Wood surface image classification using multi-attention graph Neural Networks

Abstract

This study aims to address the challenge of detecting wood surface defects with high accuracy in real-world manufacturing environments, which are affected by various conditions. The goal is to build a hybrid deep learning model that can operate effectively in real time, serving automated quality control in wood processing. The research team proposes a hybrid architecture that combines multiple models: You Only Look Once – version 10 and CenterNet for accurate and fast localization of defect areas, Graph Attention Network to exploit spatial relationships between features, and Multilayer Perceptron for the final classification stage. At the same time, the Battle Royale Optimization algorithm is applied to select the optimal feature. The model is trained and tested on a large-scale dataset of 20,275 high-resolution wood surface images, covering 10 different types of defects. Experiments show that the model achieves a classification accuracy of 92.7% and a processing speed of 40 fps, meeting the requirements of real-time industrial systems. The model also demonstrates good generalization ability when effectively detecting both obvious defects and subtle abnormalities. The proposed hybrid modeling framework not only improves the efficiency of wood surface defect detection but also provides an automated, robust, and highly scalable solution for quality control in smart manufacturing. This contributes to optimizing the use of raw materials, minimizing waste, and enhancing the competitiveness of the wood processing industry.

Introduction

The surface quality of wood products is a key determinant of customer satisfaction, market competitiveness, and product longevity in the wood processing and furniture manufacturing industries. Surface defects such as cracks, wormholes, knots, and discoloration frequently arise due to machining errors, environmental exposure, or handling during transportation. These anomalies significantly compromise the esthetic appeal and structural integrity of final products, often resulting in increased material waste and decreased product value. ¹ Traditional inspection methods such as visual assessment, tactile inspection, and manual measurements are still widely practiced and regulated (TCVN 1757:1975, TCVN 8931:2013), but these approaches are labor-intensive, prone to inconsistency, and unsuitable for high-throughput environments. The demand for scalable, real-time, and accurate defect detection solutions has grown significantly in light of Industry 4.0 trends and smart manufacturing initiatives. ²

In recent years, automated inspection systems utilizing image processing, 3D scanning, and surface roughness measurement tools have been introduced to address objectivity and repeatability issues. Nonetheless, ensuring consistent real-time performance in dynamic factory environments often characterized by varying lighting, surface textures, and mechanical vibrations remains challenging. ³ Deep learning models, especially Convolutional Neural Networks (CNNs), have shown strong potential in surface defect detection. However, single-stage detectors like You Only Look Once (YOLO) and Single Shot MultiBox Detector (SSD) offer speed at the expense of detection robustness, while two-stage models such as Faster R-CNN and Mask R-CNN provide greater accuracy but require high computational power, limiting their real-time deployment. Furthermore, many models suffer from poor generalization across different wood species and defect types, and they often struggle to capture spatial and contextual relationships between defect features.^4,5 Although some hybrid approaches using optimization algorithms (such as Particle Swarm Optimization (PSO), Battle Royale Optimization (BRO)) or graph-based models (such as Graph Attention Network (GAT)) have been explored, there remains a lack of integrated solutions that balance speed, accuracy, and adaptability in production-scale applications.

Recent studies have shown the efficacy of Artificial Intelligence (AI) in mechanical fault detection and health management in various sectors. For instance, machine learning has been effectively applied in robotic rotary vector reducers and strain wave gear systems under variable operational speeds. Deep learning-based models like attention-guided feature aggregation networks have proven useful in diagnosing servo motor bearing faults, while transfer learning has enabled robust fault detection in industrial robotics. These findings underscore the significance of intelligent, adaptable, and data-driven approaches in real-time industrial applications, including wood defect inspection.

The research addresses the following questions: (1) How to balance real-time performance with accuracy in defect detection? (2) How to generalize detection across varying defect types and wood surfaces?. To overcome these limitations, this study proposes a novel hybrid framework for wood surface defect detection that integrates multiple advanced components: YOLOv10 and CenterNet for object detection, a Multi-Attention Graph Neural Network (MAGNN) for spatial-contextual learning, a Multi-Layer Perceptron (MLP) for final classification, and the BRO algorithm for feature selection. Among many available metaheuristic algorithms, we selected the BRO algorithm due to its ability to maintain population diversity and avoid local optima through competitive survival-based search strategies. Compared to more commonly used methods like PSO or Genetic Algorithm (GA). The BRO algorithm offers a balance between convergence speed and global search accuracy, which is critical for selecting optimal features from complex wood surface datasets. The system is further enhanced by the camera to support real-time defect localization under variable lighting and surface conditions. Unlike prior work, the proposed model leverages the complementary strengths of these components to achieve high detection accuracy, fast inference speed, and improved robustness across diverse wood textures and defect types. This paper offers the following key contributions:

Novel hybrid framework

We propose the first integration of YOLOv10 and CenterNet for defect detection, GAT for spatial-context learning, MLP for classification, and BRO for feature optimization, a configuration not previously explored in wood defect detection.

Real-time industrial application

Achieves high accuracy (92.7%) with 40 Frames Per Second (FPS), validated on a dataset of 20,275 industrial images under real-world conditions, surpassing existing models like YOLOv8 and Mask R-CNN.

Enhanced generalization

Our model addresses the limitation of poor adaptability in existing models by using GAT + BRO to robustly learn spatial and contextual relationships across varying wood species and defect types.

Impactful for smart manufacturing

Designed for easy integration into Industry 4.0 production lines, reducing defect-related waste and increasing product quality with minimal computational overhead.

This framework is intended to serve as a foundation for intelligent, automated quality control systems that minimize human intervention, reduce material waste, and enhance overall production efficiency. The paper is structured as follows: Section II explores related works. Section III elaborates on the proposed method. Section IV discusses experimental outcomes. Section V offers conclusions and future research directions.

Related works

Traditional inspection methods

Traditional inspection of wood surfaces largely relies on manual visual observation and tactile evaluation. While standardized by national codes such as TCVN 1757:1975 and TCVN 8931:2013, these methods are prone to subjectivity and inconsistency, especially when detecting small or hidden defects such as internal cracks, wormholes, or discoloration.^6,7 Studies have shown that such methods are inefficient in high-throughput environments and struggle with accurate classification in complex defect scenarios.^8,9

Machine learning approaches

Machine Learning (ML) has improved defect detection in terms of speed and automation. Techniques such as Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), and Decision Trees have achieved accuracies ranging from 85% to 94%, depending on dataset size and wood type.^10,11 However, these models often fail to generalize well to new or complex data, especially with small or overlapping defects, and typically require large, labeled datasets for reliable performance.^12,13

Deep learning approaches

Deep Learning (DL), particularly Convolutional Neural Networks (CNNs), has demonstrated superior accuracy (90%–96%) in defect classification, thanks to automatic feature extraction from raw image data.^14,15 Models such as Residual Network (ResNet), YOLO, EfficientNet, and CenterNet have been widely applied. Nevertheless, DL methods face challenges such as high computational costs, the need for Graphics Processing Unit (GPU) resources, and limited effectiveness when only Red–Green–Blue (RGB) images are used, especially for detecting subtle depth-related defects. ¹⁶ Although classical CNN-based models like ResNet and VGGNet show strong performance, they often struggle with spatial generalization and computational efficiency in high-speed environments. ¹⁷ Our method improves upon these by integrating GAT for spatial attention and BRO for optimal feature selection, achieving better balance in speed and accuracy.

Hybrid models and optimization algorithms

Recent studies propose hybrid models combining detection and optimization techniques. ¹⁸ The integration of YOLO with optimization methods like PSO or ACO (Ant Colony Optimization) has yielded improved classification rates, particularly in specific wood types.^19,20 However, these models are often not scalable due to complexity and long computation times. Some recent efforts incorporate Graph Neural Network (GNN), which offer advantages in capturing spatial relationships and graph-structured data. Liu et al. ²¹ applied a CNN-GAT hybrid, improving classification performance to 99.37%. Yet, few studies have explored multi-attention GNNs combined with optimization strategies like BRO in real-time environments, highlighting the research gap this study addresses.

Proposed method

This study proposes a real-time model for identifying and classifying surface defects in wood during the final inspection stage on industrial conveyor belts. The model consists of three main components: image rule recognition, image annotation, model training, and real-time defect detection. Surface images are collected using industrial cameras and undergo preprocessing steps such as resizing, cropping, and labeling. The processed images are then used to train a deep learning model. The proposed approach combines a GAT within a GNN framework with a MLP for classification. GAT captures complex relationships between image regions, while MLP performs effective classification. This hybrid architecture allows for accurate and robust detection of wood surface defects in dynamic and uncertain manufacturing environments. Experimental results demonstrate that the proposed model significantly improves recognition performance and enables real-time application (Figure 1).

Figure 1.

Overview of the proposed method.

