🔍 Pipeline Crack Detection

A deep-learning-based system for detecting metal surface defects such as cracks, corrosion/rust, scratches, and holes on industrial pipelines. Built using YOLOv8 Nano for real-time, lightweight inference — optimized for edge deployment on devices like the Raspberry Pi.

---

📋 Table of Contents

• Overview
• Features
• Defect Classes
• Dataset
• Project Structure
• Installation
• Usage
  • Label Generation
  • Training
• Model Details
• Future Scope
• License

---

Overview

Industrial pipelines are critical infrastructure assets that require regular inspection to prevent catastrophic failures. Manual inspection is time-consuming, expensive, and error-prone. This project automates the detection of surface-level defects using computer vision, enabling faster and more reliable pipeline health assessments.

The system trains a YOLOv8 Nano object-detection model on labeled images of metal surfaces and is designed to run efficiently on resource-constrained hardware for in-field deployment.

---

Features

• Multi-class defect detection — identifies cracks, rust, scratches, holes, and normal surfaces.
• Lightweight model — YOLOv8 Nano architecture for fast inference on edge devices.
• Edge-ready — optimized image size (320×320) and small batch training for Raspberry Pi deployment.
• Automated labelling — utility script to generate YOLO-format labels from class-organized image folders.
• Roboflow integration — supplementary crack-detection dataset sourced from Roboflow.

---

Defect Classes

Class ID	Defect Type	Description
0	Crack	Fractures or fissures on the surface
1	Hole	Punctures or perforations in the metal
2	Rust	Corrosion or oxidation patches
3	Scratch	Surface abrasions or score marks
4	Normal	Defect-free surface (healthy baseline)

---

Dataset

The project uses a combination of:

1. Custom industrial defect dataset — images organized by class into train/ and val/ splits with YOLO-format bounding-box labels.
2. Roboflow crack dataset — 1,551 annotated crack images exported in YOLOv8 format (no augmentation applied).

Directory Layout

dataset/
├── images/
│   ├── train/       # Training images
│   └── val/         # Validation images
├── labels/
│   ├── train/       # Training labels (YOLO format)
│   └── val/         # Validation labels (YOLO format)
└── runs/            # Dataset-level experiment runs

Label format (YOLO): <class_id> <x_center> <y_center> <width> <height> (normalized 0–1)

---

Project Structure

Pipeline crack detection/
│
├── dataset/              # Images and labels (train/val splits)
├── roboflow_data/        # Supplementary Roboflow crack dataset
├── runs/                 # Training run outputs (weights, metrics)
│   └── detect/
│       ├── train/        # First training run
│       └── train-2/      # Second training run
│
├── labelling.py          # Auto-generates YOLO labels from folder structure
├── train.py              # Model training script
├── data.yaml             # Dataset configuration for YOLO
├── requirements.txt      # Python dependencies
├── .gitignore            # Git ignore rules
└── readme.md             # This file

---

Installation

Prerequisites

• Python 3.8+
• NVIDIA GPU with CUDA support (recommended for training)
• pip

Setup

# Clone the repository
git clone https://github.com/<your-username>/Pipeline-crack-detection.git
cd Pipeline-crack-detection

# Create and activate a virtual environment
python -m venv yoloenv
# Windows
yoloenv\Scripts\activate
# Linux / macOS
source yoloenv/bin/activate

# Install dependencies
pip install -r requirements.txt

---

Usage

Label Generation

If your images are organized by class in a folder structure, use labelling.py to auto-generate YOLO-format label files:

python labelling.py

Note: By default, the script processes images in industrial_defect_dataset/val/ and writes labels to industrial_defect_dataset/labels/val/. Update the base_path and output_labels variables in the script for your specific paths.

Training

Train the YOLOv8 Nano model using the configured dataset:

python train.py

Default training configuration:

Parameter	Value
Base model	yolov8n.pt
Epochs	50
Batch size	8
Image size	320 × 320
Workers	2
Device	GPU (device=0)
Output	runs/pipeline_defect_model

Trained weights are saved to runs/detect/pipeline_defect_model/weights/.

---

Model Details

Property	Value
Architecture	YOLOv8 Nano (yolov8n)
Task	Object Detection
Input resolution	320 × 320 px
Number of classes	5
Framework	Ultralytics
Target hardware	Raspberry Pi / edge devices

---

Future Scope

• 🎥 Real-time video inference — integrate with a live camera feed for on-site pipeline inspection.
• 📊 Performance benchmarking — evaluate mAP, precision, recall, and inference latency across hardware.
• 🔄 Data augmentation — apply mosaic, mixup, and geometric transforms to improve robustness.
• 🚀 Model export — convert to ONNX / TensorRT / TFLite for accelerated edge inference.
• 🤖 Drone / robot integration — mount the inference pipeline on autonomous inspection platforms.
• 📈 Active learning — iteratively improve the model with newly collected field data.

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

Note: The supplementary Roboflow dataset component is under a separate Private license as specified by Roboflow.

---

Built with ❤️ using Ultralytics YOLOv8