Overview

This project implements a production-style deep learning defect detection system demonstrating the complete model lifecycle, including training, versioned model registry, real-time inference, retraining pipelines, and deployment-ready APIs.

The system detects industrial surface defects using a YOLO-based object detection model and exposes a FastAPI inference service with a lightweight Streamlit visualization interface.

---

Key Highlights

• End-to-end DL lifecycle (training → registry → inference → retraining)
• Versioned model registry with dynamic model loading
• Config-driven inference behavior
• Automated retraining and batch inference pipelines
• Observability endpoints for health and version tracking
• Visualization UI for interactive defect detection
• Docker-ready inference service

---

Project Structure

defect-detection-dl-system/
├── training/
├── inference/
├── pipelines/
├── models/registry/
├── streamlit_app.py
└── system_design.md

---

Running the Project

Start backend:

uvicorn inference.app:app --reload

Launch UI:

streamlit run streamlit_app.py

---

Running with Docker

docker -compose up --build

---

Dataset

The dataset is not included due to size. Refer to external dataset and preprocessing pipeline for YOLO conversion.

---

Tech Stack

• Python
• PyTorch / YOLO
• FastAPI
• Streamlit
• Docker

---

Documentation

Detailed architecture, design decisions, lifecycle considerations, and failure analysis are available in:

👉 system_design.md

---