Defect Inspector Sample Code (YOLO + OCR Defect Inspection System)

Defect Inspector is an industrial-grade dual-screen visual inspection application developed with Python and Tkinter, specially designed to be paired with Transcend ECM series lenses/cameras. It integrates Ultralytics YOLO for object/defect detection and PaddleOCR for real-time text recognition.

The system utilizes a multiprocessing architecture to ensure smooth rendering of high-resolution camera feeds (30 FPS) while running heavy deep learning models in the background. Features include real-time Region of Interest (ROI) selection, automatic image deskewing (alignment), and one-click training data collection.

• Real-time Video Streaming: Automatically scans and connects to high-resolution video feeds, optimized for Transcend ECM series hardware.
• Dual-Panel UI Design:
  • Left Panel (40%): Live video preview, mode configuration, and ROI selection.
  • Right Panel (60%): High-resolution Gallery view, dynamic glow for defect alerts, and OCR readouts.
• Multi-Model Support: Dynamically switch between different detection modes (e.g., Sticker) without restarting the application.
• Smart Image Processing: Supports multi-target tracking, automated angle calculation, and physical deskewing of cropped images.
• Data Collection Mode: Save the current detection results with a single keystroke (Shortcut S). The system automatically generates YOLO-format annotation files (.txt) for future model training.

• OS: Jetson Linux (Download from https://developer.nvidia.com/embedded/learn/get-started-jetson-nano-devkit#intro).
• GPU: NVIDIA GPU with CUDA installed is highly recommended to significantly improve YOLO and OCR inference speeds.
• Camera: Transcend ECM series lenses/cameras are required to capture high-resolution video via the UVC protocol, ensuring optimal image clarity, precise field of view, and full hardware compatibility.

Please ensure Python 3.8 or higher is installed.

# 1. Clone the repository
git clone https://github.com/transcend-information/Jetson_Print-Defect-Detection_Demo.git
cd defect-inspector

# 2. Create a virtual environment (Highly Recommended)
python -m venv venv

# Activate on Linux:
source venv/bin/activate

# 3. Install required packages
# Note: Adjust the PyTorch installation URL based on your specific CUDA version.
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install ultralytics
pip install paddlepaddle-gpu  # Use 'paddlepaddle' if you do not have a GPU
pip install paddleocr==2.7.3 numpy==1.26.1  #Version compatibility warnings or errors can be safely ignored

Before running the application, you must manually create a models directory and place your trained YOLO models inside. You also need an application icon image.

defect-inspector/
|
|-- demo_camera.py         # Main application script
|
|-- models/                # -> You must create this folder and add models
|   |-- stick_best.pt      # YOLO model for sticker detection
|
|-- dataset_update/        # Auto-generated when pressing 'S' to save training data

(Note: If you only use one mode, simply place the corresponding .pt file in the folder.)

Once your environment is set up and models are in place, start the system with:

python demo_camera.py

Note: The application starts in full-screen mode by default. It will load the YOLO and OCR models in the background. Please wait 10-30 seconds until the status bar at the bottom left displays "Detection model ready".

Look at the CONFIG & PREVIEW panel on the left. Click the radio buttons at the top to switch between detection modes (e.g., Sticker). The system will automatically swap the YOLO model in the background.

1. Ensure the Transcend ECM camera feed is displaying correctly on the left panel.
2. Click and drag your mouse over the left video feed to draw a red bounding box (ROI).
3. Release the mouse button. The system will lock onto this area and begin continuous automatic detection.

As shown in the image below, once the target enters the user-defined red bounding box (ROI), the system automatically isolates the object, physically deskews (straightens) it, and begins defect detection.

The right panel provides a detailed Gallery Inspection view. The system provides clear visual feedback based on the detection results.

PASS State: If no defects are detected, the zoomed inspection gallery will display a green glowing border along with a clear PASS indicator.

FAIL State: If defects are found, the inspection gallery will switch to a red glowing border. The exact locations of the defects will be highlighted with red bounding boxes directly on the deskewed image.

• Mouse Wheel Zoom: Hover over the right panel and use your mouse wheel to zoom in and out for closer manual inspection.
• OCR Readout: The bottom right section automatically formats and displays the text recognized by PaddleOCR.

If you encounter a false positive or a missed detection and want to add it to your dataset for future training:

• Ensure the target object is currently displayed in the right Gallery panel.
• Press the S key on your keyboard.
• The status bar will display "Saved X sticker sample(s)...".
• The cropped images and corresponding YOLO .txt label files will be automatically saved in the dataset_update/ directory.

• Press the ESC key to toggle between full-screen and windowed modes.
• Click the X button on the window (in windowed mode) to close the application. The system will safely terminate background workers and release camera/GPU resources.

1. Q: The bottom status bar is stuck at "Detection model not found: stick".

   • A: Make sure you have created the models folder and placed the stick_best.pt file inside it.

2. Q: The screen is black, and the terminal shows a critical error about not finding a camera device.

   • A:
     1. Check your Transcend ECM camera's connection.
     2. The program scans /dev/v4l/by-path/ and /dev/video* (Linux). If your camera is being used by another application (like OBS or a web browser), close that application and restart this script.

3. Q: PaddleOCR throws an error on Linux (missing libgomp or similar libraries).

   • A: Install the required system dependencies by running: sudo apt-get install libgomp1 libglib2.0-0 libsm6 libxext6 libxrender-dev