Abstract
Surface inspection plays a critical role in the wood industry, as it helps companies to enhance product quality, improves the utilization of wood resources, and increases the value of final products. In recent years, deep learning has emerged as a promising technique in this domain, offering significant advantages over traditional methods by enabling high-precision, real-time inspection. This paper presents a comprehensive review of advancements in the field from 2021 to 2025. It begins with a brief overview of three foundational aspects: common types of wood defects, publicly available datasets, and evaluation metrics. The core of the review then examines recent deep learning applications, organized according to three computer vision tasks—classification, detection, and segmentation. The paper concludes by discussing key challenges and proposing viable directions for future research, thereby offering a clear technical roadmap.
Download PDF
Full Article
Recent Advances in Wood Surface Defect Inspection using Deep Learning (2021-2025)
Yiming Fang, Yan Ma,* Shuang Gao, and Junlei Chen
Surface inspection plays a critical role in the wood industry, as it helps companies to enhance product quality, improves the utilization of wood resources, and increases the value of final products. In recent years, deep learning has emerged as a promising technique in this domain, offering significant advantages over traditional methods by enabling high-precision, real-time inspection. This paper presents a comprehensive review of advancements in the field from 2021 to 2025. It begins with a brief overview of three foundational aspects: common types of wood defects, publicly available datasets, and evaluation metrics. The core of the review then examines recent deep learning applications, organized according to three computer vision tasks—classification, detection, and segmentation. The paper concludes by discussing key challenges and proposing viable directions for future research, thereby offering a clear technical roadmap.
DOI: 10.15376/biores.21.3.Fang
Keywords: Machine vision; Deep learning; Surface defect inspection; Wood product; Literature review
Contact information: School of Intelligent Engineering, Shaoxing University, Shaoxing 312000, China; *Corresponding author: mayan7687@163.com
INTRODUCTION
Wood remains an essential natural resource in human civilization and continues to be widely used despite technological advances and the emergence of alternative materials. However, it is prone to various defects, including knots, cracks, warping, twisting, decay, and insect damage (Ross et al. 2023). These imperfections not only compromise the aesthetic quality of wood but also adversely affect its mechanical properties, thereby reducing its practical utility and economic value. For instance, knots induce localized stress concentrations that undermine structural integrity. Such defects are also susceptible to splitting during processing, resulting in undesirable irregular fractures and surface imperfections in the final product (Fang et al. 2021). Against this backdrop, accurate detection of wood surface defects along with precise evaluation of their type, severity, and distribution can enable scientific wood grading, which in turn can greatly enhance wood utilization efficiency and economic benefits (Ehtisham et al. 2024).
Traditional defect inspection in wood primarily relies on manual methods, which are subjective and labor-intensive, failing to meet the demands of modern automated manufacturing (Chen et al. 2023b). Moreover, the reliability of manual inspection rarely exceeds 70%, as it is affected by human factors such as visual fatigue and lapses in concentration (Urbonas et al. 2019). Consequently, alternative inspection techniques have been developed, including acoustic, X-ray, and terahertz methods. Among these, vision-based systems integrated with machine learning have emerged as a promising solution for wood surface defect inspection due to their cost-effectiveness, efficiency, and adaptability to field conditions (Wang et al. 2021b).
Extensive research on this approach can be found in the literature, as exemplified by the comprehensive reviews of Kamal et al. (2017) and Urbonas et al. (2019). Although the types of input images vary across studies, most approaches follow two key steps: feature extraction and classification, as illustrated in Fig. 1a. For feature extraction, techniques such as Gray Level Co-occurrence Matrix (Savolainen 2023), Gabor filters (Hittawe et al. 2015), and Local Binary Pattern analysis (Li et al. 2019) have been employed to capture the distinctive color and texture characteristics of wood surface defects. Subsequently, classifiers including Artificial Neural Networks (Hwang et al. 2022) and Support Vector Machines (Gu et al. 2009) analyze these features and produce the final decision.
Alternative methods such as clustering (Silvén et al. 2003) and compressed sensing (Zhang et al. 2016) have also been explored. However, both feature extraction and classifier training are manually designed, making them dependent on human expertise. Furthermore, the performance of these methods degrades with changes in wood species, surface conditions, or lighting environments (He et al. 2019). Therefore, developing reliable techniques for wood surface defect inspection remains an ongoing challenge for the wood industry.
Fig. 1. The concepts of (a) traditional machine learning and (b) deep learning
Deep learning, an advanced subfield of machine learning, autonomously learns feature hierarchies from image datasets, as illustrated in Fig. 1b. Its capacity to extract discriminative features with minimal human intervention or specialized domain knowledge (Alzubaidi et al. 2021) has profoundly influenced wood surface defect inspection, positioning deep learning as a leading research trend. The number of related publications has grown substantially, with a marked acceleration during the period 2021 to 2025. This growth underscores the need for a dedicated review. Although previous reviews by An et al. (2022) and Yi et al. (2024) have discussed both traditional and deep learning methods, they did not focus specifically on the latter, thereby failing to provide a comprehensive overview of deep learning-centric research developments.
This paper summarizes advances in wood surface defect inspection, with particular emphasis on research conducted between 2021 and 2025. It covers key aspects including wood surface image datasets, network architectures, evaluation metrics, challenges, and alternative solutions. Researchers and students will gain a clear understanding of the topic, which will help identify promising directions for future work.
KEY ASPECTS IN WOOD SURFACE DEFECT INSPECTION
Main Wood Surface Defects
The categories of wood defects considered vary across industries, depending on their specific objectives. In the construction industry, defect inspection primarily focuses on assessing strength-related characteristics, with emphasis on insect damage, decay, and splits. In the furniture industry, inspections aim to eliminate both aesthetic imperfections and structural weaknesses, paying particular attention to knots, holes, and cracks. Table 1 summarizes the most frequently investigated wood surface defect types in the literature.
Table 1. The Types of Wood Surface Defects Concerned
Live knots, dead knots, and checks are consistently identified as critical defects in the literature due to their substantial impact on the quality of wood products. As shown in Fig. 2a, a live knot originates from a branch that was integrated into the trunk of a living tree. A high density of live knots complicates the wood grain pattern and reduces its ornamental value. In contrast, a dead knot (Fig. 2b) results from a deceased branch. Its fiber structure is often partially or completely detached from the surrounding wood tissue, significantly compromising the mechanical properties of solid wood panels. Checks (Fig. 2c) are fissures caused by the separation of wood fibers, typically due to external stresses. They reduce the wood’s shear strength parallel to the grain and adversely affect its overall structural integrity. Moreover, checks are susceptible to fungal infection, which can lead to rot and progressive deterioration of the wood (Ding et al. 2020).
Fig. 2. Three common types of defect images (Ding et al. 2020): (a) live knot, b) dead knot, and (c) checking
Wood Image Datasets
The construction of high-quality datasets is a foundational step in deep learning research, as such data are essential for both training and evaluating models. However, data acquisition and annotation are inherently time-consuming and labor-intensive. Consequently, the availability of high-quality datasets remains limited, with public access being even more restricted. Table 2 summarizes typical open-access wood surface image datasets.
Table 2. Typical Open-access Datasets of Wood Surface Images
In 2021, Kodytek et al. (2021) released a large-scale dataset of high-resolution sawn timber surface images collected in an industrial environment. This dataset contains 43,000 annotated images covering ten common types of wood defects. Each image is annotated with dual labels, enabling defect classification, localization, and semantic segmentation. Despite its relatively recent publication (approximately three years ago), this dataset has been widely adopted as a benchmark for validating numerous novel methodologies (Ehtisham et al. 2023; Meng and Yuan 2023; Wang et al. 2023a; Wang et al. 2023b; Zou et al. 2023; An et al. 2024; Ehtisham et al. 2024; Luo et al. 2024; Wang et al. 2024a; Wang et al. 2024b; Zheng et al. 2024b; Chen et al. 2025; Kılıç et al. 2025; Wang et al. 2025a; Wang et al. 2025b; Xi et al. 2024; Zou et al. 2025).
In addition, a limited number of research teams have shared their self-developed datasets via GitHub repositories. For instance, Tu et al. (2021) released a dataset of 1,545 sawn rubber wood images featuring four defect types: intergrown knots, dead knots, growth shakes, and inbark. Separately, Zhong et al. (2024) provided two distinct datasets: one for pine lumber and the other for rubberwood lumber. The pine dataset contains 2,686 annotated images with three typical defects (dead knots, sound knots, and missing edges), while the rubber dataset includes 1,362 annotated images with five typical defects (dead knots, sound knots, missing edges, tree cores, and cracks). All images in these datasets have a resolution of 658 × 492 pixels. Furthermore, Li et al. (2025) compiled a dataset of 950 wood panel images sourced from individual users on the Roboflow platform and shared it via a GitHub repository.
