https://github.com/hari9-9/captcha-solver

A Scalable CAPTCHA solving Solution using CNN's
https://github.com/hari9-9/captcha-solver

captcha-solver deep-neural-networks scalability tensorflow tf-lite

Last synced: 3 months ago
JSON representation

A Scalable CAPTCHA solving Solution using CNN's

• Host: GitHub
• URL: https://github.com/hari9-9/captcha-solver
• Owner: hari9-9
• Created: 2024-12-09T02:04:46.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-12-09T02:27:50.000Z (over 1 year ago)
• Last Synced: 2025-04-03T15:12:58.923Z (over 1 year ago)
• Topics: captcha-solver, deep-neural-networks, scalability, tensorflow, tf-lite
• Language: Python
• Homepage:
• Size: 65.4 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# CAPTCHA Solver: A Scalable and Efficient Approach

## Overview

This project focuses on building a scalable CAPTCHA solving system capable of identifying text-based CAPTCHAs with varying character lengths (1 to 6) and diverse fonts. Using a modular approach,implemented and trained seven deep learning models: one for predicting CAPTCHA length and six specialized for CAPTCHA classification based on length.

The project emphasizes scalability by leveraging TensorFlow Lite (TFLite) for efficient deployment on resource-constrained devices like Raspberry Pi, ensuring fast and accurate CAPTCHA classification.

---

## Features

- **Scalable Design:** Utilizes a modular seven-model architecture to handle variable CAPTCHA lengths efficiently.

- **Font Robustness:** Models trained on datasets containing diverse fonts to improve generalization.

- **Efficient Deployment:** Conversion of TensorFlow models to TFLite format ensures compatibility with edge devices like Raspberry Pi.

- **High Performance:** Employs preprocessing and optimized training configurations to achieve high accuracy and efficient resource utilization.

---

## Architecture

The system is based on a **divide-and-conquer** approach, where:

1. **Length Prediction Model:** Predicts the length of the CAPTCHA.

2. **Classification Models:** Six specialized models classify CAPTCHAs for lengths 1 through 6.

### Key Components

- **Data Preprocessing:** Explored various preprocessing techniques, with the best configuration enhancing image clarity and model performance.

- **Training Models:** Each model trained on datasets specific to its target length for optimized performance.

- **Font Handling:** Included images of all font variations in training datasets to ensure robustness.

### Training Configuration

- **Epochs:** 20

- **Batch Size:** 64

- **Early Stopping:** Enabled to prevent overfitting.

### Dataset

- Generated 64,000 images for training each model (90/10 train-test split).

- Data preprocessing included thresholding and noise removal to enhance input quality.

---

## TensorFlow Lite Conversion

To ensure scalability and compatibility with resource-constrained environments:

1. **Model Conversion:** TensorFlow models were converted to TFLite format.

2. **Quantization (Optional):** Experimented with quantized models for faster inference, though with reduced accuracy.

The TFLite models were deployed on a Raspberry Pi, achieving efficient classification within acceptable time limits.

---

## Scalability

- **Local vs. Edge Computing:** Training was performed on local machines with GPU acceleration, while inference was executed on Raspberry Pi using TFLite models.

- **Performance Metrics:** The system demonstrated high computational efficiency on edge devices:

  - **Classification Time:** ~900 seconds for 4,000 images.

  - **Resource Utilization:** Optimal use of CPU cores and memory during inference.

---

## Results

- **Score:** 2249/4000

- **Deployment Metrics:**

  - Classification time on Raspberry Pi: ~900 seconds.

  - Efficient model loading and prediction pipeline reduced resource usage and processing time.

---

## Future Work

- **Optimized Multithreading:** To enhance classification speed on Raspberry Pi.

- **Advanced Preprocessing:** Explore image segmentation and additional preprocessing techniques.

- **Quantization-aware Training:** Improve TFLite model performance without significant accuracy loss.

---

## How to Use

- Run length_model training length_model_train.py

- Run cpatha_model/process.sh to automatically generate model for all six lengths

- Pi/ folder has tf to tf_lite converter and classifier files