Surface defect detection model of laser cutting polycrystalline cubic boron nitride tool based on asymptotic fusion strategy | Scientific Reports

Nov 06, 2024

Scientific Reports volume 14, Article number: 26689 (2024) Cite this article

Metrics details

Due to the polycrystalline cubic boron nitride tool has the characteristics of high hardness, brittleness, etc., it is easy to break the tool or produce defects in the laser cutting process, which affects the cutting performance of the tool. Traditional defect detection methods can no longer meet the needs of modern manufacturing. Aiming at the problems of low accuracy and poor real-time detection of surface defects on laser-cutting polycrystalline cubic boron nitride tools, this study proposes the surface defect detection model of laser cutting polycrystalline cubic boron nitride tool based on asymptotic fusion strategy, which fills the gap in the field. In the backbone network, the C3SE module is constructed by modeling the correlation between feature channels to improve the model’s focus on key features in order to enhance the feature extraction and processing capability of the backbone network; In the neck network, adaptive spatial fusion operation and direct interaction of non-adjacent layers are utilized for multi-scale information fusion, and the asymptotic feature pyramid network for object detection (AFPN) is used instead of the FPN structure to improve the detection performance; In the head network, a soft suppression mechanism is introduced to reduce the overlapping frame score using a decay function, thus improving the detection accuracy. The experimental results based on the self-constructed laser-cutting polycrystalline cubic boron nitride tool surface defects dataset show that the average accuracy of the AFFS-YOLO model is improved by 5.6% compared with that of the YOLOv5 model, reaching 86.1%, and the detection effect is better than that of the original network and other classical target detection networks.

As a new tool material in the 21st century, polycrystalline cubic boron nitride tools are becoming an important and indispensable tool in modern manufacturing1. Polycrystalline cubic boron nitride tools, due to its high hardness and brittle characteristics, are prone to burst, cracking, ablation and other defects on the surface due to improper power parameters during laser cutting, which in turn affects its cutting performance and service life. Therefore, in the process of rapid development of modern manufacturing industry, it is particularly important to realize fast and high-precision laser cutting polycrystalline cubic boron nitride tool surface defect detection.

Traditional surface defect detection methods rely on the original characteristics of the local defects, roughly through the manual visual inspection method, mechanical device contact detection method and machine vision inspection method and other stages2. These traditional methods work in limited environmental conditions, limiting efficiency. In recent years, the research and application of deep learning convolutional neural networks have developed rapidly and achieved remarkable results in the field of computer vision, and the deep learning method has gradually taken a dominant position in the field due to the excellent performance advantages it shows when dealing with industrial images with complex backgrounds and insignificant defects3. However, samples with surface defects in industrial cutting processes are often difficult to obtain and suffer from problems such as class imbalance, making it difficult to train deep learning models. The current mainstream practice of algorithmic improvement around the surface defect detection accuracy problem of deep learning algorithms is to optimize the network structure, introduce the attention mechanism, improve the loss function and so on. In the research on defect detection based on deep learning technology, in 2015, Ren et al.4 introduced the regional proposal network as an “attention” mechanism for the first time to guide the network to locate targets, enhance feature extraction, capabilities, and improve detection accuracy; In the same year, Joseph et al.5 proposed the YOLO algorithm, which solves object detection as a regression problem. Unlike the Faster R-CNN, which requires repeated training of the RPN network and the Fast R-CNN6 network, the YOLO algorithm is trained and detected in a separate network, achieving fast detection with high accuracy. Subsequently, scholars proposed algorithms such as YOLOv2-97,8,9,10,11,12 to further improve the accuracy and detection speed of the model. YOLO algorithm with its real-time processing advantages and defect detection capabilities can meet the industrial real-time inspection needs, is widely used in various fields, and has been continuously optimized and developed. Su et al.13 improved the yolov3 model by introducing the SE attention mechanism and the spatial pyramid pooling module for detecting apparent bridge damage in complex contexts; Li et al.14 improved the yolov4 model by introducing a dual-channel feature enhancement module to ameliorate the local information loss and gradient propagation blockage caused by four-chain convolution in PANet networks for detecting fabric surface defects; Zhu et al.15 proposed the MFF-YOLO model by constructing a multi-scale feature fusion network and designing a reweighting screening method for achieving accurate detection of tunnel defects; Xu et al.16 improved the yolov7 algorithm by introducing the Le-HorBlock module to the CBS module and the Coordinate Attention module to construct a surface defect detection model for pipeline welds.

