改进YOLOv4的表面缺陷检测算法

李彬; 汪诚; 丁相玉; 巨海娟; 郭振平; 李卓越

doi:10.13700/j.bh.1001-5965.2021.0301

改进YOLOv4的表面缺陷检测算法

doi: 10.13700/j.bh.1001-5965.2021.0301

空军工程大学基础部，西安 710038

详细信息

通讯作者:
E-mail：warrant_74@126.com

中图分类号: V239；TP181
计量
- 文章访问数: 875
- HTML全文浏览量: 225
- PDF下载量: 206
- 被引次数: 0
出版历程
- 收稿日期: 2021-06-04
- 录用日期: 2021-10-11
- 网络出版日期: 2023-03-10
- 整期出版日期: 2023-03-30

Surface defect detection algorithm based on improved YOLOv4

Fundamentals Department，Air Force Engineering University，Xi’an 710038，China

More Information

Corresponding author: E-mail：warrant_74@126.com

摘要

摘要:
为解决航空发动机部件表面缺陷检测精度低、检测速度慢的问题,提出一种改进的YOLOv4算法进行智能检测。在路径聚合网络（PANet）结构中融合浅层特征与深层特征，增大特征检测尺度，同时去除自下而上的路径增强结构，提高小目标检测精度和整体检测速度；根据各类缺陷数量不同的情况，优化聚焦损失中的平衡参数，增加权重因子调节各类缺陷的损失权重，将改进后的聚焦损失代替分类误差中的交叉熵损失函数，降低样本不平衡和难易样本对检测精度的影响。实验表明：相比于原始YOLOv4算法，改进后的YOLOv4算法在测试集上的平均精度均值（mAP）为90.10%，提高了2.17%；检测速度为24.82 fps，提高了1.58 fps，检测精度也高于单发多框检测（SSD）算法、EfficientDet算法、YOLOv3算法和YOLOv4-Tiny算法。
- YOLOv4 /
- 表面缺陷检测 /
- 航空发动机 /
- 小目标检测 /
- 聚焦损失
Abstract:
In order to enhance the accuracy and speed of surface defect detection of aeroengine components, an improved YOLOv4 algorithm is proposed for intelligent detection. Firstly, shallow features and deep features were integrated into the path aggregation network (PANet) to improve the feature detection scale, and the bottom-up path augmentation structure was removed to increase the accuracy of small target detection and the overall detection speed. Then, according to the numbers of various defects, the balance parameter of the focal loss was optimized, and a weight factor was added to adjust the loss weight of various defects. The improved focal loss was used to replace the cross-entropy loss function in the classification error, thus reducing the impact of imbalanced samples and hard and easy samples on the detection accuracy. The experimental results show that the mean average precision (mAP) of the improved YOLOv4 on the test set is 90.10%, which is 2.17% higher than that of the traditional YOLOv4, and the detection speed is 24.82 fps, which is increased by 1.58 fps. The detection accuracy is also higher than other algorithms including single shot multibox detector (SSD), EfficientDet, YOLOv3 and YOLOv4-Tiny.
- YOLOv4 /
- surface defect detection /
- aeroengine /
- small target detection /
- focal loss

HTML全文

图 1 YOLOv4网络结构

Figure 1. YOLOv4 network structure

下载: 全尺寸图片幻灯片

图 2 改进的YOLOv4结构

Figure 2. Improved YOLOv4 structure

下载: 全尺寸图片幻灯片

图 3 缺陷类型

Figure 3. Defect types

下载: 全尺寸图片幻灯片

图 4 聚类中心为12的聚类结果

Figure 4. Clustering results when cluster center is 12

下载: 全尺寸图片幻灯片

图 5 不同模型的F_β对比

Figure 5. Comparison of F_βof different models

下载: 全尺寸图片幻灯片

图 6 各类缺陷样本的检测结果

Figure 6. Test results of various defect samples

下载: 全尺寸图片幻灯片

表 1 数据集中的目标数量

Table 1. Number of targets in dataset

缺陷类型	目标数量			总和
缺陷类型	训练集	验证集	测试集	总和
裂纹	1 579	580	595	2 754
缺口	848	272	270	1 390
凹坑	931	313	349	1 593
划痕	1 782	628	649	3 059

下载: 导出CSV

表 2 不同参数组合下的检测精度对比

Table 2. Comparison of detection accuracy under different parameter combinations

α₀	β₀	mAP/%	AP/%
α₀	β₀	mAP/%	裂纹	缺口	凹坑	划痕
0.20	2.50	88.02	85.04	86.91	91.73	88.38
0.20	3.00	88.24	85.94	87.74	90.66	88.60
0.20	3.50	89.01	84.90	90.62	92.18	88.32
0.25	2.50	89.09	86.40	90.92	90.07	88.95
0.25	3.00	90.10	87.09	91.21	92.65	89.45
0.25	3.50	88.49	84.55	90.09	90.27	89.06
0.30	2.50	89.00	86.62	87.66	92.29	89.41
0.30	3.00	88.62	86.86	90.65	89.30	87.68
0.30	3.50	88.34	85.98	89.53	89.91	87.94

下载: 导出CSV

表 3 不同模型的检测性能对比

Table 3. Comparison of detection performance of different models

网络结构	mAP/%	Speed /fps	AP/%
网络结构	mAP/%	Speed /fps	裂纹	缺口	凹坑	划痕
YOLOv4	87.93	23.24	86.64	92.29	86.02	86.78
YOLOv4-A	89.61	24.94	86.36	93.12	89.10	89.86
YOLOv4-B	90.10	24.82	87.09	91.21	92.65	89.45
注：fps表示帧/s。

下载: 导出CSV

表 4 不同模型的网络结构参数量对比

Table 4. Comparison of network structure parameters of different models

网络结构	总参数量	参数容量/MB
YOLOv4	63 953 841	243.96
YOLOv4-A	48 971 468	186.81
YOLOv4-B	48 971 468	186.81

下载: 导出CSV

表 5 不同算法的检测性能对比

Table 5. Comparison of detection performance of different algorithms

网络结构	mAP/%	Speed/fps
SSD	60.07	33.07
EfficientDet-D0	49.39	13.07
EfficientDet-D1	57.71	10.52
EfficientDet-D2	63.40	9.45
YOLOv3	86.25	30.37
YOLOv4-Tiny	54.25	64.55
YOLOv4	87.93	23.24
YOLOv4-B	90.10	24.82

下载: 导出CSV

参考文献(28)

