C2f-MCFF

In YOLOv8, the C2f module serves as a critical component for feature extraction, employing convolutional and bottleneck structures to process intermediate representations. However, it relies on standard convolution operations, which constrain the receptive field and limit the model’s capacity to capture features from objects of varying scales. This limitation can result in incomplete feature representation, particularly for multi-scale or fine-grained targets, thereby reducing detection accuracy.

In the context of wood board edge defect detection, defects often exhibit irregular shapes, varying sizes, and appear against complex background textures. To address these challenges and enhance the model’s ability to extract discriminative features, we propose replacing the standard Bottleneck structure within the C2f module with a newly designed Multi-Channel Feature Fusion (MCFF) module, forming an enhanced structure termed C2f-MCFF.

The C2f-MCFF module incorporates the concept of grouped convolution to enrich feature diversity. Specifically, the input feature map is first passed through a 3×3 convolution layer, followed by batch normalization and SiLU activation. The output is then split into two branches, denoted as x1 and x2, each containing half of the original channels. The x1 branch is processed through a 3×3 convolutional, while x2 undergoes a 3×1 convolution to introduce asymmetric receptive fields. The outputs from x1, x2, and the initial convolution are subsequently concatenated, followed by a 1×1 convolution for channel adjustment and a residual connection for information preservation.

By integrating convolutions with varying kernel sizes and orientations, C2f-MCFF effectively enlarges the receptive field and enhances the extraction of features across multiple scales. This design is particularly beneficial for detecting slender and irregular edge defects on wood surfaces. Furthermore, the multi-branch structure improves feature fusion while maintaining critical defect-related information. The architecture of the improved C2f-MCFF module is illustrated in S2 and S3 Figs.

SPPF-D

YOLOv8 incorporates the Spatial Pyramid Pooling – Fast (SPPF) module to expand the receptive field and facilitate global feature fusion. This module utilizes a spatial pyramid structure composed of convolutional operations followed by three consecutive max-pooling layers, aiming to capture multi-scale contextual information. However, relying solely on max pooling can result in the loss of fine-grained defect features, especially in scenarios where defect boundaries are subtle or partially occluded.

In the task of wood board edge defect detection, input images often contain noise, and defects such as gaps, chips, or residual glue typically exhibit small size and irregular morphology. Under these conditions, the original SPPF module may inadequately represent fine-scale defect features, potentially causing feature fusion loss and missed detections.

To address these limitations, we propose an enhanced module termed SPPF-D (SPPF with Dilation), which builds upon the original SPPF structure. The proposed SPPF-D introduces the Efficient Local Attention (ELA) mechanism to strengthen feature expressiveness, and incorporates dilated convolutions and average pooling to capture broader contextual dependencies while preserving spatial resolution. These additions aim to improve the module’s sensitivity to small and subtle defects, and mitigate feature loss during the pooling process.

The architectural structure of the SPPF-D module is illustrated in S4 Fig.

The input feature map is initially processed using a standard 1×1 convolution to adjust channel dimensions. Subsequently, a 3×3 dilated convolution is applied to expand the receptive field and enhance the representation of subtle and fine-grained features.

The resulting feature map is then propagated through two parallel branches. The first branch performs three consecutive max pooling operations to capture high-level semantic information and emphasize prominent features. In parallel, the second branch applies three consecutive average pooling operations to retain background context and low-frequency information.

The outputs from both branches are concatenated along the channel dimension to form a unified feature representation that integrates multi-scale contextual cues. A final 1×1 convolution is employed to fuse and compress the concatenated features, producing the output feature map.

The overall process can be mathematically formulated as follows:

ELA Attention

In wood board edge defect detection, background clutter and wood grain textures often introduce irrelevant information, which may mislead the model and reduce detection accuracy. To address this challenge, an attention mechanism is introduced to enhance the model’s focus on salient features, thereby improving its overall detection performance. In this study, the Efficient Local Attention (ELA) module is integrated directly before the input of the SPPF-D module, serving as a lightweight spatial attention gate.

By placing ELA at the beginning of SPPF-D, the network is guided to emphasize informative regions in the feature map before multi-scale pooling operations. This design allows the downstream pooling layers to extract more discriminative spatial features, thereby enhancing overall feature fusion. Importantly, ELA achieves this without incurring significant computational overhead.

