Page 37 - FDMAsia Sep/Oct 2026

TECHNOLOGY 35
www.fdmasia.com | FDM ASIA SEP/OCT 2026

Class Imbalance Representative models, including MFWL-DETR, GBCD-
Due to the differing causes of defect formation, the YOLO, and YOLOv8-OCHD, have been reported to effectively
prevalence of various wood surface defect types varies inspect wood surface images with fewer parameters and lower
significantly, and this imbalance is consistently observed computational demands.
in related datasets. A typical case is MFWL-DETR, which, when applied to
Such imbalanced sample distribution can bias deep learning detect defects on water-based wood paint surfaces, achieved

models toward features from the majority classes, thereby a 40.5 percent decrease in computation, a 40.2 percent
compromising the reliability of the outcomes. reduction in parameters, and a 36.1 percent reduction in
To address this issue, data augmentation techniques model size compared to the baseline.
have been widely employed to enrich existing datasets. As Quantisation and pruning represent alternative approaches,
another prominent generative approach, generative adversarial although their adoption in this field has remained relatively
networks (GANs) have been extensively used to synthesise limited to date.
defect samples and alleviate data imbalance.
These methods, which can be categorised as data- More Efficient Algorithms
level approaches, mitigate the effects of class imbalance Despite the remarkable progress reported in existing research,

to some extent. Furthermore, a range of algorithm-level most developed methods still face barriers to real-world
and feature-level methods have recently emerged in other practical deployment.
industrial domains. Fortunately, wood surface defect inspection remains a
Such techniques also hold significant potential for application relatively specialised subdomain within the broader deep
in wood surface defect inspection. learning research landscape.
Advances in surface defect detection for other materials
Real-time Capabilities can offer valuable insights and methodological references for

For real-time operation, the inspection process must keep wood defect inspection, and it is foreseeable that emerging
pace with manufacturing or image acquisition speeds. This innovative algorithms will further narrow the gap between
is often challenging for deep learning-based models, as their academic research and industrial application.
complex and deep architectures frequently lead to reduced First, a continuous stream of novel CNN architectures
inference speeds. has been proposed, such as EfficientNetV2, ConvNeXt, and
In industrial applications, however, both high accuracy MobileNetV3.
and real-time performance are indispensable. Consequently, Incorporating these advanced lightweight and high-
enhancing real-time capability remains a critical issue for wood performance models is expected to facilitate low-cost, high-
surface defect inspection models, even when high accuracy accuracy inspection of wood surface defects.

has been achieved. Second, transfer learning has evolved from a conventional
Lightweight models have been extensively employed to technique merely targeting accuracy enhancement into an
effectively balance accuracy and computational efficiency. integrated methodological framework that simultaneously
Various techniques, such as factorised convolutions, group optimises computational efficiency, model trustworthiness,
convolution, depthwise separable convolution, bottleneck data privacy, and interpretability.
design, and neural architecture search, have been explored In the context of wood surface defect inspection, privacy-
to simplify model structures and reduce complexity. preserving mechanisms embedded in trustworthy transfer