artstyle-classifier

AI project to classify artworks by painting style.

🎨 ArtStyle Classifier

A machine learning project leveraging AI and deep learning (CNNs) to classify paintings into 5 distinct artistic styles.
Built by Ana Menkshi and Fatjona Murrani, this system also includes an interactive web app for real-time predictions.

Features

• Classifies paintings into 5 major artistic styles
• Uses a fine-tuned ResNet50 backbone with custom layers
• Includes an easy-to-use web app (Gradio / Streamlit)
• Deployed on Hugging Face Spaces for public use

Try the App: ArtStyle Classifier on Hugging Face The username:art ,the pasword:art.

Dataset

• Dataset: painter-by-numbers.zip
• Images: .jpg files of paintings
• Metadata: .csv file with painting info
• Filtered to 5 distinct styles to improve model learning

---

Preprocessing & Augmentation

• Images resized to 224x224 pixels
• Normalized pixel values [0, 1]
• Augmentation techniques:
  • Rotation (±10°)
  • Shear (5%)
  • Horizontal & Vertical Flips
  • Rescaling
• Dataset split: 80% training / 20% validation

---

Model Architecture

• Base Model: ResNet50 (include_top=False)
• Custom Layers:
  • Flatten ➔ Dense(512, He init) ➔ Dropout ➔ BatchNorm ➔ ReLU
  • Dense(16) ➔ Dropout ➔ BatchNorm ➔ ReLU
  • Output Layer: Softmax (5 classes)

Training Config:

• Optimizer: Adam (lr=0.0001)
• Loss: Categorical Crossentropy
• Batch size: 32
• Metrics: Accuracy

---

Web App

• Built with Gradio
• Features:
  • Upload a painting ➔ Get style prediction
  • User-friendly and responsive interface
• Deployed via Hugging Face Spaces

---

Installation

1. Clone this repository:

git clone https://github.com/AnaMenkshi/artstyle-classifier.git
cd artstyle-classifier