The core mechanism of ELA involves applying strided average pooling along the spatial dimensions (horizontal and vertical) for each channel, resulting in two feature descriptors that encode directional dependencies. This narrow kernel pooling strategy enables the module to capture long-range spatial correlations while suppressing background noise and irrelevant textures. The resulting horizontal and vertical descriptors are each passed through lightweight transformations to compute attention weights, which are then fused via element-wise multiplication to modulate the original features.

This process enables the model to focus more accurately on defect-related regions, improving localization precision while maintaining computational efficiency. The detailed architecture of the ELA module is illustrated in S5 Fig.

The input feature map is processed through a one-dimensional convolution, followed by group normalization and Sigmoid function to generate two positional attention weights y^h and y^w. y^h is the horizontal positional attention weight, y^w is the vertical positional attention weight, is Sigmoid function, F_h and F_w is one-dimensional convolution, G_n is group normalization.The calculation formula is as follows Eq (1), Eq (2), Eq (3).

(1)

(2)

(3)

The final output can be expressed as the element-wise multiplication of the original input feature map with the two distinct positional attention weights F_h and F_w. x_c is the original input feature map.

In wooden board edge defect detection, a single defect type may appear in diverse morphological forms, posing challenges for accurate identification. To address this, the Efficient Local Attention (ELA) mechanism is introduced to enhance the model’s ability to capture both spatial-semantic information and global contextual features. Owing to its lightweight design, ELA facilitates efficient extraction of defect-relevant features while enriching the overall feature representation. This ultimately contributes to improved detection accuracy, particularly for subtle or irregular edge defects.

Within the SPPF-D module, the input feature map is first refined by ELA to highlight informative regions. Subsequently, dilated convolutions are employed to enhance the representation of small-scale defects, followed by the fusion of multi-scale features through the combination of global max pooling and average pooling. This design enables more effective integration of contextual information across different spatial resolutions, thereby strengthening the model’s ability to detect and localize edge defects with complex distributions. As a result, the SPPF-D module significantly improves the robustness and reliability of detection under real-world conditions.

WIoU

YOLOv8 adopts Complete IoU (CIoU) as its bounding box regression loss function, which considers factors such as the overlap area and the distance between the centers of the predicted and ground truth boxes. However, the aspect ratio penalty term in CIoU does not accurately capture the differences in width and height with respect to the actual confidence of the predicted boxes. This can lead to suboptimal optimization similarity, potentially hindering precise localization.

In real-world production environments, wood panel images are often collected under variable lighting conditions and equipment setups, leading to datasets that contain a substantial number of low-quality samples. In such scenarios, CIoU tends to place excessive emphasis on high-quality bounding boxes during regression. As a result, the model may become biased, reducing its generalization capability and limiting its robustness in detecting diverse defects under inconsistent imaging conditions.

To address the impact of variable-quality samples on bounding box regression, this study adopts WIoU v3 as the regression loss function. WIoU introduces a dynamic, non-monotonic focusing mechanism and replaces IoU with a novel outlier-aware score to evaluate anchor box quality. The outlier degree is calculated as shown in Eq (4):

(4)

A smaller outlier degree indicates higher-quality anchor boxes. To mitigate the influence of low-quality samples, WIoU assigns lower gradient gains to samples with higher outlier scores. Through this gradient gain allocation strategy, the harmful impact of low-quality samples is significantly reduced.

WIoU v3 constructs a non-monotonic focusing factor using the outlier score to modulate the regression loss. Specifically, when the , r = 1, the gradient gain reaches its maximum. This ensures that high-quality samples receive more attention during optimization, while the effect of noisy or poorly labeled samples is down-weighted.

The WIoU loss is computed as shown in Eq (5), Eq (6), Eq (7):

(5)

(6)

(7)

WIoU IoU as shown in S6 Fig, is the outlier degree. is the mean of the loss function. is the moving average of momentum m. R_WIoU is attention mechanism. x, y, x_gt, y_gt are the x and y coordinates of the center points of the predicted and ground-truth boxes. W_g and H_g are the width and height of the minimum enclosing bounding box. In order to prevent R_WIoU from generating gradients that hinder convergence. Here, the asterisk * denotes that the corresponding term is detached from the computational graph during backpropagation. That is, gradients are not propagated through W_g and H_g. and are hyperparameters that are empirically selected based on validation performance. They play a crucial role in the optimization process by dynamically re-weighting the gradient contributions, thereby focusing training on samples with more beneficial IoU characteristics. This mechanism improves the localization accuracy and facilitates better convergence behavior of the model.

