🎨 Gallery Genie - AI Art Classifier & Recommendation System

An intelligent art classification and recommendation system that combines computer vision and similarity search to help users discover artworks. Built with a fine-tuned VGG16 model and FAISS vector search on a massive 500k artwork dataset.

Features

Art Classification

• Style Recognition: Classifies artwork into different artistic styles and movements
• Top-5 Predictions: Returns the 5 most likely style classifications with confidence scores
• Real-time Processing: Fast inference using optimized YOLO and VGG models

Smart Recommendations

• Visual Similarity: Finds artworks visually similar to uploaded images
• Feature Extraction: Uses VGG16 block5_pool features with Global Average Pooling
• FAISS Search: Lightning-fast similarity search across 500k+ artworks
• Style Filtering: Recommends artworks from similar artistic styles

Performance

• Scalable: Handles large-scale artwork databases efficiently
• Fast API: RESTful API built with FastAPI for production deployment
• Optimized Models: Custom VGG16 architecture fine-tuned for art classification

Architecture

Models Used

• Classification: Custom YOLO model fine-tuned on artwork dataset
• Feature Extraction: Modified VGG16 with Global Average Pooling
• Similarity Search: FAISS index for efficient nearest neighbor search

Data Pipeline

1. Image Upload: Users upload artwork images via API
2. Classification: YOLO model predicts artistic style
3. Feature Extraction: VGG16 extracts visual features
4. Similarity Search: FAISS finds most similar artworks
5. Recommendations: Returns curated list of similar artworks

Installation

Requirements

• Python 3.8+
• FastAPI
• TensorFlow/Keras
• OpenCV
• FAISS
• Ultralytics YOLO
• pandas, numpy

Setup

Install required packages and download the pre-trained models:

• YOLO classification model: best-87810.pt
• VGG16 feature extractor: models_20240613-101029.keras
• FAISS similarity index: faiss_index.bin
• Metadata: meta.csv

API Endpoints

• Used OpenAI api to fetch information about all of the art pieces, year of creation, author, and meaning behind.
• Used Streamlit API to host the website.

POST /upload_image

Upload an artwork image for classification and recommendations.

Request: Multipart form data with image file

Response:

• pred_label: Primary style classification
• top_5_names: Top 5 style predictions
• most_similar: List of similar artworks with metadata and distances

GET /

Health check endpoint returning system status.

Dataset

Trained on a comprehensive 500k artwork dataset featuring:

• Multiple artistic styles and movements
• Diverse time periods and cultural backgrounds
• High-quality artwork images
• Rich metadata including artist, style, and period information

Technical Details

Feature Extraction

• Uses VGG16 block5_pool layer output
• Global Average Pooling for dimensionality reduction
• 512-dimensional feature vectors

Similarity Search

• FAISS index for efficient nearest neighbor search
• Cosine similarity for artwork comparison
• Top-k retrieval with configurable k value

Model Architecture

• Fine-tuned VGG16 backbone
• Custom classification head for art styles
• Optimized for both accuracy and inference speed

Use Cases

• Art Discovery: Help users find new artworks similar to their preferences
• Style Analysis: Classify and categorize artwork collections
• Museum Applications: Digital art recommendation systems
• Art Education: Learn about different artistic styles and movements
• Collection Management: Organize and search large art databases

Deployment

The system is production-ready with:

• FastAPI backend with CORS support
• Async image processing
• Scalable architecture
• RESTful API design

Performance Metrics

• Classification Accuracy: High accuracy on art style recognition
• Search Speed: Sub-second similarity search on 500k+ database
• Scalability: Handles concurrent requests efficiently
• Model Size: Optimized for production deployment