A review of recent deep learning applications in wood surface defect identification

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 14, No. 3, June 2025, pp. 1696~1707
ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i3.pp1696-1707  1696
Journal homepage: http://ijai.iaescore.com
A review of recent deep learning applications in wood surface
defect identification
Martina Ali1
, Ummi Raba’ah Hashim1
, Kasturi Kanchymalay1
, Aji Prasetya Wibawa2
,
Lizawati Salahuddin1
, Rahillda Nadhirah Norizzaty Rahiddin1
1
Centre for Advanced Computing Technology, Faculty of Information and Communication Technology, Universiti Teknikal Malaysia Melaka,
Melaka, Malaysia
2
Department of Electrical Engineering and Informatics, Faculty of Engineering, Universitas Negeri Malang, Malang, Indonesia
Article Info ABSTRACT
Article history:
Received Mar 28, 2024
Revised Nov 19, 2024
Accepted Jan 27, 2025
Wood is widely used in construction, art, and home applications due to its
aesthetic appeal and favorable mechanical properties. However,
environmental factors significantly affect the growth and preservation of
wood, often leading to defects that can reduce its performance and
ornamental value. Researchers have introduced machine vision and deep
learning methods to address the challenges of high labor costs and
inefficiencies in identifying wood defects. Deep learning has shown great
success in image recognition tasks, yielding impressive results. This paper
reviews previous work on deep-learning strategies for identifying wood
surface defects. It also discusses data augmentation techniques to address
limited defect data and explores transfer learning to enhance classification
accuracy on small datasets. Finally, the paper examines the potential
limitations of deep learning for defect identification and suggests future
research directions.
Keywords:
Automated inspection
Deep learning
Defect identification
Transfer learning
Wood surface defects
This is an open access article under the CC BY-SA license.
Corresponding Author:
Ummi Raba’ah Hashim
Centre for Advanced Computing Technology, Faculty of Information and Communication Technology
Universiti Teknikal Malaysia Melaka
St. Hang Tuah Jaya, 76100 Durian Tunggal, Melaka, Malaysia
Email: ummi@utem.edu.my
1. INTRODUCTION
Since the Paris Agreement, endorsed by the United Nations in New York in 2016, countries
worldwide have strived to preserve resources, use energy efficiently, and reduce carbon emissions. In this
context, conserving forest resources and optimizing wood utilization have become increasingly important.
Historically, wood has been one of the most abundant and valuable materials. In ancient times, it was
essential for shelter, tools, fuel, and weaponry. As civilizations advanced, alternative materials began to
replace wood for specific uses. However, there has been a resurgence in the production of new and diverse
wood-based products. With rapid global economic growth, the demand for wood and wood products has
surged as people seek an improved quality of life. Unfortunately, the current wood storage and processing
capacity is insufficient to meet this demand. Limited wood supply and low utilization rates have hindered the
development of the wood industry. Therefore, a comprehensive evaluation of log and board processing
quality is necessary to improve utilization rates and enhance the quality of wood products.
Wood is widely used in manufacturing due to its strength, durability, and versatility, making it
suitable for various applications. However, it is rare to find logs with flawless surfaces in wood manufacturing.
As a natural biological material, wood is vulnerable to microorganisms that can damage its structure, leading
to defects. These defects, such as irregular tissue formations and structural damage, can affect the quality and

Int J Artif Intell ISSN: 2252-8938 
A review of recent deep learning applications in wood surface defect identification (Martina Ali)
1697
lifespan of wood products [1], [2]. Wood defects can be categorized into three types: growth defects (caused
by physiological factors), pest damage defects (due to pathological factors), and processing defects (resulting
from human error) [3]. Moreover, surface defects significantly impact the quality of wood finishes, influencing
strength and aesthetic appeal [4]. Defects can be classified as permissible or non-permissible, depending on
their severity. Permissible defects are minor and allowable, while non-permissible defects render wood
unsuitable for products that require high structural integrity or visual perfection [5]. Wood or its components
can be sorted based on the type, size, and number of defects, which helps determine whether it can be repaired,
recycled, or discarded. Identifying defective wood surfaces was essential for maintaining high production
quality and safety [6]–[9]. In wood manufacturing, defects reduced wood yield by an average of 10% [10].
Therefore, the early identification of defective items on the production line was critical to maintaining overall
quality [11]. A thorough inspection was necessary to ensure that wood met the specifications for its intended
use by implementing robust quality control processes [12], [13]. Quality control in wood manufacturing
ensured that products met performance, safety, and aesthetic standards. Early identification of defects enabled
manufacturers to take corrective action, helping to maintain product quality and consistency [8], [12].
Before the introduction of automated visual inspection (AVI), the wood industry relied on manual
inspection methods, where human operators physically inspected wood surfaces to identify defects. This
traditional approach was widely used in primary and secondary wood industries and did not require complex
technical setups. However, it offered limited potential for future development due to frequent changes in
standard operating procedures as new defects were identified. The manual inspection methods were often
inaccurate, inefficient, and restricted production volumes, with accuracy rates of only about 70%, raising
concerns about their reliability [14]–[16]. Additionally, manual inspections were costly, as training personnel
required significant time and resources, and visual fatigue often led to misidentified defects [17]. Human
error also varies depending on the workers' experience, skills, and alertness [18]. Furthermore, high
production volumes and repetitive tasks over prolonged periods negatively impacted human operators,
leading to exhaustion, stress, and reduced inspection quality [19]. As a result, manual inspections were not
only slower but also less accurate compared to automated methods [20]–[22].
To overcome the limitations of manual inspection, the wood industry increasingly embraced
technology integrated with intelligent algorithms [23]. These algorithms significantly enhanced defect
identification processes, which became a primary focus within the industry [24], [25]. Researchers actively
explored advanced techniques, such as deep learning to improve accuracy and efficiency. As a result, there
was growing interest in machine vision systems, which offered faster and more precise approaches to defect
identification [26]. AVI systems, equipped with artificial intelligence, gained traction as a solution to
enhance quality control in the wood-based industry [5], [15], [27]. With advancements in artificial
intelligence and computer vision technologies, deep learning emerged as a highly effective method for
identifying wood defects [28]–[30]. Furthermore, due to their simplicity and affordability, AVI systems
became popular for ensuring higher accuracy and production volumes by eliminating human limitations,
improving reliability, and maintaining quality standards [14]. Despite advancements, the wood industry still
requires solutions to improve processing efficiency and increase yield without compromising product quality.
AVI has been highlighted for its role in ensuring consistent product reliability and addressing yield losses due
to limitations in manual inspections. Research indicates that AVI offers 25% higher identification accuracy
than traditional methods, leading to a 5.3% increase in yield and considerable cost savings for the average
rough mill [31]. Furthermore, automated grading has shown greater precision and consistency than
conventional inspections, which often struggle to optimize wood resources [32]. Studies also demonstrate
that AVI outperforms human inspectors in identifying defects and plays a crucial role in maintaining quality
control, ultimately benefiting the secondary wood industry by improving yields and production quality [26].
Moreover, one significant advancement in deep learning for wood defect identification is using data
augmentation and transfer learning techniques [33]. Data augmentation addresses the challenge of limited
datasets by enabling models to generalize better and improve classification accuracy. On the other hand,
transfer learning leverages pre-trained models, such as residual network (ResNet) and Inception, fine-tuning
them for specific tasks like wood defect classification, reducing the need for retraining from scratch. As deep
learning techniques evolve, further exploration of sophisticated augmentation methods and advanced transfer
learning frameworks will be crucial for enhancing defect identification systems. This combination of
techniques overcomes the limitations of manual inspection and significantly improves wood production
efficiency, quality assurance, and resource utilization.
2. WOOD DEFECT IDENTIFICATION APPROACHES USING DEEP LEARNING
Detecting and identifying defects are crucial in manufacturing to maintain controlled and efficient
production processes [34]. Traditionally, these tasks relied on skilled human operators. However, the
introduction of AVI has enhanced the autonomy of manufacturing operations. AVI-based quality control has