In the task of wood panel edge defect detection, defect samples exhibit varying levels of quality and considerable variation in anchor box scales. By adopting WIoU as the regression loss function, the model’s convergence capability is significantly enhanced. This leads to more accurate localization of defects on the wood panel edges and effectively improves the overall detection performance of the model.

Test indicators

To accurately evaluate the model’s performance in detecting wood board edge defects, this study adopts the following metrics: Precision (P), Recall (R), mean Average Precision (mAP50), and mAP50:95.

Precision (P) refers to the proportion of correctly predicted positive samples among all predicted positive samples. The calculation formula for precision is as shown in Eq (8):

(8)

Recall (R) refers to the proportion of actual positive samples that are correctly predicted as positive. The calculation formula for recall is as shown in Eq (9):

(9)

In the above formulas,TP is the number of correctly predicted positive samples,FN is the number of actual positive samples incorrectly predicted as negative,FP is the number of negative samples incorrectly predicted as positive. mAP (mean Average Precision) is one of the most commonly used evaluation metrics in object detection. It takes into account the detection performance across different classes and consolidates them into a single evaluation metric.In object detection, a higher mAP value indicates better model performance. The general formula is shown in Eq (10), Eq (11):

(10)

(11)

Attention mechanism comparison experiments

To validate the effectiveness of the ELA attention mechanism within the proposed algorithm, we conducted comparative experiments using several attention mechanisms, including CAFM, CBAM, SimAM, and EMA. The experimental results are summarized in S3 Table.

As shown in S3 Table the incorporation of the ELA attention mechanism results in the most substantial and consistent improvement across all evaluation metrics. ELA achieves the highest mAP50 (0.732) and competitive mAP50:95 (0.394), while also maintaining a favorable balance between precision and recall. These results indicate that ELA effectively enhances the model’s ability to detect edge defects on wood panels.

Although the EMA module records a slightly higher mAP50:95 (0.395), it underperforms compared to ELA in terms of both mAP50 and recall, suggesting weaker overall detection capability. Similarly, the SimAM module yields a marginal improvement in Recall (by 0.4%), but this comes at the cost of a 0.2% drop in mAP50 and a 1.3% decrease in Precision, reflecting a trade-off that may not be ideal for the task at hand.

The CAFM module demonstrates a relatively higher recall (0.678), but exhibits a notable decline in both mAP50 (0.728) and mAP50:95 (0.391), indicating unstable regression performance. CBAM shows moderate results overall but does not surpass ELA in whole individual metric.

These findings highlight that among the attention mechanisms evaluated, ELA offers the most balanced enhancement to detection accuracy and consistency. Its superior performance validates its integration into the proposed YOLO-MDEW architecture for industrial edge defect detection applications.

Comparative study of loss functions

YOLOv8 employs the Complete IoU (CIoU) loss function, which has been observed to exhibit relatively slow convergence and suboptimal bounding box regression performance when applied to edge defect detection tasks. To evaluate the effectiveness of Wise-IoU (WIoU) in improving bounding box regression for such fine-grained defects, a comparative study was conducted against several mainstream IoU-based loss functions. The results are presented in S4 Table.

The results demonstrate that WIoU consistently outperforms CIoU and other loss functions across all evaluation metrics. Compared to CIoU, WIoU improves precision, recall, mAP50, and mAP50:95 by 0.9%, 0.8%, 0.6%, and 0.4%, respectively. While other IoU Loss falls behind WIoU in terms of both mAP metrics. These findings indicate that WIoU provides a more balanced enhancement in both localization accuracy and classification stability.

To further investigate the impact of WIoU’s hyperparameters— and —on performance, a series of experiments were conducted with different combinations. The results are shown in S5 Table.

Among the tested configurations, the combination of and yielded the best overall performance, achieving the highest mAP50 and mAP50:95 values. Deviations from this setting resulted in slight declines in performance, indicating that the values of and play a crucial role in controlling the non-monotonic focusing factor of the WIoU loss and, consequently, the gradient distribution during training. In summary, WIoU demonstrates superior capability in reducing bounding box regression errors and improving overall detection accuracy when compared with existing IoU-based losses. Additionally, appropriate tuning of its hyperparameters further enhances its effectiveness, underscoring its suitability for complex defect detection tasks in industrial scenarios.

