Page 38 - FDMAsia Sep/Oct 2026

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

learning and the multi-task paradigm of cascaded transfer The second method incorporates padding to transform
learning are particularly promising, as they can provide viable images into a square aspect ratio by adding background
solutions for the industrial deployment of inspection models. regions. However, this increases computational overhead
Finally, Vision Transformer (ViT)-based approaches have and substantially slows inspection speed, thereby rendering
gained widespread attention in image processing tasks. real-time online inspection unfeasible.
Different from standard CNNs, ViT frameworks leverage the The third approach involves resising non-conforming image

self-attention mechanism to adaptively assess the importance regions into square dimensions, which leads to pixel loss. This
of individual image patches, thereby offering innovative compromises the integrity of image details and exacerbates
perspectives for visual detection research. the difficulty of detecting small defects.
Although the application of ViT in wood surface defect
detection is still in its early stage, such methods have yielded New Applications
encouraging results in other forestry image processing scenarios. With the advancement of research, the application of deep
Furthermore, hybrid architectures that combine Transformer learning has expanded beyond its initial focus on sawn timber
and classic CNN structures, represented by Rep-MobileViT and wooden boards to a broader range of wood products,
and MFWL-DETR, have been widely acknowledged as a including glued panels, wooden flooring, wood paint, and

promising direction to achieve superior detection performance. particleboard.
Furthermore, a greater variety of imaging approaches, such
Real Image Inspection as spectral and thermal imaging, are being employed for the
Most existing deep learning models require input images to surface characterisation of wood products.
be standardised to a fixed size with an aspect ratio of 1:1. In these emerging applications, defects exhibit diverse
Consequently, in current studies on wood surface defect colours, textures, and geometric characteristics, necessitating
detection, the images employed are often not of entire wood substantial further efforts in areas such as dataset construction,

products but rather square patches extracted from authentic specialised model design, and extensive testing.
images. Second, a growing number of previously challenging
In practical settings, however, wood products typically application scenarios are expected to become technically
exhibit elongated and narrow dimensions, resulting in surface feasible.
images with high width-to-height ratios. A typical representative is automated wood grading, which
Therefore, to facilitate the application of these research categorises timber into distinct quality grades according to
findings in real-world scenarios, it is imperative to address the severity and distribution of surface defects.
the discrepancy between the characteristics of authentic Such grading enables rational allocation of timber resources
wood surface images and the input requirements of deep to their most suitable end applications, thereby minimising

learning models. material waste and overall resource utilisation efficiency.
Three primary approaches have been explored to mitigate Another promising scenario is intelligent wood sawing. By
this issue. The first involves designing feature maps of varying preplanning cutting strategies based on the spatial distribution
sizes tailored to different product dimensions. of wood defects, production processes can effectively avoid
While this method appears to offer an ideal solution, it defective regions and substantially improve the yield of high-
significantly increases model complexity and lacks generalisability grade timber. FDM
across products of differing sizes. ENQUIRY NO. 5201