In contrast to these efforts, VNWoodKnot is a recently released publicly available image dataset comprising 1,515 high-resolution wood surface images. It includes three categories: live knots, dead knots, and knot-free surfaces. As such, it serves as a critical benchmark for developing learning models for industrial-grade wood defect inspection (Tran et al. 2025). Many researchers have also constructed proprietary datasets for their investigations. As summarized in Table 3, these self-developed datasets are typically small in scale. The largest among them comprises only about 5,000 images, which is far fewer than the dataset shared by Kodytek et al. (2021). The smallest contains merely 310 samples. Although studies using these datasets have reported satisfactory results, with average recognition rates exceeding 90%, their practical utility in real-world industrial settings remains significantly limited.
Table 3. Some Typical Self-developed Datasets
Evaluation Metrics
Different applications impose distinct requirements on wood surface defect inspection methods, necessitating the use of diverse evaluation criteria. Existing research primarily focuses on two aspects: accuracy and efficiency.
Accuracy evaluation metrics
Accuracy metrics are used to quantify the predictive performance of algorithms in classifying, detecting, and segmenting wood surface defects. As summarized in Table 4, commonly adopted metrics include accuracy, precision, recall, F1-score, the receiver operating characteristic (ROC) curve, intersection over union (IoU), average precision (AP), and their derivatives such as mean average precision (mAP) and mean IoU (mIoU). Together, these metrics form a comprehensive evaluation framework that provides critical insights for algorithmic refinement and advancement.
Efficiency evaluation metrics
Efficiency metrics assess the processing speed and computational complexity of algorithms. As summarized in Table 5, commonly used efficiency metrics include the number of parameters, floating point operations (FLOPs), inference time, and frames per second (FPS). These indicators are vital for enabling fast and accurate identification of wood surface defects.
DEEP LEARNING MODELS FOR VISUAL INSPECTION OF WOOD SURFACE DEFECT
Similar to general computer vision tasks, deep learning techniques for wood surface defect inspection can be broadly classified into three categories: classification, detection, and segmentation. As illustrated in Fig. 3a, defect classification methods identify the presence and type of defects in wood surface images. Defect detection methods extend beyond classification by localizing defects using rectangular bounding boxes. Current detection algorithms are mainly divided into one-stage and two-stage approaches. Representative one-stage algorithms include the You Only Look Once (YOLO) series and the Single Shot Multibox Detector (SSD). Over the years, the YOLO architecture has evolved significantly, with successive versions from YOLOv1 to YOLOv13 introducing consistent improvements in accuracy, speed, and efficiency. Representative two-stage algorithms belong to the region-based CNN (R-CNN) family, which includes R-CNN, Fast R-CNN, and Faster R-CNN. The developmental trajectories of these two model families are shown in Figs. 3b and 3c, respectively. Figure 3d illustrates defect segmentation methods, which perform pixel-level segmentation of defect regions from the background, thereby providing both geometric information and defect categorization.
Table 4. Common Accuracy Evaluation Metrics for Different Inspection Methods
Table 5. Common Efficiency Evaluation Metrics
Fig. 3. Visual inspection of wood surface defects. Wood defect (a) classification, (b) detection using one-stage algorithm, (c) detection using two-stage algorithm, and (d) segmentation
Wood Surface Defect Classification
In the early stages of this field, several comparative studies were conducted on various wood surface image datasets. Hacıefendioğlu et al. (2022) reported that ResNet-50 and Xception achieved over 90% accuracy in classifying damage to wooden building elements, outperforming VGG-16, VGG-19, and Inception-V3. Chun et al. (2022) found that ResNet50 performed better than ShuffleNet, AlexNet, MobileNetV2, NASNetMobile, and GoogLeNet, attaining a best accuracy of 94.59% in detecting timber defects across four Malaysian timber species. Ehtisham et al. (2023) evaluated ten pre-trained CNN models-ResNet18, ResNet50, ResNet101, GoogLeNet, ShuffleNet, Inception-V3, MobileNet-V2, Xception, Inception-ResNet-V2, and NASNetMobile-on a dataset of 9,000 wood surface images. Among these, Inception-V3 demonstrated superior performance in defect classification based on accuracy, processing speed, and overall efficacy.
Several researchers have improved classification accuracy by modifying existing deep learning architectures. Zou et al. (2023) integrated the Convolutional Block Attention Module (CBAM) and the Cross-Stage Partial Network (CSPNet) into ResNet-50, and replaced the Adam optimizer with the Ranger optimizer during training. These modifications significantly enhanced accuracy while reducing training time. Xie and Ling (2023) introduced two modifications to the RegNet architecture: incorporating an attention mechanism and adding the Ghost module into the bottleneck to replace some group convolution structures. The modified RegNet achieved a high classification accuracy of 96.58% with a smaller weight file. Wang modified the cross-entropy loss in a standard CNN classifier by adding Var Loss to the main loss function. The modified loss made the classifier more sensitive to hard samples regardless of batch size, while Var Loss reduced prediction confidence fluctuations, improving model robustness for practical use (Wang 2025). Yeh et al. (2025) proposed a Gabor Convolutional Network (GCN) that integrates CNNs with Gabor filters for wood defect recognition. Using the Taguchi method to optimize key hyperparameters, the optimized GCN achieved 98.92% accuracy on the MVTec wood dataset, outperforming a baseline Taguchi-optimized CNN by 2.73%.
The development of self-designed deep learning architectures has also been a focus of recent research. Wang (2022) proposed a novel dual-parallel network to classify light and dark wood defects, enhancing robustness against color diversity. Compared with complex CNN architectures, this design offered higher inference speeds, making it more suitable for industrial applications. Zhang et al. (2025a) introduced a novel network for particleboard defect detection based on a Densely Connected Dual Attention mechanism (DC-DACB). This architecture integrates three core modules-defect detection, classification, and feature enhancement-with the feature enhancement module embedded within the former two to synergistically boost overall performance.
Wood Surface Defect Detection
One-stage models
YOLO v3 and v4. The YOLO series is widely used in object detection due to its lightweight architecture, which entails a reduced number of parameters and lower computational complexity, thus achieving fast inference speed. However, a frequently cited limitation is its substantial computational demand, which hinders widespread deployment. Moreover, the conventional cross-entropy loss function often provides suboptimal guidance for model training. To address these issues, Wang et al. (2021a) replaced the backbone (i.e., the core feature extraction network) of YOLOv3 with a Ghost module, a lightweight convolutional module that generates features through low-cost linear operations, creating a more lightweight architecture that increased online detection speed for wood surface defects to 28 FPS. Musa et al. (2023) introduced a large-scale dataset of 3,500 annotated wood surface images across seven defect categories and reported a mean Average Precision (mAP) of 67.3% using YOLOv4-tiny.
In a separate approach, Akhyar et al. (2022) simplified YOLOv4 by reducing residual block components, backbone channels, and the residual network within the Path Aggregation Network (PAN). Experiments on pine and rubber lumber datasets showed that these modifications improved FPS by 10.8 compared to the original YOLOv4. Tu et al. (2021) proposed GC-YOLOv3 for accurate and fast detection of sawn lumber defects, using a Gaussian function to determine prediction box coordinates and localization uncertainty, and incorporating a focal IoU loss that considers overlap area, center points, and aspect ratio. These innovations effectively enhanced detection accuracy. Lim et al. (2023) integrated Efficient Channel Attention (ECA) modules into YOLOv4-Tiny to boost detection performance with minimal impact on model size, followed by an aggressive iterative channel pruning strategy that drastically reduced model size, achieving an 88% reduction in parameters.
YOLO v5, and X. As an enhanced iteration, YOLOv5 has demonstrated superior performance over YOLOv3-SPP in knot detection for sawn timbers in terms of detection accuracy, training speed, and model size (Fang et al. 2021). YOLOv5 also proved effective for automated inspection of small defects on hardwood flooring in production line settings (Truong et al. 2023). Several task-specific modifications have been made to YOLOv5 and YOLOX for wood surface defect detection. Yu et al. (2022) replaced the original PANet in YOLOv5 with a Bidirectional Feature Pyramid Network (BiFPN). Cheng et al. (2022) substituted the ReLU activation function with SiLU and incorporated a Squeeze-and-Excitation (SE) attention block to enhance inter-channel feature extraction. Meng and Yuan (2023) implemented three key changes: a Semi-Global Network (SGN) to replace the C3 module, an Extended Efficient Layer Aggregation Network (E-ELAN) embedded into the backbone, and the Efficient IoU (EIOU) loss function. Xu et al. (2023) developed YOLOv5-C3Ghost for detecting five wood defects, reporting a maximum improvement of 1.6% in mAP@0.5:0.95. Zheng et al. (2024b) proposed GBCD-YOLO, an enhanced YOLOv5s variant, achieving a 13.45% increase in mAP@0.5 (reaching 88.72%), an 11.95% increase in mAP@0.5:0.95, a 6.25% increase in FPS, and a 15.49% reduction in parameters.
