Page 34 - FDMAsia Sep/Oct 2026

32 TECHNOLOGY Wood Surface Defect Segmentation
SEP/OCT 2026 FDM ASIA | www.fdmasia.com

the Creative Commons Attribution-Share Alike 3.0 Unported In the domain of computer vision, image segmentation

techniques are predominantly categorised into two paradigms:
semantic segmentation and instance segmentation.
Semantic segmentation operates at the pixel level, assigning
each pixel to a predefined defect category, such as live knots,

dead knots, or resin pockets.
Instance segmentation extends this capability by not only
performing pixel-wise classification but also differentiating
between distinct individual instances within the same category.

Semantic segmentation
fibres, typically due to external stresses. They reduce the Mohsin et al. introduced a novel framework for real-time
wood’s shear strength parallel to the grain and adversely detection of wood planks on a high-speed conveyor system,
affect its overall structural integrity. Moreover, checks are integrating a CNN network for simultaneous surface defect

susceptible to fungal infection, which can lead to rot and segmentation.
progressive deterioration of the wood. The proposed architecture comprises three core components:
a backbone network for feature extraction, a detection algorithm
Deep Learning Models for plank identification, and a segmentation module for real-
Similar to general computer vision tasks, deep learning time defect analysis.
techniques for wood surface defect inspection can be broadly Zhong et al. proposed a deep Gaussian attention segmentation
classified into three categories: classification, detection, and network for lumber surface defect detection.

segmentation. The network first employs a self-attention mechanism-
Defect classification methods identify the presence and incorporating both position and channel attention modules-
type of defects in wood surface images. Defect detection to aggregate contextual information. It then integrates two
methods extend beyond classification by localising defects Gaussian modules into these attention mechanisms to facilitate
using rectangular bounding boxes. the fusion of enhanced and salient features.
Current detection algorithms are mainly divided into one- Additionally, Gaussian attention modules (GAMs) are
stage and two-stage approaches. incorporated into the Deeplabv3+ architecture, operating
Representative one-stage algorithms include the You in parallel with the Atrous Spatial Pyramid Pooling (ASPP)
Only Look Once (YOLO) series and the Single Shot Multibox module, to effectively merge multiscale, minor, and highly

Detector (SSD). Over the years, the YOLO architecture has discriminative features.
evolved significantly, with successive versions from YOLOv1 Zhao et al. proposed the YOLOv5-Seg-Lab-4 model, which
to YOLOv13 introducing consistent improvements in accuracy, integrates object detection with semantic segmentation to ensure
speed, and efficiency. real-time performance while enhancing detection accuracy.
Representative two-stage algorithms belong to the region- Deployed in a particleboard factory, the model successfully
based CNN (R-CNN) family, which includes R-CNN, Fast automated the classification and grading of boards containing
R-CNN, and Faster R-CNN. defects such as sand leakage, big shavings, glue spots, oil