Datathon 2023 Case Competition

Host: KU Leuven & LStat

Team: The Curse of Dimensionality

Topic: The code in this repository investigates possible factors that may help an existing GAN model discern between real and artificially generated artworks. The project was awarded Best Modeling Datathon 2023.

Project overview

1. Preprocessing

   • Image preprocessing
     • Remove censored AI generated images
     • Remove WikiGallery border
     • Watermark cropping
   • Tabular data
     • Feature inspection

2. Feature engineering

   • Artist & artwork clustering (Neo4j)
   • Number of recommendations per artwork as a proxy for popularity (Neo4j)
   • Image feature extraction (brightness, contrast, edge, saturation)

3. Image classification

   • assuming:
     • equal misclassification costs
     • equal number of AI generated and traditional artworks
   • Final models:
     • GAN classifier (ResNet50 NoDown)
     • Image feature classifier (logistic regression)
     • GAN + image feature classifier (logistic regression)

4. Model inspection using logistic regressions and XGBoost

Data information

The datasets, provided by KU Leuven, can be found in the FilesCSVFormat folder and include:

• Data scraped from Wikipedia and meta-data of artists and artworks
• Links to WikiGallery images for each artwork
• Dataframe of AI generated images using a text-to-image model
• Information and Entity-Relation diagram on the datasets provided

Documentation

• Data Processing: tabular processing and image preprocessing of scripts and dataframes
• GAN Detection: outputs using an existing GAN model based on resnet50
• Image Features: outputs of feature extractions of images
• Data Analysis: exploratory analyses using logistic regression and XGBoost
• Pitch: slides and images