Beyond YOLOv5, Li et al. (2022) adapted YOLOX for detecting four types of lumber defects by integrating an ECA attention mechanism and an ASSF module for multi-feature adaptive fusion into the feature fusion network, and replacing BCELoss and IoU loss with FocalLoss and EIoU loss. This model achieved 96.42% mAP@0.5 and a detection speed of 46.7 FPS. Han et al. (2023) introduced STC-YOLOv5 by integrating Coordinate Attention, a Transformer Encoder, and a Swin Transformer module into YOLOv5, resulting in a 3.1% mAP improvement over the baseline.
YOLO v7 and v8. YOLOv7 and YOLOv8 offer significant advancements in inference speed and detection accuracy. Afaf et al. (2024) evaluated YOLOv5, YOLOv5-tiny, and YOLOv7 on a wood image dataset, finding that YOLOv7 achieves higher accuracy in classifying defects including live knots, dead knots, knots with cracks, missing knots, cracks, resin, blue stain, and marrow. Fan et al. (2024) employed YOLOv7 to inspect lumber surfaces and used the detection results to guide a bilateral sawing strategy for targeted defect removal, reportedly increasing sawn timber volume yield by 12.3%. Fang et al. (2024) compared YOLOv8, Faster R-CNN, and Mask R-CNN for wood defect inspection, demonstrating that YOLOv8 excels in both detection speed and accuracy, with strong robustness on complex backgrounds and multiple defect types.
Recent research has focused on modifying YOLOv7 and YOLOv8 to further enhance speed, accuracy, and robustness. Table 6 summarizes the latest work, showing that YOLOv8 is most frequently selected as the baseline architecture, likely due to the short six-month interval between the releases of YOLOv7 and YOLOv8. Common enhancement strategies include integrating attention mechanisms, modifying feature extraction or fusion networks, and optimizing loss functions. All enhanced models have demonstrated superior detection performance over their baselines, with reported mAP improvements ranging from at least 3.8% (Zhang et al. 2024) to as high as 9% (Wang et al. 2023b). Most optimized models also exhibit faster inference speeds; for example, Chen et al. (2025) reported a 21.6% reduction in parameter count and a 15.6% decrease in computational complexity.
Table 6. Improved Models Using YOLOv7 or YOLOv8
Other recent YOLO versions. Since 2024, several new YOLO versions (YOLOv9 to v12) have been released. However, their application in wood surface defect detection remained limited by December 2025 due to their recency. Comparative studies have shown that these generic YOLO networks underperform relative to purpose-designed algorithms for this task (Guo et al. 2025; Hoang et al. 2025; Zhao et al. 2025). In contrast, modified versions based on these new models have demonstrated superior performance. Jiang et al. (2025) proposed a lightweight enhanced model based on YOLOv11n to improve crack identification in wooden components under complex backgrounds. He et al. (2025) developed LE-YOLO, which optimizes YOLOv11n in terms of loss function, backbone, neck, and detection head. On the Chipboardv1.0 dataset, this algorithm achieved a 4–6% increase in mAP, a 12.69% improvement in inference speed, and an 18.6% reduction in parameters.
Other one-stage models. Beyond the YOLO series, the SSD framework has also been used for wood surface defect detection. Yang et al. (2021) replaced the VGG network in SSD with ResNet to refine input features for bounding box regression and classification, reporting an average detection time of 90 ms and 89.7% accuracy. Other customized architectures have shown considerable promise. Liu et al. (2025b) introduced MFWL-DETR, a lightweight framework with multi-scale feature fusion for detecting water-based paint defects (blisters, cracks, holes, scratches) on wooden surfaces. They also proposed MSIEGNet for identifying wood finish defects, achieving 90.8% mAP@50 and 97.4% accuracy (Liu et al. 2025a). Luo et al. (2024) developed two I2GF-Net variants: a deeper version (I2GF-Net-d) that outperforms 13 state-of-the-art methods with fewer false positives and missed detections, and a shallower version (I2GF-Net-s) prioritizing inference speed at 83.1 FPS.
Two-stage models
Two-stage object detection models have seen limited adoption for wood panel surface inspection compared to their success in other domains. Faster R-CNN has been applied in only a few instances. Fang et al. (2024) compared YOLOv8, Faster R-CNN, and Mask R-CNN for detecting knots, oil streaks, black dots, and cracks, finding that Faster R-CNN achieved the fastest convergence and best training results within the first 24 epochs.
Several modifications have been made to the Faster R-CNN framework to improve its performance in wood defect detection. Zou et al. (2025) refined three core components-the backbone network, the loss function, and soft non-maximum suppression (NMS), thereby achieving a higher mAP than classical networks. In a study on timber knot detection, Chen et al. (2023a) observed that the standard Faster R-CNN occasionally generated multiple detections for a single knot, leading to a high false-positive rate. To address this, they introduced an overlapping bounding box filter, which increased detection precision from 90.9% to 97.5%. An improved Faster R-CNN model incorporating a ResNet-50 backbone, focal loss, and soft NMS elevated mAP by 4.38% and reduced mean detection time by 3.6% compared to the baseline (Zou et al. 2025). Tong et al. (2025) proposed an enhanced Faster R-CNN with a dual attention mechanism to tackle challenges including few-shot sample scarcity, defect diversity, and complex background interference. Key enhancements include a Wood-Region Proposal Network (WRPN) and a Wood-Feature Reconstruction Head (WFRH), which improve candidate box quality and feature reconstruction. Experimental results demonstrate state-of-the-art performance on benchmark datasets, with significant gains in AP50 and AP75 for 17 wood defect types under few-shot conditions.
In a distinct approach, Pan et al. (2021) developed an elliptical detection and localization method by integrating Faster R-CNN with a Gaussian Proposal Network (GPN). This framework outputs detected ellipses parameterized by (cx, cy, rx, ry, θ), representing the center coordinates, semi-diameters, and counterclockwise rotation angle of each ellipse. Experimental results on sawn lumber images showed that the proposed method improved the mean IoU by 10%.
Wood Surface Defect Segmentation
In the domain of computer vision, image segmentation techniques are predominantly categorized into two paradigms: semantic segmentation and instance segmentation. Semantic segmentation operates at the pixel level, assigning each pixel to a predefined defect category, such as live knots, dead knots, or resin pockets. Instance segmentation extends this capability by not only performing pixel-wise classification but also differentiating between distinct individual instances within the same category (Urtans et al. 2022).
Semantic segmentation
Mohsin et al. (2023) introduced a novel framework for real-time detection of wood planks on a high-speed conveyor system, integrating a CNN network for simultaneous surface defect segmentation. The proposed architecture comprises three core components: a backbone network for feature extraction, a detection algorithm for plank identification, and a segmentation module for real-time defect analysis. Zhong et al. (2024) proposed a deep Gaussian attention segmentation network for lumber surface defect detection. The network first employs a self-attention mechanism-incorporating both position and channel attention modules-to aggregate contextual information. It then integrates two Gaussian modules into these attention mechanisms to facilitate the fusion of enhanced and salient features. Additionally, Gaussian attention modules (GAMs) are incorporated into the Deeplabv3+ architecture, operating in parallel with the Atrous Spatial Pyramid Pooling (ASPP) module, to effectively merge multiscale, minor, and highly discriminative features. Zhao et al. (2022) proposed the YOLOv5-Seg-Lab-4 model, which integrates object detection with semantic segmentation to ensure real-time performance while enhancing detection accuracy. Deployed in a particleboard factory, the model successfully automated the classification and grading of boards containing defects such as sand leakage, big shavings, glue spots, oil pollution, and soft areas. The system achieved an accuracy of 98.2%, with processing times ranging from 183 to 208 ms per image.