Although deep learning methods have achieved a large number of research results in the field of target detection, not much research has been done in the field of defect detection in laser-cut polycrystalline cubic boron nitride tools. Therefore, it is prospective to investigate the application of deep learning in defect detection of laser-cut polycrystalline cubic boron nitride tools.

This study aims at the problem of low detection accuracy of laser cutting polycrystalline cubic boron nitride tool defect detection. Firstly, a proprietary dataset is produced through electron microscope acquisition; Second, in terms of data processing, the AFFS-YOLO model is proposed. The model can effectively improve the accuracy and reliability of polycrystalline cubic boron nitride tool surface defect detection, which is of positive academic significance for advancing the development of the fusion of deep learning and laser cutting polycrystalline cubic boron nitride tool defect detection.

The main contributions of this paper are as follows:

(1) Compiled a proprietary dataset of laser-cut polycrystalline cubic boron nitride tool defects, which can meet the detection of common surface defects;

(2) In the backbone network, the C3SE module is constructed by modeling the correlation between feature channels to enhance the feature extraction and processing capability of the backbone network;

(3) In the neck network, the adaptive spatial fusion operation of the AFPN network structure as well as the direct interaction of disjoint layers are used for multi-scale information fusion to improve the detection performance;

(4) In the head network, the soft suppression mechanism is introduced to reduce the overlapping frame scores using the attenuation function to improve the detection accuracy.

The rest of the article is structured as follows: In the Methods section, the process of building the AFFS-YOLO model is described; In the Experimental section, the details of the self-constructed defect dataset and the configurations and hyperparameters used in the experimental training are presented; In the Results and discussion section, the evaluation performance of the model, the results obtained and the interpretation of the detailed data from the ablation experiments are presented in detail; In the Conclusion section, the key role of the experiments in the detection of surface defects on laser-cut polycrystalline cubic boron nitride tools, including the existing limitations, is discussed and summarized in the last section.

The Yolo family of target detection algorithms is a single-stage end-to-end real-time target detection method, and the Yolov5 target detection network is highly regarded in the field of target detection due to its accuracy and fast detection capability. Yolov5s is the network with the shallowest network depth and smallest network width among the official 4 versions, and its network structure consists of four parts: Input, Backbone17, Neck18 and Head19, which is shown in Fig. 1.

Yolov5 detection model structure diagram.

Backbone network adopts the CSPDarknet53 structure, introduces the CSP20 structure to enhance the information flow through cross-stage connections, and employs residual connections to effectively solve the gradient vanishing and gradient explosion problems in deep networks through constant mapping. Neck network with the help of Feature Pyramid Network (FPN) architecture, the network is able to fuse multi-scale feature information by means of shared feature maps, enabling the model to utilize both high-level semantic information and low-level detail information. In addition, Head network partially follows the YOLOv3 like detection head structure, which includes three different scales of detection layers to accurately predict the class and location of defects at different scales, enabling the model to perform well in the target detection task. This is the reason why we chose to improve the yolov5s model for defect detection on the surface of the laser-cut Polycrystalline Cubic Boron Nitride Tool after integrating real-time and accuracy considerations.

After introducing the basic model of yolov5s, we will describe the advantages of each improved module over the original module in the Improved methodology.

