CNN Captcha Solver

This repository contains a Convolutional Neural Network (CNN) based captcha solver using OpenCV (cv2) library. The solver is trained to recognize and classify characters in captcha images. The model is trained using a dataset obtained from Kaggle.

Dataset

The dataset used for training the model can be found on Kaggle link to dataset. It consists of a collection of Captcha images along with their corresponding labels.

Dependencies

The following dependencies are required to run the code:

• Python 3.x
• TensorFlow 2.x
• OpenCV (cv2)
• NumPy
• Matplotlib
• look at the code for more

Network Architecture

The CNN model used for solving the Captcha images has the following architecture:

Input: (40, 20, 1) image
Conv2D (16 filters, 3x3 kernel, ReLU activation) -> MaxPooling2D (2x2 pool)
Conv2D (32 filters, 3x3 kernel, ReLU activation) -> MaxPooling2D (2x2 pool)
Conv2D (128 filters, 3x3 kernel, ReLU activation) -> MaxPooling2D (2x2 pool)
Conv2D (128 filters, 3x3 kernel, ReLU activation) -> MaxPooling2D (2x2 pool)
Flatten
Dense (1500 units)
Dense (19 classes, softmax)
Compiled with categorical cross-entropy loss and Adam optimizer.

This architecture was found to be effective in solving the Captcha images and achieved good results(kinda).
Trained and saved as captcha_model.h5.

Results

After training the model on the dataset, the following results were obtained:

On Training Data(at last epoch):

• Accuracy: 100%
• Loss: 2.0184 x 10^-6 or 2.0184e-06

On Validataion Data(at last epoch):

• val_accuracy: 89.5%
• val_loss: 130% or 1.3061

On Test Data .evaluate():

• Accuracy: 87.8%
• Loss: 165% or 1.6563

On 5 random images .predict():

• 5/5

These results demonstrate the effectiveness of the trained model in solving Captcha images.