Luo et al. (2025) proposed a layer-wise adapter module (LAM) to adapt visual foundation models for wood surface defect segmentation. The LAM integrates three key components: an instance-linking token module (ILTM) to enhance sensitivity to ambiguous boundaries and improve instance-level feature representation; a feature disentanglement module (FDM) to reduce feature redundancy and increase inter-class feature independence; and a layer switch module (LSM) to dynamically activate feature refinement for optimal layer-specific adaptation. By addressing challenges such as high inter-class similarity and fuzzy boundaries, LAM significantly improved segmentation performance on both Rubber Wood and Pine Wood datasets. Zhu et al. (2024) proposed a U-Net-based multi-source data fusion network for wood break detection. Their approach employs an improved ResNet34 backbone to extract multi-level features from image and depth data using depthwise separable convolutions (DSC) and dilated convolutions (DC), which reduce computational cost and feature redundancy. Features from the two modalities are then optimally integrated through an adaptive interacting fusion module (AIF), generating accurate feature representations for broken defects. Based on Attention U-Net, Dong et al. (2025) proposed a new model called IECAU-Net. This was achieved by incorporating CBAM and ECA modules, replacing the optimizer with AdamW, and using a weighted fusion multi-loss function. The resulting IECAU-Net outperformed other models in 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 challenges of defect variety, obscure boundaries, and size and shape variations in lumber surface defect segmentation (Zhang et al. 2025b).
Beyond fully supervised methods, researchers have also addressed the challenge of acquiring costly, high-quality datasets for training deep segmentation models by introducing generators capable of synthesizing pixel-level annotations. Tsai et al. (2021) proposed a two-stage deep learning method for pixel-wise defect detection on textured surfaces without manual annotations. Their approach employs two Cycle-Consistent Adversarial Networks (CycleGANs) to automatically synthesize defective images and generate corresponding pixel-wise annotations. The output from this synthesis stage is used to create a large-scale synthetic dataset, which subsequently trains a U-Net model for precise defect detection. This framework requires only a small set of real defect samples and eliminates labor-intensive human labeling, offering both practical implementability and computational efficiency in manufacturing environments. Zheng et al. (2024a) proposed a semi-supervised defect detection approach based on generative adversarial networks, comprising a generator and a discriminator. The generator produces pixel-level segmentation results for wood defect images. To improve segmentation performance, the discriminator is trained adversarially against the generator by assessing prediction quality and providing supervised signals for unlabeled images.
Instance segmentation
To detect defects on wood surfaces, researchers have adopted Mask R-CNN, an instance segmentation framework that simultaneously predicts a pixel-level mask and the corresponding class label for each defective region. Al-Zubi and Plapper (2022) provided a proof-of-concept by applying Mask R-CNN to synthetic images of wood panels to detect and segment drilled holes. Despite pronounced variations in color, texture, and hole appearance, the model yielded satisfactory results.
To address the challenges of modeling irregular defects and extracting contextual information, Li et al. (2021) proposed a layered deformable Mask R-CNN. By establishing layered connections between residual modules, a larger receptive field was achieved at each network layer, and deformable convolution was employed to better fit defect shapes.
Some researchers have argued that Mask R-CNN does not address the problem of calculating defect size (e.g., area and diameter). In response, Zhong et al. (2025) proposed a parallel structure fusion approach. This method incorporates two dedicated branches: one for identifying veneer knot defect types using the Inception V3 network, and the other leveraging an improved K-means clustering algorithm from traditional computer vision to localize defects and determine their real-world dimensions. Ehtisham et al. (2024) presented a similar work, where Inception-ResNet-V2 is first used to classify images into three categories: 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. (2025) 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 also demonstrating robust zero-shot transferability on unseen datasets.
CHALLENGES
Deep learning has yielded significant results in the visual inspection of wood surface defects, with some laboratory studies reporting accuracy rates approaching 100%. However, numerous challenges persist in practical applications.
Small Sample Problem
The small sample problem is a common challenge in applying deep learning algorithms, as it can easily lead to overfitting during training. An overfitted model exhibits poor generalization and struggles to recognize targets in unseen data. A primary cause of this issue is the scarcity of accurately labeled wood images. Raw wood images typically lack annotations and thus cannot be used directly for supervised training. The labeling process relies on domain experts and is often tedious and time-consuming; consequently, only a limited number of experts are available for such tasks. According to Kryl et al. (2021), most research on wood surface defect inspection utilizes datasets ranging from only 250 to 5,200 images.
To address data scarcity, researchers have increasingly adopted data augmentation (Gao et al. 2021a,c; Li et al. 2021; Chun et al. 2022; Wang 2023; Sujatha et al. 2024) and transfer learning (Gao et al. 2021a,b; Chun et al. 2022; Hacıefendioğlu et al. 2022; Ehtisham et al. 2023 ). However, these methods cannot guarantee capturing the full variability of real-world defects, particularly when such variability is potentially limitless (Kodytek et al. 2021). Therefore, the introduction and adaptation of more advanced techniques-such as self-supervised, semi-supervised, and weakly supervised learning-are anticipated to address various practical challenges in wood defect inspection under limited data conditions (Safonova et al. 2023).
Class Imbalance
Due to the differing causes of defect formation, the prevalence of various wood surface defect types varies significantly, and this imbalance is consistently observed in related datasets. For instance, Li et al. (2021) reported that knots constitute approximately 60% of all defects, with other categories collectively accounting for the remaining 40%. A more extreme example is provided by Cheng et al. (2022), whose dataset contained over 18,000 samples of live knots but nearly zero instances of blue stain, as illustrated in Fig. 4.
Such imbalanced sample distribution can bias deep learning models toward features from the majority classes, thereby compromising the reliability of the outcomes (Bai et al. 2024). To address this issue, data augmentation techniques have been widely employed to enrich existing datasets (Urbonas et al. 2019; Li et al. 2021; Riana et al. 2021; Chun et al. 2022). As another prominent generative approach, generative adversarial networks (GANs) have been extensively used to synthesize defect samples and alleviate data imbalance (Hu et al. 2020; Tsai et al. 2021; Deng et al. 2024; Zheng et al. 2024a). These methods, which can be categorized as data-level approaches, mitigate the effects of class imbalance to some extent. Furthermore, a range of algorithm-level and feature-level methods have recently emerged in other industrial domains. Such techniques also hold significant potential for application in wood surface defect inspection (Bai et al. 2024).
Fig. 4. The number of defects in various types of wood (Cheng et al. 2022)
Real-time Capabilities
For real-time operation, the inspection process must keep pace with manufacturing or image acquisition speeds. This is often challenging for deep learning-based models, as their complex and deep architectures frequently lead to reduced inference speeds. In industrial applications, however, both high accuracy and real-time performance are indispensable. Consequently, enhancing real-time capability remains a critical issue for wood surface defect inspection models, even when high accuracy has been achieved (Tu et al. 2021).
Lightweight models have been extensively employed to effectively balance accuracy and computational efficiency. Various techniques – such as factorized convolutions, group convolution, depthwise separable convolution, bottleneck design, and neural architecture search – have been explored to simplify model structures and reduce complexity. Representative models, including MFWL-DETR (Liu et al. 2025b), GBCD-YOLO (Zheng et al. 2024b), and YOLOv8-OCHD (Chen et al. 2025), have been reported to effectively inspect wood surface images with fewer parameters and lower computational demands. A typical case is MFWL-DETR, which, when applied to detect defects on water-based wood paint surfaces, achieved a 40.5% decrease in computation, a 40.2% reduction in parameters, and a 36.1% reduction in model size compared to the baseline (Liu et al. 2025b).
Quantization and pruning represent alternative approaches, although their adoption in this field has remained relatively limited to date. Lim et al. (2023) introduced a pruned model in which YOLOv4-Tiny was substantially compressed via an iterative pruning and recovery procedure. This process reduced model parameters by 88%, thereby enabling near-real-time wood defect detection on a general-purpose embedded processor without external hardware accelerators. Similarly, Mohsin et al. (2023) refined a non-quantized deep learning model through quantization-aware training, wherein both activations and weights are quantized to lower precision. This approach reduced the number of parameters by 40% and increased inference speed by approximately threefold without sacrificing overall accuracy.
FUTURE DIRECTIONS
More Efficient Algorithms
Despite the remarkable progress reported in existing research, most developed methods still face barriers to real-world practical deployment. Fortunately, wood surface defect inspection remains a relatively specialized subdomain within the broader deep learning research landscape. Advances in surface defect detection for other materials can offer valuable insights and methodological references for wood defect inspection, and it is foreseeable that emerging innovative algorithms will further narrow the gap between academic research and industrial application.
First, a continuous stream of novel CNN architectures has been proposed, such as EfficientNetV2, ConvNeXt, and MobileNetV3 (Rakesh et al. 2026). Incorporating these advanced lightweight and high-performance models is expected to facilitate low-cost, high-accuracy inspection of wood surface defects.
Second, transfer learning has evolved from a conventional technique merely targeting accuracy enhancement into an integrated methodological framework that simultaneously optimizes computational efficiency, model trustworthiness, data privacy, and interpretability. In the context of wood surface defect inspection, privacy-preserving mechanisms embedded in trustworthy transfer learning and the multi-task paradigm of cascaded transfer learning are particularly promising, as they can provide viable solutions for the industrial deployment of inspection models (Plested et al. 2026).