In the backbone network of YOLOv5s, the C3 module is the key component, which extracts useful information from the input feature map through multiple convolutional layers and utilizes the Bottleneck module and residual connections to achieve effective fusion of features at different levels. However, the C3 module pays equal attention to different feature channels and does not consider the correlation between different feature channels sufficiently.

C3SE module structure diagram.

In this study, the Squeeze-and-Excitation (SE)21 attention mechanism is combined with the C322 module to construct the C3SE module as in Fig. 2. On the one hand, the core of the SE attention mechanism is to learn and redistribute the weights among the feature channels, and when combined with the C3 module, the model is able to adaptively adjust the feature weights of the different channels and learn the correlations among the different channels by explicitly modeling the interdependencies among the feature channels, thus enhancing the feature extraction and processing capabilities of the backbone network. On the other hand, the SE attention mechanism is embedded into the C3 module, and when the input image reaches the C3SE through multiple convolutional layers, the Bottleneck module of the original C3 module is utilized to solve the gradient vanishing and exploding problems, and then the depth of the network is extended by stacking the convolutional layers, and the idea of cross-layer connectivity is adopted to capture more contextual information, and then two operations of squeezing and excitation are carried out immediately to improve model expressiveness of the model.

The flowchart for performing the two operations of squeezing and excitation is shown in Fig. 3.

\(\:{F}_{\text{tr}}:X\to\:U\), After \(\:{F}_{\text{tr}}\) feature map X becomes feature map U. The formula is defined as in Eq. 2. \(\:{F}_{\text{sq}}\)is to use the global average pooling of channels to directly compress the W × H × C feature map containing global information into a 1 × 1 × C feature vector Z. The channel features of the C feature maps are all compressed into a single value, which makes the generated channel-level statistics Z contain contextual information and alleviates the problem of channel dependency, where \(\:{Z}_{c}\) in Eq. 2 is the cth element of Z.

In order to utilize the information aggregated in the compression operation, the channel dependencies are fully captured by the Excitation operation, which yields the weight s of Eq. 3:

Finally, the attentional weights obtained earlier are weighted to the features of each channel by the Scale operation.

Squeeze-and-Excitation module.

In the yolov5s algorithm, the neck network adopts a combination of SPPF and CSP-PAN structures, which effectively improves the model’s detection performance for targets of different scales. However, this structure may adopt a more fixed method in feature fusion and cannot adaptively adjust the fusion strategy according to the scale of the target, and at the same time, this structure may not be able to balance the computational efficiency well while pursuing the performance improvement.

The introduced AFPN23 network structure is shown in Fig. 4, which improves the accuracy and robustness of target detection by adaptively adjusting the feature fusion method and adopting an incremental fusion strategy to incrementally fuse low-level features, high-level features, and top-level features to the neck network.

The AFPN progressively integrates Low-Level, High-Level, and top-level features in the bottom-up feature extraction process of Backbone network. In the multi-level feature fusion process, ASFF is utilized to assign different spatial weights to features at different Levels, which enhances the importance of key Levels and mitigates the effect of contradictory information from different targets.

The key idea of ASFF is to adaptively learn the fusion spatial weights for each scale feature map in two steps: constant scaling and adaptive fusion. Its feature vector is represented as:

In Fig. 4, the links from level1, 2, 3, 4 to ASFF1, 2, 3, 4 are fully connected. The feature map is resized before fusion. the process of fusing level1, 2, 3, 4 into ASFF-3 is to perform up-sampling, down-sampling, and pooling operations on level1, 2, 3, and 4 to make the size of level1, 2, 3, and 4 the same before fusion operation. \(\:{x}_{ij}^{n\to\:l}\) represents the feature vector at (i, j) after resize from level n to level l. The fused ASFF-L needs to satisfy the following constraints:

This constraint is satisfied by \(\:{x}_{ij}^{1\to\:l},\,{x}_{ij}^{2\to\:l},\,{x}_{ij}^{3\to\:l}\), after 1∗1 convolution to get \(\:{\lambda\:}_{\alpha\:}^{l},\,{\lambda\:}_{\beta\:}^{l},\,{\lambda\:}_{\gamma\:}^{l}\) and then soft-max to output \(\:{y}^{1},\,{y}^{2},\,{y}^{3}\).

