Wood Veneer Defect Detection Based on Multiscale DETR with Position Encoder Net.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Wood veneers are crucial building materials, but defects lead to significant resource waste.
• Current defect detection methods (manual, photoelectric) are inefficient, subjective, or costly.
• Computer vision offers a promising alternative for automated defect detection.

Purpose of the Study:

• To develop an efficient and accurate deep learning pipeline for wood veneer defect detection.
• To address limitations of existing methods, particularly for small defect identification.
• To optimize the detection process for reduced wood resource waste.

Main Methods:

• Collected over 16,380 defect images using a custom device and applied mixed data augmentation.
• Designed a defect detection pipeline based on the DEtection TRansformer (DETR) model.
• Developed a novel position encoding network with multiscale feature maps and redefined the loss function for stable training.

Main Results:

• The proposed method achieves comparable accuracy to existing approaches with increased speed when using a lightweight feature mapping network.
• With a complex feature mapping network, the method demonstrates superior accuracy while maintaining similar speed.
• The pipeline effectively handles the challenges of small object detection in veneer defect identification.

Conclusions:

• The developed deep learning pipeline offers a faster and/or more accurate solution for wood veneer defect detection.
• This approach contributes to reducing wood resource waste through improved defect identification.
• The optimized DETR-based method shows significant potential for industrial application in wood product manufacturing.