Page 35 - FDMAsia Sep/Oct 2026

TECHNOLOGY 33
www.fdmasia.com | FDM ASIA SEP/OCT 2026

pollution, and soft areas. The system achieved an accuracy challenges of defect variety, obscure boundaries, and size
of 98.2 percent, with processing times ranging from 183 to and shape variations in lumber surface defect segmentation.
208 ms per image. Beyond fully supervised methods, researchers have also
Luo et al. proposed a layer-wise adapter module (LAM) addressed the challenge of acquiring costly, high-quality
to adapt visual foundation models for wood surface defect datasets for training deep segmentation models by introducing
segmentation. generators capable of synthesising pixel-level annotations.

The LAM integrates three key components: an instance- Tsai et al. proposed a two-stage deep learning method
linking token module (ILTM) to enhance sensitivity to ambiguous for pixel-wise defect detection on textured surfaces without
boundaries and improve instance-level feature representation; manual annotations. Their approach employs two Cycle-
a feature disentanglement module (FDM) to reduce feature Consistent Adversarial Networks (CycleGANs) to automatically
redundancy and increase inter-class feature independence; and synthesise defective images and generate corresponding
a layer switch module (LSM) to dynamically activate feature pixel-wise annotations.
refinement for optimal layer-specific adaptation. The output from this synthesis stage is used to create
By addressing challenges such as high inter-class similarity a large-scale synthetic dataset, which subsequently trains a
and fuzzy boundaries, LAM significantly improved segmentation U-Net model for precise defect detection.

performance on both Rubber Wood and Pine Wood datasets. This framework requires only a small set of real defect
Zhu et al. proposed a U-Net-based multi-source data samples and eliminates labour-intensive human labelling,
fusion network for wood break detection. Their approach offering both practical implementability and computational
employs an improved ResNet34 backbone to extract multi- efficiency in manufacturing environments.
level features from image and depth data using depthwise Zheng et al. proposed a semi-supervised defect detection
separable convolutions (DSC) and dilated convolutions (DC), approach based on generative adversarial networks, comprising
which reduce computational cost and feature redundancy. a generator and a discriminator.

Features from the two modalities are then optimally The generator produces pixel-level segmentation results for
integrated through an adaptive interacting fusion module (AIF), wood defect images. To improve segmentation performance,
generating accurate feature representations for broken defects. the discriminator is trained adversarially against the generator
Based on Attention U-Net, Dong et al. proposed a new by assessing prediction quality and providing supervised
model called IECAU-Net. This was achieved by incorporating signals for unlabelled images.
CBAM and ECA modules, replacing the optimiser with AdamW,
and using a weighted fusion multi-loss function.
The resulting IECAU-Net outperformed other models in rawpixel.com
semantic segmentation of sawn wood surface cracks. Based

on U Net, DBDFCNet incorporates several key innovations. It
introduces a multiscale atrous spatial pyramid pooling (MSASPP)
module for enhanced multiscale feature extraction, a dual
branch decoder (DBD) with binary and semantic branches,
and a discriminative feature cross attention module (DFCAM)
to increase inter-class distances.
Together, these improvements effectively address the