Set $G$ include ( $n+m$ ) surface image of wood with $n$ image surface of the wood that $G_1=\left\{d_1\right\}$ and $m$ is a new image surface of wood that $G_2=\left\{d_j\right\}$ , with $G_1\cup G_2=G$ and $G_1\cap G_2=\varnothing$ and $D_1=\{G_x,G_y\}$ is interactive between $G_x\in G_1$ with $G_y\in G_1$ and $D_2=\{G_x,G_z\}$ is interactive between $G_x\in G_1$ and $G_z\in G_2$ . Each type of image is represented as a structural diagram node and the author represents them as a graph $G_i(V_i,E_i)$ with $V_i=\{V_{i1},V_{i2},\dots ,V_{\mathrm{ip}}\}$ is a set of nodes representing the surface shapes of wood ( $d_i$ ). $E_i=\{(V_{\mathrm{is}},V_{\mathrm{it}}{)}_{s,t=1}^p\}$ is a presentation of images of edges in the connection of the surface of wood shapes ( $d_i$ ). $H_i^{(0)}=(h_{i1}^0,h_{i2}^0,\dots ,h_{\mathrm{ip}}^0{)}^T$ is the initial characterization of the node matrix $p$ in the $G_1$ parameter of the surface of the wood image ( $d_i$ ). The task of this research is to identify candidates in the surface of wood images that are not subject to management in the surface of wood image dataset, and this research presents two scenarios used to predict the surface of wood images as follows:

In this study, each wood surface image is represented as a graph $G_i(V_i,E_i)$ , where nodes $V_i$ correspond to defect-related feature regions extracted after detection using YOLOv10 and CenterNet, and edges $E_i$ represent spatial or contextual relationships between these regions.The set $G_1$ denotes labeled training images, while $G_2$ represents newly observed or testing images during inference. The interaction sets $D_1$ and $D_2$ model relationships either within the training set or between training and unseen images. The mapping function $F:G\times G\to \{0,1\}$ is introduced to learn whether two graph representations exhibit meaningful relational similarity, thereby enabling the Graph Attention Network (GAT) to capture structural dependencies between defect regions.Instead of performing classification directly on individual images, the proposed GAT-based framework leverages graph representations to model spatial and contextual relationships among defect features. The resulting latent embedding vector $Z_i$ encodes both local defect characteristics and their structural interactions. This embedding is subsequently fed into the MLP classifier to perform final defect category prediction. Therefore, the graph formulation enhances relational feature learning and improves classification robustness, rather than redefining the core classification task.

Data input

The input dataset used in this study consists of 20,275 high-resolution wood surface images (2800 × 1024 pixels) collected under real industrial conditions using a JAI SW-4000TL-PMCL line-scan camera at 66.4 fps. ²² Each image is annotated with bounding box labels and semantic segmentation maps identifying 10 common wood defects, including live knots, dead knots, cracks, resin, and blue stains. On average, each image contains 2.2 defects, with some having up to 16. Preprocessing steps include grayscale conversion, noise reduction, image binarization, Sobel edge detection, and histogram-based filtering to remove empty conveyor images. Cropping is performed to eliminate unnecessary backgrounds and reduce computation time. The dataset is well-structured with both pixel-level and object-level annotations, enabling applications in object detection, classification, and segmentation. Its quality and diversity make it suitable for training deep learning models in automated wood inspection systems.

Preprocessing

The testing and processing of wood images with defects involves a series of operations to rectify the issues. The simulated image of wood defects is processed using a digital image card, dividing the image into small pixels for efficient computer processing. The processed wood image, a digital lattice of 512 × 512 pixels, can be transformed by a computer to determine the gray value and brightness of each pixel. The image digitization process involved collecting and sampling wood defect images, resulting in an N × N array, which was then used for further analysis (equation (1)).

$$\begin{array}{c}f(x,y)=\left[\begin{array}{cccc}f(0,0)\hfill & f(0,1)\hfill & \cdots \hfill & f(0,N-1)\hfill \\ \multicolumn{1}{c}{f(1,0)}& f(1,1)\hfill & \cdots \hfill & f(1,N-1)\hfill \\ \multicolumn{1}{c}{\cdots }& \cdots \hfill & \cdots \hfill & \cdots \hfill \\ \multicolumn{1}{c}{f(M-1,0)}& f(M-1,1)\hfill & \cdots \hfill & f(M-1,N-1)\hfill \end{array}\right]\hfill \end{array}$$

(1)

Median filter processing

The basic principle of median filter processing involves determining the value of a point in a digital image by analyzing the middle value of the surrounding area.

Calculate the central value of the array ( $y$ ) in the data series $a_1,a_2,\dots ,a_n$ (equation (2)).

$$\begin{array}{c}\{\begin{array}{c}y=a_{\frac{n+1}{2}},\mathrm{if}n\mathrm{is}\mathrm{odd}\mathrm{number}\hfill \\ \multicolumn{1}{c}{y=\frac{1}{2}(a_{\frac{n+1}{2}}+a_{\frac{n}{2}}),\mathrm{if}n\mathrm{is}\mathrm{even}\mathrm{number}}\end{array}\hfill \end{array}$$

(2)

The median filter, a one-dimensional sliding window, determines the specific length of a point by adjusting the gray value of the middle pixel within the window. Let the digital image’s gray values be expressed as pixels, forming a two-dimensional median filter with a window of A. Calculate the value (y) in the array of two dimensions when setting $\{a_{\mathrm{ij}},(i,j)\in I^2\}$ (equation (3)). The median filter in the wood window, which is sharp, can be categorized into square, cycle, and cross.

$$\begin{array}{c}{y}_{\mathrm{ij}}=\mathrm{Med}\left\{a_{\mathrm{ij}}\right\}=\mathrm{Med}\{a_{(i+r),(j+s),(r,s)}\in A,(i,j)\in I^2\}\hfill \end{array}$$

(3)

The binary processing of a gray image in wood involves converting it into a specific threshold (value between 0 and 255), converting it into a binary image with only two gray levels (0 value: black and 255 value: white), allowing for clearer change law representation. Edge points in an image are differentiated using a characteristic that identifies the most violent gray changes, thereby confirming edge pixels in the image through second-order differential analysis. Image processing employs differential operations for derivative operations, enhancing the universality of the operation by detecting specific direct edges. Calculate the gradient value of the surface of the wood image (equation (4)).

$$\begin{array}{c}G\left[F(j,k)\right]={[{\left(\frac{\partial F}{\partial j}\right)}^2+{\left(\frac{\partial F}{\partial k}\right)}^2]}^{\frac{1}{2}}\hfill \end{array}$$

(4)

With: $F(j,k)$ : the gray value of the image.

The image gradient’s gray change is primarily observed in the gradient direction, with common operators like Sobel, Robert, Prewitt, Log, and Sobel assisting in edge detection, resulting in significant gray change. ²³

The wood image processing process consists of two main stages: preprocessing and edge detection. First, the wood image is divided into small pixels (usually a 512 × 512 matrix) for easy analysis. Next, a Median Filter is applied to remove noise, which helps smooth the image without losing important information. The image is then binarized by setting a fixed threshold to convert it into two gray levels (black – 0 and white – 255), which helps highlight the defect area from the background image. In the edge detection stage, the gradient value is first calculated, measuring the change in brightness intensity between pixels to identify the area with the strongest variation, usually the defect border. Then, Sobel Edge Detector edge detection filters are applied to identify clear boundaries of the defects on the wood surface. The final result helps to clearly distinguish cracks, wood eyes, borers, or color changes, effectively supporting the process of checking wood quality with artificial intelligence. Theorem-type environments (including propositions, lemmas, corollaries, and more) can be formatted as follows:

YOLOv10

YOLOv10 integrates features from the Task-Aligned Learning (TAL) model and introduces dual label assignments to eliminate the need for Non-Maximum Suppression (NMS), improving training efficiency. ²⁴ The model addresses the limitations of traditional NMS by applying a dimension gluing method and dual-task learning. Dual labeling helps maintain accuracy and convergence while reducing prediction errors. The two-headed architecture supports efficient optimization across varied wood surface types, lowering training costs and time. Overall, this combined approach significantly enhances automated wood quality inspection performance. Calculate the value of the consistent matching metric (equation (5)) to evaluate the level of agreement between the predicted value and the actual value when performing the training feature.

$$\begin{array}{c}m(\alpha ,\beta )=s\times p^{\alpha }\times \mathrm{IoU}\times {(\hat{b},b)}^{\beta }\hfill \end{array}$$

(5)

With: $p$ : classification score, $\hat{b}$ : prediction of the bounding box, $b$ : Instance of the bounding box, $s$ : spatial prior indicating, $\alpha ,\beta$ : two important hyperparameters.

Calculate the value that affects the labeling of the surface of the wood image and the monitoring of the two information heads. Calculate the objective function realization value for the classification operation to ensure accuracy in predicting hits that match the original targets. Calculate the distance value between two branches in supervision using the 1-Wasserstein distance to achieve the classification goal.

The YOLOv10 model features a lightweight classification head, spatial-channel decoupled down-sampling, and rank-guided block design. It enhances detection accuracy through large-kernel convolution and partial self-attention (PSA). Input and kernel sizes are carefully selected to balance processing speed and detail capture, with flexibility to adjust based on model requirements like CenterNet. YOLOv10 is highly suitable for wood image recognition due to its real-time, accurate detection of multiple defect types, including small and hard-to-spot flaws. It clearly distinguishes defect boundaries, ensuring precise classification. The model is optimized for deployment on resource-constrained devices, reducing hardware costs. This makes it ideal for integration into automated wood inspection systems, improving production efficiency and product quality.

Center net model

CenterNet ²⁵ is an efficient object detection model that replaces traditional RoI pooling by using triplets of key pointscenter, and corners for accurate object localization. It builds on CornerNet by introducing center key points to refine and filter bounding boxes, improving detection precision. Bounding boxes are generated from top-k corner predictions and validated using center key point scores and positions. A central region is defined for each box, and only those containing valid center points are retained, with scores averaged from the three key points. The architecture includes calculating central weights, measuring mapping deviation, verifying centroid alignment, and maintaining bounding boxes based on deviation limits. The model’s performance is sensitive to the size of the central region; smaller areas reduce recall, while larger ones decrease precision, highlighting the importance of a scale-aware central region. CenterNet is well-suited for detecting defects in wood surface images, especially in handling variations in size and position.

