Page 36 - FDMAsia Sep/Oct 2026

34 TECHNOLOGY SEP/OCT 2026 FDM ASIA | www.fdmasia.com

www.glazingsystems.co.uk ResNet-V2 is first used to classify images into three categories:
Ehtisham et al. presented a similar work, where Inception-

knots, cracks, and undamaged sections.
Subsequently, image processing techniques are employed
to determine key morphological characteristics of the defects,
such as width, length, angle, and spatial extent.

Luo et al. proposed a layer-wise adapter module (LAM) built
upon large-scale visual foundation models, which integrates an
instance-linking token module (ILTM), a feature disentanglement
module (FDM), and a layer switch module (LSM).
This approach achieved significant improvements in both
segmentation accuracy and efficiency over ten state-of-the-
art methods on rubber wood and pine wood datasets, while
Instance segmentation also demonstrating robust zero-shot transferability on unseen
To detect defects on wood surfaces, researchers have datasets.

adopted Mask R-CNN, an instance segmentation framework
that simultaneously predicts a pixel-level mask and the Challenges
corresponding class label for each defective region. Deep learning has yielded significant results in the visual
Al-Zubi and Plapper provided a proof-of-concept by inspection of wood surface defects, with some laboratory
applying Mask R-CNN to synthetic images of wood panels studies reporting accuracy rates approaching 100 percent.
to detect and segment drilled holes. Despite pronounced However, numerous challenges persist in practical applications.
variations in colour, texture, and hole appearance, the model Small Sample Problem

yielded satisfactory results. The small sample problem is a common challenge in
To address the challenges of modelling irregular defects applying deep learning algorithms, as it can easily lead to
and extracting contextual information, Li et al. proposed a overfitting during training.
layered deformable Mask R-CNN. By establishing layered An overfitted model exhibits poor generalisation and
connections between residual modules, a larger receptive struggles to recognise targets in unseen data. A primary cause
field was achieved at each network layer, and deformable of this issue is the scarcity of accurately labelled wood images.
convolution was employed to better fit defect shapes. Raw wood images typically lack annotations and thus
Some researchers have argued that Mask R-CNN does cannot be used directly for supervised training. The labelling
not address the problem of calculating defect size (e.g., area process relies on domain experts and is often tedious and

and diameter). In response, Zhong et al. proposed a parallel time-consuming; consequently, only a limited number of
structure fusion approach. experts are available for such tasks.
This method incorporates two dedicated branches: one Therefore, the introduction and adaptation of more advanced
for identifying veneer knot defect types using the Inception techniques-such as self-supervised, semi-supervised, and
V3 network, and the other leveraging an improved K-means weakly supervised learning-are anticipated to address various
clustering algorithm from traditional computer vision to localise practical challenges in wood defect inspection under limited
defects and determine their real-world dimensions. data conditions.