Optimizing Machine Vision System Using Machine Learning for Defect Detection in a Furniture Production Line

Abstract

Furniture manufacturing industry in Malaysia has been heavily relaying on manual visual inspection for defect detection such as surface defects, scratches and dents. This approach is easily prone to human error leading to inconsistent quality control and intensive labour. In this study, a dual-stage machine learning framework is proposed to automate the detection and classification of surface defects in wooden materials using image analysis. This framework combines YOLO-World and Mask R-CNN models. Mask R-CNN was first integrated for pixel-level segmentation and detailed verification of defect regions, while YOLO-World model was subsequently used to recover any missed detections as well as categorize the corresponding wood types. Four common defect types were studied which are edge defects, scratches, surface anomalies, and discoloration. The proposed system was evaluated on a custom dataset comprising annotated images of various wood types and defects, achieving a mean Average Precision (mAP@50) of 95.6% and a (mAP@50-95) of 92.3% for bounding box detection and yielded an (mAP@50) of 95.6% for instance segmentation. This study demonstrates the feasibility and effectiveness of integrating dual-stage object detection and segmentation techniques to support automated visual inspection for different wood types. The proposed framework contributes toward improved quality control and operational efficiency in wood-based manufacturing.