Center pooling is a technique that captures richer, more recognizable visual patterns in geometric centers. This method, known as Enriching Center, involves finding the maximum value in both horizontal and vertical directions to determine if a pixel is a center key point, improving the detection of these key points. Corner Net utilizes the principle of cascade corner pooling to efficiently manage objects with outside appearance features, enhancing their overall functionality. The principle of pooling involves determining maximum values on boundary directions, but corners are sensitive to edges, necessitating the extraction of features from central regions. Cascade corner pooling utilizes a cascade approach to determine the maximum value along a boundary, subsequently adding it to obtain visual patterns (equation (6)).

$$\begin{array}{c}\multicolumn{1}{c}{\{\begin{array}{c}\multicolumn{1}{c}{{\mathrm{ctl}}_x=\frac{(n+1){\mathrm{tl}}_x+(n-1){\mathrm{br}}_x}{2n}}\\ \multicolumn{1}{c}{{\mathrm{ctl}}_y=\frac{(n+1){\mathrm{tl}}_y+(n-1){\mathrm{br}}_y}{2n}}\\ \multicolumn{1}{c}{{\mathrm{cbr}}_x=\frac{(n-1){\mathrm{tl}}_x+(n+1){\mathrm{br}}_x}{2n}}\\ \multicolumn{1}{c}{{\mathrm{cbr}}_y=\frac{(n-1){\mathrm{tl}}_x+(n+1){\mathrm{br}}_y}{2n}}\end{array}}\end{array}$$

(6)

With: ${\mathrm{tl}}_x,{\mathrm{tl}}_y$ : top-left corner of i. ${\mathrm{br}}_x,{\mathrm{br}}_y$ : bottom-right corner of i. ${\mathrm{ctl}}_x,{\mathrm{ctl}}_y$ : top-left corner of j. ${\mathrm{cbr}}_x,{\mathrm{cbr}}_y$ : the bottom-right corner of j. $n$ : scale of the central region j.

The robust model is trained from scratch using a high-resolution input image, minimizing training loss through data augmentation strategies. Single-scale testing involves detecting bounding boxes in images with original resolutions. In multi-scale testing, images with varying resolutions are input. Bounding boxes are detected using heatmaps and redundant ones. CenterNet is highly effective in wood image processing due to its ability to accurately detect the center and shape of small or unclear defects such as cracks, knots, and bugs. Its precision in identifying small, complex objects makes it ideal for industrial wood quality inspection. When combined with YOLOv10, the system enhances both defect detection and wood surface feature extraction. YOLOv10 detects large, obvious defects, while CenterNet focuses on smaller or ambiguous ones, allowing for comprehensive analysis. Together, they extract features like size, shape, position, and texture variations, which are crucial for classifying defects and assessing wood quality. These features support automated decision-making in inspection, predict future issues such as rot or cracking, and optimize production processes. This combined model significantly improves efficiency and accuracy in wood monitoring systems.

Battle Royale Optimization Algorithm (BRO)

The Battle Royale Optimization Algorithm (BRO) ²⁶ effectively enhances feature selection for wood images after extraction by YOLOv10 and CenterNet. It identifies the most relevant features while eliminating redundant ones, reducing overfitting, and improving model generalization. BRO boosts computational efficiency, lowers system load, and increases processing speed for real-time wood quality monitoring. It also highlights critical features like cracks, knots, and beetles, improving classification accuracy. Integrating BRO with detection models enhances overall recognition performance and optimizes the wood inspection process (Shown in Figure 2).

Figure 2.

Apply BRO in the image feature selection.

BRO is a competition-based game where players must survive and be the last man standing. The game’s objective is to eliminate opponents to win. Players must survive in a smaller and smaller play area, and resurrect if injured (equation (7)).

$$\begin{array}{c}{S}_i^g=\{S_{i,1}^g,S_{i,2}^g,\dots ,S_{i,d}^g\},i=1,2,\dots ,N_p\hfill \end{array}$$

(7)

With: $N_p$ : population size, $g=0$ : initial population, $d$ : problem’s dimensionality.

Calculate the first initial value (equation (8)).

$$\begin{array}{c}{S}_{i,d}^0={\mathrm{lb}}_d+\mathrm{rand}(0,1)*({\mathrm{ub}}_d-{\mathrm{lb}}_d)\hfill \end{array}$$

(8)

With: $\mathrm{rand}(0,1)$ : Distributed random variable, ${\mathrm{ub}}_d$ : upper bounds, ${\mathrm{lb}}_d$ : Lower bounds.

The fitness level of neighboring solutions is compared, with the winner having the highest fitness level, resulting in a change in the position of the damaged solutions. The damage level of a solution in the search area is determined by its threshold, with the most damaged solution resetting to zero. The Battle Royale game utilizes the BRO algorithm to reduce the lower limits of d-dimensional space, ensuring a smaller playing area for optimal gameplay. Calculate the first iteration value. Calculate the reduction value in the spatial region. Calculate the upper bound and lower bound zone values.

The optimal outcome in a d-dimensional space is determined by updating the boundaries and standard deviation with the current location being the optimal outcome, and the best player/soldier considered an elite (equation (9)).

$$\begin{array}{c}{S}_{i,d}^{g+1}=\{\begin{array}{c}{S}_{\mathrm{dam},d}^g,\mathrm{if}f\left(S_{\mathrm{dam},d}^g\right)\le f\left(S_{i,d}^g\right)\hfill \\ \multicolumn{1}{c}{S_{i,d}^g,\mathrm{others}}\end{array}\hfill \end{array}$$

(9)

With: $f(S_{\mathrm{dam},d}^g)\le f(S_{i,d}^g)$ : evaluations of solutions on the objective function.

The selection process involves identifying the best candidate solution, known as the winner. The fitness function minimizes the optimal solutions, thereby facilitating the pursuit of optimal solutions. The pseudo-code of BRO and the hyperparameter are shown in Table 1.

Table 1.

Hyperparameter analysis.

Hyperparameter	Value	Description
Population size (N)	30	Number of agents
Max cycle	200	Maximum iterations
Shrink rate	ceil(log(MaxCycle))	Domain reduction per cycle
Fitness function	Classification accuracy	Objective to optimize
Threshold (elimination)	0.5	Minimum survival score

Details of pseudo-code
1: Randomly initialize a population $(\overrightarrow{x})$
2: Initialize all parameters
3:  $\mathrm{shrink}=\mathrm{ceil}({\log }_{10}(\mathrm{MaxCicle}))$
4:  $\Delta =\mathrm{round}(\mathrm{MaxCicle}/\mathrm{Shrink})$
5:  $\mathrm{Iter}=0$
6: While the termination criterion is not met, do
7:   $\mathrm{Iter}=\mathrm{Iter}+1$
8:  For $i=1$ : population size //compare $i^{\mathrm{th}}$ soldier with the nearest one ( $i^{\mathrm{th}}$ )
9:    $\mathrm{dam}=j$
10:    $\mathrm{vic}=i$
11:   If $f\left(x_i\right)<f(x_j)$
12:     $\mathrm{dam}=i$
13:     $\mathrm{vic}=j$
14:   End if
15:   If $x_{\mathrm{dam}.\mathrm{damge}}<\mathrm{Threshold}$
16:    For $d=1:\mathrm{Dimension}$
17:     Change the position of damaged solder based on;
18:      $x_{\mathrm{dam},d}=r(\max (x_{\mathrm{dam},s},x_{\mathrm{best},d})-\min (x_{\mathrm{dam},d},x_{\mathrm{best},d}))+\max (x_{\mathrm{dam},d},x_{\mathrm{best},d})$
19:    End for
20:     $x_{\mathrm{dam}.}\mathrm{damage}=x_i.\mathrm{damage}+1$
21:     $x_{\mathrm{vic}.}\mathrm{damage}=0$
22:   else
23:    For $d=1:\mathrm{Dimension}$
24:      $x_{\mathrm{dam},d}=r({\mathrm{ub}}_d-{\mathrm{lb}}_d)+{\mathrm{lb}}_d$
25:    End for $d$
26:    Update $f(x_{\mathrm{dam}})$
27:     $x_{\mathrm{dam}.}\mathrm{damage}=0$
28:   End if
29:   If $\mathrm{Iter}\ge \Delta$
30:    Update ( ${\mathrm{ub}}_d-{\mathrm{lb}}_d$ )
31:     $\mathrm{\Delta }=\mathrm{\Delta }+\mathrm{round}\left[\frac{\Delta }{2}\right]$
32:   End if
33:  The ${\mathrm{ub}}_d\mathrm{or}{\mathrm{lb}}_d$ exceeds the original upper or lower bound then it set to the original ${\mathrm{ub}}_d\mathrm{or}{\mathrm{lb}}_d$
34: End while
35: Select the best soldier as the solution

Graph attention Convolutional Network (GAT)

To enhance wood image recognition, a five-layer GAT model is used to learn important features such as texture, color, and geometric patterns characteristic of wood. GAT captures spatial relationships between pixels through graph-based node interactions, assigning attention weights to focus on relevant visual cues. ²⁷ To optimize feature selection, the BRO algorithm is applied to evaluate and refine initial features like wood grain, brightness, and contrast. BRO calculates the fitness of each feature using an objective function, replacing low-performing features and adjusting search boundaries through repositioning and randomization. Features are selected based on accuracy and contribution to classification performance. This iterative process ensures that only the most relevant features are retained for training. As a result, the GAT model achieves better accuracy and robustness in wood recognition and quality analysis tasks.