Ablation experiments

To evaluate the effectiveness of each proposed module, ablation experiments were conducted and the results are presented in S6 Table. The baseline model (Model 1) corresponds to the original YOLOv8n, which achieves an mAP50 of 0.721 ± 0.007 and mAP50:95 of 0.390 ± 0.002 with 2.7M parameters and 6.9 GFLOPs.

In Model 2, the integration of the SPPF-D module leads to a performance gain, increasing mAP50 by 1.1% and mAP50:95 by 0.4%, while maintaining computational efficiency (2.9M parameters and 7.1 GFLOPs). This confirms that the SPPF-D structure effectively enhances multi-scale feature fusion and improves defect localization.

Model 3 builds upon Model 2 by incorporating the C2f-MCFF module. This addition brings further improvements, with mAP50 reaching 0.734 ± 0.007 and mAP50:95 rising to 0.396 ± 0.003. Compared to the baseline, this represents an overall gain of 0.2% in mAP50 and 0.2% in mAP50:95, demonstrating that MCFF enhances the model’s capacity to extract discriminative features.

Model 4 introduces the Wise-IoU (WIoU) loss function on top of Model 3. While mAP50 slightly increases to 0.740 ± 0.003, mAP50:95 also improves to 0.400 ± 0.003. Importantly, the inclusion of WIoU results in a noticeable refinement in localization performance, as indicated by the consistent gains across all metrics. The model complexity remains nearly unchanged compared to Model 3, with 3.2M parameters and 7.5 GFLOPs.

Comparative experiments

To further validate the performance of the proposed improved algorithm, comparative experiments were conducted between the proposed method and the latest YOLO series algorithms on a custom-built dataset. The training parameters and experimental settings were kept consistent with the initial configuration. The experimental results are presented in S7 Table

The results clearly demonstrate that YOLO-MDEW achieves the highest detection accuracy among all compared models. Specifically, it attains a mean mAP50 of 0.740 and mAP50:95 of 0.400, outperforming YOLOv8n by 1.9% and YOLOv9t, YOLOv10n, YOLOv11n, and YOLOv12n by 2.3%, 5.2%, 2.7%, and 2.7%, respectively, in terms of mAP50. The mAP50:95 also shows consistent improvement, with increases of 1.0%, 1.9%, 3.1%, 1.4%, and 2.3% over the respective baselines.

In addition to overall accuracy, YOLO-MDEW achieves the highest mean precision (0.756) and competitive recall (0.668), with relatively low standard deviations across multiple runs, indicating both strong performance and training stability. Compared to YOLOv8n, the precision and recall improve by 3.5% and 0.2%, respectively. Consistent gains are also observed over other YOLO-based baselines, including YOLOv10n, where improvements of 5.0% in precision and 4.6% in recall are achieved.

Although YOLO-MDEW introduces a slightly higher computational cost (3.2M parameters and 7.5 GFLOPs) compared to YOLOv8n (2.7M parameters and 6.9 GFLOPs), the increase remains marginal relative to the performance gains. In terms of inference efficiency, the proposed model achieves 119 FPS, which, although slightly lower than YOLOv8n (131 FPS) and YOLOv10n (133 FPS), is still far beyond the real-time requirement for industrial inspection (2.5–3 FPS), ensuring practical deployment feasibility.

To further evaluate the generalization capability of the proposed method beyond YOLO-based detectors, a representative transformer-based model (RT-DETR) is included for comparison. As shown in S8 Table, RT-DETR achieves competitive performance (mAP50 of 0.723 and mAP50:95 of 0.395), but at a significantly higher computational cost and lower inference speed (28 FPS). In contrast, YOLO-MDEW achieves better accuracy while maintaining substantially higher efficiency, making it more suitable for real-time industrial applications.

The mAP50 curves of YOLO-MDEW and YOLOv8n are presented in S8 Fig to illustrate training dynamics. The proposed method exhibits more stable convergence behavior, indicating improved optimization efficiency.

Overall, the experimental results demonstrate that YOLO-MDEW achieves a balanced improvement in accuracy, efficiency, and stability. These characteristics make it well-suited for practical industrial defect detection scenarios.

Scale-wise performance analysis

To further evaluate the effectiveness of the proposed method in handling objects of different sizes, scale-specific recall metrics are analyzed for small, medium, and large defect instances. As shown in S9 Table, YOLO-MDEW consistently outperforms the baseline YOLOv8n across all object scales. Specifically, recall improves from 65.6% to 67.8% for small objects, from 74.1% to 76.8% for medium objects, and from 67.3% to 72.9% for large objects.

