Skip to content

Latest commit

 

History

History
60 lines (41 loc) · 3.2 KB

README.md

File metadata and controls

60 lines (41 loc) · 3.2 KB

Captcha

The purpose of this project was to stay sharp and train my skills in Python and Deep Learning. This repository contains the implementation of a PyTorch-based deep learning model, CaptchaModel, designed for solving CAPTCHA tasks. The model predicts a sequence of characters (e.g., 6 characters per image) from grayscale CAPTCHA images of fixed dimensions.

Model Architecture

The CaptchaModel is a convolutional neural network (CNN) combined with fully connected layers, built to extract features from images and predict multiple characters. Below is a detailed description of the architecture:

Components

  1. Input Dimensions:

    • Input: A single-channel (grayscale) image with size (height, width).

  2. Convolutional Layers:

    • Conv1: 16 filters, kernel size 3x3, stride 1, padding 1.
    • Conv2: 32 filters, kernel size 3x3, stride 1, padding 1.
    • Conv3: 32 filters, kernel size 3x3, stride 1, padding 1.

  3. Pooling:

    • A MaxPool2d layer with a kernel size of 2x2 and stride 2 is applied after each convolutional layer to reduce the spatial dimensions by half.

  4. Batch Normalization:

    • A BatchNorm2d layer is applied after the third convolutional layer to normalize feature maps.

  5. Flatten:

    • The 2D output of the convolutional and pooling layers is flattened into a 1D vector for the fully connected layers.

  6. Fully Connected Layers:

    • The model has a fully connected network for each character to be predicted:
      • First dense layer: 64 neurons with ReLU activation.
      • Dropout: 50% dropout to prevent overfitting. (30% or 80% has been tried)
      • Output layer: Number of unique characters in the CAPTCHA (e.g., 36 for alphanumeric) with a Sigmoid activation function.

  7. Output:

    • The model predicts a list of outputs, each representing the probability distribution over all possible characters for one position in the CAPTCHA.

Forward Pass

  1. The input image passes through three convolutional and pooling layers.
  2. The output is flattened into a 1D vector.
  3. For each character position in the CAPTCHA, the flattened vector is passed through a fully connected subnetwork to predict the character.

Parameters

  • size_image: Tuple specifying the dimensions of the input image (height, width).
  • n_characters_per_image: Number of characters to predict in each CAPTCHA.
  • n_characters_total: Number of unique possible characters (e.g., 36 for [0-9-A-Z]).
  • dropout_val: The dropout randomly selects a subset of neurons to be deactivated with a predefined probability.

Dataset

If you want to try it out, you can use the following link: https://www.kaggle.com/datasets/nasironeeb/captcha-dataset-6-character/data to get the data

the file model_full_1.pth is the model after training for 20 epochs with 0.001 learning rate and 0.8 dropout. the file model_full_2.pth is the model after training for 20 epochs with 0.001 learning rate and 0.2 dropout. the file model_full_3.pth is the model after training for 20 epochs with 0.001 learning rate and 0.5 dropout. the file model_full_4.pth is the model after training for 40 epochs with 0.001 learning rate and 0.8 dropout. The best model is the model_full_4.pth