 ISSN: 2252-8938
Int J Artif Intell, Vol. 14, No. 3, June 2025: 1696-1707
1698
gained popularity in the secondary wood industry due to its ability to improve inspection accuracy, boost
production rates, and lower labor costs. An AVI system typically involves several stages: image acquisition,
image enhancement, segmentation, feature extraction, and feature classification [35]. Efficient material
handling ensures smooth logistics, maintaining steady material flow, reducing vibrations, and regulating
speed during timber imaging. Critical subsystems like sensors and lighting are essential for capturing,
digitizing, and storing image data. The inspection process begins with defect detection, pinpointing the
location of flaws in the wood. This is followed by defect identification, where defects' type, size, and
frequency are analyzed. These components also offer guidelines for optimized timber cutting based on the
identified defects. Finally, the collected data is used in the grading phase, where the timber is classified
according to specific production standards.
In the wood industry, one of the most effective methods for identifying defects involves processing
and analyzing images of wood surfaces. Numerous studies have explored using AVI systems, employing
traditional image processing, specialized techniques, and artificial intelligence methods [36]. Deep learning
represents a significant advancement within machine learning, characterized by the construction of multi-
layered neural networks that emulate the complex processes of the human brain. It enables the recognition of
diverse inputs such as images, sounds, and texts [37]. Within these neural networks, algorithms in each layer
continuously perform calculations and predictions, progressively enhancing accuracy over time. This
methodology adheres to data-driven principles, representing observed images in various forms, whether as pixel
density value vectors or abstract features like edge sequences [38]. While deep networks often include more
hidden layers than traditional neural networks, the effectiveness of learning outcomes does not depend solely on
layer depth. Successful deep learning design hinges on determining the optimal number of hidden layers tailored
to specific tasks [3]. This evolution has propelled deep learning into a leading approach in academic research
and industry applications. Deep learning algorithms automated the feature extraction process, eliminating the
need for manual input from experts. Convolutional neural networks (CNNs) exemplified this approach,
extracting features directly from images in a fully automated manner [29]. The goal of applying deep learning
was to independently process new data, make accurate decisions, and provide reliable recommendations based
on thousands of calculations while significantly reducing the potential for human error.
Various advanced deep learning frameworks have proven particularly effective for wood surface
defect detection, simultaneously learning feature extraction and classification during training. This innovative
approach highlights the adaptability and robustness of deep learning methodologies in addressing real-world
challenges. Table 1 summarizes significant studies from 2019 to 2024, showcasing the evolution of deep
learning architectures for wood defect identification. These studies demonstrate the growing application of
deep learning techniques, which have significantly improved defect detection accuracy in the wood industry.
Table 1. Overview of deep learning models used in wood surface defect identification
Reference Year Approach Defect type
[39] 2024 Inception-ResNet-V2 CNN model Cracks, knots, and undamaged
[40] 2024 Improved ResNet-50 model Brown stain, blue stain, knot, borer holes, rot, bark
pocket, wane, and split
[41] 2023 Inception-V3 model Cracks, knots, and undamaged
[42] 2023 Improved ResNet-50 model Knots, cracks, and color-related defects
[43] 2023 Improved RegNet model Wormhole, dead knot, and live knot
[33] 2022 ResNet50 model Brown stain, blue stain, knot, borer holes, rot, bark
pocket, wane, and split
[11] 2022 Faster R-CNN model Knot, wane, edge, stain, and branch.
[44] 2022 Inception-ResNet-V2 CNN model Edge-glued
[45] 2022 Derived ResNet-v2 model Wormhole, live joint, and dead joint.
[46] 2021 TL-ResNet34 model Decay knots, dry knots, edge knots, encased knots,
horn knots, leaf knots, and sound knots.
[47] 2021 Based bilinear fine-grained (BLNN) Knot
[48] 2021 CNN model Defective wood
[49] 2021 Deep convolutional neural network
(DCNN) model
Not specified
[29] 2020 Multi-channel mask R-CNN Dead knots, live knots, and cracks.
[30] 2020 DCNN model Crack, knot, and mildew
[50] 2020 Mask R-CNN model Dead knots, live knots, and insect holes
[51] 2019 Mix-FCN model Dead knot, live knot, blue stain, crack, brown stain,
and pitch streak
[52] 2019 AlexNet model External defects
[53] 2019 ResNet18 model Not specified
[14] 2019 Faster R-CNN model Split, core, branch, and stain
[22] 2019 Faster R-CNN model Knots and holes
[54] 2019 CNN model Knots
[55] 2019 VGG-16 model Not specified

1699
Recent studies on wood surface defect identification have led to the development of several CNN
architectures. In 2024, Ehtisham et al. [39] developed an automated method for assessing defects in wooden
structures using CNNs and image processing techniques to enhance inspection efficiency. Traditional manual
inspections, though effective, are time-consuming and costly. The study employed the Inception-ResNet-V2
CNN model, trained with a dataset of 9000 images divided into cracks, knots, and undamaged sections. The
model achieved 92% classification accuracy on new test images, with minimal errors in defect quantification.
This automated approach provided significant practical benefits for industry professionals by reducing labor
costs, speeding up inspections, and improving accuracy, making it a valuable tool for maintaining the safety and
durability of wooden structures. Next, in the same year, Chun et al. [40] investigated ways to enhance the
classification accuracy of timber defect identification using deep learning, particularly ResNets. The focus is
refining these networks by increasing depth and incorporating multi-level features to build a more effective
framework for identifying different defect types. A series of ablation experiments were performed, testing the
performance of three new architectures (R1, R2, and R3) against the standard ResNet50 model. While the R1
architecture achieved slight improvements over ResNet50, particularly with adding an extra layer called
"ConvG," ResNet50 still showed better overall performance. Similarly, the R2 framework provided some
improvements but still fell behind R1. The most significant results were achieved with the R3 architecture, which
integrated fully pre-activation functions, leading to a 14.18% increase in classification accuracy compared to
ResNet50. The R3 architecture also performed well across various timber species, though it showed slightly
reduced accuracy for rubberwood. Nonetheless, R3 outperformed both ResNet50 and the other proposed models,
highlighting the effectiveness of deeper networks and pre-activation functions. Statistical analyses, including
independent t-tests and one-way ANOVA, confirmed that the improvements were significant across multiple
species, demonstrating the potential of the R3 architecture for timber defect classification.
Next, in 2023, Zou et al. [42] proposed an improved modeling method based on ResNet-50 to
enhance the accuracy and efficiency of wood defect identification, which was crucial for ensuring furniture
manufacturing quality. Recognizing the complexities involved in data and the need for high efficiency, the
study introduced a new optimization scheme that combined a convolutional block attention module with a
cross-stage partial network (CSPNet) tailored for the ResNet-50 architecture. The research explored how
varying the cross-stage parameters in CSPNet affected classification performance, revealing that the default
parameters were not always optimal. Additionally, the author proposed the ranger optimizer, which showed
superior performance over the traditional Adam optimizer in terms of training efficiency and prediction
accuracy. Experiments conducted using a dataset of real-world wood surface defects demonstrated that the
proposed model achieved an impressive overall detection accuracy of 86.25% in identifying defects such as
knots, cracks, and color-related issues. These results underlined the effectiveness and feasibility of the new
modeling approach in improving wood defect identification compared to existing state-of-the-art methods.
Next, in 2022, Ling and Xie [45] developed a derived model based on ResNet-v2 to address the
limitations of traditional manual detection methods for wood defects, which are often time-consuming,
inefficient, and inaccurate. This new model aims to accurately identify defect types such as wormholes, live
joints, and dead joints on wooden surfaces, significantly improving classification accuracy while reducing labor
requirements. The study highlights that the ResNet-v2-derived model demonstrates better recognition capabilities
and enhanced generalization abilities than traditional CNN. Experimental results indicate that the classification
accuracy of the ResNet-v2-derived network exceeds 80% for various configurations, with a peak accuracy of
97.27% under certain conditions. These findings suggest that the proposed model offers a more efficient and
reliable solution for wood defect identification, making it a valuable tool in the industry for improving quality
control and operational efficiency. In the same year, Mohsin et al. [11] proposed a novel method for automatic
real-time defect detection and classification on wood surfaces. This approach utilized a DCNN framework called
Faster R-CNN for detection, combined with MobileNetV3 as the backbone network for feature extraction. The
key innovation of this method lies in its ability to efficiently detect knots and other defects while performing
classification in real-time from input video frames. The study focused on achieving speed and accuracy, which
were crucial for industrial quality control and inspection tasks requiring defect detection in real-time, particularly
on computationally limited processing units or commodity processors. The model was designed to be
lightweight, making it suitable for deployment on mobile and edge devices. MobileNetV3 was pre-trained on a
large image dataset to enhance feature extraction, while Faster R-CNN was employed for defect detection and
classification. The system processed input video frames at an average speed of 37 per second using a low-cost,
low-memory graphics processing unit (GPU). The method achieved an overall accuracy of 99% in detecting and
classifying defects, demonstrating its effectiveness for practical applications in wood surface inspection.
In 2021, Gao et al. [46] developed the TL-ResNet-34 model, a transfer residual neural network, to
enhance the identification of wood knot defects with speed and accuracy. This model improved accuracy by
over 0.78% through careful extraction of structural defect features and optimization of training parameters and
datasets. By incorporating transfer learning, the model benefited from a pre-training phase that allowed for
more effective learning from the limited available data. Experimental results demonstrated that TL-ResNet-34

