AI Defect Inspector — NCR System

A Raspberry Pi-powered system that uses a camera and AI to inspect manufactured parts for surface defects. When a defect is detected, a Non-Conformance Report (NCR) form auto-populates — saving time and reducing human error.

---

How It Works

Camera captures image
        ↓
AI model analyzes for defects (98.6% accuracy)
        ↓
  Defect found? ──No──→ PASS (green)
        ↓ Yes
NCR form auto-fills with:
  - Defect type (scratch, dent, crack, stain, chip)
  - Confidence score
  - Inspection ID + timestamp
        ↓
Operator reviews and submits NCR

---

Features

• Real-time defect detection using MobileNetV2 (TensorFlow Lite)
• 6 defect categories: scratch, crack, dent, stain, chip, good
• Web-based UI — open on any device on the same network
• NCR form auto-fills from AI detection results
• Upload mode for testing without a camera
• NCR history with severity tracking
• 82ms inference time on Raspberry Pi 4

---

Project Structure

├── app.py                  # Flask web server
├── detect.py               # TFLite inference engine + camera
├── train_model.py          # MobileNetV2 training → TFLite
├── capture_training.py     # Pi camera training data collection
├── prepare_dataset.py      # Dataset preparation script
├── templates/
│   └── index.html          # Web UI
├── requirements.txt
├── PROJECT_GUIDE.md        # Full beginner setup guide
└── TRAINING_LOG.md         # Step-by-step learning notes

---

Quick Start

1. Clone and install

git clone https://github.com/crackcode09/ai-defect-identify.git
cd ai-defect-identify
pip install -r requirements.txt

2. Prepare dataset

Download the NEU Surface Defect Database from Kaggle, extract to archive/, then:

python prepare_dataset.py

3. Train the model

python train_model.py

Training takes ~10 min on a laptop GPU, ~60 min on Raspberry Pi 4.

4. Run the app

# With camera (Raspberry Pi)
python app.py

# Without camera (upload mode, for testing on PC)
python app.py --no-camera

Open http://localhost:5000 in your browser.

---

Hardware (Raspberry Pi deployment)

Item	Purpose
Raspberry Pi 4 (4GB+)	Main computer
Pi Camera Module v2 / HQ	Image capture
MicroSD Card (32GB+)	Storage
USB-C Power Supply (5V/3A)	Power

See PROJECT_GUIDE.md for full hardware setup and Pi deployment steps.

---

Tech Stack

• AI Model: MobileNetV2 (transfer learning) → TensorFlow Lite
• Backend: Python + Flask
• Frontend: Vanilla HTML/CSS/JS
• Camera: picamera2 (Pi) / OpenCV (webcam fallback)
• Training dataset: NEU Surface Defect Database

---

Model Performance

Metric	Value
Validation accuracy	98.6%
Inference time (Pi 4)	~83ms
Training images	1,440
Test images	360
Categories	6