Archaeological Artefact Classification using Convolutional Neural Network (CNN) and Vision Transformer

Known platform support

• Linux Manjaro
• Linux Ubuntu
• Windows WSL
• Windows 11

Hardware requirements

Nvidia GPU:

• PyTorch's primary support is for CUDA, which is Nvidia's parallel computing platform. To leverage GPU acceleration and parallelization tasks, an Nvidia GPU is required.
• Ensure your Nvidia GPU is CUDA-enabled, as this is essential for PyTorch operations.

CUDA Toolkit:

• Install the appropriate version of the CUDA Toolkit that is compatible with your Nvidia GPU and PyTorch. This toolkit is essential for parallel computing on Nvidia GPUs.

Overview

This project aims to classify archaeological artefacts using two state-of-the-art models: Convolutional Neural Network (CNN) and Vision Transformer (ViT). The primary goal is to explore strengths and weaknesses of each approach in the context of archaeological artefact classification. Additionally, in the future we want to focus on object regeneration i.e. the possibilities of reconstructing damaged artefacts.

Convolutional Neural Network (CNN)

Implementation details

• Input layer: RGB images of archaeological artefacts.
• Convolutional Layers: Capture local features.
• Pooling layers: Reduce spatial dimensions.
• Fully Connected Layers: Make classification decisions.

Vision Transformer (Vit)

Implementation details

• Input Embedding: Splits the image into patches.
• Positional Encoding: Adds spatial information to the patches.
• Add learnable class/token parameter to the input sequence.
• Transformer Body: Apply self-attention mechanism (Multi-head attention) for feature connections.
• Classification Head: Makes final predictions.

Dataset

The dataset is a diverse collection of archaeological artefacts recieved from the DIME database maintained and provided by Moesgaard Museum Aarhus, Denmark. Annotation for image classes is provided. Check out the DIME database here.

Paper

We have worked out a first draft paper for the solution in classifying coin objects in images from the DIME database. You can find the paper here. ADD LINK!

Installation and Launch

This script automates the installation of dependencies and the launches the project in Jupyter lab.

Table of Contents

• Introduction
• Installation
• Usage
• Troubleshooting

Introduction

This installation process works for: Linux Manjaro and Ubuntu.

Installation

To install and setup the project run the following commands:

First clone the project from Github:

	git clone git@github.com:5kj41d/Classification-and-3D-reconstruction-of-archaeological-artifacts.git

Locate to the project folder:

	cd Classification-and-3D-reconstruction-of-archaeological-artifacts

To install the dependencies for the Jupyter project run:

	./installDependenciesScriptLinux.sh

Note: Check out the requirements.txt file for full overview of installed Python packages.

Usage

To run the project use the following command in the root project folder:

	./launchJupyterProjectLinux.sh

Troubleshooting

If you encounter any issues during installation or launch, consider the following:

Make sure you are located in the root project folder Classification-and-3D-reconstruction-of-archaeological-artifacts

• Issue: If you you cannot enter the project folder with provided cd command:
  • Solution: Locate the project folder where you installed it and open a terminal inside.

	sudo find / -type d -name Classification-and-3D-reconstruction-of-archaeological-artifacts

• Issue: If the virtual enviroment wont activate:
  • Solution: Enter the linuxPythonVenv/bin folder and run

	source activate

Then try to install again.

• Issue: If the installation script is finding a virtual enviroment but installing on the system path can be handled bu removing the linuxPythonVenv folder and create it again:
  • Solution: Try following commands: Remove:

	rm -rf linuxPythonVenv 

Create:

	python -m venv linuxPythonVenv 

Activate:

	source linuxPythonVenv/bin/activate

• Issue: If the error of exernal management is occuring may be due to missing Python Venv package.
  • Solution: If the script do not install this, this can be done manually. Then try install process again.

Sources

A non-exhaustive list of sources used for the project and inspirational sources for potential aspirants to work on simular projects.

• An Introduction to Convolutional Neural Networks
  • Introduction to the basic theory behind the Convolutional Neural Network (CNN) model and architecture. Starting point in learning and using the CNN model for classification tasks.
• A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects
  • A more in-depth theory and analysis of what and how CNN works.
• AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE
  • This source provided background on using Vision Transformers for image classification.
• Attention Is All You Need
  • As background knowledge on Transformer model for NLP tasks. More extensive theory of the work behind such models.