This research paper proposes a GNN model to predict the surface of wood images through two stages: image feature extraction and the surface of wood image prediction stage. The first stage module of the proposed model uses a five-layer GAT model containing functional properties $f_e(G_i)$ of latent vectors $Z_i$ of each type of separate image structure $G_i(V_i,E_i)$ of surface images of wood, with $Z_i\in R^{1\times k}$ . The two latent vectors of the surface of wood ( $Z_i\mathrm{and}{Z}_j$ ) are linked together to form a feature vector ( $Z_{\mathrm{ij}}$ ) of the pair of surfaces of wood images. The five-layer GAT model of the GNN model encodes the structural graphs of each surface of the wood image into feature vectors $H_i^{(k)}=(h_{i1}^{\left(k\right)},h_{i2}^{\left(k\right)},\dots ,h_{\mathrm{ip}}^{\left(k\right)}{)}^T$ , The feature structure of the surface of the wood image is very transparent. The surface of the wood image feature vector nodes ( $H_i^{(k)}$ ) of each latent layer and the surface of the wood image features ( $H_{G_i}^{\left(k\right)}$ ) are connected. together with the feature vector ( $Z_{G_i}$ ) of the surface of the wood image. The implicit characteristics of the surface of wood ( $Z_{G_i}\mathrm{and}{Z}_{G_j}$ ) are included in the MLP model to perform the estimation task. Extract the surface of wood image features at each layer of convolutions aggregated in neighboring regions according to each of their elements, and node features are updated according to each element of the structural elements image of the surface of the wood. The properties of the wood surface at each convolutional layer are captured through information, which is an important operation. The second stage of the module identifies classes that predict interaction probabilities of the surface of wood image features using input vectors ( $Z_{\mathrm{ij}}$ ) of the input surface of wood images. GNN algorithm flow chart in the surface of wood image recognition.

The five-layer GAT model uses a hidden size of 128, ReLU activation, a dropout rate of 0.3, and is trained with the Adam optimizer (learning rate = 0.001) for 500 epochs. Each layer aggregates information from neighboring nodes using learned attention weights to capture spatial patterns between defects.

Feature extractor

The graph convolutional network utilizes an information aggregation function and an update function to generate the surface of wood informative representations, utilizing the surface of wood structure graph for learning and updating. Due to the different importance of neighbor nodes, weighted aggregation can yield more effective representations for the surface of the wood in a five-layer GAT model, generating embedding representations for each atomic node without manual feature engineering.

Information aggregation and update

The feature extractor uses convolutional operations to aggregate information and update node features in each layer of the surface of a wood structure graph, capturing crucial features of the surface of a wood functional group within each structure. To enhance readability and ensure consistent notation, we present a simplified description of the GAT mechanism applied in this study. Given a graph $G_i(V_i,E_i)$ , each node is associated with a feature vector $h_i^{\left(0\right)}$ derived from the defect detection stage. At layer $k$ , Node embeddings are updated using attention-weighted aggregation of neighboring nodes (equation (10)).

$$\begin{array}{c}{h}_i^{(k+1)}=\sigma (\sum _{j\in N_i}{\alpha }_{\mathrm{ij}}.W^{\left(k\right)}.h_j^{\left(k\right)}+W^{\left(k\right)}.h_j^{\left(k\right)})\hfill \end{array}$$

(10)

With: $N_i$ : atomic node neighbors, $h_j^{\left(k\right)}$ : input feature vector, $W^{\left(k\right)}$ : trainable weight matrix. $\sigma$ : non-linear element-wise activation function. ${\alpha }_{\mathrm{ij}}$ : aggregation weight (equation (11)).

$$\{\begin{array}{c}\begin{array}{c}{\alpha }_{\mathrm{ij}}=\mathrm{softmax}\left(e_{\mathrm{ij}}\right)=\frac{\exp \left(e_{\mathrm{ij}}\right)}{\sum _{k\in N_i}\exp \left(e_{\mathrm{ik}}\right)}\hfill \\ \multicolumn{1}{c}{e_{\mathrm{ij}}=\mathrm{LeakyReLU}\left({\overrightarrow{a}}^T(w_{\mathrm{att}}.h_i^{\left(k\right)}|w_{\mathrm{att}}.h_j^{\left(k\right)})\right)}\end{array}\hfill \end{array}$$

(11)

With: ${\overrightarrow{a}}^T\in R^{2F^{\prime }}$ : weight vector. $\mathrm{LeakyReLU}$ : activation function.

The attention mechanism enables the model to selectively emphasize structurally important defect regions while suppressing irrelevant or noisy connections. After multiple GAT layers, a global pooling operation produces the graph-level embedding $Z_i$ , which encodes both local defect features and their structural relationships. The embedding $Z_i$ is then fed into the MLP classifier to perform final defect category prediction.

Pooling of atomic feature vectors

Extracting the feature elements and initial point elements of the surface of the wood image forms the latent vector characteristic Z of each surface of the wood image of each multi-layer convolutional network node. At each latent layer, the information of neighboring elements $v_{\mathrm{xi}}$ in the graph of the surface of the wood image $G_x$ is continuously gathered according to the $h_i^{(k)}$ feature of the surface of the wood image at the $v_{\mathrm{xi}}$ nodes. The updated feature vector matrix $H_i^{(k)}\in R^{p\times k}$ of each atom of the surface of the wood image $G_x$ and the size of each latent layer $k$ . The first layer feature matrix is obtained according to the image feature extraction surface of the wood of the first module. The $H_i^{(k)}$ A feature matrix of each element of the surface of the wood image is used to predict the interaction between pairs of the surface of wood images and converted into a vector to set the surface of the wood $H_{G_i}^{(k)}$ image extraction feature. Calculate the convolution values in each hidden layer when performing these conversions.

Feature aggregation for the surface of wood pairs

The feature vectors of the latent layers $H_{G_x}^k$ of the 5 GAT layers of the surface of the wood image and the feature vectors of the neighbors of the latent layers are different due to the size of the surface of the wood image structure elements being different. Extracting the surface of wood image features requires synthesizing information from neighboring regions in its element graphs. The two-step convolution operation of image feature extraction used to obtain fixed dimensions causes substructures to become incomplete and may give rise to cases that are too small or result in overstructures. be redundant. The latent feature vector function $H_{G_x}^k$ . Each type of surface of wood image is combined to represent the final features for performing the prediction task. The feature vector value of the latent vectors of the surface of wood images is formed into a feature vector representing each surface of the wood image (equation (12)), and the $h(G_x,G_y)$ value is used as the input value. for the MLP model to predict the interaction probability value between two surfaces of wood images.

$$\begin{array}{c}h(G_x,G_y)=[Z_{G_x},Z_{G_y}]\hfill \end{array}$$

(12)

The five-layer GAT model offers significant advantages in wood image classification by effectively capturing spatial relationships between pixels. Through its attention mechanism, GAT focuses on important features such as wood grain, color, and geometric patterns, improving detection accuracy. The model learns complex representations across layers, enhancing its ability to distinguish different types of wood. GAT also models correlations between features in graph space, enabling more robust classification. It minimizes the impact of noise, lighting variations, and viewing angles, maintaining high accuracy in challenging conditions. This makes it ideal for applications in material identification, wood quality analysis, and automated inspection in the wood manufacturing industry.

Multi-layer Perceptron (MLP) prediction

Multi-Layer Perceptron (MLP) prediction, ²⁸ the GNN model performs the task of predicting the surface of wood images for the task of binary classification. The MLP model is found to be an excellent performance model in the classification task with five prediction classes. At the first 4 layers, the MLP model uses the ReLU model as the activation function, and at the last layer, the SoftMax function is used. The output mapping function in the range of 0–1 serves as the ability to detect latent classes of the surface of the wood image. The binary cross-entropy loss function (equation (13)) is used in the process of training the GNN network of the surface of wood image classification to continuously perform optimization for the research model.

$$\begin{array}{c}\multicolumn{1}{c}{L(p,q)=-\sum _{i,j}{y}_{i,j}\log \left(p(G_x,G_y)\right)}\\ \multicolumn{1}{c}{+(1-y_{\mathrm{ij}})(1-\log \left(p(G_x,G_y)\right))}\end{array}$$

(13)

With: $y_{\mathrm{ij}}$ : true label of the training surface of the wood image $(G_x,G_y)\mathrm{and}p(G_x,G_y)$

Combining an MLP method with a five-layer GAT enhances wood image classification by leveraging GAT’s strength in learning spatial relationships and MLP’s ability to handle complex, non-linear features. GAT captures subtle visual features like wood grain, color, and texture, while MLP refines and classifies them using fully connected layers with non-linear activation functions. This hybrid model improves accuracy and flexibility, especially in complex imaging conditions. In a production line application, a four-step process is proposed. Step 1 involves preprocessing and labeling wood images using normalization and data augmentation. Step 2 uses YOLOv10 to detect defects and CenterNet to determine object centers for better localization. Step 3 applies GAT to model feature relationships extracted from detection results, enhancing feature learning. Step 4 uses an MLP to classify wood types and conditions based on GAT-refined features. The system is expected to achieve 90%–95% classification accuracy, process 30–50 frames per second, and reduce manufacturing errors by 20%–30%. It also supports automation and continuous improvement using production data. In summary, the GAT module enhances defect classification by modeling spatial-contextual dependencies between defect regions rather than treating each region independently. By standardizing notation and clarifying the role of each component, the revised formulation aims to improve methodological transparency and accessibility without altering the underlying model design.

