Softmax Output: Predicts the characters in the CAPTCHA.

4. Training

The model is trained on 80% of the dataset (264,000 images) and validated on the remaining 20%. Early stopping is used to avoid overfitting, with validation loss monitored over 5 epochs.

Project Demonstration

Below is a video demonstrating the project execution, including its computational efficiency and output:

And here is the resulting CAPTCHA prediction:

Key Highlights

• Optimized image preprocessing for consistent input dimensions.
• Robust CNN architecture with pooling and dense layers.
• Scalable design supporting large datasets and parallel processing.
• Accuracy achieved across multiple font families.

GitHub Repository

Access the complete code, dataset, and documentation here: GitHub Repository

Developed with passion. Check out more projects on GitHub.