AI Vision Academy

Access the code of this tutorial, computer vision courses and an exclusive community on AI Vision Academy

• Access to over 50+ source codes from Pysource.com/blog
• Dedicated video courses about computer vision
• Access to an exclusive community of professionals
• Real-World AI Projects – Get hands-on experience building practical AI Computer Vision solutions with a structured path.
• Monthly Coaching Calls – get support and any of your questions answered

Subscribe to our newsletter to learn more

❌

In the fast-paced world of manufacturing and production, real-time defect detection is not just a luxury—it’s a necessity. Whether you’re working with conveyor belt systems or automating quality control, integrating computer vision can drastically enhance efficiency and accuracy.

In this guide, I’ll walk you through how to build a defect detection system from scratch using practical, beginner-friendly methods. This tutorial is based on a real-world implementation and includes all the essentials you need to replicate the system on your own.

Why Real-Time Defect Detection Matters

Automating defect detection helps:

• Maintain consistent product quality.
• Reduce human error in quality checks.
• Save costs associated with manual inspection.

Step-by-Step Overview

1. Data Collection and Preparation

The first step involves collecting video samples of your product on a conveyor belt. Capture videos that include:

• Products in ideal condition.
• Products with simulated or real defects (e.g., scratches, missing labels, wrong caps).

Even if you don’t have real defective items, you can create defects manually (e.g., using a marker or removing parts) to begin building your training dataset.

Tips:

• Record under consistent lighting conditions.
• Use a stable camera setup.
• Ensure the product is clearly visible.

2. Object Detection Setup

Start by isolating the product from its background. This significantly reduces noise and increases model accuracy.

Steps:

• Extract frames from the recorded video.
• Use annotation tools (e.g., Roboflow, MakeSense.ai) to draw bounding boxes around each object of interest (e.g., bottle, label, cap).
• Assign class labels to each object.

You will train a YOLO (You Only Look Once) model using these annotated images. YOLO is effective due to its speed and ability to process images in real-time.

3. Training the Detection Model

Using your labeled dataset, train a YOLO object detection model.

Tools:

• Jupyter Notebook (for reproducibility and clarity).
• Ultralytics’ YOLOv8 (or the latest version).

Process:

• Load your annotated dataset into the notebook.
• Select the detection project type.
• Train the model and monitor performance metrics.

This produces a .pt file (e.g., best.pt), which is the trained model ready for deployment.

4. Real-Time Video Analysis with OpenCV

Load your YOLO model into a Python script using OpenCV to process your video in real time.

Core steps:

• Read the video frame-by-frame.
• Apply the YOLO model to detect objects.
• Draw bounding boxes around detected items.
• Display frames for visual verification.

This confirms that your model correctly identifies the product components under real-world conditions.

5. Implement Object Tracking

Detection alone isn’t sufficient. You need to track objects across frames to:

• Assign unique IDs to products.
• Avoid repeated inspections of the same item.

Method:

• Use a Multi-Object Tracking algorithm (e.g., SORT, Deep SORT).
• Update object states in real-time using bounding box data.
• Maintain consistency of object IDs across frames.

6. Frame Extraction and ROI Handling

To inspect specific parts, define a virtual line (e.g., mid-screen). When an object crosses this line:

• Crop the object’s region of interest (ROI).
• Save this image for further classification.

Advantages:

• Simplifies classification by reducing input variability.
• Ensures inspection occurs when the object is fully visible.

7. Classification of Detected Components

Train separate classification models for each component (e.g., label, cap, bottle) to identify defect types.

Preparation:

• Sort images into folders by class (e.g., good, scratch, wrong label).
• Train classification models using the same Jupyter interface.

Classification Output:

• Label each ROI as good or defective.
• Use multiple categories for detailed defect types.

8. System Integration and Feedback Loop

Combine detection, tracking, and classification into a unified pipeline.

Actions:

• Display green boxes for good items and red for defective ones.
• Log classification results for quality reports.
• Optionally integrate with automation systems (e.g., trigger alerts, stop conveyor belts).

Why Join Our Skool Community?

This guide gives a surface-level view of what’s possible. Inside our AI Vision Academy, you’ll get:

• Access to full, in-depth courses on computer vision and AI.
• Exclusive tools and Jupyter Notebooks to simplify implementation.
• Support and feedback from me and other practitioners.
• Real-world projects and collaborative learning opportunities.

Whether you’re a beginner or looking to refine your skills, our academy equips you with the practical know-how to build robust AI systems from scratch.

Start Building Today
Don’t wait for the perfect dataset or ideal conditions. Begin with what you have. As shown, even a few annotated images and basic video footage are enough to kickstart your journey in defect detection systems.

👉 Join the AI Vision Academy Today