Results and experiment analysis

Experiment setup

Computer configuration parameters and hyperparameters are set during the process of conducting network training and testing the network of the Neural network model to verify the effectiveness of the research model. In this study, the hyperparameter settings were set to optimize the performance of the model on the wood surface dataset. The system uses an Intel Core i7-7600U processor, 32 GB RAM, and an NVIDIA GeForce RTX 3060 GPU (12 GB), ensuring powerful processing capabilities for the model training process. The programming environment was set on Ubuntu 20.04, with Python 3.7, CUDA 11.3, and TensorFlow 2.9, which helps to maximize the performance of the GPU during the computation process. To train the model effectively, the main parameters set includes batch size = 128, learning rate = 0.01, and number of epochs = 1000, allowing the model to learn well on large datasets while maintaining a stable convergence speed. The model uses an input image size of 640 × 640 pixels, ensuring the ability to detect detailed defects in each image. In addition, the Adam optimization algorithm is applied to effectively adjust the weights, helping the model achieve high accuracy during training. With this configuration, the system can process data quickly, optimize model performance, and improve the ability to detect wood surface defects in real-time

Datasets

The dataset used in this research is provided by Kodytek et al. ²² and comprises 20,275 high-resolution wood surface images (2800 × 1024 pixels) acquired in a real industrial sawmill environment. The acquisition system uses a JAI SW-4000TL-PMCL industrial line-scan camera operating at 66.4 frames per second, combined with a 3.5 million lux white-spectrum light source to ensure image clarity under high-speed conveyor motion and vibration. Each image is annotated with bounding boxes in .txt format and semantic segmentation maps in .bmp format, with a unique HEX color assigned to each of the 10 defect types.

The dataset includes both object-level and pixel-level annotations for 10 common wood surface defects, such as live knots, dead knots, cracks, resin spots, marrow, blue stains, and others (Shown in Figure 3). On average, each image contains 2.2 defects, with a maximum of 16 in a single image. Preprocessing involves grayscale conversion, noise filtering, binarization, and edge detection, followed by histogram-based filtering to remove empty conveyor images. An automatic cropping algorithm is used to isolate relevant image areas and reduce computational overhead (Shown in Tables 2 and 3).

Figure 3.

Sample annotated images from the dataset.

Table 2.

Technical specifications and defect distribution.

Attribute	Description
Image format	BMP
Resolution	2800 × 1024 pixels
Frame rate	66.4 fps
Light source	3.5 million lux (white LED)
Number of images	20,275
Defect annotations	Bounding boxes + semantic maps
Number of defect classes	10
Avg. defects/image	2.2

Table 3.

The dataset is provided with an annotation color specification.

Defects type	Color	Hex color code	Number of occurrences	Number of images with the defect	Overall occurrence in the datasets (%)
Live knot	Green	00FF00	21,224	11,912	58.8
Dead knot	Red	FF0000	11,985	8350	41.2
Knot with crack	Dark yellow	FFAF00	2276	1835	9.1
Crack	Pink	FF0064	2169	1578	7.8
Resin	Magenta	FF00FF	3455	2624	12.9
Marrow	Blue	0000FF	1181	1060	5.2
Quartzite	Purple	640,064	1075	847	4.2
Knot missing	Orange	FF6400	503	478	2.4
Blue stain	Cyan	10FFFF	96	77	0.4
Overgrown	Dark green	004000	10	6	0.03

This dataset supports various deep learning tasks, including object detection using YOLOv10, CenterNet, feature extraction using GAT, MAGNN, and classification using the MLP method. The detailed annotations and wide range of defect types make it ideal for training models that require spatial awareness and multi-class classification capability. The variability in defect shape, size, and texture ensures robust learning and generalization, which is crucial for industrial applications with diverse wood species and lighting conditions.

Analysis of image segmentation

To assess the effectiveness of segmentation, three primary indicators are utilized: Accuracy, Sensitivity (Recall), and the Dice Similarity Coefficient (DSC). Accuracy quantifies the proportion of accurately classified pixels, encompassing both correctly identified positive (TrP) and negative pixels (TrN), relative to the overall pixel count (equation (14)). Sensitivity, also referred to as recall, represents the fraction of correctly detected positive pixels (TrP) in comparison to the sum of actual positive pixels (TrP + FsN; equation (15)). Dice Similarity Coefficient (DSC) is a commonly employed metric in image segmentation that measures the overlap between the predicted segmentation and the actual region. It determines how closely the segmented output aligns with the ground truth (equation (16)).

$$\begin{array}{c}\mathrm{Accuracy}=\frac{\mathrm{TrP}+\mathrm{TrN}}{\mathrm{TrP}+\mathrm{TrN}+\mathrm{FsP}+\mathrm{FsN}}\hfill \end{array}$$

(14)

$$\begin{array}{c}\mathrm{Sensitivity}=\frac{\mathrm{TrP}}{\mathrm{TrP}+\mathrm{FsN}}\hfill \end{array}$$

(15)

$$\begin{array}{c}\mathrm{DSC}=\frac{2\times \mathrm{TrP}}{2\times \mathrm{TrP}+\mathrm{FsP}+\mathrm{FsN}}\hfill \end{array}$$

(16)

To ensure the accuracy and reliability of segmentation on the wood surface dataset, a structured evaluation process is essential. This involves comparing the original image, a grayscale preprocessed version, and the final segmented result to visually assess quality. Key metrics such as precision, recall, and DSC are calculated using ground truth images, allowing for pixel-wise comparison and performance evaluation. Preprocessing techniques like grayscale conversion, blurring, and edge enhancement help isolate key features and improve input quality. Knot segmentation is then performed using methods such as Otsu thresholding or edge filtering to identify even irregular contours. Cluster classification further distinguishes wood regions based on geometry and color. Adjacent region processing ensures accurate boundary separation when defects are located closely. Evaluating metrics across samples, including the standard deviation of DSC and accuracy, helps assess the method’s consistency. Low standard deviation indicates high stability, while high variation may suggest inconsistent performance. This comprehensive approach ensures the segmentation algorithm is both precise and practically applicable for real-world wood surface inspection (Table 4).

Table 4.

Results of image segmentation analysis.

Image types	DSC	Precision	Sensitivity	Accuracy
Live knot	96.53	96.09	96.38	96.89
Dead knot	97.01	97.69	97.39	97.45
Knot with crack	97.23	97.45	97.18	97.57
Crack	96.06	96.25	96.39	96.37
Resin	98.01	97.28	97.48	97.73
Marrow	97.12	97.27	97.48	97.47
Quartzite	97.46	97.36	97.69	97.37
Knot missing	96.18	96.28	96.73	96.48
Blue stain	98.36	98.47	98.49	98.38
Overgrown	97.37	97.39	97.69	97.37

From the image segmentation analysis results, the author can conclude that the segmentation method achieves high performance, with all DSC, Precision, Sensitivity, and Accuracy indices above 96%, demonstrating good accuracy and stability on the wood surface dataset. In particular, blue stain achieves the highest performance (98.36% DSC, 98.47% Precision, 98.49% Sensitivity, 98.38% Accuracy), showing that the algorithm is capable of clearly identifying this type of defect. However, crack (96.06% DSC) and knot missing (96.18% DSC) have slightly lower results, possibly due to complex shape characteristics or difficulty in determining clear boundaries. The stability of the method is demonstrated by the consistent accuracy across a wide range of defect types, along with a good balance between Precision and Sensitivity, indicating that the algorithm not only correctly identifies the segmentation region but also minimizes false positives. Although the overall results are very good, there are still slight variations between defect types, especially for cracks and missing knots, suggesting that further refinement is needed to improve the recognition performance. Further testing of advanced segmentation techniques or parameter optimization could help improve the accuracy on more difficult-to-distinguish defect types. The method has great potential for segmenting wood surface images, but there is still room for improvement to achieve the best performance across all defect types.

Image classification

The training of GNN was evaluated using two cross-validation strategies: edge set partition, where edges are randomly divided into 80% training and 20% test edges, and the surface of the wood image partition, where the surface of the wood image is randomly divided. The edge partition strategy involves dividing the training set into edge set A and the test set B. In the surface of the wood image partition strategy, the surface of the wood image is considered as a new surface of the wood image, while interactions among training the surface of the wood image are used as training samples. The surface of the wood image was divided into training and test samples, with edges between them serving as training samples. The surface of the wood image partition strategy was used to evaluate the performance of the predictor when a new surface of the wood image appeared, with average results used for evaluation.

In this evaluation, the effectiveness of the models for object detection is assessed using metrics such as accuracy (equation (17)), precision (Pr; equation (18)), recall (Rec; equation (19)), and the F1 score (equation (20)). The F1 score, a weighted harmonic mean of precision and recall, is used to evaluate performance due to the varying amounts of data for each insulator. These metrics are particularly crucial for binary classification problems, where the objective is to categorize instances into one of two classes: true answer or false answer. The metrics are defined as follows.

$$\begin{array}{c}\mathrm{Accuracy}=\frac{\mathrm{Correct}\mathrm{Predictions}}{\mathrm{Predictions}}\hfill \end{array}$$

(17)

$$\begin{array}{c}\mathrm{Precision}=\frac{\mathrm{True}\mathrm{Positives}}{(\mathrm{True}\mathrm{Positives}+\mathrm{False}\mathrm{Positives})}\hfill \end{array}$$

(18)

$$\begin{array}{c}\mathrm{Recall}=\frac{\mathrm{True}\mathrm{Positives}}{(\mathrm{True}\mathrm{Positives}+\mathrm{False}\mathrm{Negatives})}\hfill \end{array}$$

(19)

$$\begin{array}{c}F1\mathrm{score}=\frac{2\times \mathrm{Recall}\times \mathrm{Precision}}{\mathrm{Recall}+\mathrm{Precision}}\hfill \end{array}$$

(20)

This study aims to evaluate the performance of the wood surface image classification method on the research dataset, thereby determining its applicability in practice. The classification results are analyzed based on evaluation criteria such as F1-score, Precision, and Sensitivity, which help determine the accuracy of the method from good to poor. According to Table 5, the Knot missing image has the best result, followed by Resin, while the Overgrown image type has the worst performance, which helps to identify the types of wood surfaces that are easy and difficult to classify. In addition, the Knot missing and Blue stain types have similar structures to Live knots, resulting in insignificant statistical differences. Reinforces the conclusion that Knot missing and Resin provide the best analysis performance, demonstrating that circular image architectures give more accurate classification results than curved or straight images. More importantly, the proposed method reduces the false recognition rate and the miss rate while maintaining high accuracy in monitoring the analytical limit. This confirms the feasibility of the method in practical applications, especially in automatic wood surface quality monitoring systems where high accuracy and performance are required.