Finally, Vision Transformer (ViT)-based approaches have gained widespread attention in image processing tasks. Different from standard CNNs, ViT frameworks leverage the self-attention mechanism to adaptively assess the importance of individual image patches, thereby offering innovative perspectives for visual detection research. Although the application of ViT in wood surface defect detection is still in its early stage (Xie and Ling 2023), such methods have yielded encouraging results in other forestry image processing scenarios (Kılıç 2025; Zhang et al. 2023). Furthermore, hybrid architectures that combine Transformer and classic CNN structures, represented by Rep-MobileViT (Duanmu et al. 2025) and MFWL-DETR (Liu et al. 2025b), have been widely acknowledged as a promising direction to achieve superior detection performance.
Real Image Inspection
Most existing deep learning models require input images to be standardized to a fixed size with an aspect ratio of 1:1. Consequently, in current studies on wood surface defect detection, the images employed are often not of entire wood products but rather square patches extracted from authentic images, as illustrated in Fig. 2. In practical settings, however, wood products typically exhibit elongated and narrow dimensions, resulting in surface images with high width-to-height ratios. Therefore, to facilitate the application of these research findings in real-world scenarios, it is imperative to address the discrepancy between the characteristics of authentic wood surface images and the input requirements of deep learning models.
Three primary approaches have been explored to mitigate this issue. The first involves designing feature maps of varying sizes tailored to different product dimensions. While this method appears to offer an ideal solution, it significantly increases model complexity and lacks generalizability across products of differing sizes. The second method incorporates padding to transform images into a square aspect ratio by adding background regions. However, this increases computational overhead and substantially slows inspection speed, thereby rendering real-time online inspection unfeasible. The third approach involves resizing non-conforming image regions into square dimensions, which leads to pixel loss. This compromises the integrity of image details and exacerbates the difficulty of detecting small defects.
New Applications
With the advancement of research, the application of deep learning has expanded beyond its initial focus on sawn timber and wooden boards to a broader range of wood products, including glued panels (Chen et al. 2022), wooden flooring (Truong et al. 2023), wood paint (Liu et al. 2025a,b), and particleboard (Zhao et al. 2022; Zhang et al. 2025a). Furthermore, a greater variety of imaging approaches – such as spectral and thermal imaging – are being employed for the surface characterization of wood products. In these emerging applications, defects exhibit diverse colors, textures, and geometric characteristics, necessitating substantial further efforts in areas such as dataset construction, specialized model design, and extensive testing.
Second, a growing number of previously challenging application scenarios are expected to become technically feasible. A typical representative is automated wood grading, which categorizes timber into distinct quality grades according to the severity and distribution of surface defects (Sarnaghi and Kuilen 2019). Such grading enables rational allocation of timber resources to their most suitable end applications, thereby minimizing material waste and overall resource utilization efficiency. Another promising scenario is intelligent wood sawing. By preplanning cutting strategies based on the spatial distribution of wood defects, production processes can effectively avoid defective regions and substantially improve the yield of high-grade timber (Fan et al. 2024).
CONCLUSIONS
Accurate and efficient defect inspection is critical for maximizing the utilization and economic value of wood products. This review has presented a comprehensive survey of state-of-the-art deep learning-based methods for wood surface inspection. It began with a concise introduction to the foundational components of an effective inspection system, including common wood surface defects, publicly available image datasets, and standard evaluation metrics. The review then provided a taxonomy of research advances during 2021-2025, categorizing them into classification, object detection, and segmentation tasks. Finally, it outlined persistent challenges and promising future research directions. This work is expected to serve as a valuable reference and guide for researchers in the field.
ACKNOWLEDGMENTS
The authors are grateful for the support of the MOE (Ministry of Education in China) Project of Humanities and Social Sciences (Grant No. 22YJAZH017), the National Natural Science Foundation of China (Grant No. 52301379), and the Natural Science Foundation of Zhejiang Province (Grant No. LQN25E070003).
REFERENCES CITED
Afaf, S., Kalam, A. A. E., and Bouslimani, Y. (2024). “Wood surface defects detection based on AI algorithm Yolov7,” in: International Conference on Advanced Intelligent Systems for Sustainable Development (AI2SD’2023), M. Ezziyyani, M. Ezziyyani, and V. E. Balas (eds.), Springer, Cham, Switzerland, pp. 9-17. https://doi.org/10.1007/978-3-031-54288-6_2
Akhyar, F., Novamizanti, L., Putra, T., Furqon, E. N., Chang, M. C., and Lin, C. Y. (2022). “Lightning YOLOv4 for a surface defect detection system for sawn lumber,” in: 2022 IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR), IEEE, CA, USA, pp. 184-189. https://doi.org/10.1109/MIPR54900.2022.00039
Al-Zubi, M., and Plapper, P. (2022). “Segmentation of drilled holes in textured wood panels using deep learning framework,” in: 2022 IEEE 9th International Conference on Computational Intelligence and Virtual Environments for Measurement Systems and Applications (CIVEMSA), H. Haario (ed.), Chemnitz, Germany, pp. 1-4. https://doi.org/10.1109/CIVEMSA53371.2022.9853682
Alzubaidi, L., Zhang, J., Humaidi, A. J., Al‑Dujaili, A., Duan, Y., Al‑Shamma, O., Santamaría, J., Fadhel, M. A., Al‑Amidie, M., and Farhan, L. (2021). “Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions,” Journal of Big Data 8, 53. https://doi.org/10.1186/s40537-021-00444-8
An, H., Liang, Z., Qin, M., Huang, Y., Xiong, F., and Zeng, G. (2024). “Wood defect detection based on the CWB-YOLOv8 algorithm,” Journal of Wood Science 70, 26. https://doi.org/10.1186/s10086-024-02139-z
An, H., Liu, K., Liang, Z., Qin, M., Huang, Y., and Xiong, F. (2022). “Research on wood defect detection based on deep learning,” in: IEEE 2nd International Conference on Data Science and Computer Application (ICDSCA), IEEE, Dalian, China, pp. 106-109. https://doi.org/10.1109/ICDSCA56264.2022.9988006
Bai, D., Li, G., Jiang, D., Yun, J., Tao, B., Jiang, G., Sun, Y., and Ju, Z. (2024). “Surface defect detection methods for industrial products with imbalanced samples: A review of progress in the 2020s,” Engineering Applications of Artificial Intelligence 130, article 107697. https://doi.org/10.1016/j.engappai.2023.107697
Chen, L. C., Pardeshi, M. S., Lo, W. T., Sheu, R. K., Pai, K. C., Chen, C. Y., Tsai, P. Y., and Tsai, Y. T. (2022). “Edge‑glued wooden panel defect detection using deep learning,” Wood Science and Technology 56, 477-507. https://doi.org/10.1007/s00226-021-01316-3
Chen, W., Liu, J., Fang, Y., and Zhao, J. (2023a). “Timber knot detector with low false-positive results by integrating an overlapping bounding box filter with faster R-CNN algorithm,” BioResources 18(3), 4964-4976. https://doi.org/10.15376/biores.18.3.4964-4976
Chen, Y., Sun, C., Ren, Z., and Na, B. (2023b). “Review of the current state of application of wood defect recognition technology,” BioResources 18(1), 2288-2302. https://doi.org/10.15376/biores.18.1.Chen