 ISSN: 2252-8938
1700
achieved an impressive recognition rate of 99.22% on the training dataset, with a low training loss of 2.83% on
the validation dataset while identifying seven distinct wood knot defects. Overall accuracy reached 98.69%,
with minimal fluctuations in the loss and accuracy curves during testing, indicating the model's stability and
reliability. Significantly, this method minimized the need for extensive image preprocessing and feature
extraction, making it efficient and accurate in both the training and testing phases. The findings suggested that
TL-ResNet-34 held significant potential for applications in wood non-destructive testing and wood defect
identification, allowing for quick and accurate assessments of collected wood knot defects. This innovation
could significantly improve wood utilization and resource conservation in the industry.
3. DISCUSSION
Based on the findings from recent studies on wood defect identification, several critical discussions
can be drawn. The research highlights the growing importance of utilizing deep learning techniques,
particularly CNNs, to enhance the accuracy and efficiency of defect identification in wood. The evolution of
CNN architectures, such as Inception-ResNet-V2 and various ResNet models, has shown promising results in
improving classification accuracy. For instance, the automated method developed in 2024 achieved a
classification accuracy of 92%, demonstrating how deep learning can significantly outperform traditional
manual inspection methods, which are often time-consuming and costly [39], [42], [46]. Next, as seen in the
Faster R-CNN and MobileNetV3 framework, the emphasis on real-time detection and classification
capabilities addresses practical needs in industrial settings where quick decision-making is essential. The
ability to process video frames at 37 frames per second with an accuracy of 99% illustrates the potential of
these advanced models to enhance quality control in wood processing [11], [46]. Moreover, integrating
optimization strategies, such as convolutional block attention modules and transfer learning techniques, has
improved model performance, addressing challenges like limited labeled datasets. These advancements
promise substantial benefits, including reduced labor costs and enhanced safety, and they also optimize
resource utilization in timber manufacturing, demonstrating a significant shift toward automation in wood
defect identification. Furthermore, introducing the ranger optimizer and adjusting the CSPNet parameters
contributed to better accuracy, underscoring the necessity of continual improvement in the model training
process [39], [42], [46]. The studies also highlighted challenges, particularly the limited availability of labeled
datasets in wood defect detection. Models like TL-ResNet-34, which effectively leverage transfer learning,
show that even with fewer training examples, it is possible to achieve high recognition rates, emphasizing the
importance of innovative approaches to data scarcity [46]. In addition, implementing these advanced detection
systems in real-world scenarios promises significant benefits, including reduced labor costs, improved safety,
and enhanced durability of wooden structures. By automating defect detection, industries can optimize their
processes, leading to better resource conservation and more efficient use of timber [39], [42], [46].
The analysis of wood defect identification methods from 2019 to 2024 reveals a predominant
reliance on CNN-based models, highlighting their effectiveness in identifying surface defects in wood. CNNs
enhance defect identification through hierarchical feature learning, significantly reducing the time and effort
associated with manual inspections [39]. Various architectures, including ResNet, Faster R-CNN, and
Inception-ResNet, showcase the versatility and dominance of CNNs in this field. Their popularity stems from
two main factors: high performance and ease of training [56]. The modular design of CNNs, comprising
convolutional bases and fully connected layers, allows for robust feature extraction and accurate defect
classification. This capability improves the precision of defect identification by capturing intricate features
within high-dimensional data [57]–[59]. However, challenges persist, notably the substantial computational
power required for effective CNN operation, often necessitating expensive GPUs and large datasets [60]. To
overcome these barriers, research is needed to develop compact, efficient CNN models that can function on
low-power computing resources, such as CPUs or FPGAs, making them more accessible for industrial
applications. Furthermore, creating and sharing large, high-quality datasets of surface defects is crucial for
enhancing model training and performance. Addressing these gaps will facilitate the broader adoption of
CNNs in automated surface inspection systems, balancing computational demands with cost-effectiveness
and data availability. Overall, CNNs play a vital role in advancing wood defect identification, offering
superior feature extraction and classification capabilities essential for improving automation in the industry.
Furthermore, the analysis of models presented in Table 1 indicates significant advancements in the
sophistication of CNN architectures, with notable models like Inception-ResNet-V2 and ResNet-50 demonstrating
enhanced performance across various defect categories. These advancements highlight the robustness and
adaptability of CNNs in real-time defect detection, as seen with models such as Faster R-CNN, which effectively
identify defects like knots, holes, and stains in industrial applications. In 2024, the focus has shifted towards
refining advanced models, particularly Inception-ResNet-V2 and improved versions of ResNet-50, underscoring
ongoing innovation in defect identification methodologies. The previous years, 2022 and 2023, also reflected