AFPN module structure diagram.

The target detection model generates multiple possible target bounding boxes, each of which has a corresponding confidence level indicating the probability that the bounding box contain the target, and a large number of overlapping boxes are generated because the target may be detected multiple times at different scales or different locations.

SOFT-NMS module flow chart.

The role of NMS is to filter these overlapping bounding boxes, removing those with lower confidence and retaining only those with the highest confidence, thus obtaining more accurate target detection results.

Traditional NMS methods may incorrectly remove some bounding boxes with high confidence due to the use of hard thresholds when dealing with overlapping target boxes, leading to a decrease in detection accuracy. In contrast, Soft-NMS24 selects the best bounding boxes more efficiently by introducing a decay function that reduces the scores of overlapping bounding boxes instead of deleting them directly, as shown in the flowchart in Fig. 5.

The functional relationship is expressed as:

The structural features of the improved model are shown in Fig. 6. The model improvement has three main parts:

First, the C3SE structure of Backbone network, which enhances the feature extraction and processing capability of the backbone network;

Second, the AFPN structure of Neck network, which utilizes adaptive spatial fusion operation and direct interaction of non-adjacent layers for multi-scale information fusion;

Third, the introduction of Soft-NMS soft suppression mechanism, which utilizes the attenuation function to reduce the overlapping frame score and improve the detection accuracy.

AFFS-YOLO module structure diagram.

The laser cutting polycrystalline cubic boron nitride tool dataset was constructed by collecting contour images of polycrystalline cubic boron nitride tools processed by Covington fiber laser cutting machine using an AO-HD208C electron microscope. Table 1 is the experimental environment parameters for data set production.

The product samples in this dataset are all from the same batch of products processed by the same processing equipment under the same factory environmental conditions.

Combining the data collected in the experiment with expert experience and industrial demand, we made a proprietary data set for the 4 identified defects by manual annotation, and conducted the defect detection algorithm research by offline processing. The Ablation defect is caused by the local melting, evaporation or chemical decomposition of the surface or sub-surface of the material when the high temperature laser beam interacts with the material; Burst defects are the material breaking phenomena caused by the sudden release of energy in the laser cutting process due to the uneven or excessive internal stress of the material; Adulterant is the presence of unintended ingredients or foreign substances in the material; Cracks are linear defects on the surface or inside the material due to factors such as thermal stress or material inhomogeneity. Figure 7 illustrates some typical examples of these defects.

The final nearly 3000 images obtained by Labelimg software labeling comprise the dataset of this study, and the defect images are categorized into four categories such as AB (ablation), BU (burst), AD (adulterant), and CR (crack), which denote four types of defects, namely, ablation, burst, adulterant, and crack, respectively. Table 2 demonstrates the distribution of the dataset species defect images.

Examples of typical defects.

The deep learning simulation experiments are built on a Linux system with a deep learning framework using Python and PyTorch, and the hardware setup shown in Table 3 includes components such as CPU, GPU, memory, and storage. The experiment consists of 50 epoches with a learning rate of 0.01, a weight decay factor set to 0.0005, and a gradient optimization algorithm of SGD. The default image size is 640 × 640 and the batch size is 16 threads.