Notably, the most significant improvement is observed in large-scale defects, with a gain of 5.6%, suggesting that the proposed feature fusion strategy effectively enhances the model’s ability to capture more comprehensive contextual information. Meanwhile, the consistent improvements in small and medium objects indicate that the model maintains stable sensitivity to fine-grained and intermediate-scale features.

These results demonstrate that YOLO-MDEW achieves consistent performance improvements across different object scales, providing strong quantitative evidence for the effectiveness of the proposed multi-scale feature enhancement design.

Confusion matrix comparison

Beyond scale-wise performance, a more detailed class-level analysis is conducted using confusion matrices. As shown in S9 Fig, the confusion matrices for YOLOv8 (left) and the proposed YOLO-MDEW model (right) reveal clear differences in classification behavior. In a confusion matrix, higher values along the diagonal indicate greater classification accuracy. Compared to YOLOv8, the YOLO-MDEW model demonstrates an overall improvement in defect classification accuracy. Notably, the accuracy for the Edge banding tackless category increases by 6%, and Edge banding longer improves by 3%. Except for a slight decline in the Board gap category, classification performance across the remaining categories is either comparable or improved. Additionally, YOLO-MDEW reduces confusion between similar defects: while YOLOv8 exhibits a 1% misclassification rate for Board gap, Tape residual, and Edge banding dirty,the improved model further reduces these errors, indicating enhanced reliability and robustness in edge defect recognition.

To further validate these improvements, we report per-class Precision, Recall, and mAP50 for both models in S10 and S11 Tables. YOLO-MDEW achieves notable gains across most defect categories. For example, in the Glue seam category, Precision and Recall increase by 1.7% and 1.1%, respectively, with a 4.5% gain in mAP50. In the Edge banding longer category, improvements of 1.9% in Precision and 3.9% in Recall are observed. Minor drops in Precision are seen in categories like Tape residue and Short edge banding, though these are often offset by increases in Recall. One notable exception is the Board gap category, where Recall decreases by 1.3% and mAP50 decreases by 1.9% despite an increase in Precision.

Further analysis suggests that this performance degradation is likely attributed to the inherent difficulty of the Board gap category. Specifically, some instances exhibit extremely small and weakly distinguishable visual cues, which can be easily confused with normal edges under complex lighting and textured backgrounds. In addition, Board gap defects often co-occur with other edge-related imperfections such as glue residue and seam cracks, leading to inter-class similarity and spatial ambiguity.

Meanwhile, a secondary contributing factor may stem from the multi-scale feature learning process. Although the proposed C2f-MCFF and SPPF-D modules enhance feature representation across scales, although overall performance is consistently improved across all scales, minor variations may still appear under different evaluation conditions due to IoU-based matching sensitivity due to optimization trade-offs and IoU-based matching sensitivity during evaluation.

Overall, these factors jointly explain the minor performance variation in the Board gap category, while the overall robustness of YOLO-MDEW remains unaffected.

Comparison of detection results

To intuitively demonstrate the performance difference between YOLOv8 and the proposed YOLO-MDEW model in edge defect detection for wooden boards, we compared the detection results of both models across various defect types, as illustrated in S10 Fig. On the left is shown the detection output of YOLOv8n, while right presents the results from YOLOv8-MDEW.

In the first three sets of images, YOLOv8n exhibits notable instances of missed detections. Specifically, in the third set, where multiple defect types co-exist, YOLOv8n fails to detect the glue line at the bottom of the image. In contrast, YOLOv8-MDEW successfully identifies all present defects, including glue line, edge chipping, and residual tape.

In the following four image sets, YOLOv8n struggles to accurately detect continuous defects of the same type, often resulting in incomplete localization. YOLOv8-MDEW, on the other hand, demonstrates a superior ability to capture the full extent of such defects with higher accuracy.In the final two image sets, YOLOv8n produces several false positives and incorrect classifications, whereas YOLOv8-MDEW maintains better precision and reliability.

Overall, the experimental results clearly indicate that the proposed YOLOv8-MDEW model outperforms the baseline YOLOv8n in terms of detection accuracy, robustness, and its ability to recognize and localize diverse defect types on wooden board edges, thereby demonstrating its practical value in industrial defect inspection scenarios.