An Improved YOLOv 11 Model for Image Processing to Detect Wood Deformation

Ng, Kah Fei (2025) An Improved YOLOv 11 Model for Image Processing to Detect Wood Deformation. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text NG KAH FEI_FULL TEXT.pdf Restricted to Registered users only Download (7MB)

Abstract

This research presents a machine vision-based methodology for detecting fine and lowcontrast surface defects on black-painted wood, such as cracks, dents, and gesso lines. Manual inspection in the wood industry is time-consuming, error-prone, and inconsistent, especially for subtle defects. To address this, a comprehensive detection pipeline was developed, beginning with high-resolution image acquisition using a line scan camera under controlled lighting. Preprocessing techniques included tile cropping, Real-ESRGAN-based image enhancement, and data augmentation to increase defect visibility and dataset diversity. A YOLOv11n deep learning model was used as the base detector, selected for its efficient performance and adaptability to small object detection. Enhancements were made using pseudo-labelling to handle class imbalance, CBAM attention modules to improve feature focus, and a custom loss function combining Focal Loss and SIoU to better localize subtle anomalies. The model was trained using PyTorch and evaluated using mAP metrics. Results demonstrated improved accuracy in detecting rare and finegrained defects, particularly gesso lines, with significant gains in mAP@0.5 and mAP@0.5:0.95. The proposed framework offers a scalable and accurate solution for automatedwood quality control and shows potential for deployment in real industrial environments.