Chen, Z., Feng, J., Zhu, X., and Wang, B. (2025). “YOLOv8-OCHD: A lightweight wood surface defect detection method based on improved YOLOv8,” IEEE Access 13, 84435-84450. https://doi.org/10.1109/ACCESS.2025.3569175
Cheng, D., Cheng, G., and Wang, X. (2022). “Real-time detection method of wood defects based on deep learning,” in: 2022 IEEE 8th International Conference on Computer and Communications (ICCC), IEEE, Chengdu, China, pp. 2192-2197. https://doi.org/10.1109/ICCC56324.2022.10066017
Chun, T. H., Hashim, U. R. a., Ahmad, S., Salahuddin, L., Choon, N. H., and Kanchymalay, K. (2022). “Efficacy of the image augmentation method using CNN transfer learning in identification of timber defect,” International Journal of Advanced Computer Science and Applications 13(5), 107-114. https://doi.org/10.14569/IJACSA.2022.0130514
Cui, W., Li, Z., Duanmu, A., Xue, S., Guo, Y., and Ni, C. (2024). “CCG-YOLOv7: A wood defect detection model for small targets using improved YOLOv7,” IEEE Access 12, 10575-10585. https://doi.org/10.1109/ACCESS.2024.3352445
Deng, F., Luo, J., Fu, L., Huang, Y., Jianle Chen1, N. L., Zhong, J., and Lam, T. L. (2024). “DG2GAN: Improving defect recognition performance with generated defect image sample,” Scientific Reports 14, article 14787. https://doi.org/10.1038/s41598-024-64716-y
Ding, F., Zhuang, Z., Liu, Y., Jiang, D., Yan, X., and Wang, Z. (2020). “Detecting defects on solid wood panels based on an improved SSD algorithm,” Sensors 20(18), article 5315. https://doi.org/10.3390/s20185315
Dong, Y., He, C., Xiang, X., Cui, Y., Kang, Y., Ding, A., Duo, H., and Wang, X. (2025). “IECAU-Net: A wood defects image segmentation network based on improved attention U-Net and attention mechanism,” BioResources 20(2), 3545-3556. https://doi.org/10.15376/biores.20.2.3545-3556
Dou, W., and You, J. (2025). “A novel wood surface defect detection model based on improved Yolov8,” BioResouces 20, 5709-5730. https://doi.org/10.15376/biores.20.3.5709-5730
Duanmu, A., Xue, S., Li, Z., Zhang, Y., and Ni, C. (2025). “Rep-MobileViT: Texture and color classification of solid wood floors based on a re-parameterized CNN-Transformer hybrid model,” IEEE Access 13, 39950-39963. https://doi.org/10.1109/ACCESS.2025.3545645
Ehtisham, R., Qayyum, W., Camp, C. V., Plevris, V., Mir, J., Khan, Q.-u. Z., and Ahmad, A. (2024). “Computing the characteristics of defects in wooden structures using image processing and CNN,” Automation in Construction 158, article 105211. https://doi.org/10.1016/j.autcon.2023.105211
Ehtisham, R., Qayyum, W., Camp, C. V., Plevris, V., Mir, J., Khan, Q. Z., and Ahmad, A. (2023). “Classification of defects in wooden structures using pre-trained models of convolutional neural network,” Case Studies in Construction Materials 19, article e02530. https://doi.org/10.1016/j.cscm.2023.e02530
Fan, C., Zhuang, Z., Liu, Y., Yang, Y., Zhou, H., and Wang, X. (2024). “Bilateral defect cutting strategy for sawn timber based on artificial intelligence defect detection model,” Sensors, 24(20), article 6697. https://doi.org/10.3390/s24206697
Fang, Y., Guo, X., Chen, K., Zhou, Z., and Ye, Q. (2021). “Accurate and automated detection of surface knots on sawn timbers using YOLO-v5 model,” BioResources 16(3), 5390-5406. https://doi.org/10.15376/biores.16.3.5390-5406
Fang, Y., Huang, W., Zheng, C., Huang, X., Lin, M. T., and Yang, J. K. (2024). “Comparative analysis of object detection algorithms for wood defect detection,” in: 2024 IEEE 7th International Conference on Electronic Information and Communication Technology (ICEICT), Xi’an, China, pp. 991-996. https://doi.org/10.1109/ICEICT61637.2024.10670842
Gao, M., Qi, D., Mu, H., and Chen, J. (2021a). “A transfer residual neural network based on ResNet-34 for detection of wood knot defects,” Forests 12(2), article 212. https://doi.org/10.3390/f12020212
Gao, M., Song, P., Wang, F., Liu, J., Mandelis, A., and Qi, D. (2021b). “A novel deep convolutional neural network based on ResNet-18 and transfer learning for detection of wood knot defects,” Journal of Sensors 2021, article 4428964. https://doi.org/10.1155/2021/4428964
Gao, M., Wang, F., Song, P., Liu, J., and Qi, D. (2021c). “BLNN: Multiscale feature fusion-based bilinear fine-grained convolutional neural network for image classification of wood knot defects,” Journal of Sensors 2021, article 8109496. https://doi.org/10.1155/2021/8109496
Ge, Y., Ji, H., and Liu, X. (2025). “Wood surface defect detection based on improved YOLOv8,” Signal, Image and Video Processing 19, 663. https://doi.org/10.1007/s11760-025-04226-0
Gu, I. Y. H., Andersson, H., and Vicen, R. (2009). “Wood defect classification based on image analysis and support vector machines,” Wood Science and Technology 44(4), 693-704. https://doi.org/10.1007/s00226-009-0287-9
Guo, H., Chai, Z., Dai, H., Yan, L., Cheng, P., and Yang, J. (2025). “PBD-YOLO: Dual-strategy integration of multi-scale feature fusion and weak texture enhancement for lightweight particleboard surface defect detection,” Applied Sciences-Basel 15, article 4343. https://doi.org/10.3390/app15084343
Guo, Y., and Cao, W. (2024). “Wood surface defect detection using improved deep learning algorithm: FRCE-YOLO,” in: 2024 6th International Conference on Electronic Engineering and Informatics (EEI), Chongqing, China, pp. 324-328. https://doi.org/10.1109/EEI63073.2024.10696054
Hacıefendioğlu, K., Ayas, S., Başağa, H. B., Toğan, V., Mostofi, F., and Can, A. (2022). “Wood construction damage detection and localization using deep convolutional neural network with transfer learning,” European Journal of Wood and Wood Products 80, 791-804. https://doi.org/10.1007/s00107-022-01815-5
Han, S., Jiang, X., and Wu, Z. (2023). “An improved YOLOv5 algorithm for wood defect detection based on attention,” IEEE Access 11, 71800-71810. https://doi.org/10.1109/ACCESS.2023.3293864
He, C., Kang, Y., Ding, A., Jia, W., and Duo, H. (2025). “LE-YOLO: A lightweight and enhanced algorithm for detecting surface defects on particleboard,” BioResources 20(3), 7179-7193. https://doi.org/10.15376/biores.20.3.7179-7193
He, T., Liu, Y., Xu, C., Zhou, X., Hu, Z., and Fan, J. (2019). “A fully convolutional neural network for wood defect location and identification,” IEEE Access 7, 123453-123462. https://doi.org/10.1109/ACCESS.2019.2937461
Hittawe, M. M., Sidibé, D., and Mériaudeau, F. (2015). “A machine vision based approach for timber knots detection,” in: 12th International Conference on Quality Control by Artificial Vision, International Society for Optics and Photonics, 95340L. https://doi.org/10.1117/12.2182770
Hoang, V. T., Le, V.-T., Dinh, N., Tran-Trung, K., Bay Nguyen Van, Hong, H. D. T., and Huong, T. H. (2025). “Deep learning-based faulty wood detection with area attention,” Computers, Materials and Continua 85(1), 1495-1514. https://doi.org/10.32604/cmc.2025.066506
Hu, K., Wang, B., Shen, Y., Guan, J., and Cai, Y. (2020). “Defect identification method for poplar veneer based on progressive growing generated adversarial network and MASK R-CNN model,” BioResources 15(2), 3041-3052. https://doi.org/10.15376/biores.15.2.3041-3052
Hwang, S.-W., Lee, T., Kim, H., Chung, H., Choi, J. G., and Yeo, H. (2022). “Classification of wood knots using artificial neural networks with texture and local feature-based image descriptors,” Holzforschung 76(1), 1-13. https://doi.org/10.1515/hf-2021-0051
Jiang, Q., Zhu, X., Wang, H., Chong, X., Li, L., and Li, W. (2025). “YOLOv11n-CBP: A lightweight model for crack detection in wooden components under complex backgrounds,” Structures 80, article 110077. https://doi.org/10.1016/j.istruc.2025.110077
Kamal, K., Qayyum, R., Mathavan, S., and Zafar, T. (2017). “Wood defects classification using laws texture energy measures and supervised learning approach,” Advanced Engineering Informatics 34, 125-135. https://doi.org/10.1016/j.aei.2017.09.007
