Real-Time Fabric Defect Detection on Jetson Orin Nano

Project Overview

This project implements a real-time fabric defect detection system using purely classical computer vision techniques, deployed and optimized on the Jetson Nano. The system avoids deep learning entirely and relies on multi-method classical analysis combined with IoU-based bounding box fusion for robust and interpretable defect localization. The solution is designed for low-power edge devices and is suitable for industrial textile inspection where cost, explainability, and real-time performance are critical.

---

Problem Statement

Manual fabric inspection is:

• Time-consuming
• Inconsistent
• Prone to human error

While deep learning–based solutions exist, they:

• Require large labeled datasets
• Demand high computational resources
• Are expensive to deploy

Objective:
Develop a low-cost, real-time, and interpretable fabric defect detection system using classical computer vision, deployable on an edge device without any training or labeled data.

---

Methodology

The system uses multiple independent classical detectors, each capturing different defect characteristics. Their outputs are fused using an IoU-based strategy to produce stable final detections.

Classical Techniques Used

• GLCM Texture Analysis – detects texture irregularities
• FFT Frequency Analysis – detects disruptions in periodic patterns
• Gabor Wavelets – captures directional and repetitive textures
• Statistical Local Variance – highlights abrupt anomalies
• Background Subtraction – detects stains and fading
• Edge Detection + Hough Transform – detects linear defects

Fusion Strategy

• Bounding boxes from all detectors are merged using Intersection over Union (IoU)
• Size-based and border-based filtering removes false positives

---

System Pipeline

Camera Frame
↓
Preprocessing (Grayscale, Normalization)
↓
Parallel Classical Detectors
↓
Bounding Box Extraction
↓
IoU-Based Box Fusion
↓
Post-Processing Filters
↓
Final Defect Localization

---

Jetson Orin Nano Deployment

• Real-time processing using live USB camera feed
• Resolution optimized to 640×480 for performance
• No disk I/O during runtime
• Lightweight classical algorithms ensure stable FPS

This makes the system suitable for on-device industrial inspection.

---

Performance Highlights

• Fully real-time execution on Jetson Nano
• No model training or dataset labeling required
• Low power consumption
• High interpretability and explainability

---

Industrial Applications

• Textile manufacturing quality control
• Automated inspection on production lines
• Low-cost alternatives to GPU-based vision systems
• Edge-based visual monitoring systems

Future Work

• Detector confidence voting
• Further runtime optimization
• Optional deep learning verifier
• Integration with industrial conveyor systems

---

Demo

https://github.com/user-attachments/assets/fac8b5cb-90fb-4048-a1f4-c30167beda15

https://github.com/user-attachments/assets/c4d4eb7a-723e-45d2-abae-a23ca4ae641c

License

This project is licensed under the MIT License.

---

Author

Awais Asghar Electrical Engineering | Computer Vision | Edge AI

⭐ If you find this project useful, consider giving it a star! ⭐