面向刨花板表面多尺度缺陷正偏态分布的目标检测方法

doi:10.11707/j.1001-7488.LYKX20250130

摘要/Abstract

摘要：

目的: 针对刨花板表面缺陷尺度变化大、多尺度缺陷共存及缺陷数量随缺陷尺度呈正偏态分布导致检测精度低的问题，提出一种具备自适应感受野能力的目标检测方法（PBDNet），用于提升刨花板表面缺陷检测的精度和速度。方法: PBDNet通过引入空间分割与通道融合策略（SPDConv）作为下采样方法，将特征张量在空间上分割并在通道上拼接，减少下采样过程中的信息损失，为正偏态分布中高峰一侧的缺陷保留更多的细粒度特征，提高了检测模型对缺陷数量随缺陷尺度呈正偏态分布时的检测能力。此外，在PBDNet中提出的特征提取模块（C2f_SD），通过将可切换空洞卷积与差分卷积添加到C2f特征提取模块中，显著提高了模型对不同尺度缺陷的检测能力。结果: 对比试验及消融试验结果表明，PBDNet在mAP₅₀值和Recall值上均优于主流缺陷检测算法。相较于YOLOv8s，PBDNet的mAP₅₀值和Recall值分别提升4.8%和6.4%，达到0.881与0.840，同时在保持推理速度小于3 ms的情况下模型参数量减少42.2%。结论: PBDNet检测方法能够满足刨花板表面多尺度缺陷正偏态分布时的检测需求，为实时精确检测的需求刨花板表面缺陷检测产业化应用提供了一种高效、精准、可边缘部署的自动化检测方法。

关键词: 刨花板表面缺陷, 目标检测, 多尺度, 正偏态分布, YOLO

Abstract:

Objective: To address the low detection accuracy caused by the large scale changes in particleboard surface defects, the coexistence of multi-scale defects, and the positively skewed distribution of defect quantities with respect to defect sizes, an object detection method with adaptive receptive field capability (PBDNet) was proposed in this study. This method was designed with an adaptive receptive field capability to improve the accuracy and efficiency in particleboard surface defect detection. Method: By introducing the spatial splitting and channel fusion strategy (SPDConv) as the downsampling method, PBDNet spatially split feature tensors and concatenated them in channels, thereby reducing information loss during downsampling, and preserving more fine-grained features for defects on the high-frequency side of the positively skewed distribution. This method enhanced the detection ability of the detection model when the number of defects follows a positively skewed distribution with defect scale. Additionally, the feature extraction module (C2f_SD) proposed in PBDNet significantly improved the model's ability to detect defects of different scales by incorporating switchable atrous convolution and differential convolution into the C2f feature extraction module. Result: The comparative and ablation experiments demonstrated that the PBDNet outperformed mainstream defect detection algorithms in terms of both mAP₅₀ and Recall. Compared with YOLOv8s, PBDNet achieved improvements of 4.8% and 6.4% in mAP₅₀ and Recall, reaching 0.881 and 0.840, respectively. Furthermore, the parameter count was reduced by 42.2% while nearly maintaining the inference speed under 3 ms. Conclusion: The PBDNet detection method can meet the requirements for detection of the positively skewed distribution of multi-scale defects on particleboard surface. It provides an efficient, accurate, and edge-deployable automated solution for real-time precision detection, thereby facilitating industrial applications on particleboard surface defect detection.

Key words: particleboard surface defect, object detection, multi-scale, positively skewed distribution, YOLO

中图分类号:

S784
TP391.4

刘恒,郭浩盟,戴蕙泽,柴哲明,李春育,杨建华. 面向刨花板表面多尺度缺陷正偏态分布的目标检测方法[J]. 林业科学, 2025, 61(12): 164-176.

Heng Liu,Haomeng Guo,Huize Dai,Zheming Chai,Chunyu Li,Jianhua Yang. Object Detection Method for Positively Skewed Distribution of Multi-Scale Defects on Particleboard Surface[J]. Scientia Silvae Sinicae, 2025, 61(12): 164-176.

