Skip to content

Latest commit

 

History

History
129 lines (86 loc) · 4.34 KB

README.md

File metadata and controls

129 lines (86 loc) · 4.34 KB
Raw

License Version

Overview

This repository contains a Convolutional Neural Network (CNN) designed to identify art drawn by Randall Munroe and Nicholas Gurewitch. The model leverages deep learning techniques to achieve high accuracy in distinguishing between these two distinct comic styles.

The repository includes a carefully curated dataset, straightforward preprocessing scripts, and user-friendly model training and evaluation pipelines. Comprehensive documentation is provided to facilitate reproducibility and encourage further exploration in the field of machine learning for art analysis

Image Sample

confidence_1

confidence_2

Repository Contents

  • Trained model
  • Training script
  • Classification script
  • Jupyter notebook for visualization
  • Curated dataset

Table of Contents

  • Introduction
  • Model Architecture
  • Dataset
  • Training Procedure
  • Results
  • Usage
  • Contributing
  • License

Introduction

Comic strips are a unique form of artistic expression, each with its own style and thematic elements. This project aims to utilize a CNN to automate the classification of comics, providing insights into the stylistic differences between the works of Randall Munroe, creator of xkcd and artist Nicholas Gurewitch.

Model Architecture

The CNN architecture consists of two convolutional layers and pooling layers, followed by fully connected layers for classification. The model is designed to be lightweight and efficient, allowing for rapid training and inference.

Key Features:

  • Convolutional Layers: Extracts features from comic images.
  • Pooling Layers: Reduces dimensionality while retaining important information.
  • Fully Connected Layers: Classifies the extracted features into the respective comic categories.

Dataset

The dataset comprises a collection of drawings from both XKCD and Perry Bible Fellowship, sampled randomly from their respective archives. Each comic is labeled according to its source, allowing the model to learn the distinguishing features of each artist's style.

  • Total Images: 1017
  • Training: 907
  • Test Set: 110

Training Procedure

The model is trained using a simple yet effective approach, achieving satisfactory performance in just a few epochs. This rapid training process is a testament to the model's efficiency and the quality of the dataset.

Hyperparameters

  • Epochs: 7
  • Batch Size: 16
  • Optimizer: Stochastic Gradient Descent
  • Loss Function: Binary Cross-Entropy with Logits Loss

Training Steps

  1. Load and preprocess the dataset.
  2. Define the CNN architecture.
  3. Compile the model with the chosen optimizer and loss function.
  4. Train the model on the training set while validating on the validation set.

Results

The model demonstrates solid performance in classifying the comic strips. Below are the performance metrics and visualizations of the results.

Performance Metrics

Accuracy: 87%
Loss: 0.01937

Visualization

Test chart Accuracy Chart Figure 1: Test Accuracy Over Epochs

Usage

To use the model for classifying new comic strips, follow these steps:

Clone the repository:

git clone https://github.com/kratss/artist_discriminator

cd artist_discriminator

Install the required dependencies (available via pip, conda, and most distribution repositories):

install matplotlib numpy pytorch

Run the classification script:

  • Place your folder of images in ./images or use the provided images
  • Run python classify.py

Reproduce the model with provided data or your own dataset:

  • Find the original dataset in ./data/ or provide a custom dataset in ./data
  • Run python train.py
  • Receive trained model in working directory

Data Structure

 data
 └─ comics
    ├─ artist_1
    │  ├── img1.jpg
    │  └── img2.jpg
    └─ artist_2
       ├── img1.jpg
       └── img2.jpg

Contributing

Contributions are welcome! If you have suggestions for improvements or additional features, please open an issue or submit a pull request.

License

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