In this study, the experiments use precision (P), recall (R), and mean accuracy (mAP) to evaluate the effectiveness of the model in detecting defects in polycrystalline cubic boron nitride cutting tools. mAP can reflect the stability and consistency of the model under different thresholds, so as to more accurately evaluate the performance of the model. In this experiment, two indicators, [email protected]% and [email protected]:0.95%, are used as the performance reference indexes, in which the recall rate indicates the proportion of all positive cases that are detected as positive cases, and the average accuracy value indicates the proportion of samples that are detected as positive cases that are truly positive cases, which can be computed from the Precision and the Recall rate. The formulae for each indicator are as follows:

The Fig. 8 shows the confusion matrix generated by the AFFS-YOLO model in the laser-cut polycrystalline cubic boron nitride tool surface defects dataset, and the results show that the detection accuracies of CR and BU reach 91% and 92%, respectively, while the detection accuracies of AD and AB are 64% and 77%, respectively, and [email protected] reaches 86.1%, and the precision-recall curve is shown in Fig. 9. In practical scenarios, laser cutting is supposed to utilize high-energy laser beam irradiation, which will form a texture on the material surface. During image acquisition using an electron microscope, the texture features produced by cutting are easily confused with ablation defects, which means that the model shows limitations in recognizing multi-scale objects as well as image details.

The confusion matrix generated by the AFFS-YOLO model.

PR curve.

The study used the AFFS-YOLO model for the surface defect detection of laser-cutting polycrystalline cubic boron nitride tools, and successfully detected a variety of types and sizes of defects, and some of the results are shown in Fig. 10.

Defect detection results.

The training data is represented as a scatter plot to obtain the curve shown in Fig. 11, which is the loss function curve. It can be seen through observation that: in the 20th round, the loss function begins to converge and then tends to be stable, the decline of the loss function value indicates that the model is gradually optimized during the training process, while the convergence and stability of the loss function value indicates that the model has reached the optimal state.

Comparison of loss function curves of each frame.

In order to test the effectiveness of each model for various types of defect detection to highlight the importance of each module, the study has counted the results of different types of accuracies as shown in Fig. 12; Table 4. The four parts in Fig. 12 represent four types of defects, and different modules are added to each part from left to right, where the original model is shown in blue and the improved AFFS-YOLO model is shown in red. The results show that the improved model proposed in this paper outperforms the original model in all types, which further validates the effectiveness of the model in detecting defects in laser-cut polycrystalline cubic boron nitride tools.

The accuracy of each framework is compared in the absence of defect types.

In order to evaluate the performance of the AFFS-YOLO model proposed in this paper more accurately, the study conducted ablation tests and obtained the experimental results as shown in Fig. 13; Table 4. The results show that the model AFFS-YOLO achieves 86.1% accuracy.

Performance comparison of typical target detection frameworks.

The study in Table 5 shows that the mAP, Recall and Gflops of the model are significantly improved by adding AFPN through the deep fusion extraction of feature information.

After the model adds to reprocess the prediction frames, the accuracy is improved by 5.6% compared to the base model, although the Gflops is slightly decreased compared to adding AFPN.

In this study, when the three modules are combined together to construct the AFFS-YOLO model, the result is a synergistic effect that outperforms the effect of each module acting individually and raises the upper limit of the total performance of the model.

In the field of deep learning, there are differences and uniqueness of target detection models for different scenarios. In this study, the AFFS-YOLO model is proposed for the specific scenario of defect detection in laser-cut polycrystalline cubic boron nitride tools. After simulation experiments, it is verified that the improved model can meet the current laser-cutting polycrystalline cubic boron nitride tool defect detection requirements, and the average accuracy rate reaches 86.1%. Compared with the yolov5s base model and other inspection models, the accuracy of this method is higher.

Although the dataset provided in this paper contains four defect types, there are some defect types that are not included. The defect dataset is difficult to obtain during industrial processing, resulting in a small number of samples. Therefore, the proposed defect detection method still has some limitations.

In our future work, we will try to investigate small-sample defect detection techniques to effectively improve the effect of small dataset on detection accuracy. In addition, we note that in the industrial field mostly deployed with edge mobile devices, and deep learning models usually have a large number of parameters, which makes the models complex and difficult to deploy to resource-limited devices, so we will also consider using methods such as knowledge distillation to reduce the model size and explore lightweight target detection algorithms.