Kılıç, K. (2025). “Categorization of microscopic wood images with transfer learning approach on pretrained vision transformer models,” BioResources 20, 6394-6405. https://doi.org/10.15376/biores.20.3.6394-6405
Kılıç, K., Kılıç, K., Doğru, İ. A., and Özcan, U. (2025). “WD Detector: Deep learning-based hybrid sensor design for wood defect detection,” European Journal of Wood and Wood Products 83, 50. https://doi.org/10.1007/s00107-025-02211-5
Kodytek, P., Bodzas, A., and Bilik, P. (2021). “A large-scale image dataset of wood surface defects for automated vision-based quality control processes,” F1000Research 10, 581. https://doi.org/10.12688/f1000research.52903.2
Kryl, M., Danys, L., Jaros, R., Martinek, R., Kodytek, P., and Bilik, P. (2020). “Wood recognition and quality imaging inspection systems,” Journal of Sensors 2020, article 3217126. https://doi.org/10.1155/2020/3217126
Li, D., Xie, W., Wang, B., Zhong, W., and Wang, H. (2021). “Data augmentation and layered deformable Mask R-CNN-based detection of wood defects,” IEEE Access 9, 108162-108174. https://doi.org/10.1109/ACCESS.2021.3101247
Li, D., Zhang, Z., Wang, B., Yang, C., and Deng, L. (2022). “Detection method of timber defects based on target detection algorithm,” Measurement 203, article 111937. https://doi.org/10.1016/j.measurement.2022.111937
Li, R., Zhong, S., and Yang, X. (2025). “Wood panel defect detection based on improved YOLOv8n,” BioResources 20(2), 2556-2573. https://doi.org/10.15376/biores.20.2.2556-2573
Li, S., Li, D., and Yuan, W. (2019). “Wood defect classification based on two-dimen-sional histogram constituted by LBP and local binary differential excitation pattern,” IEEE Access 7, 145829-145842. https://doi.org/10.1109/ACCESS.2019.2945355
Lim, W.-H., Bonab, M. B., and Chua, K. H. (2023). “An aggressively pruned CNN model with visual attention for near real-time wood defects detection on embedded processors,” IEEE Access 11, 36834-36848. https://doi.org/10.1109/ACCESS.2023.3266737
Liu, C., Chen, K., Wang, N., Pei, Y., Jia, N., and Zhang, Y. (2025a). “MSIEGNet: A multiscale information-enhanced global network focusing on wood finish defects,” Signal, Image and Video Processing 19, 717. https://doi.org/10.1007/s11760-025-04317-y
Liu, C., Chen, K., Wang, N., Shi, W., and Jia, N. (2025b). “A lightweight multi-scale feature fusion method for detecting defects in water-based wood paint surfaces,” Measurement 253, article 117505. https://doi.org/10.1016/j.measurement.2025.117505
Long, Y., and Lin, W. (2025). “Surface defect detection of ultrathin fiberboard based on improved YOLOv8x,” Journal of Nondestructive Evaluation 44, 58. https://doi.org/10.1007/s10921-025-01196-8
Luo, Q., Xu, W., Su, J., Yang, C., Gui, W., and Silvén, O. (2024). “I²GF-Net: Interlayer information guidance feedback networks for wood surface defect detection in complex texture backgrounds,” IEEE Transactions on Instrumentation and Measurement 73, 5023013. https://doi.org/10.1109/TIM.2024.3413144
Luo, Q., Xu, W., Su, J., Yang, C., Gui, W., and Silvén, O. (2025). “Efficient adaptation of visual foundation models for wood defect segmentation via instance linking and feature disentanglement,” IEEE Transactions on Instrumentation and Measurement 74, 5031913. https://doi.org/10.1109/TIM.2025.3565348
Meng, W., and Yuan, Y. (2023). “SGN-YOLO: Detecting wood defects with improved YOLOv5 based on semi-global network,” Sensors 23(21), article 8705. https://doi.org/10.3390/s23218705
Mohsin, M., Balogun, O. S., Haataja, K., and Toivanen, P. (2023). “Convolutional neural networks for real-time wood plank detection and defect segmentation,” F1000Research 12, 319. https://doi.org/10.12688/f1000research.131905.1
Musa, A. M. b., Momin, M. A., Khairuddin, A. S. M., Khairuddin, U., Ahmad, A., and Rosli, N. R. (2023). “Automated wood surface defects recognition system using Yolov4-tiny model,” in: 2023 Innovations in Power and Advanced Computing Technologies (i-PACT), IEEE, Kuala Lumpur, Malaysia, pp. 1-5. https://doi.org/10.1109/i-PACT58649.2023.10434717
Pan, S., Fan, S., Wong, S. W. K., Zidek, J. V., and Rhodin, H. (2021). “Ellipse detection and localization with applications to knots in sawn lumber images,” in: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), IEEE, Waikoloa, HI, USA, pp. 3891-3900. https://doi.org/10.1109/WACV48630.2021.00394
Plested, J., Phiri, M., and Gedeon, T. (2026). “Deep transfer learning for image classification: A survey,” Artificial Intelligence Review 59, article 100. https://doi.org/10.1007/s10462-026-11491-z
Rakesh, V. K., Mazumdar, S., Samanta, T., Pandey, H. K., Das, A., and Pal, S. (2026). “Analysis of hyperparameter optimization effects on lightweight deep models for real-time image classification,” Scientific Reports 2026, 1-18. https://doi.org/10.1038/s41598-026-42748-w
Riana, D., Rahayu, S., Hasan, M., and Anton. (2021). “Comparison of segmentation and identification of swietenia mahagoni wood defects with augmentation images,” Heliyon 7, article e07417. https://doi.org/10.1016/j.heliyon.2021.e07417
Ross, R. J., Wang, X., and Senalik, C. A. (2023). “Nondestructive evaluation of wood products,” in: Handbook of Wood Science and Technology, P. Niemz, A. Teischinger, and D. Sandberg (eds.), Springer, Cham, Switzerland, pp. 991-1017. https://doi.org/10.1007/978-3-030-81315-4_19
Safonova, A., Ghazaryan, G., Stiller, S., Main-Knorn, M., Nendel, C., and Ryo, M. (2023). “Ten deep learning techniques to address small data problems with remote sensing,” International Journal of Applied Earth Observation and Geoinformation 125, article 103569. https://doi.org/10.1016/j.jag.2023.103569
Savolainen, J. (2023). “Matching method for mutated veneer sheet images using gray‑level co‑occurrence matrix features,” European Journal of Wood and Wood Products 81, 1021-1031. https://doi.org/10.1007/s00107-023-01946-3
Shi, J., Li, Z., Zhu, T., Wang, D., and Ni, C. (2020). “Defect detection of industry wood veneer based on NAS and multi-channel mask R-CNN,” Sensors 20(16), article 4398. https://doi.org/10.3390/s20164398
Silvén, O., Niskanen, M., and Kauppinen, H. (2003). “Wood inspection with non-supervised clustering,” Machine Vision and Applications 13, 275-285. https://doi.org/10.1007/s00138-002-0084-z
Sujatha, R., Srisakthi, K., Sanjana, S., and Supritha, P. (2024). “DenseNet framework for wood quality assessment: Predicting defects in wood images,” in: 2024 5th International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV), IEEE, Tirunelveli, India, pp. 432-438. https://doi.org/10.1109/ICICV62344.2024.00074
Tong, X., Liang, Z., Qin, M., Liu, F., JiayuYang, Xiao, H., and Da, W. (2025). “DAM-Faster RCNN: Few-shot defect detection method for wood based on dual attention mechanism,” Sci. Reports 15, 22860. https://doi.org/10.1038/s41598-025-05479-y
Tran, V., Lam, D., and Le, T. (2025). “VNWoodKnot: A benchmark image dataset for wood knot detection and classification,” Data in Brief 62, article 112039. https://doi.org/10.1016/j.dib.2025.112039
Truong, V. D., Xia, J., Jeong, Y., and Yoon, J. (2023). “An automatic machine vision-based algorithm for inspection of hardwood flooring defects during manufacturing,” Engineering Applications of Artificial Intelligence 123, article 106268. https://doi.org/10.1016/j.engappai.2023.106268
Tsai, D. M., Fan, S. K. S., and Chou, Y. H. (2021). “Auto-annotated deep segmentation for surface defect detection,” IEEE Transactions on Instrumentation and Measurement 70, 5011410. https://doi.org/10.1109/TIM.2021.3087826
Tu, Y., Ling, Z., Guo, S., and Wen, H. (2021). “An accurate and real-time surface defects detection method for sawn lumber,” IEEE Transactions on Instrumentation and Measurement 70, 2501911. https://doi.org/10.1109/TIM.2020.3024431
Urbonas, A., Raudonis, V., Maskeliunas, R., and Damaševiˇcius, R. (2019). “Automated identification of wood veneer surface defects using faster region-based convolutional neural network with data augmentation and transfer learning,” Applied Sciences-Basel 9, article 4898. https://doi.org/10.3390/app9224898