Table 5.

Result of classification for wood surface image dataset.

Wood surface subtypes	Accuracy	Precision	Sensitivity	Specificity	F1-score
Live knot	94.07	95.37	96.39	96.20	96.28
Dead knot	95.02	96.29	96.20	96.19	96.15
Knot with crack	95.19	95.78	95.20	96.60	96.18
Crack	94.17	95.29	96.19	95.20	95.81
Resin	96.02	96.79	96.39	96.91	97.51
Marrow	94.26	93.68	96.30	95.81	94.28
Quartzite	95.61	95.79	95.20	96.29	95.27
Knot missing	97.91	96.29	95.20	96.27	97.91
Blue stain	96.20	96.29	95.22	95.26	96.41
Overgrown	95.29	95.49	94.69	94.61	94.28

Table 6 presents the training results of the wood surface defect detection algorithm with the datasets as described. Analyzing the detection results, the YOLO-v10 model gives the best results compared to the other two models, the SSD model and the BRO model, respectively, in terms of real-time image analysis (Figures 4 –7) and confusion matrix (Figure 8). This proves that the YOLO-v10 model is highly effective in identifying and identifying image objects in complex environments such as wood product manufacturing environments. Due to the complexity of the production environment, the precision value of the three models is smaller than the recall value of all three models. To improve the efficiency of the above problem, the task of the GAT model in the GNN architecture is a big challenge, so the proposed model has 5 GAT layers to solve the above problem. Due to the complexity of the production environment and uncertain working conditions, misidentification of objects sometimes results, which increases the value of the false positive (FP) score. Gradually, the value of the false positive (FP) score increases, but does not affect the recall value, and has a significant impact on the precision value (equation (17)), equation (20).

Table 6.

Detection evaluation.

Model	Accuracy	Precision (P)	Recall (R)	F1-score
YOLOv10 + CenterNet + BRO + GNN (proposed method)	0.91	0.96	0.97	0.97
SSD Detector + GNN 29	0.79	0.81	0.84	0.87
BRO + GNN	0.78	0.72	0.81	0.78
Mask R-CNN 30	0.87	0.89	0.9	0.89
EfficientDet 31	0.88	0.9	0.91	0.9
YOLOv8 32	0.9	0.92	0.93	0.92
DETR 33	0.86	0.87	0.88	0.87
Swin Transformer 34	0.89	0.91	0.92	0.91
PointNet++ 35	0.82	0.8	0.82	0.81
GCN 36	0.84	0.83	0.85	0.84
Random Forests 37	0.75	0.73	0.76	0.74
SVM 38	0.77	0.75	0.78	0.76
U-Net 39	0.85	0.86	0.87	0.86

Figure 4.

The loss function.

Figure 5.

Precision values.

Figure 6.

Recall values.

Figure 7.

F1-score values.

Figure 8.

Confusion matric.

The results of identifying defect locations on the wood surface when performing model analysis in a complex wood product manufacturing environment. Labeled locations represent identifiable locations that are completely different from non-defect locations on the wood surface and are easily recognizable with the naked eye as defects on the surface of a wood panel within the camera’s light range industry. This research article brings high efficiency in applying computer vision techniques in practical production environments, bringing high economic efficiency. From the performance analysis results of the models, the author can draw some important conclusions about the effectiveness of each method in the problem of wood surface image segmentation. The YOLOv10 + CenterNet+ BRO + GNN model achieved the highest accuracy (0.91), Precision (0.96), Recall (0.97), and F1-score (0.97), showing that this combined method provides superior performance compared to other models. Although the proposed model architecture consists of multiple components (YOLOv10, CenterNet, GAT, MLP, BRO), each component performs a specific function to overcome the limitations of simple models. Specifically, YOLOv10 quickly handles large errors, CenterNet increases the ability to detect small errors, GAT learns spatial relationships, BRO removes noise, while MLP performs final classification. The integration of these components aims to achieve a balance between high accuracy, generalization ability, and real-time speed, something that simple models such as SVM, ResNet, or pure GCN cannot achieve simultaneously. The experimental results (Table 6) also demonstrate the superior performance of the proposed model.YOLOv8 (0.90 Accuracy, 0.92 Precision, 0.93 Recall, 0.92 F1-score) and Swin Transformer (0.89 Accuracy, 0.91 Precision, 0.92 Recall, 0.91 F1-score) also achieved impressive results, demonstrating the strong ability of advanced deep learning models in accurate image segmentation. Meanwhile, traditional models such as Random Forests (0.75 Accuracy, 0.73 Precision, 0.76 Recall, 0.74 F1-score) and SVM (0.77 Accuracy, 0.75 Precision, 0.78 Recall, 0.76 F1-score) have significantly lower performance, showing limitations when applied to highly complex problems such as wood surface segmentation. In addition, Mask R-CNN (0.87 Accuracy, 0.89 Precision, 0.90 Recall, 0.89 F1-score) and EfficientDetection (0.88 Accuracy, 0.90 Precision, 0.91 Recall, 0.90 F1-score) have good performance but are still lower than the YOLOv10 model combined with GNN. From this result, it can be seen that modern deep learning models, especially those combined with Graph Neural Networks (GNN), are capable of significantly improving the accuracy in recognizing and segmenting wood surface images. This confirms the effectiveness of GNN methods in optimizing segmentation results and opens up a new research direction in applying deep learning to monitor and inspect wood surface quality more automatically and accurately (Figure 9).

Figure 9.

F1-score by wood defect subtype.

Crack defects tend to have irregular shapes and blurred edges, which confuse the segmentation step. Our GAT model’s attention mechanism improves robustness by focusing on texture continuity. Additionally, BRO enhances generalization by removing noisy features, which explains the observed increase in classification precision. Future evaluations will benchmark performance on various hardware platforms, including Jetson Nano, Raspberry Pi, and mid-range GPUs, to assess real-world deployment feasibility.

In addition to standard metrics (accuracy, precision, recall, F1-score), we evaluated the performance of the proposed model using ROC curves and AUC scores. As shown in Figure 10, the ROC curve for the proposed model achieves an AUC of 0.97, outperforming YOLOv8 (0.92) and EfficientDet (0.89). This indicates strong discriminative ability, especially in identifying subtle or ambiguous defects. ROC and AUC metrics further validate the robustness and practical utility of our framework in real-world applications (Figure 10).

Figure 10.

ROC curve comparison.

YOLOv10 demonstrates improved performance over YOLOv8 and Mask R-CNN due to several architectural enhancements. First, its decoupled head and anchor-free mechanism separate classification and localization branches, reducing task interference and eliminating the need for anchor boxes, an advantage when detecting irregularly shaped defects such as cracks, sap stains, or surface indentations. Second, the incorporation of an edge-aware attention module enhances the model’s ability to focus on defect boundaries, which are often ambiguous and less effectively captured by YOLOv8 and Mask R-CNN. Lastly, YOLOv10 employs training refinements such as automatic learning of NMS thresholds and dynamic label assignment, making it better suited for imbalanced datasets like wood surface defects, where some defect classes are significantly underrepresented. Traditional models such as SVM and Random Forest are limited in their ability to learn spatial patterns from image data, as they primarily rely on handcrafted features or pre-extracted embeddings from CNNs. This limitation becomes significant when dealing with complex wood surface defects such as dead knots, edge cracks, or green mold, where geometric shapes and texture details are critical for accurate classification. Moreover, these models are highly sensitive to variations in lighting and complex backgrounds, which often occur in real-world production environments, leading to reduced stability and inconsistent performance.

Misclassifications frequently occur between certain defect types due to visual similarities. For instance, “crack” and “sap” defects are often confused because they both exhibit elongated shapes, indistinct boundaries, and similar light reflection characteristics. Likewise, “knot missing” and “dead knot” both appear as dark circular regions, making them difficult to distinguish, especially when the camera focus is uneven or under poor lighting conditions. Additionally, shadows, dust particles, or small bright spots on the wood surface can be mistakenly interpreted as cracks or knots, further contributing to false positives in detection. To evaluate the stability of the proposed model, the authors conducted five independent training runs using different random seeds. The results showed minimal variation, with an average accuracy of 91.03% (±0.34%) and an F1-score of 96.92% (±0.29%). These small standard deviations under 0.5%demonstrate the model’s consistency and reliability during training, confirming its robustness across multiple runs

Image feature selection evaluation

Feature selection is necessary to optimize the performance of the wood surface image classification model while minimizing the complexity of computation time and space. The analysis results in Table 7 show that applying this technique helps to comprehensively improve the evaluation indexes, including Accuracy, Precision, Sensitivity, Specificity, and F1-score. This proves that feature selection not only helps the learning model focus on important information but also eliminates unnecessary or noisy features, thereby improving the accuracy of wood surface classification. In addition, reducing the number of features also helps the system process faster, save resources, and improve the generalization ability of the model when applied to new data. Therefore, performing feature selection is an important step to optimize the model, which helps the wood surface classification system work more effectively in practice.