In summary, the improved AFFS-YOLO model in this study performs well in detecting defects in laser-cut polycrystalline cubic boron nitride tools, providing new ideas and methods for research in this field. We hope that these works can provide useful references and guidance for practical applications and future research.

Data supporting the results of this study are available from the corresponding author upon reasonable request.

Jarosz, K. et al. Laser ablation and processing of polycrystalline cubic boron nitride cutting tool material. Int. J. Adv. Manuf. Technol. 118, 785–800 (2021).

Article Google Scholar

Pang, S., Zhao, W., Wang, S., Zhang, L. & Wang, S. Permute-MAML: exploring industrial surface defect detection algorithms for few-shot learning. Complex. Intell. Syst. 10 (1), 1473–1482 (2024).

Article Google Scholar

Chen, X., Lv, J., Fang, Y. & Du, S. Online detection of Surface defects based on improved YOLOV3. Sensors. 22 (3), 817. https://doi.org/10.3390/s22030817 (2022).

Article ADS PubMed PubMed Central Google Scholar

Ren, S., He, K., Girshick, R. & Sun, J. Faster r-cnn: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39 (6), 1137–1149 (2017).

Article PubMed Google Scholar

Redmon, J. et al. You Only Look once: Unified, Real-Time Object Detection ( Computer Vision & Pattern Recognition IEEE, 2016).

Girshick, R. R-CNN[J].Computer Science, 2015. https://doi.org/10.1109/ICCV.2015.169.

Redmon, J., Farhadi, A. & Recognition, P. IEEE Conference on Computer Vision and YOLO9000: Better, Faster, Stronger, (CVPR), Honolulu, HI, USA, 2017, pp. 6517–6525, doi: (2017). https://doi.org/10.1109/CVPR.2017.690

Redmon, J. & Farhadi, A. Yolov3: An Incremental Improvement (arXiv e-prints, 2018).

Bochkovskiy, A., Wang, C. Y. & Liao, H. Y. M. .YOLOv4: optimal speed and accuracy of object Detection[J]. DOI: (2020). https://doi.org/10.48550/arXiv.2004.10934

Li, C. et al. YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications. ArXiv, abs/2209.02976. (2022).

Wang, C. Y., Bochkovskiy, A. & Liao, H. Y. M. YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors, 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 2023, pp. 7464–7475, https://doi.org/10.1109/CVPR52729.2023.00721

Wang, C., Yeh, I. & Liao, H. YOLOv9: learning what you want to learn using programmable gradient information. Preprint at https://arxiv.org/abs/2402.13616 (2024).

Su, H. et al. Enhanced YOLO v3 for precise detection of apparent damage on bridges amidst complex backgrounds. Sci. Rep. 14, 8627 (2024).

Article ADS PubMed PubMed Central Google Scholar

Li, Y. et al. Research on fabric surface defect detection algorithm based on improved Yolo_v4. Sci. Rep. 14, 5537. https://doi.org/10.1038/s41598-023-50671-7 (2024).

Article ADS PubMed PubMed Central Google Scholar

Zhu, A., Wang, B., Xie, J. & Ma, C. MFF-YOLO: an Accurate Model for detecting tunnel defects based on Multi-scale Feature Fusion. Sensors. 23 (14), 6490. https://doi.org/10.3390/s23146490 (2023).

Article ADS PubMed PubMed Central Google Scholar

Xu, X. & Li, X. Research on surface defect detection algorithm of pipeline weld based on YOLOv7. Sci. Rep. 14, 1881. https://doi.org/10.1038/s41598-024-52451-3 (2024).

Article ADS PubMed PubMed Central Google Scholar

Huang, H. et al. Small object detection method with shallow feature fusion network for chip surface defect detection. Sci. Rep. 12, 3914. https://doi.org/10.1038/s41598-022-07654-x (2022).

