Industrial Part Surface Defect Inspection and Analysis System

A factory-dashboard application for metal surface defect localization and technician-ready quality diagnosis. The system combines a PyTorch/YOLO-style detector, an OpenAI-compatible LLM report generator, and a Streamlit web interface.

Architecture

data/               Dataset scripts and processing utilities
models/             Deep learning architecture and optional weights
  defect_detector.py
llm/                Prompt templates and OpenAI-compatible API client
  report_generator.py
app.py              Streamlit dashboard
requirements.txt    Python dependencies
run.bat             Windows one-click launcher
run.sh              Linux/macOS one-click launcher

Features

• Safe CUDA initialization with automatic CPU fallback.
• YOLOv8-compatible inference path when local weights are available.
• Deterministic mock inference when weights are missing, returning high-confidence metal-surface defects.
• OpenAI-compatible LLM API integration through LLM_API_KEY, LLM_BASE_URL, and LLM_MODEL.
• Factory-dashboard style Streamlit interface with upload, bounding-box visualization, metrics, and quality report output.

Installation

Windows

run.bat

Linux/macOS

chmod +x run.sh
./run.sh

Both launchers create .venv, install dependencies from requirements.txt, and start:

streamlit run app.py

Optional LLM Configuration

Set these environment variables before launch if you want live LLM reports:

export LLM_API_KEY="your-api-key"
export LLM_BASE_URL="https://your-openai-compatible-endpoint/v1"
export LLM_MODEL="your-model-name"

If no API key is configured, the application uses an offline industrial quality report template.

Optional Real Detector Weights

Place trained YOLO weights at:

models/weights/best.pt

If weights are unavailable or fail to load, the detector automatically uses mock mode so the application remains runnable for demos and grading.