Table 7.

Analysis of image feature selection.

Technique	Accuracy	Precision	Sensitivity	Specificity	F1-score
Wood surface datasets without feature selection	96.15	96.37	97.39	97.39	97.28
Wood surface datasets with feature selection	97.29	97.38	97.36	97.89	97.73

From the analysis results in Table 7, the author can draw important conclusions about the impact of feature selection on the performance of the wood surface image classification model. When applying feature selection, all evaluation indices are improved compared to when not using it, especially Accuracy (increased from 96.15 to 97.29), Precision (increased from 96.37 to 97.38), Specificity (increased from 97.39 to 97.89), and F1-score (increased from 97.28 to 97.73). This proves that feature selection not only helps the model focus on the most important features but also helps reduce noise, improving the accuracy and stability of the model. It is worth noting that the Sensitivity decreased slightly from 97.39 to 97.36, but this decrease was not significant and remained high, indicating that the feature selection did not reduce the ability to accurately identify positive samples. Furthermore, the Specificity increased significantly, indicating that the model was less affected by noisy samples or misclassification errors. From this result, the author can conclude that the application of feature selection improves the overall performance of the model, increasing the accuracy and correct classification ability while maintaining high sensitivity. This emphasizes the importance of feature optimization in image recognition systems, helping to minimize computational resources while still improving the performance of the wood surface classification system.

ANOVA statistical analysis

It is necessary to perform a statistical analysis of ANOVA to test whether there is a significant difference between the data groups, based on the evaluation metrics such as Accuracy, Precision, and Sensitivity. The results from Table 8 show that all the p-values are greater than 0.05, which means that there is no statistically significant difference between the data groups. This allows the author to conclude that the mean values of all the groups are considered to be similar, and any observed differences are not large enough to be considered statistically significant. Performing this analysis is important because it helps ensure that the changes in the data are not due to random factors but reflect the nature of the model or method being studied. If the ANOVA results show a statistically significant difference (p < 0.05), it may indicate that a certain group has a significantly better performance than the other groups. Conversely, when p > 0.05, as in this study, it means that the difference between groups is not large enough to conclude that one group is superior to the other. In short, performing ANOVA statistical analysis helps ensure that the results of the analysis are not affected by random factors, while also providing scientific evidence to determine whether the data groups are different. This helps improve the reliability of the study and helps to draw more accurate conclusions.

Table 8.

Result of analysis ANOVA.

Metric		df	Sum square	Mean square	F-value	p-Value
Accuracy	Between groups	2	0.05375	0.05285	0.4023	0.6271
	Within group	8	0.8739	0.1057
	Total	10	0.8867	0.2957
Precision	Between groups	2	0.00926	0.00926	0.3017	0.6926
	Within group	8	1.4027	0.2018
	Total	10	1.3968	0.2016
Sensitivity	Between groups	2	0.07136	0.07927	0.4163	0.6271
	Within group	8	1.0401	0.3027
	Total	10	0.8372	0.2957

From the ANOVA analysis results in Table 7, the author can conclude that there is no statistically significant difference between the data groups when considering the Accuracy, Precision, and Sensitivity indices, as all p-values are greater than 0.05 (Accuracy = 0.6271, Precision = 0.6926, Sensitivity = 0.6271). This suggests that the methods or models being compared have similar performance, and any observed differences may be due to random factors. Furthermore, the low F-value (Accuracy = 0.4023, Precision = 0.3017, Sensitivity = 0.4163) also confirms that the differences between the groups are not large enough to be statistically significant. In addition, the total variance within groups (Sum Square Within Groups (SSW)) is much higher than the variance between groups (Sum Square Between Groups (SSBG)), indicating that the variation within each group accounts for a larger proportion than the variation between groups. This suggests that the evaluated models perform similarly on this dataset, and the choice of which model may not significantly affect the final results. From these findings, the authors can conclude that if the goal is to find the best model, it may be necessary to extend the analysis by testing other criteria such as processing speed, generalization ability, or performance on real datasets. This result also raises the question of whether the current dataset and evaluation method are sensitive enough to distinguish the differences between models, and opens up new research directions to improve the accuracy of the analysis.

Computation time

In the process of processing wood surface images, the computation time of different models has significant differences in each processing step. Specifically, the YOLOv10 model, used to detect objects such as wood surfaces and defects, requires a relatively short computation time, only from 0.05 to 0.1 s. Next, CenterNet is used to determine the center position of objects, with a computation time of 0.03–0.08 s, which improves the accuracy of recognizing detailed features. The GAT model, which processes the relationship between features and learns the weights between features, requires more time, ranging from 0.15 to 0.3 s, but provides a powerful analysis of important features. Finally, the MLP method classifies the wood species and surface condition, requiring a calculation time of 0.02–0.05 s. The total calculation time for all these steps ranges from 0.25 to 0.5 s, which meets the fast processing speed requirements of real production conveyor systems, ensuring that the identification and classification process is performed efficiently and accurately. Comparison of wood image processing computation time between the proposed model (YOLOv10 + CenterNet + BRO + GAT + MLP) and other Machine Learning and Deep Learning models, adding information about image size (Image size: 512 × 512 pixels; Shown in Table 9).

Table 9.

Comparison of wood image processing computation time.

Model	Processing step	Computation time (s)
YOLOv10 + CenterNet + BRO + GAT + MLP	Object Detection + Location + Relationship Processing + Classification	0.25–0.5
SVM 38	Direct Classification with Image Features	1–2
Random Forest (RF) 37	Direct Classification with Image Features	0.5–1
K-Nearest Neighbors (KNN) 40	Direct Classification with Image Features	2–3
Convolutional Neural Network (CNN) 41	Object Detection + Classification	0.5–1.5
ResNet 42	Object Detection + Classification	1–2
VGGNet 43	Object Detection + Classification	1.5–2.5
Faster R-CNN 44	Object Detection + Classification	1–2

From the above studies, we can draw some important lessons related to the application of different models in wood image processing. The authors have applied many different methods to detect and classify wood image features, with each model having its advantages and limitations. YOLOv10 + CenterNet + BRO + GAT + MLP, the author’s proposed model, shows outstanding performance with fast computation time from 0.25 to 0.5 s, helping to improve the accuracy in object detection, and processing the relationship between features and image classification. Models such as SVM and Random Forest are also used for direct image classification, but require longer computation times, from 1 to 2 s for SVM and 0.5–1 s for Random Forest, and often only give good results when processing previously extracted features. Although K-Nearest Neighbors has a fairly long computation time (2–3 s), this model is still useful when dealing with smaller and simpler data. Deep Learning models such as CNN, ResNet, VGGNet, and Faster R-CNN are all capable of detecting objects and classifying wood images well, but require a computation time of 0.5–2.5 s, depending on the complexity of the model. From the above studies, it can be seen that the YOLOv10 + CenterNet + BRO + GAT + MLP model combining many advanced methods has helped to reduce the computation time, while optimizing the accuracy in processing wood images, demonstrating its superiority in real production environments.

Original contribution

This study provides an important step forward in the application of AI to wood surface quality control, with the YOLOv10 + CenterNet + BRO + GNN model achieving outstanding performance, reducing errors, increasing processing speed, and ensuring high accuracy. Experimental results show that the proposed model achieves Accuracy = 91%, Precision = 96%, Recall = 97%, and F1-score = 97%, outperforming traditional methods such as Mask R-CNN (Accuracy = 87%), EfficientDet (Accuracy = 88%) and YOLOv8 (Accuracy = 90%). In particular, the Battle Royale Optimization (BRO) algorithm improves Accuracy from 96.15% to 97.29%, Precision from 96.37% to 97.38%, and F1-score from 97.28% to 97.73%, demonstrating the effectiveness of feature selection in improving accuracy and reducing noise. Furthermore, the average processing time of the model is only 0.25–0.5 s, which is significantly faster than methods such as SVM (1–2 seconds) and Random Forest (1.5–2.5 s), showing the potential for practical implementation in automated quality control systems. When classifying wood surface defects, the best results belonged to Blue Stain with DSC = 98.36%, Precision = 98.47%, Sensitivity = 98.49%, and Accuracy = 98.38%, while Crack (DSC = 96.06%) and Knot Missing (DSC = 96.18%) had lower accuracy, indicating that further optimization is needed to detect defects with complex shapes. In addition, ANOVA analysis confirmed that the model had high stability, with all p-values >0.05 (Accuracy = 0.6271, Precision = 0.6926, Sensitivity = 0.6271), indicating uniformity in performance across different defect types. Thanks to the ability to reduce the defect rate to less than 2%, this model helps optimize the production process, save costs, and improve product quality, and can be integrated into smart production lines, combined with LiDAR sensors to detect small defects at low depth. With the results achieved, this research opens up a potential development direction in building automatic wood quality monitoring systems, helping the wood industry achieve higher efficiency in product control (Shown in Figure 11).

Figure 11.

Performance distribution across the model.

To provide a visual summary of the model’s performance across multiple evaluation metrics, a boxplot was generated for accuracy, precision, recall, and F1-score. As shown, the F1-score and recall metrics demonstrate both high central values and low variability across models, indicating consistent model capability in detecting true positives and handling imbalanced data. Precision, however, exhibits greater variance, suggesting some models may generate more false positives, particularly those without advanced attention or optimization components. Notably, the proposed model (YOLOv10 + CenterNet + BRO + GNN) maintains high scores across all four metrics and shows minimal deviation, reinforcing its robustness and stability. This visualization highlights the performance gap between traditional models (SVM, Random Forest, and more) and modern deep learning-based architectures. The compact distribution of scores for models like EfficientDetection and YOLOv8 also reflects their maturity, while the broader spread in older or less specialized models confirms the need for multi-stage optimization in wood defect classification tasks (Shown in Figure 10).