Article ADS PubMed PubMed Central Google Scholar

Liu, Z. & Lv, H. YOLO_Bolt: a lightweight network model for bolt detection. Sci. Rep. 14, 656. https://doi.org/10.1038/s41598-023-50527-0 (2024).

Article ADS PubMed PubMed Central Google Scholar

Wang, J. et al. Toward surface defect detection in electronics manufacturing by an accurate and lightweight YOLO-style object detector. Sci. Rep. 13, 7062. https://doi.org/10.1038/s41598-023-33804-w (2023).

Article ADS PubMed PubMed Central Google Scholar

Wang, C. et al. CSPNet: A New Backbone that can Enhance Learning Capability of CNN. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 1571–1580. (2019).

Hu, J., Shen, L. & Sun, G. Squeeze-and-Excitation Networks, 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, pp. 7132–7141, doi: (2018). https://doi.org/10.1109/CVPR.2018.00745

Park, H. et al. C3: Concentrated-Comprehensive Convolution and its Application to Semantic Segmentation (Computer Vision and Pattern Recognition, 2018).

Yang, G. et al. AFPN: Asymptotic Feature Pyramid Network for Object Detection. 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2184–2189. (2023).

Bodla, N., Singh, B., Chellappa, R. & Davis, L. S. — Improving Object Detection with One Line of Code. 2017 IEEE International Conference on Computer Vision (ICCV), 5562–5570. (2017).

Hou, Q., Zhou, D. & Feng, J. Coordinate Attention for Efficient Mobile Network Design. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13708–13717. (2021).

Woo, S., Park, J., Lee, J. & Kweon, I. CBAM: Convolutional Block Attention Module. (2018). arXiv:1807.06521.

Wang, Q., Wu, B., Zhu, P., Li, P. & Hu, Q. ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE. (2020).

Xiao, W., Xu, M. & Lin, Y. Global Feature Pyramid Network. arXiv:2312.11231. (2023).

Tan, M., Pang, R., Le, Q. V. & Recognition, P. EfficientDet: Scalable and Efficient Object Detection, 2020 IEEE/CVF Conference on Computer Vision and (CVPR), Seattle, WA, USA, pp. 10778–10787, doi: (2020). https://doi.org/10.1109/CVPR42600.2020.01079

Download references

North China University of Water Resources and Electric Power, Zhengzhou, 450045, China

Anfu Zhu, Jiaxiao Xie, Heng Guo, Jie Wang, Zilong Guo, Lei Xu, SiXin Zhu & Bin Wang

Science and Technology Research Institute of China Railway Zhengzhou Group Co., Ltd, Zhengzhou, 450045, China

Zhanping Yang

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

Conceptualization, A.F.Z.; Methodology, J.X.X.; Software, J.X.X. and H.G.; Validation, J.X.X. and B.W.; Formal analysis, J.W.; Investigation, L.X. and S.Z.; Resources, H.G.; Data curation, J.W.; Writing—original draft, J.X.X.; Writing—review & editing, J.X.X.; Supervision, A.F.Z. and Z.Y.; Project administration, H.G. and A.F.Z. All authors have read and agreed to the published version of the manuscript.

Correspondence to Anfu Zhu.

The authors declare no competing interests. The corresponding author is responsible for submitting a competing interests statement on behalf of all authors of the paper.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

Zhu, A., Xie, J., Guo, H. et al. Surface defect detection model of laser cutting polycrystalline cubic boron nitride tool based on asymptotic fusion strategy. Sci Rep 14, 26689 (2024). https://doi.org/10.1038/s41598-024-77676-0

Download citation

Received: 25 May 2024

Accepted: 24 October 2024

Published: 04 November 2024

DOI: https://doi.org/10.1038/s41598-024-77676-0

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative