August 5, 2021

Overview

Recognize Any Artwork AI began as an experiment and evolved into a production-grade artwork recognition system. Designed to run efficiently on CPU-only environments, the model uses pickled weights for sub-second inference while maintaining over 95% accuracy. Inspired by the Humpback Whale Identification Kaggle Competition, the system was adapted for art authentication and deployed on China-based servers for Artshield, powering a real-world WeChat Mini Program.

Key Features

• Optimized CPU Inference – The model loads instantly from memory via pickled weights, enabling fast recognition without GPUs.
• Scalable Deployment – Built with FastAPI and Celery, containerized in Docker, and orchestrated via Kubernetes on a serverless Alibaba Cloud infrastructure.
• Dataset Augmentation – Enhanced artwork recognition using fast.ai to address data scarcity and improve generalization.
• WeChat Mini Program Integration – Real-time inference was seamlessly integrated with a Mini Program for user-friendly interaction.

Technologies Used

• PyTorch, fast.ai, Pickle – Training, augmentation, and model storage.
• OpenCV, Pillow – Image preprocessing.
• FastAPI, Celery – Backend for asynchronous inference.
• Docker, Kubernetes, Terraform – Scalable cloud deployment.
• Alibaba Cloud – Serverless hosting for production inference.

Outcome

This system proved that AI-powered artwork recognition can be both fast and cost-efficient, running entirely on CPU without sacrificing accuracy. Deployed via a WeChat Mini Program, it successfully scaled in production, making AI-driven authentication accessible to real users. See the repo.