🎨 Art2Artist Recognizer

An end-to-end deep learning image classifier that identifies the master behind any painting — spanning data collection, cleaning, model training, deployment, and API integration.

Art2Artist Recognizer — Web Interface

📦 Dataset Preparation

Data Collection was sourced from two places:

• DuckDuckGo — images scraped by artist name
• Kaggle WikiArt Dataset — curated art repository

Stage	Image Count
Raw collected	~20,000
After cleaning	~16,000

DataLoader — Built with the fastai DataBlock API.

Data Augmentation — fastai's default GPU-accelerated augmentation pipeline:

• Random resized crops (min_scale=0.5)
• Multiple augmentations with 2.0 multiplier

See notebooks/data_prep.ipynb for full details.

---

🏋️ Training & Data Cleaning

Training

• Transfer learning with pretrained ResNet architectures
• Fine-tuned over 3 cycles × 5 epochs each

Data Cleaning

The most time-consuming phase of the project. Images scraped from DuckDuckGo required extensive manual review and cleaning before being merged with the WikiArt dataset to produce high-quality, relevant training examples.

---

📊 Model Comparison

Model	Architecture	Dataset Size	Accuracy
Model 1	ResNet34	16,000 images	88%
Model 2	ResNet50	16,000 images	90% ✅

ResNet50 was selected as the final model — its deeper architecture enables richer feature extraction, capturing the subtle stylistic signatures that distinguish each artist.

---

🚀 Deployment

The model is deployed as a Gradio app on Hugging Face Spaces.

Gradio App — Deployed on Hugging Face Spaces

🔗 Live app: huggingface.co/spaces/goldphish2209/art2artist-recognizer

Implementation details in the deployment/ folder.

---

🌐 API Integration (GitHub Pages)

The Gradio model API is integrated into a GitHub Pages website for browser-based access without any local setup.

Art2Artist — GitHub Pages Live Demo

🔗 Website: naawshin.github.io/Art2Artist-Recognizer

Implementation details in the docs/ folder.

---

🛠️ Tech Stack

Component	Technology
Deep learning framework	fastai + PyTorch
Model architectures	ResNet34, ResNet50
Data sources	DuckDuckGo scraping, Kaggle WikiArt
Deployment	Hugging Face Spaces + Gradio
Web integration	GitHub Pages
Training strategy	Transfer learning + fine-tuning

---

✨ Summary

• 🧠 Built with fastai and PyTorch
• 🔁 Transfer learning with ResNet34 and ResNet50
• 🖼️ Curated and cleaned 20k+ images from DuckDuckGo and Kaggle WikiArt
• 🎯 Achieved up to 90% accuracy in artist classification
• ☁️ Deployed on Hugging Face Spaces with Gradio
• 🌐 Integrated via API on GitHub Pages for web-based access

---

👩‍💻 Author

Nowshin Tabasum — AI Engineer