[1]	李华. 基于孔探图像分析的航空发动机故障诊断专家系统研究[D]. 南京: 南京航空航天大学, 2015. LI H. Research on aeroengine fault diagnosis expert system based on endoscopic image analysis[D]. Nanjing: Nanjing University of Aeronautics and Astronautics, 2015(in Chinese).
[2]	关玉璞, 陈伟, 高德平. 航空发动机叶片外物损伤研究现状[J]. 航空学报, 2007, 28(4): 851-857. doi: 10.3321/j.issn:1000-6893.2007.04.014 GUAN Y P, CHEN W, GAO D P. Present status of investigation of foreign object damage to blade in aeroengine[J]. Acta Aeronautica et Astronautica Sinica, 2007, 28(4): 851-857(in Chinese). doi: 10.3321/j.issn:1000-6893.2007.04.014
[3]	何嘉辉, 张栋善, 赵成, 等. 航空发动机叶片裂纹检测技术及应用分析[J]. 内燃机与配件, 2020(15): 151-152. doi: 10.19475/j.cnki.issn1674-957x.2020.15.065 HE J H, ZHANG D S, ZHAO C, et al. Detection technology and application analysis of aero-engine blade crack[J]. Internal Combustion Engine & Parts, 2020(15): 151-152(in Chinese). doi: 10.19475/j.cnki.issn1674-957x.2020.15.065
[4]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: Single shot multibox detector[C]//Proceedings of European Conference on Computer Vision. Berlin: Springer, 2016: 21-37.
[5]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2016: 779-788.
[6]	REDMON J, FARHADI A. YOLO9000: Better, faster, stronger[C]//IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 6517-6525.
[7]	REDMON J, FARHADI A. YOLOv3: An incremental improvement[C]//IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2018: 1804-0276.
[8]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: Optimal speed and accuracy of object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2020: 10934.
[9]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[10]	DAI J F, LI Y, HE K M, et al. R-FCN: Object detection via region-based fully convolutional networks[C]//Proceedings of the 30th International Conference on Neural Information Processing Systems. New York: ACM, 2016: 379-387.
[11]	HE K M, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]//IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2017: 2980-2988.
[12]	王玺坤, 姜宏旭, 林珂玉. 基于改进型YOLO算法的遥感图像舰船检测[J]. 北京航空航天大学学报, 2020, 46(6): 1184-1191. doi: 10.13700/j.bh.1001-5965.2019.0394 WANG X K, JIANG H X, LIN K Y. Remote sensing image ship detection based on modified YOLO algorithm[J]. Journal of Beijing University of Aeronautics and Astronautics, 2020, 46(6): 1184-1191(in Chinese). doi: 10.13700/j.bh.1001-5965.2019.0394
[13]	陈科山, 郝宇, 何泓波, 等. 基于R-D SSD模型航空发动机安装工位检测算法[J]. 北京航空航天大学学报, 2021, 47(4): 682-689. CHEN K S, HAO Y, HE H B, et al. Detection algorithm of aeroengine installation station based on R-D SSD model[J]. Journal of Beijing University of Aeronautics and Astronautics, 2021, 47(4): 682-689(in Chinese).
[14]	旷可嘉. 深度学习及其在航空发动机缺陷检测中的应用研究[D]. 广州: 华南理工大学, 2017. KUANG K J. Research on deep learning and its application on the defects detection for aero engine[D]. Guangzhou: South China University of Technology, 2017(in Chinese).
[15]	李浩. 基于图像识别的航空发动机叶片裂纹检测研究[D]. 成都: 电子科技大学, 2019. LI H. Research on the blade crack detection of aero-engine based on image recognition[D]. Chengdu: University of Electronic Science and Technology of China, 2019(in Chinese).
[16]	陈为, 梁晨红. 基于改进SSD的航空发动机目标缺陷检测[J]. 控制工程, 2021, 28(12): 2329-2335. doi: 10.14107/j.cnki.kzgc.cpcc2019-063 CHEN W, LIANG C H. Aeroengine target defect detection based on improved SSD[J]. Control Engineering of China, 2021, 28(12): 2329-2335(in Chinese). doi: 10.14107/j.cnki.kzgc.cpcc2019-063
[17]	WANG C Y, MARK L H Y, WU Y H, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway: IEEE Press, 2020: 1571-1580.
[18]	MISRA D. Mish: A self regularized non-monotonic neural activation function[EB/OL]. (2020-08-13) [2021-06-01].
[19]	MAAS A L, HANNUN A Y, NG A Y. Rectifier nonlinearities improve neural network acoustic models[C]//Proceedings of Interational Conference on Machine learning, 2013: 1-6.
[20]	HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition.[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916. doi: 10.1109/TPAMI.2015.2389824
[21]	LIU S, QI L, QIN H F, et al. Path aggregation network for instance segmentation[C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2018: 8759-8768.
[22]	LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 2117-2125.
[23]	YU J H, JIANG Y N, WANG Z Y, et al. UnitBox: An advanced object detection network[C]//Proceedings of the 24th ACM International Conference on Multimedia. New York: ACM, 2016: 516-520.
[24]	REZATOFIGHI H, TSOI N, GWAK J, et al. Generalized intersection over union: A metric and a loss for bounding box regression [C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2019: 658-666.
[25]	ZHENG Z, WANG P, LIU W, et al. Distance-IoU loss: Faster and better learning for bounding box regression[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2020, 34(7): 12993-13000.
[26]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(2): 318-327.
[27]	TAN M X, PANG R M, LE Q V. EfficientDet: Scalable and efficient object detection[C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2020: 10778-10787.
[28]	TAN M X, LE Q V. EfficientNet: Rethinking model scaling for convolutional neural networks [EB/OL].(2020-09-11) [2021-06-01].