Urtans, E., Bumanis, K., Vecins, V., Ancans, M., Andrijanova, A., and Upenieks, M. T. (2022). “Detection of knots in oak wood planks: Instance versus semantic segmenta-tion,” in: 2022 IEEE 5th International Conference on Big Data and Artificial Intelligence (BDAI), IEEE, Fuzhou, China, 163-168. https://doi.org/10.1109/BDAI56143.2022.9862633
Wang, B., Wang, R., Chen, Y., Yang, C., Teng, X., and Sun, P. (2025a). “FDD-YOLO: A novel detection model for detecting surface defects in wood,” Forests 16(2), article 308. https://doi.org/10.3390/f16020308
Wang, B., Yang, C., Ding, Y., and Qin, G. (2021a). “Detection of wood surface defects based on improved YOLOv3 algorithm,” BioResources 16(4), 6766-6780. https://doi.org/10.15376/biores.16.4.6766-6780
Wang, M., Li, M., Cui, W., Xiang, X., and Duo, H. (2023a). “TSW-YOLO-v8n: Optimization of detection algorithms for surface defects on sawn timber,” BioResources 18(4), 8444-8457. https://doi.org/10.15376/biores.18.4.8444-8457
Wang, R., Chen, Y., Liang, F., Wang, B., Mou, X., and Zhang, G. (2024a). “BPN-YOLO: A novel method for wood defect detection based on YOLOv7,” Forests 15(7), article 1096. https://doi.org/10.3390/f15071096
Wang, R., Chen, Y., Zhang, G., Liang, F., Mou, X., and Jin, H. (2025b). “DRR-YOLO: A multiscale wood surface defect detection method based on improved YOLOv8,” IEEE Sensors Journal 25(10), 16702-16719. https://doi.org/10.1109/JSEN.2025.3552848
Wang, R., Liang, F., Wang, B., and Mou, X. (2023b). “ODCA-YOLO: An omni-dynamic convolution coordinate attention-based YOLO for wood defect detection,” Forests 14(9), article 1885. https://doi.org/10.3390/f14091885
Wang, R., Liang, F., Wang, B., Zhang, G., Chen, Y., and Mou, X. (2024b). “An efficient and accurate surface defect detection method for wood based on improved YOLOv8,” Forests 15(7), article 1176. https://doi.org/10.3390/f15071176
Wang, X. (2022). “An optimized wood knots recognition scheme based on double detection,” in: 2022 International Conference on Computer Engineering and Artificial Intelligence (ICCEAI), IEEE, Shijiazhuang, China, pp. 84-88. https://doi.org/10.1109/ICCEAI55464.2022.00026
Wang, X. (2023). “Detection of natural wood defects with large color differences based on branched network,” Multimedia Tools and Applications 82, 44719-44739. https://doi.org/10.1007/s11042-023-15487-7
Wang, X. (2025). “Research on wood defects feature imbalance optimization and recognition,” IEEE Access 13, 23841-23850. https://doi.org/10.1109/ACCESS.2025.3538285
Wang, Y., Zhang, W., Gao, R., Jin, Z., and Wang, X. (2021b). “Recent advances in the application of deep learning methods to forestry,” Wood Science and Technology 55, 1171-1202. https://doi.org/10.1007/s00226-021-01309-2
Xi, H., Wang, R., Liang, F., Chen, Y., Zhang, G., Zhang, G., and Wang, B. (2024). “SiM-YOLO: A wood surface defect detection method based on the improved YOLOv8,” Coatings 14(8), article 1001. https://doi.org/10.3390/coatings14081001
Xia, B., Luo, H., and Shi, S. (2022). “Improved Faster R-CNN based surface defect detection algorithm for plates,” Computational Intelligence and Neuroscience 2022(1), article 3248722. https://doi.org/10.1155/2022/3248722
Xie, Y., and Ling, J. (2023). “Wood defect classification based on lightweight convolu-tional neural networks,” BioResources 18(4), 7663-7680. https://doi.org/10.15376/biores.18.4.7663-7680
Xu, J., Yang, H., Wan, Z., Mu, H., Qi, D., and Han, S. (2023). “Wood surface defects detection based on the improved YOLOv5-C3Ghost with SimAm module,” IEEE Access 11, 105281-105287. https://doi.org/10.1109/ACCESS.2023.3303890
Yang, Y., Wang, H., Jiang, D., and Hu, Z. (2021). “Surface detection of solid wood defects based on SSD improved with ResNet,” Forests 12(10), article 1419. https://doi.org/10.3390/f12101419
Yeh, M.-F., Luo, C.-C., and Liu, Y.-C. (2025). “Optimization of Gabor convolutional networks using the Taguchi method and their application in wood defect detection,” Applied Sciences-Basel 15, article 9557. https://doi.org/10.3390/app15179557
Yi, L. P., Akbar, M. F., Wahab, M. N. A., Rosdi, B. A., Fauthan, M. A., and Shrifan, N. H. M. M. (2024). “The prospect of artificial intelligence-based wood surface inspection: A review,” IEEE Access 12, 84706-84725. https://doi.org/10.1109/ACCESS.2024.3412928
Yu, C., Bian, E., and Wang, Y. (2022). “Research on surface defect detection method of solid wood board based on improved YOLOv5s,” in: 2022 IEEE 4th International Conference on Civil Aviation Safety and Information Technology (ICCASIT), Dali, China, pp. 959-963. https://doi.org/10.1109/ICCASIT55263.2022.9986834
Zhang, C., Wang, C., Zhao, L., Qu, X., and Gao, X. (2025a). “A method of particleboard surface defect detection and recognition based on deep learning,” Wood Material Science & Engineering 20(1), 50-61. https://doi.org/10.1080/17480272.2024.2323579
Zhang, Q., Liu, L., Yang, Z., Yin, J., and Jing, Z. (2024). “WLSD-YOLO: A model for detecting surface defects in wood lumber,” IEEE Access 12, 65088-65098. https://doi.org/10.1109/ACCESS.2024.3395623
Zhang, S., Ling, Z., Peng, P., Zhong, Y., and Wen, H. (2025b). “Dual-branch discrimi-native feature cross-attention network for lumber surface defect segmentation,” IEEE Transactions on Instrumentation and Measurement 74, 5049013. https://doi.org/10.1109/TIM.2025.3617405
Zhang, Y., Liu, S., Cao, J., Li, C., and Yu, H. (2016). “Wood board image processing based on dual-tree complex wavelet feature selection and compressed sensing,” Wood Science and Technology 50, 297-311. https://doi.org/10.1007/s00226-015-0776-y
Zhao, Z., Ge, Z., Jia, M., Yang, X., Ding, R., and Zhou, Y. (2022). “A particleboard surface defect detection method research based on the deep learning algorithm,” Sensors 22(20), article 7733. https://doi.org/10.3390/s22207733
Zhao, Z., Han, Y., Wang, M., Yang, X., and Wu, X. (2025). “ERTD: A multi-scale fusion and feature enhancement network for efficient and real-time detection of surface defects in particleboard,” AIP Advances 15, 125101. https://doi.org/10.1063/5.0307743
Zheng, K., Zheng, S., Xu, Z., Lin, Y., and Zhu, Y. (2024a). “Semi-supervised learning with GAN for semantic segmentation of wood defect image,” in: China Automation Congress (CAC), IEEE, Qingdao, China, 6760-6764. https://doi.org/10.1109/CAC63892.2024.10865192
Zheng, Y., Wang, M., Zhang, B., Shi, X., and Chang, Q. (2024b). “GBCD-YOLO: A high-precision and real-time lightweight model for wood defect detection,” IEEE Access 12, 12853-12868. https://doi.org/10.1109/ACCESS.2024.3356048
Zhong, L., Dai, Z., Zhang, Z., Sun, Y., Cao, Y., and Wang, L. (2025). “Identification and localization of veneer knot defects based on parallel structure fusion approach,” European Journal of Wood and Wood Products 82, 1301-1317. https://doi.org/10.1007/s00107-024-02086-y
Zhong, Y., Ling, Z., Liu, L., and Zhang, S. (2024). “Deep Gaussian attention network for lumber surface defect segmentation,” IEEE Transactions on Instrumentation and Measurement 73, 5015512. https://doi.org/10.1109/TIM.2024.3381269
Zhu, Y., Xu, Z., Lin, Y., Chen, D., Ai, Z., and Zhang, H. (2024). “A multi-source data fusion network for wood surface broken defect segmentation,” Sensors 24(5), article 1635. https://doi.org/10.3390/s24051635
Zou, X., Wu, C., Liu, H., and Yu, Z. (2023). “Improved ResNet-50 model for identifying defects on wood surfaces,” Signal, Image and Video Processing 17, 3119-3126. https://doi.org/10.1007/s11760-023-02533-y
Zou, X., Wu, C., Liu, H., Yu, Z., and Kuang, X. (2025). “An accurate object detection of wood defects using an improved Faster R-CNN model,” Wood Material Science & Engineering 20(2), 413-419. https://doi.org/10.1080/17480272.2024.2352605
Article submitted: April 5, 2026; Peer review completed: April 24, 2026; Revised version received and accepted: May 8, 2026; Published: May 20, 2026.
DOI: 10.15376/biores.21.3.Fang