1701
significant improvements in ResNet-50, emphasizing efforts to optimize accuracy and efficiency. While the
models in Table 1 effectively address a range of defects from common issues like cracks and knots to specific
problems such as wormholes and insect damage they often concentrate on a limited number of defect types. Most
studies focus on knots, significantly impacting structural strength and product quality. This trend underscores the
necessity for more comprehensive deep learning models capable of multiclass identification, allowing for a broader
scope of wood defect detection. Moreover, several studies labeled defect types as "not specified," indicating a
general focus on wood surface defects without detailing specific types. This presents an opportunity for future
research to explore and document various defects, enhancing quality control in the wood industry.
In addition, the quality of the training dataset is crucial for effectively training deep learning models.
Deep neural networks typically contain many trainable parameters, making them highly prone to overfitting,
especially when training data is limited. Various regularization techniques have been developed to address
this issue to enhance the model's generalization capability. One effective method is data augmentation, which
directly impacts the training dataset. Data augmentation involves applying various transformations to input
images to expand the training dataset artificially. It is essential to carefully select transformations appropriate
for a specific problem, as these transformations should maintain the semantic meaning of the labels. Data
augmentation is considered a standard approach in deep learning, especially when fewer data samples are
available. It is widely used to improve the adaptability of neural networks for classification tasks and to
mitigate overfitting issues. Numerous studies have explored the effectiveness of data augmentation in wood
defect identification, demonstrating its potential to enhance accuracy and efficiency. Table 2 summarizes
various data augmentation techniques employed in past research, highlighting the specific methods used by
different researchers to improve the diversity and effectiveness of their training datasets.
Table 2. Techniques for data augmentation in wood defect identification
Researchers Data augmentation techniques
[33] Morphological transformations: x-reflection and rotation
[14] Geometric transformations: flip, rotation transformation, and resize transformation.
[61] Brighten change and Gaussian noise.
[30] Rotating the training images at 90°, 180°, and 270°, randomly cropping the training images, flipping the training
images horizontally, vertically, and diagonally, adjusting the brightness, contrast, saturation, or hue of the
training images, and adding Gaussian and salt-and-pepper noises to the images
Research by Chun et al. [33] highlighted the efficacy of data augmentation in enhancing the
performance of CNNs for classifying timber defects in various species. They employed morphological
transformations, such as x-reflection and rotation, to expand the training dataset, which addressed the challenges
posed by limited data availability. The results demonstrated a significant improvement in classification
performance, with the best model, ResNet50, achieving an accuracy of 94.59%. Statistical analyses, including
T-tests, confirmed that the enhancements in classification accuracy due to data augmentation were significant
for most CNN algorithms tested, except for AlexNet. Furthermore, the authors observed that simply increasing
the number of training epochs and the learning rate did not necessarily yield better classification precision,
underscoring the importance of carefully tuning hyperparameters for optimal results. Overall, the study
concluded that data augmentation was a vital technique for overcoming the limitations of small datasets and
improving the classification capabilities of CNNs in the context of timber defect identification.
Next, Urbonas et al. [14] discussed the importance of data augmentation in enhancing the performance
of AVI systems for identifying defects on wood veneer surfaces. They highlighted that traditional visual quality
inspections in the lumber and wood processing industry were often performed by human operators, leading to
potential errors due to the tedious nature of the task. To improve accuracy and speed, the authors proposed using
a Faster R-CNN alongside data augmentation and transfer learning techniques. The authors specifically
mentioned employing a synthetically augmented dataset to train their models, which included pre-trained
networks like AlexNet, VGG16, BNInception, and ResNet152. Their experiments showed that data
augmentation significantly contributed to the overall classification accuracy of defect identification, achieving
an average accuracy of 80.6% with ResNet152 and an impressive 96.1% when combining all defect classes.
This indicated that data augmentation effectively enhanced the robustness of the models, making them suitable
for real-time industrial applications. Moreover, the authors acknowledged that while data augmentation and
transfer learning improved defect detection, the method's reliance on manually labeled images for training
presented a limitation, as these labels might not always be accurate. They suggested that future work could
explore more complex data augmentation techniques and apply these methods to analyze surface defects in
other types of wood panels, thereby broadening the scope of their approach. Overall, the authors emphasized
that data augmentation was crucial in developing effective automated inspection systems for wood processing,
demonstrating its applicability to various industrial materials.

 ISSN: 2252-8938
1702
Then, Romanovskis et al. [61] discussed the importance of data augmentation in optimizing the use
of oak wood during industrial processing. The author trained two different Mask R-CNN models using the
instance segmentation method. The first model was trained solely on the original data obtained from the
scanner. In contrast, the second model included the original images and additional artificially adjusted images
to expand the dataset. The authors employed two data augmentation methods “brighten change” and
“Gaussian noise” which were selected because they did not alter the physical location of objects within the
images. All new images were generated based on labeled images, eliminating the need for relabeling, as the
original image JSON file was attached to the artificially generated images. This approach saved significant
time in the marking process. The results indicated that the original model achieved an average accuracy of
73%. However, the model that utilized data augmentation showed an improvement of approximately 16%,
reaching an average accuracy of around 89%. The authors concluded that the augmented model successfully
identified and localized defects specific to oak wood, demonstrating that data augmentation is a valuable
technique for enhancing model performance in the context of wood processing.
Lastly, He et al. [30] discussed the role of data augmentation in their proposed method for detecting
wood features and classifying defects from images collected using a laser scanner. In their approach, the
authors implemented data augmentation as a strategy to avoid overfitting during the training of the DCNN
model. They collected a dataset comprising images of 600 pieces of red and camphor pine wood, which
included specific defect categories such as knots, cracks, and mildew stains. The dataset was divided into
training, validation, and testing subsets, with data augmentation applied to enhance the size and diversity of
the training data. The results demonstrated that the DCNN model achieved an overall accuracy of 99.13%,
with only 1.12 seconds required for detection, including image preprocessing and identification. Data
augmentation significantly contributed to the model’s performance, allowing it to recognize wood defects
more accurately and effectively than conventional methods.
By incorporating transfer learning, the model benefited from a pre-training phase that facilitated
more effective learning from the limited available data. In the rapidly evolving field of deep learning, transfer
learning has emerged as a pivotal technique, particularly for tasks involving limited labeled data. Numerous
pre-trained CNN models are readily available in the literature, providing significant advantages in learning
comprehensive visual feature representations. These models have been extensively trained on large datasets,
allowing them to capture complex patterns and intricate features, which is especially beneficial when
working with small, task-specific training datasets [62]. Leveraging transfer learning enables researchers to
adapt pre-trained models to meet their specific needs, offering a pragmatic and efficient approach for training
deep learning models in scenarios with limited labeled data. This methodology allows practitioners to modify
and fine-tune pre-trained models using domain-specific information, significantly accelerating the learning
process for image classification and identification tasks. Instead of starting the training process from scratch,
transfer learning utilizes existing models trained on larger datasets and adjusts their weights to align with the
intended task. This technique enhances both accuracy and effectiveness across various applications [33].
Training deep learning models from scratch can be computationally expensive; therefore, utilizing
models already published in the literature can save time and resources while achieving comparable performance in
computer vision tasks [63]. Several deep neural network models, such as Xception, Inception, AlexNet, VGGNet,
GoogLeNet, and ResNet, have been proposed for object classification over the past few decades. Moreover,
transfer learning strategies have been widely applied in wood surface defect classification. Additionally, transfer
learning has proven effective across various domains, including manufacturing, medical diagnostics, and baggage
screening [64]–[66]. This approach eliminates the need for extensive datasets and minimizes the prolonged training
periods typically associated with developing deep learning algorithms from scratch [67], [68].
Several studies effectively utilized transfer learning to identify wood defects. Urbonas et al. [14]
discussed transfer learning as a critical technique in developing an AVI system for detecting defects on wood
veneer surfaces. By employing pre-trained neural network models such as AlexNet, VGG16, BNInception,
and ResNet152, the researchers aimed to enhance the accuracy of the faster R-CNN model for this specific
task. Transfer learning allowed them to leverage the learned features from large datasets, resulting in the best
average accuracy of 80.6% using ResNet152 and a remarkable 96.1% accuracy when combining all defect
classes. Additionally, data augmentation was used to increase the diversity of the training dataset, further
improving model performance. The findings suggested that these methods could apply to other industrial
materials, showcasing the versatility of transfer learning. However, the reliance on manually labeled images
for training posed a limitation, as labeling errors could impact results. The authors planned to explore more
complex transfer learning and data augmentation techniques in future research and extend their methods to
analyze surface defects in other types of wood panels.
Next, Gao et al. [69] discussed transfer learning as an integral component of their proposed model,
ResNet-18, for detecting wood knot defects. In this study, transfer learning was leveraged to enhance the
performance of deep learning techniques in the context of wood defect detection, which had historically