图/表 19

图1

图2

图3

图4

图5

图6

图7

图8

图9

图10

表1

图11

图12

表2

表3

表4

表5

图13

图14

参考文献 0

	关淯尹, 王文财, 沈　彤, 等. 基于改进YOLOv5s模型的刨花板表面缺陷检测方法. 智能制造, 2024, 31 (5): 104- 112.
	Guan Y, Wang W, Shen T, et al. Surface defect detection method of particleboard based on improved YOLOv5s model. Intelligent Manufacturing, 2024, 31 (5): 104- 112.
	唐忠荣, 李素珍. 刨花板外观质量缺陷分析及对策. 林产工业, 2005, 42 (6): 26- 28.
	Tang Z R, Li S Z. External defects of particleboard and their countermeasures. China Forest Products Industry, 2005, 42 (6): 26- 28.
	王文财, 党亚光, 朱　翔, 等. 基于Lite-YOLOv5s模型的刨花板表面缺陷检测方法. 木材科学与技术, 2023, 37 (3): 58- 67.
	Wang W, Dang Y, Zhu X, et al. Detecting method for particleboard surface defects based on the Lite-YOLOv5s model. Chinese Journal of Wood Science and Technology, 2023, 37 (3): 58- 67.
	杨　龙, 毛顺国, 陆铜华, 等. 我国刨花板工业技术进步回顾与展望. 木材科学与技术, 2024, 38 (1): 1- 12.
	Yang L, Mao S, Lu T et al. Review and outlook of technology development in China’s particleboard industry. Chinese Journal of Wood Science and Technology, 2024, 38 (1): 1- 12.
	杨　渊. 机器视觉在刨花板表面缺陷检测中的应用现状与建议. 中国人造板, 2024, 31 (5): 33- 37.
	Yang Y. Application of machine vision in particleboard surface defect detection: status and recommendations. China Wood-Based Panels, 2024, 31 (5): 33- 37.
	由　佳, 胡生辉, 金　枝. 浅析我国人造板产业简况. 绿色中国, 2024, 21 (16): 56- 61.
	You J, Hu S, Jin Z. A brief analysis of China’s artificial board industry. Green China, 2024, 21 (16): 56- 61.
	查　健, 陈先中, 王文财, 等. 基于改进的YOLOv5s刨花板表面小目标缺陷检测算法. 计算机工程与应用, 2024, 60 (17): 158- 166.
	Zha J, Chen X, Wang W, et al. Small defect detection algorithm of particle board surface based on improved YOLOv5s. Computer Engineering and Applications, 2024, 60 (17): 158- 166.
	朱　豪, 周顺勇, 曾雅兰, 等. 基于改进YOLOv5s的木材表面缺陷检测模型. 木材科学与技术, 2023, 37 (2): 8- 15.
	Zhu H, Zhou S, Zeng Y et al. Detection model of wood surface defects based on improved YOLOv5s. Chinese Journal of Wood Science and Technology, 2023, 37 (2): 8- 15.
	Bany Muhammad M, Yeasin M. Eigen-CAM: visual explanations for deep convolutional neural networks. SN Computer Science, 2021, 2 (1): 47. doi: 10.1007/s42979-021-00449-3
	He K, Gkioxari G, Dollár P, et al. 2017. Mask R-CNN. 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy, 2980–2988.
	Hu W, Wang T, Wang Y, et al. LE–MSFE–DDNet: a defect detection network based on low-light enhancement and multi-scale feature extraction. The Visual Computer, 2022, 38 (11): 3731- 3745. doi: 10.1007/s00371-021-02210-6
	Jiao L, Zhang F, Liu F, et al. A survey of deep learning-based object detection. IEEE Access, 2019, 7, 128837- 128868. doi: 10.1109/ACCESS.2019.2939201
	Lin T Y, Goyal P, Girshick R, et al. 2017. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy, 2999–3007.
	Liu W, Anguelov D, Erhan D, et al. 2016. SSD: single shot multibox detector. Computer Vision–ECCV 2016, 14th European Conference, Amsterdam, The Netherlands, Proceedings, Part I 14, Springer, 21−37.
	Luo W, Li Y, Urtasun R, et al. Understanding the effective receptive field in deep convolutional neural networks. Advances in Neural Information Processing Systems, 2016, 29, 4898- 4906.
	Lyu C, Zhang W, Huang H, et al. 2022. RTMDet: an empirical study of designing real-time object detectors. arXiv preprint arXiv, 2212.07784.
	Mehta S, Rastegari M. 2021. MobileViT: light-weight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv: 2110.02178.
	Peng C, Li X, Wang Y. TD-YOLOA: an efficient YOLO network with attention mechanism for tire defect detection. IEEE Transactions on Instrumentation and Measurement, 2023, 72, 1- 11.
	Redmon J, Divvala S, Girshick R, et al. 2016. You only look once: unified, real-time object detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Barcelona, Spain, 779–788.
	Ren S, He K, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39 (6): 1137- 1149. doi: 10.1109/TPAMI.2016.2577031
	Sunkara R, Luo T. 2023. No more strided convolutions or pooling: a new CNN building block for low-resolution images and small objects. Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 443–459.
	Wang A, Chen H, Liu L, et al. YOLOv10: real-time end-to-end object detection. Advances in Neural Information Processing Systems, 2024, 37, 107984- 108011.
	Zhang C, Wang C, Zhao L, et al. A method of particleboard surface defect detection and recognition based on deep learning. Wood Material Science & Engineering, 2025, 20 (1): 50- 61.
	Zhang Y, Li X, Wang F, et al. 2021. A comprehensive review of one-stage networks for object detection. 2021 IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC). Xi’an, Shaanxi, China, 1–6.
	Zhao C, Shu X, Yan X, et al. RDD-YOLO: a modified YOLO for detection of steel surface defects. Measurement, 2023, 214, 112776. doi: 10.1016/j.measurement.2023.112776
	Zhao Y, Lv W, Xu S, et al. 2024. DETRs beat YOLOs on real-time object detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle WA, USA, 16965–16974.
	Zhao Z, Yang X, Zhou Y, et al. Real-time detection of particleboard surface defects based on improved YOLOV5 target detection. Scientific Reports, 2021, 11 (1): 21777. doi: 10.1038/s41598-021-01084-x
	Zhou W, Li C, Ye Z, et al. An efficient tiny defect detection method for PCB with improved YOLO through a compression training strategy. IEEE Transactions on Instrumentation and Measurement, 2024, 73, 1- 14.
	Zou Z, Chen K, Shi Z, et al. Object detection in 20 years: a survey. Proceedings of the IEEE, 2023, 111 (3): 257- 276. doi: 10.1109/JPROC.2023.3238524

