Publication: Enhanced yolo-based nondestructive inspection of wood defects and grading system
Loading...
Date
2024-12-01
Authors
Liew, Pei Yi
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Wood defect inspection is crucial for maintaining structural integrity and
aesthetics in the construction and furniture industries, reducing wastage, and ensuring
product quality. Current machine vision systems using Machine Learning (ML)
algorithms require complex preprocessing, while Deep Learning (DL) algorithms face
issues such as lack of feature prioritization, fixed receptive fields, and poor aggregation.
Therefore, this research develops an efficient defect inspection system using the You
Only Look Once version 8 (YOLOv8) algorithm integrated, followed by a wood
grading system based on its output. A custom wood defect dataset with cracks, dead
knots, pinholes, and undersized defects is prepared and enhanced with data
augmentation techniques. The YOLOv8 is modified by integrating a Convolutional
Block Attention Module (CBAM_spatial), Reparametrized Convolution with Channel
Shuffle and One-Shot Aggregation (RCS-OSA), and the Weighted Intersection over
Union (WIoUv3) loss function, achieving 97% precision, 97.3% recall, 98.9%
mAP0.5, 83.1% mAP0.5:0.95, and 84.87FPS, outperforming other state-of-the-art
algorithms by improving the ability of feature learning and aggregation. Additionally,
an automated wood visual grading system is developed, accurately classifying wood
samples into Grades A, B, and C, achieving 93% accuracy, 7.11% Mean Absolute
Percentage Error (MAPE) and 89.79% average Intersection over Union (IoU). This
research provides a new strategy to enhance wood defect detection and grading
efficiency and precision, significantly improving operational effectiveness and product
quality in the wood industry.