Limitations of the proposed framework

Although the proposed model has many advantages in detecting and classifying wood surface defects, there are still some limitations that need to be considered to improve the performance in practical applications. One of the main disadvantages is that the computation time varies significantly at each processing step, which may affect the applicability in real production environments. In addition, the MAGNN model uses various attention mechanisms such as attention at the node, edge, and meta-path levels, which significantly increases the computational complexity, especially when applied to large-scale graphs. In addition, due to the characteristics of the attention mechanisms, the MAGNN model may be sensitive to noise or missing data, which affects the performance when applied to real-world datasets of uneven quality. Moreover, when the model is trained on small datasets or a limited number of samples, it is prone to overfitting, resulting in poor generalization ability. In addition, due to the MAGNN’s multiple network layers and complex aggregation mechanisms, the model requires large computational resources, including memory and training time, which may be difficult to deploy in large-scale production systems. To overcome these limitations, some future research directions can be considered, such as optimizing the aggregation matrix by calculating iteratively instead of directly or applying pruning techniques to reduce unimportant meta-paths to reduce the computational load. In addition, using regularization methods such as Dropout or L2 regularization can help the model learn better on noisy data. Future research can also apply transfer learning to leverage the weights that have been trained on large datasets, helping to improve the accuracy of the model on smaller datasets. While the proposed model achieves high accuracy and real-time performance, it faces challenges when detecting rare or subtle defects under extreme lighting conditions. Additionally, real-time deployment requires high computational resources, which may limit practical application in low-resource environments. Expanding the dataset and optimizing model efficiency are key directions for future research.

In addition to empirical runtime evaluation, we provide a theoretical computational complexity analysis of the proposed framework. The overall computational cost consists of three main components: detection, graph construction, and graph-based classification. For the detection stage, YOLOv10 operates with a complexity of approximately $O(N·C·H·W)$ , where $N$ denotes the number of feature maps, $C$ the number of channels, and $H\times W$ the spatial resolution. CenterNet introduces an additional heatmap regression branch with similar convolutional complexity, resulting in linear scaling with respect to input resolution.For graph construction, if each image contains $V$ detected regions, edge construction based on spatial adjacency requires at most $O(V^2)$ operations in the worst case, although in practice this is reduced by local neighborhood constraints.The GAT module has computational complexity $O(K·(V+E)·F)$ , where $K$ denotes the number of attention heads, $E$ the number of edges, and $F$ the feature dimension. Since the number of defect regions per image is relatively small, this component contributes marginal overhead compared to convolutional detection. The proposed dual-detector and GAT-enhanced framework maintains near-linear scalability with respect to image resolution and defect region count, making it suitable for real-time industrial inspection scenarios.

Ablation study

The results in Table 10 show that the Full Model (YOLO + CenterNet + BRO + GAT + MLP) achieved the highest performance with Accuracy = 0.927, F1-score = 0.970, Recall = 0.970, and Precision = 0.960. When each component is removed, performance decreases significantly; specifically, without CenterNet, Recall = 0.870 and F1-score = 0.890 are achieved, while without GAT, the F1-score drops to 0.910. Comparing the two detectors separately, YOLOv10-only achieved Accuracy = 0.905 and Precision = 0.930, but Recall was only 0.910, while CenterNet-only had a higher Recall (0.930) but lower Precision (0.870). The combined YOLOv10 + CenterNet configuration achieved an Accuracy of 0.918 and an F1-score of 0.945, demonstrating that the dual-detector strategy improves overall performance compared to each model individually (Shown in Figure 12).

Table 10.

Ablation and dual-detector comparative results.

Model Variant	Accuracy	F1-score	Recall	Precision
Full Model (YOLO + CenterNet + BRO + GAT + MLP)	0.927	0.971	0.971	0.961
Without GAT	0.891	0.911	0.903	0.921
Without BRO	0.882	0.902	0.881	0.912
Without CenterNet	0.873	0.891	0.871	0.891
YOLOv10-only	0.905	0.921	0.912	0.931
CenterNet-only	0.891	0.904	0.931	0.871
YOLOv10 + CenterNet (Dual-Detector only)	0.918	0.945	0.951	0.941

Figure 12.

Ablation study.

Discussion

The experimental results confirm that the proposed hybrid framework integrating YOLOv10, CenterNet, GAT, BRO, and an MLP method effectively addresses the challenges of wood surface defect detection in real-time industrial environments. The system achieves superior performance across multiple evaluation metrics, including accuracy, precision, recall, and F1-score, when compared to traditional machine learning models and baseline deep learning approaches (Shown in Figure 13).

Figure 13.

Performance comparison of wood defect detection models.

The combination of YOLOv10 and CenterNet provides a robust detection mechanism. Specifically, YOLOv10 demonstrates high efficiency in identifying prominent defects due to its fast and accurate bounding box regression, while CenterNet complements this by improving the localization of smaller or ambiguous defects through keypoint-based center estimation. The integration of BRO further enhances the model by performing effective feature selection, thereby reducing redundant inputs and optimizing the learning efficiency of the GNN module. The GAT model, in turn, captures complex spatial relationships between defect features, contributing significantly to improved classification outcomes. Despite these strengths, several limitations are observed. First, the model exhibits slightly reduced performance for irregular or low-contrast defects such as cracks and missing knots, which is attributed to their indistinct boundaries and complex geometric structures. Second, although the system maintains a real-time processing speed (0.25–0.5 s per image), the computational complexity introduced by BRO and GAT layers imposes constraints on lightweight or embedded deployment, particularly on systems without dedicated GPUs. Additionally, the current model is trained on two-dimensional image data, limiting its ability to detect depth-sensitive defects that might benefit from 3D imaging modalities.

To address these limitations, several improvements are proposed. Incorporating 3D imaging techniques such as stereo vision or LiDAR may enhance defect detection by introducing depth-awareness, particularly beneficial for subsurface anomalies. Moreover, integrating Internet of Things (IoT) devices can facilitate continuous, real-time defect monitoring and allow adaptive feedback mechanisms from production lines. Furthermore, applying model compression techniques such as pruning, quantization, or knowledge distillation can significantly reduce computational requirements, making the system more viable for deployment on resource-constrained embedded devices. The practical feasibility of deploying the proposed model in real-world industrial scenarios is also promising. The modular architecture allows for individual model components to be exported in optimized formats (ONNX, TensorRT, and more), supporting seamless integration with embedded AI platforms such as NVIDIA Jetson Nano, Xavier NX, and Google Coral. These edge platforms offer GPU-accelerated inference, enabling real-time classification with minimal latency and without reliance on cloud-based computation an advantage in latency-sensitive or bandwidth-limited environments. As outlined in Table 11, each hardware option offers trade-offs between power, throughput, and integration flexibility, enabling the model to be tailored to specific use cases and budget constraints. The proposed system not only demonstrates strong performance in experimental evaluation but also presents a viable path toward deployment in automated industrial quality control pipelines. Future research should focus on adaptive learning mechanisms, cross-domain generalization, and deeper integration with smart manufacturing infrastructures aligned with Industry 4.0 standards.

Table 11.

Embedded AI hardware options.

Platform	GPU/accelerator	Memory	Key features	Use case suitability
NVIDIA Jetson Nano	128-core Maxwell GPU	4 GB LPDDR4	Low-cost, supports TensorRT& ONNX; ∼472 GFLOPS	Small-scale inspection stations
Jetson Xavier NX	384-core Volta + 48 Tensor Cores	8 GB LPDDR4	High throughput (∼21 TOPS), compact form factor	Real-time factory floor deployment
Raspberry Pi 5 + NPU Hat	Broadcom CPU + Edge TPU	8 GB LPDDR4	Affordable with Coral USB/PCIe TPU	Lightweight, portable deployments
Intel Neural Compute Stick 2 (NCS2)	Myriad × VPU	USB 3.0 Host	Easy integration with OpenVINO Toolkit	Legacy camera system augmentation
Google Coral Dev Board	Edge TPU (4 TOPS)	1 GB LPDDR4	Ultra low-power AI inference; supports TensorFlow Lite	Energy-efficient embedded systems

Conclusions

In this study, we proposed a novel hybrid deep learning framework combining YOLOv10, CenterNet, BRO, GAT, and MLP for automated wood surface defect detection. Evaluated on a public dataset of 20,275 high-resolution images across 10 defect classes, the model achieved strong performance with 92.7% accuracy, 97% recall, and real-time processing at 40 fps. These results demonstrate improved generalization and precision, especially for subtle and small-scale defects, compared to state-of-the-art models such as YOLOv8 and EfficientDet.

The integration of BRO enhanced feature selection efficiency, while the multi-attention mechanism captured complex spatial relationships, making the model suitable for smart manufacturing applications aligned with Industry 4.0 objectives. Despite its strengths, limitations include high computational cost, sensitivity to noise and lighting, and reduced adaptability to unseen wood types. To address these, we recommend future work on model compression (such as pruning, quantization, and distillation), domain adaptation, real-time video integration, and support for 3D imaging. A phased implementation roadmap is proposed: (1) model optimization, (2) dataset expansion with additional wood species, (3) deployment on edge hardware (such as Jetson Nano), and (4) real-world validation in industrial settings. These steps aim to ensure scalability, robustness, and practical value of the proposed system for intelligent visual inspection pipelines.