Artistic Style Classification and Style Transfer

This project combines classical machine learning and deep learning techniques to perform artistic style classification and neural style transfer. It was created as an academic portfolio piece, practical studying and demonstration of applying computer vision to visual arts.

---

Project Goals

• Build a robust image classifier that can distinguish between six different painting styles:

  • Baroque
  • Cubism
  • Impressionism
  • Northern Renaissance
  • Post-Impressionism
  • Romanticism

• Apply style transfer to real-world photos, recreating them in the style of one of the above artistic genres.

• Explore explainability techniques (e.g., Grad-CAM) to understand model predictions.

---

Dataset

The dataset consists of thousands of paintings labeled with artist names and styles. To improve data quality:

• I filtered the dataset to only include artists with a single, dominant style.
• I selected the six most represented styles and balanced the dataset by limiting samples per class.
• Images were resized and organized into folders for training.

---

Data Analysis and Preprocessing

• Exploratory Data Analysis (EDA) was performed to understand the distribution of artists, styles, and image availability.
• I visualized class imbalance, style frequency, and created a clean folder-based structure for classification.
• Applied basic augmentations and feature extraction (e.g., HOG) for traditional models.

---

Models and Training

Classical Models:

• Random Forest Regressor: ~42% accuracy
• SVM: ~43% accuracy

Deep Learning Models:

• Custom CNN: ~59% accuracy
• ResNet18: ~75% accuracy
• ResNet50: ~74%
• EfficientNetB2: ~75%

The best-performing model was ResNet18, which was saved and used for final inference.

---

Explainability

To interpret the model's decisions, I implemented Grad-CAM, a technique that highlights the image regions the CNN focuses on during classification.

Example: On a graffiti photo, Grad-CAM showed attention on expressive regions, leading the model to (correctly) classify the style as Post-Impressionism.

---

Neural Style Transfer

I implemented the original Neural Style Transfer algorithm (Gatys et al.) using a frozen VGG19 network. This allowed me to generate new images that preserve the structure of a photo while imitating the texture and color style of classical paintings.

Examples include:

• Graffiti stylized as Post-Impressionism

---

Project Structure

art-style-classification/ ├── images/ ├── models/ # Saved models (e.g., resnet18.pt) ├── notebooks/ # Jupyter notebooks for training & analysis └── README.md

Hope you like my work there :)