模型名称 Model Name	mAP₅₀	mAP_50-95	召回率 Recall	参数量 Parameters (M)	推理时间 Time/ms
Faster R-CNN	0.702	0.445	0.541	84.70	25.10
YOLOv8s	0.833	0.619	0.776	11.14	2.49
RTMDet-s	0.836	0.566	0.676	39.00	54.20
RT-DETR	0.793	0.583	0.776	42.78	8.10
YOLOv10s	0.801	0.575	0.684	8.07	3.05
PBDNet	0.881	0.655	0.840	6.43	2.46

模型 Model	通道裁剪 Slim	SPDConv	C2f_SD	mAP₅₀	mAP_50-95	召回率 Recall	参数量 Parameters（M）	推理时间 Time/ms
A	—	—	—	0.833	0.619	0.776	11.14	2.49
B	—	√	—	0.828	0.61	0.747	10.11	2.51
C	—	—	√	0.849	0.653	0.818	14.42	4.98
D	—	√	√	0.85	0.636	0.791	13.60	5.12
E	√	—	—	0.845	0.634	0.785	8.48	2.39
F	√	√	—	0.857	0.623	0.823	5.50	1.65
G	√	—	√	0.867	0.648	0.804	6.94	2.34
H	√	√	√	0.881	0.655	0.840	6.40	2.46

缺陷类别 Defect class	mAP₅₀
缺陷类别 Defect class	A	B	C	D	E	F	G	H
斑状缺陷 Spot-like defect	0.814	0.832	0.815	0.831	0.822	0.825	0.850	0.863
刨花 Shavings	0.833	0.780	0.859	0.842	0.88	0.818	0.893	0.883
油污 Oil pollution	0.462	0.530	0.523	0.541	0.484	0.589	0.543	0.618
崩边 Edge breakage	0.973	0.958	0.971	0.943	0.971	0.969	0.972	0.958
粉笔记号 Chalk mark	0.902	0.919	0.913	0.934	0.911	0.966	0.951	0.968
划痕 Scratch	0.890	0.811	0.910	0.910	0.895	0.879	0.893	0.926
裂纹 Crack	0.954	0.968	0.953	0.950	0.954	0.953	0.964	0.953

缺陷类别 Defect class	召回率Recall
缺陷类别 Defect class	A	B	C	D	E	F	G	H
斑状缺陷 Spot-like defect	0.721	0.721	0.754	0.754	0.803	0.787	0.820	0.816
刨花 Shavings	0.764	0.739	0.857	0.810	0.832	0.810	0.810	0.857
油污 Oil pollution	0.402	0.325	0.500	0.400	0.359	0.566	0.428	0.575
崩边 Edge breakage	0.967	0.933	0.967	0.933	0.967	0.945	0.967	0.933
粉笔记号 Chalk mark	0.786	0.821	0.872	0.893	0.821	0.879	0.856	0.893
划痕 Scratch	0.845	0.743	0.829	0.800	0.771	0.829	0.800	0.861
裂纹 Crack	0.944	0.944	0.944	0.944	0.944	0.944	0.944	0.944

模型 Model	mAP₅₀	mAP_50-95	召回率 Recall	模型 Model	mAP₅₀	mAP_50-95	召回率 Recall
A	0.456	0.303	0.492	E	0.468	0.309	0.542
B	0.385	0.246	0.429	F	0.627	0.371	0.700
C	0.474	0.297	0.517	G	0.559	0.353	0.567
D	0.472	0.296	0.521	H	0.705	0.427	0.746