1703
faced challenges such as long training times, low recognition accuracy, and the need for manual feature
extraction. Integrating transfer learning with the "squeeze-and-excitation" (SE) module into the residual basic
block structure, the ResNet-18 model improved feature extraction in the channel dimension and effectively
fused features across multiple scales. The experimental results indicated that the model achieved an
impressive accuracy of 99.02% in detecting wood knot defects, significantly higher than the classical
ResNet-18 model's accuracy of 90.83%. Additionally, the proposed approach reduced training time and
eliminated the need for extensive image preprocessing and manual feature extraction, thereby greatly
enhancing recognition efficiency. Overall, the author concluded that the ResNet-18 model, through its
transfer learning, provided a promising solution for wood nondestructive testing and defect identification,
paving the way for more efficient wood knot defect detection in future applications.
In addition, Hu et al. [50] discussed transfer learning as a crucial component of their proposed method
for identifying wood defects in poplar veneer using a combination of a progressive growing generative
adversarial network (PGGAN) and the Mask R-CNN model. In this study, transfer learning was employed to
enhance the performance of the Mask R-CNN model, specifically by integrating it with a classifier layer. This
approach allowed the model to learn from a pre-trained network, improving its ability to identify and classify
defects such as dead knots, live knots, and insect holes in the veneer. The experimental results indicated that the
model achieved high accuracy rates of 99.05% for live knots, 97.05% for dead knots, and 99.10% for insect
holes, demonstrating the effectiveness of the transfer learning strategy in achieving precise defect detection. The
author highlights that the classification accuracy of the Mask R-CNN model, which was based on ResNet50 and
further enhanced through transfer learning, reached 98.4% when tested on an expanded dataset. Additionally, the
use of PGGAN for data augmentation helped to improve the diversity of defect images and balance sample
distribution, contributing to the overall performance of the detection model. The findings suggest that transfer
learning can significantly enhance detection accuracy, especially when working with smaller datasets. The author
concludes that this technology has potential applications in wood processing equipment, particularly in wood
classification systems, indicating a promising avenue for future implementation.
Also, Ding et al. [1] discussed transfer learning as a key strategy in their approach to detecting wood
defects, specifically in solid wood boards. This study applied the transfer learning method to the single-shot
multi-box detector (SSD), a target detection algorithm, using the DenseNet network to enhance performance.
The results indicated a mean average precision of 96.1% for detecting the three types of defects, showcasing
the effectiveness of transfer learning in improving detection accuracy. The authors emphasized the potential
of transfer learning and advanced neural network architectures in enhancing wood defect detection methods.
They also intended to extend their model to identify other wood surface defects, such as wormholes and
discoloration, and to improve target location precision in future research.
Then, Ehtisham et al. [41] discussed transfer learning as a crucial element in their study of
identifying and classifying defects in wooden structures. The research evaluated ten pre-trained CNN models
ResNet18, ResNet50, ResNet101, ShuffleNet, GoogLeNet, Inception-V3, MobileNet-V2, Xception,
Inception-ResNet-V2, and NASNet-Mobile and each model was further trained and validated, demonstrating
the effectiveness of transfer learning in adapting these pre-trained models for specific tasks. The findings
indicated that the Inception-V3 model performed the best, achieving an overall accuracy rate of 97.3% and a
training time of 97 minutes. This model’s architecture, featuring 48 deep layers, allowed it to classify defects
in wooden structures effectively. The study highlighted the efficiency of using pre-trained CNNs, as they
significantly improved classification accuracy and reduced the time required for defect identification
compared to traditional manual inspection methods.
Furthermore, Chun et al. [33] discussed transfer learning as a significant component in their study of
identifying timber defects across four Malaysian timber species. They highlighted using the ResNet50
algorithm, based on transfer learning, outperforming other CNN models, achieving a classification accuracy
of 94.59% when combined with data augmentation techniques. The research indicated that data augmentation
addressed issues related to limited datasets and improved the classification performance of CNNs by 5.78%.
The study evaluated the effectiveness of transfer learning in conjunction with various hyperparameter
settings, such as learning rates and epochs, demonstrating that while these adjustments were important,
higher values did not always lead to better accuracy in CNN classification. Overall, the findings suggested
that data augmentation and transfer learning were effective strategies for enhancing the identification of
timber defects, and the authors noted potential for future research in exploring more complex augmentation
techniques and applying deep learning to analyze different types of timber defects across various species.
4. CONCLUSION
The growing importance of deep learning methods is evident, as recent research highlights their
effectiveness in enhancing the accuracy and efficiency of wood defect identification, significantly
outperforming traditional inspection methods and leading to improved quality control in the timber industry.

 ISSN: 2252-8938
1704
Advancements in CNN architectures, such as Inception-ResNet-V2 and various ResNet models, have
demonstrated remarkable improvements in classification accuracy, reflecting a shift towards more sophisticated
and practical techniques for defect identification. Furthermore, implementing frameworks like Faster R-CNN
and MobileNetV3 showcases the ability of these advanced models to provide real-time detection and
classification, which is essential for industrial applications where quick decision-making is crucial. Despite
these advancements, challenges remain, particularly regarding the limited availability of labeled datasets.
However, innovative approaches like transfer learning and data augmentation have proven effective in
enhancing model performance, even when working with smaller datasets. The economic and practical benefits
of integrating these advanced detection systems are substantial, including reduced labor costs, improved safety,
and more efficient resource utilization in timber manufacturing, ultimately leading to significant process
optimizations in the wood industry. While current studies predominantly focus on specific defect types, such as
knots, there is an urgent need for more comprehensive models capable of multiclass identification to improve
the detection of a broader range of defects and enhance overall quality control in wood processing. Future
research should prioritize the creation of larger, high-quality datasets and developing compact, efficient CNN
models that can function on low-power computing resources. Exploring more complex data augmentation
techniques and applying transfer learning across various wood types and defect categories could further propel
advancements in this field. In summary, the research emphasizes that significant strides have been made in
utilizing deep learning for wood defect identification; however, ongoing innovation and adaptation are crucial to
overcoming existing challenges and maximizing the benefits of these technologies in real-world applications.
FUNDING INFORMATION
This research is supported by the Ministry of Higher Education (MOHE), Malaysia, through the
Fundamental Research Grant Scheme (FRGS/1/2022/ICT02/UTEM/02/2) and Fakulti Teknologi Maklumat
dan Komunikasi, Universiti Teknikal Malaysia Melaka, Melaka, Malaysia.
AUTHOR CONTRIBUTIONS STATEMENT
This journal uses the Contributor Roles Taxonomy (CRediT) to recognize individual author
contributions, reduce authorship disputes, and facilitate collaboration.
Name of Author C M So Va Fo I R D O E Vi Su P Fu
Martina Ali ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Ummi Raba’ah Hashim ✓ ✓ ✓ ✓ ✓ ✓ ✓
Kasturi Kanchymalay ✓ ✓ ✓
Aji Prasetya Wibawa ✓ ✓
Lizawati Salahuddin ✓ ✓
Rahillda Nadhirah
Norizzaty Rahiddin
✓ ✓
C : Conceptualization
M : Methodology
So : Software
Va : Validation
Fo : Formal analysis
I : Investigation
R : Resources
D : Data Curation
O : Writing - Original Draft
E : Writing - Review & Editing
Vi : Visualization
Su : Supervision
P : Project administration
Fu : Funding acquisition
CONFLICT OF INTEREST STATEMENT
Authors state no conflict of interest.
DATA AVAILABILITY
Data availability is not applicable to this paper as no new data were created or analyzed in this study.
REFERENCES
[1] F. Ding, Z. Zhuang, Y. Liu, D. Jiang, X. Yan, and Z. Wang, “Detecting defects on solid wood panels based on an improved SSD
algorithm,” Sensors, vol. 20, no. 18, pp. 1–17, 2020, doi: 10.3390/s20185315.
[2] U. R. Hashim, S. Z. Hashim, and A. K. Muda, “Image collection for non-segmenting approach of timber surface defect
detection,” International Journal of Advances in Soft Computing and its Applications, vol. 7, no. 1, pp. 16–34, 2015.

1705
[3] Y. Chen, C. Sun, Z. Ren, and B. Na, “Review of the current state of application of wood defect recognition technology,”
BioResources, vol. 18, no. 1, pp. 2288–2302, 2023, doi: 10.15376/biores.18.1.Chen.
[4] M. Budakçı and H. Çınar, “Colour effects of stains on wood with knots, cracks and rots,” Progress in Organic Coatings, vol. 51,
pp. 1–5, 2004.
[5] U. R. Hashim, S. Z. Hashim, and A. K. Muda, “Automated vision inspection of timber surface defect: A review,” Jurnal
Teknologi, vol. 77, no. 20, pp. 127–135, 2015, doi: 10.11113/jt.v77.6562.
[6] Y. Tan, X. Yang, X. Bai, H. Dong, J. Liu, and L. Zhang, “Inspection and evaluation of wood components of ancient buildings in
the south-three courts of the forbidden city,” BioResources, vol. 17, no. 1, pp. 962–974, 2022, doi: 10.15376/biores.17.1.962-974.
[7] P. Xu et al., “Application of nondestructive testing technologies in preserving historic trees and ancient timber structures in
China,” Forests, vol. 12, no. 3, 2021, doi: 10.3390/f12030318.
[8] C. L. Todoroki, E. C. Lowell, and D. Dykstra, “Automated knot detection with visual post-processing of Douglas-fir veneer
images,” Computers and Electronics in Agriculture, vol. 70, no. 1, pp. 163–171, 2010, doi: 10.1016/j.compag.2009.10.002.
[9] W. Zhou, M. Fei, H. Zhou, and K. Li, “A sparse representation based fast detection method for surface defect detection of bottle
caps,” Neurocomputing, vol. 123, pp. 406–414, 2014, doi: 10.1016/j.neucom.2013.07.038.
[10] C. Clément, T. Lihra, R. Gazo, and R. Beauregard, “Maximizing lumber use: The effect of manufacturing defects on yield, a case
study,” Forest Products Journal, vol. 56, no. 1, pp. 60–65, 2006.
[11] M. Mohsin, O. S. Balogun, K. Haataja, and P. Toivanen, “Real-time defect detection and classification on wood surfaces using
deep learning,” IS and T International Symposium on Electronic Imaging Science and Technology, vol. 34, no. 10, 2022, doi:
10.2352/EI.2022.34.10.IPAS-382.
[12] Y. Fang, L. Lin, H. Feng, Z. Lu, and G. W. Emms, “Review of the use of air-coupled ultrasonic technologies for nondestructive
testing of wood and wood products,” Computers and Electronics in Agriculture, vol. 137, pp. 79–87, 2017, doi:
10.1016/j.compag.2017.03.015.
[13] B. Yuce, E. Mastrocinque, M. S. Packianather, D. Pham, A. Lambiase, and F. Fruggiero, “Neural network design and feature
selection using principal component analysis and Taguchi method for identifying wood veneer defects,” Production and
Manufacturing Research, vol. 2, no. 1, pp. 291–308, 2014, doi: 10.1080/21693277.2014.892442.
[14] A. Urbonas, V. Raudonis, R. Maskeliunas, and R. Damaševičius, “Automated identification of wood veneer surface defects using
faster region-based convolutional neural network with data augmentation and transfer learning,” Applied Sciences, vol. 9, no. 22,
2019, doi: 10.3390/app9224898.
[15] M. Kryl, L. Danys, R. Jaros, R. Martinek, P. Kodytek, and P. Bilik, “Wood recognition and quality imaging inspection systems,”
Journal of Sensors, vol. 2020, 2020, doi: 10.1155/2020/3217126.
[16] I. Y. H. Gu, H. Andersson, and R. Vicen, “Automatic classification of wood defects using support vector machines,” in Computer
Vision and Graphics, 2009, pp. 356–367, doi: 10.1007/978-3-642-02345-3_35.
[17] T. H. Chun et al., “Identification of wood defect using pattern recognition technique,” International Journal of Advances in
Intelligent Informatics, vol. 7, no. 2, pp. 163–176, 2021, doi: 10.26555/ijain.v7i2.588.
[18] N. D. Abdullah, U. R. Hashim, S. Ahmad, and L. Salahuddin, “Analysis of texture features for wood defect classification,”
Bulletin of Electrical Engineering and Informatics, vol. 9, no. 1, pp. 121–128, 2020, doi: 10.11591/eei.v9i1.1553.
[19] H. A. Huber, C. W. McMillin, and J. P. McKinney, “Lumber defect detection abilities of furniture rough mill employees,” Forest
Products Journal, vol. 35, no. 11–12, pp. 79–82, 1985.
[20] D. Yadav, Akanksha, and A. K. Yadav, “A novel convolutional neural network based model for recognition and classification of
apple leaf diseases,” Traitement du Signal, vol. 37, no. 6, pp. 1093–1101, 2020, doi: 10.18280/TS.370622.
[21] X. Zhu, M. Zhu, and H. Ren, “Method of plant leaf recognition based on improved deep convolutional neural network,” Cognitive
Systems Research, vol. 52, pp. 223–233, 2018, doi: 10.1016/j.cogsys.2018.06.008.
[22] J. Fan, Y. Liu, Z. Hu, Q. Zhao, L. Shen, and X. Zhou, “Solid wood panel defect detection and recognition system based on Faster
R-CNN,” Journal of Forestry Engineering, vol. 4, no. 3, pp. 112–117, 2019, doi: 10.13360/j.issn.2096-1359.2019.03.017.
[23] T. Gergel, M. Sedliak, T. Bucha, M. Oravec, M. Slamka, and M. Pástor, “Prediction model of wooden logs cutting patterns and its
efficiency in practice,” Applied Sciences, vol. 10, no. 9, 2020, doi: 10.3390/app10093003.
[24] A. R. D. Geus, A. R. Backes, A. B. Gontijo, G. H. Q. Albuquerque, and J. R. Souza, “Amazon wood species classification: a
comparison between deep learning and pre-designed features,” Wood Science and Technology, vol. 55, no. 3, pp. 857–872, 2021,
doi: 10.1007/s00226-021-01282-w.
[25] K. Zhang, B. Yuan, and Y. Li, “Efficiency analysis of wood processing industry in China during 2006-2015,” IOP Conference
Series: Materials Science and Engineering, vol. 322, no. 5, 2018, doi: 10.1088/1757-899X/322/5/052062.
[26] R. Qayyum, K. Kamal, T. Zafar, and S. Mathavan, “Wood defects classification using GLCM based features and PSO trained
neural network,” in 2016 22nd International Conference on Automation and Computing (ICAC), 2016, pp. 273–277, doi:
10.1109/IConAC.2016.7604931.
[27] R. N. N. Rahiddin, U. R. Hashim, N. H. Ismail, L. Salahuddin, N. H. Choon, and S. N. Zabri, “Classification of wood defect
images using local binary pattern variants,” International Journal of Advances in Intelligent Informatics, vol. 6, no. 1, pp. 36–45,
2020, doi: 10.26555/ijain.v6i1.392.
[28] Y. Huang, J. Jing, and Z. Wang, “Fabric defect segmentation method based on deep learning,” IEEE Transactions on
Instrumentation and Measurement, vol. 70, pp. 1–15, 2021, doi: 10.1109/TIM.2020.3047190.
[29] J. Shi, Z. Li, T. Zhu, D. Wang, and C. Ni, “Defect detection of industry wood veneer based on NAS and multi-channel mask
R-CNN,” Sensors, vol. 20, no. 16, pp. 1–17, 2020, doi: 10.3390/s20164398.
[30] T. He, Y. Liu, Y. Yu, Q. Zhao, and Z. Hu, “Application of deep convolutional neural network on feature extraction and detection
of wood defects,” Measurement: Journal of the International Measurement Confederation, vol. 152, 2020, doi:
10.1016/j.measurement.2019.107357.
[31] U. Buehlmann and R. E. Thomas, “Relationship between lumber yield and board marker accuracy in rip-first rough mills,”
European Journal of Wood and Wood Products: Holz als Roh- und Werkstoff, vol. 65, no. 1, pp. 43–48, 2007, doi:
10.1007/s00107-006-0099-7.
[32] D. E. Kline, C. Surak, and P. A. Araman, “Automated hardwood lumber grading utilizing a multiple sensor machine vision
technology,” Computers and Electronics in Agriculture, vol. 41, no. 1–3, pp. 139–155, 2003, doi: 10.1016/S0168-1699(03)00048-6.
[33] T. H. Chun, U. R. Hashim, S. Ahmad, L. Salahuddin, N. H. Choon, and K. Kanchymalay, “Efficacy of the image augmentation
method using CNN transfer learning in identification of timber defect,” International Journal of Advanced Computer Science and
Applications, vol. 13, no. 5, pp. 107–114, 2022, doi: 10.14569/IJACSA.2022.0130514.
[34] M. H. Harun et al., “The investigation on defect recognition system using Gaussian smoothing and template matching approach,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 18, no. 2, pp. 812–820, 2020, doi:
10.11591/ijeecs.v18.i2.pp812-820.

 ISSN: 2252-8938
1706
[35] Y. Zhang, S. Liu, W. Tu, H. Yu, and C. Li, “Using computer vision and compressed sensing for wood plate surface detection,”
Optical Engineering, vol. 54, no. 10, 2015, doi: 10.1117/1.oe.54.10.103102.
[36] P. M. Bhatt et al., “Image-based surface defect detection using deep learning: a review,” Journal of Computing and Information
Science in Engineering, vol. 21, no. 4, 2021, doi: 10.1115/1.4049535.
[37] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015, doi: 10.1038/nature14539.
[38] Y. Xin et al., “Machine learning and deep learning methods for cybersecurity,” IEEE Access, vol. 6, pp. 35365–35381, 2018, doi:
10.1109/ACCESS.2018.2836950.
[39] R. Ehtisham et al., “Computing the characteristics of defects in wooden structures using image processing and CNN,” Automation
in Construction, vol. 158, 2024, doi: 10.1016/j.autcon.2023.105211.
[40] T. H. Chun, U. R. Hashim, and S. Ahmad, “Timber defect identification: Enhanced classification with residual networks,”
International Journal of Advanced Computer Science and Applications, vol. 15, no. 4, pp. 665–671, 2024, doi:
10.14569/IJACSA.2024.0150468.
[41] R. Ehtisham et al., “Classification of defects in wooden structures using pre-trained models of convolutional neural network,”
Case Studies in Construction Materials, vol. 19, 2023, doi: 10.1016/j.cscm.2023.e02530.
[42] X. Zou, C. Wu, H. Liu, and Z. Yu, “Improved ResNet-50 model for identifying defects on wood surfaces,” Signal, Image and
Video Processing, vol. 17, no. 6, pp. 3119–3126, 2023, doi: 10.1007/s11760-023-02533-y.
[43] Y. Xie and J. Ling, “Wood defect classification based on lightweight convolutional neural networks,” BioResources, vol. 18,
no. 4, pp. 7663–7680, 2023, doi: 10.15376/biores.18.4.7663-7680.
[44] L. C. Chen et al., “Edge-glued wooden panel defect detection using deep learning,” Wood Science and Technology, vol. 56, no. 2,
pp. 477–507, 2022, doi: 10.1007/s00226-021-01316-3.
[45] J. Ling and Y. Xie, “Research on wood defects classification based on deep learning,” Wood Research, vol. 67, no. 1,
pp. 147–156, 2022, doi: 10.37763/WR.1336-4561/67.1.147156.
[46] M. Gao, J. Chen, H. Mu, and D. Qi, “A transfer residual neural network based on resnet‐34 for detection of wood knot defects,”
Forests, vol. 12, no. 2, pp. 1–16, 2021, doi: 10.3390/f12020212.
[47] M. Gao, F. Wang, P. Song, J. Liu, and D. W. Qi, “BLNN: Multiscale feature fusion-based bilinear fine-grained convolutional
neural network for image classification of wood knot defects,” Journal of Sensors, vol. 2021, 2021, doi: 10.1155/2021/8109496.
[48] G. Ramesh, T. Siddhartha, K. Sivaraman, and V. Subramani, “Identification of timber defects using convolution neural network,”
in 2021 6th International Conference on Communication and Electronics Systems (ICCES), 2021, pp. 1641–1647, doi:
10.1109/ICCES51350.2021.9489136.
[49] F. Wu, R. Gazo, E. Haviarova, and B. Benes, “Wood identification based on longitudinal section images by using deep learning,”
Wood Science and Technology, vol. 55, no. 2, pp. 553–563, 2021, doi: 10.1007/s00226-021-01261-1.
[50] K. Hu, B. Wang, Y. Shen, J. Guan, and Y. Cai, “Defect identification method for poplar veneer based on progressive growing
generated adversarial network and MASK R-CNN model,” Bioresources, vol. 15, no. 2, pp. 3041–3052, 2018.
[51] T. He, Y. Liu, C. Xu, X. Zhou, Z. Hu, and J. Fan, “A fully convolutional neural network for wood defect location and
identification,” IEEE Access, vol. 7, pp. 123453–123462, 2019, doi: 10.1109/ACCESS.2019.2937461.
[52] A. P.-Taraseviciene, K. Sutiene, and L. Pipiras, “Wooden dowels classification using convolutional neural network,” Proceedings of
the Romanian Academy Series A - Mathematics Physics Technical Sciences Information Science, vol. 20, no. 4, pp. 401–409, 2019.
[53] J. Hu, W. Song, W. Zhang, Y. Zhao, and A. Yilmaz, “Deep learning for use in lumber classification tasks,” Wood Science and
Technology, vol. 53, no. 2, pp. 505–517, 2019, doi: 10.1007/s00226-019-01086-z.
[54] H. Kim et al., “Visual classification of wood knots using k-nearest neighbor and convolutional neural network,” Journal of the
Korean Wood Science and Technology, vol. 47, no. 2, pp. 229–238, 2019, doi: 10.5658/WOOD.2019.47.2.229.
[55] N. Rudakov, T. Eerola, L. Lensu, H. Kälviäinen, and H. Haario, “Detection of mechanical damages in sawn timber using
convolutional neural networks,” in Pattern Recognition, 2019, pp. 115–126, doi: 10.1007/978-3-030-12939-2_9.
[56] B. Jena, G. K. Nayak, and S. Saxena, “Convolutional neural network and its pretrained models for image classification and object
detection: A survey,” Concurrency and Computation: Practice and Experience, vol. 34, no. 6, 2022, doi: 10.1002/cpe.6767.
[57] P. Simon and V. Uma, “Review of texture descriptors for texture classification,” Advances in Intelligent Systems and Computing,
vol. 542, pp. 159–176, 2018, doi: 10.1007/978-981-10-3223-3_15.
[58] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. Cambridge, Massachusetts: MIT Press, 2016.
[59] P. Simon and V. Uma, “Deep learning based feature extraction for texture classification,” Procedia Computer Science, vol. 171,
pp. 1680–1687, 2020, doi: 10.1016/j.procs.2020.04.180.
[60] Y. Huang, C. Qiu, X. Wang, S. Wang, and K. Yuan, “A compact convolutional neural network for surface defect inspection,”
Sensors, vol. 20, no. 7, 2020, doi: 10.3390/s20071974.
[61] G. Romanovskis, K. Bumanis, G. Kulikovskis, and P. Rivza, “Location and identification of oak wood defect by deep learning,”
Engineering for Rural Development, vol. 20, pp. 1837–1842, 2021, doi: 10.22616/ERDev.2021.20.TF402.
[62] P. Rawat, A. Pandey, and A. K. Panaiyappan, “Rice leaf diseases classification using deep learning techniques,” in 2023
International Conference on Networking and Communications (ICNWC), 2023, pp. 1–8, doi:
10.1109/ICNWC57852.2023.10127315.
[63] H. C. Shin et al., “Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics
and transfer learning,” IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1285–1298, 2016, doi:
10.1109/TMI.2016.2528162.
[64] S. Akçay, M. E. Kundegorski, M. Devereux, and T. P. Breckon, “Transfer learning using convolutional neural networks for object
classification within X-ray baggage security imagery,” International Conference on Image Processing, ICIP, pp. 1057–1061,
2016, doi: 10.1109/ICIP.2016.7532519.
[65] S. Christodoulidis, M. Anthimopoulos, L. Ebner, A. Christe, and S. Mougiakakou, “Multisource transfer learning with
Convolutional Neural Networks for lung pattern analysis,” IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 1,
pp. 76–84, 2017, doi: 10.1109/JBHI.2016.2636929.
[66] H. Yang, S. Mei, K. Song, B. Tao, and Z. Yin, “Transfer-learning-based online mura defect classification,” IEEE Transactions on
Semiconductor Manufacturing, vol. 31, no. 1, pp. 116–123, 2018, doi: 10.1109/TSM.2017.2777499.
[67] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10,
pp. 1345–1359, 2010, doi: 10.1109/TKDE.2009.191.
[68] N. Tajbakhsh et al., “Convolutional neural networks for medical image analysis: full training or fine tuning?,” IEEE transactions
on medical imaging, vol. 35, no. 5, pp. 1299–1312, 2016.
[69] M. Gao, P. Song, F. Wang, J. Liu, A. Mandelis, and D. Qi, “A novel deep convolutional neural network based on ResNet-18 and
transfer learning for detection of wood knot defects,” Journal of Sensors, vol. 2021, 2021, doi: 10.1155/2021/4428964.

1707
BIOGRAPHIES OF AUTHORS
Martina Ali completed her undergraduate studies in computer science, concentrating
on information security and assurance, at Universiti Sains Islam Malaysia in 2015. Following
this, she held the position of lecturer at the School of Computing Technology and Engineering,
International College of Yayasan Melaka, from 2015 to 2023. She is pursuing her master’s degree
in information technology at Universiti Teknikal Malaysia Melaka, a program that began in 2023.
She can be contacted at email: m032220016@student.utem.edu.my.
Ummi Raba’ah Hashim received her bachelor’s degree in computer science from
Universiti Teknologi Malaysia in 2000, master of multimedia from Multimedia University in
2008, and Ph.D. in computer science from Universiti Teknologi Malaysia in 2015. She is
currently a senior lecturer in the Department of Software Engineering, Faculty of Information and
Communication Technology, Universiti Teknikal Malaysia Melaka (UTeM). Her research
interests include pattern recognition, machine vision, and automated defect inspection. She is also
a member of the Center for Advanced Computing Technology, UTeM. She can be contacted at
email: ummi@utem.edu.my.
Kasturi Kanchymalay received her bachelor’s degree in computer science (Hons)
from Universiti Malaya in 1997, master’s in IT (computer science) from Universiti Kebangsaan
Malaysia in 2005, and Ph.D. in computer science from Universiti Teknologi Malaysia in 2020.
She is a senior lecturer in the Department of Software Engineering, Faculty of Information and
Communication Technology, Universiti Teknikal Malaysia Melaka (UTeM). Her research
interests include machine learning, deep learning, and time series prediction. She can be
contacted at email: kasturi@utem.edu.my.
Aji Prasetya Wibawa is an esteemed academic figure who holds the position of
distinguished professor at Universitas Negeri Malang (UM) and is a successful editor-in-chief in
knowledge engineering and data science. The individual obtained a doctor of philosophy degree
in 2015 from the University of South Australia, where they developed and refined their
knowledge and skills in artificial intelligence, natural language processing, and social informatics.
He is vice-chair of the Association for Scientific Computing Electronics and Engineering
(ASCEE), showcasing his dedication to promoting computational science and engineering
progress. His extensive expertise and unwavering dedication to pioneering research have
consistently resulted in substantial contributions to the academic sphere and the broader scientific
community. He can be contacted at email: aji.prasetya.ft@um.ac.id.
Lizawati Salahuddin is a senior lecturer at the Department of Software Engineering,
Fakulti Teknologi Maklumat dan Komunikasi, Universiti Teknikal Malaysia Melaka (UTeM),
Malaysia. She received a bachelor’s degree in computer science (software engineering) from
Universiti Teknologi Malaysia (UTM) in 2005 and an M.Sc. in biosystems from Korea Advanced
Institute of Science and Technology (KAIST), South Korea, in 2008. She obtained a Ph.D. in
information systems from Universiti Teknologi Malaysia (UTM) in 2016. Her research areas of
interest include information systems, health information technology, mobile health, and
technology adoption. She has published over 50 articles in high-impact journals and proceedings
and was granted more than RM100K on competitive research calls. She can be contacted at
email: lizawati@utem.edu.my.
Rahillda Nadhirah Norizzaty Rahiddin is a Ph.D. candidate in information and
communication technology at the Universiti Teknikal Malaysia Melaka (UTeM). Her research
interests are pattern recognition and machine learning. Before pursuing her studies, she received
her diploma in information and communication from UTeM in 2012, her bachelor of computer
science (software development) with honours in 2016, and her master of science in information
and communication technology in 2022 from UTeM. She has gained 8 years of experience as a
software programmer, data engineer, SQL developer, and data warehouse (ETL) support.
Furthermore, she is also a member of the Malaysia Board of Technologists (MBOT). She can be
contacted at email: p032120002@student.utem.edu.my.

A review of recent deep learning applications in wood surface defect identification

More Related Content

More from IAESIJAI (20)

Recently uploaded (20)

A review of recent deep learning applications in wood surface defect identification