基于改进U-net的沥青路面图像裂缝分割方法

张涛; 王金; 刘斌; 许牛琦

doi:10.3963/j.jssn.1674-4861.2023.06.010

基于改进U-net的沥青路面图像裂缝分割方法

doi: 10.3963/j.jssn.1674-4861.2023.06.010

张涛^1,,
王金^{1, 2, ,},
刘斌²,
许牛琦²

1.
北京工业大学交通工程北京市重点实验室北京 100124
2.
北京工业大学北京市城市交通运行保障工程技术研究中心北京 100124

基金项目:

北京市自然科学基金项目 8232005

北京市自然科学基金-丰台轨道交通前沿研究联合基金项目 L221026

详细信息

作者简介:
张涛（1998—），硕士研究生. 研究方向：路面检测与养护管理. E-mail: zhangtaowoo@emails.bjut.edu.cn

通讯作者:
王金（1984—），博士，副教授. 研究方向：时空数据下道路基础设施表观巡检及安全分析. E-mail: j.wang@bjut.edu.cn

中图分类号: U416.2
计量
- 文章访问数: 232
- HTML全文浏览量: 128
- PDF下载量: 32
- 被引次数: 0
出版历程
- 收稿日期: 2023-07-03
- 网络出版日期: 2024-04-03

Crack Segmentation of Asphalt Pavement Images Based on Improved U-net

ZHANG Tao^1
,,
WANG Jin^{1, 2
, ,},
LIU Bin²,
XU Niuqi²

1.
Beijing Key Laboratory of Traffic Engineering, Beijing University of Technology, Beijing 100124, China
2.
Beijing Engineering Research Center of Urban Transport Operation Guarantee, Beijing University of Technology, Beijing 100124, China

摘要

摘要: 为提高基于图像的沥青路面裂缝分割精度，基于U-net架构提出了strip-attention-u-net（SAU）网络。该网络采用ResNeSt50作为特征提取网络，能有效地捕捉图像中的语义信息和局部细节；在编解码跳跃连接阶段、解码器上采样阶段分别引入通道增强条形池化（channel enhanced strip pooling，CESP）模块、卷积块注意力（convolutional block attention，CBA）模块，该模块能有效减少通道压缩导致的特征丢失情况，更好地保留裂缝特征；结合Dice Loss和Focal Loss的损失函数可以使模型关注像素占比小、难以分割的细长裂缝。为测试SAU网络的性能，使用EdmCrack600公共数据集和BJCrack600实验数据集开展了模块消融实验，并与典型图像分割模型（FCN、PSPNet、DeepLabv3、U-net、Attention U-net和U-net++）进行了对比。结果表明：在EdmCrack600公共数据集上的对比实验中，SAU网络的裂缝分割效果更佳，裂缝交并比（intersection over union，IoU）和F1分数分别为50.89%和83.59%；在BJCrack600实验数据集上进行网络训练和对比实验中，表明SAU网络在沥青路面裂缝分割上的性能更优，裂缝IoU和F1分数分别为69.69%和90.90%，可为道路养护提供更为智能化、高效的决策支持。
- 道路工程 /
- 裂缝分割 /
- 编解码网络 /
- 深度学习 /
- 语义分割
Abstract: To improve the segmentation accuracy of image-based asphalt pavement cracks, this paper proposes a strip-attention-u-net (SAU) network based on U-net. The network uses ResNeSt50 as a core feature learning struc-ture to effectively capture semantic information and local details. A channel enhanced strip pooling (CESP) module in the encode-decode skip connection is investigated to enhance the ability of learning crack features and better utilize residual connections. A convolutional block attention (CBA) module in the up sampling stage of the decoder is developed to mitigate feature losses caused by channel compression and preserve crack features. A loss function comprised by a Dice Loss and a Focal Loss function is performed to attract thin and small crack features. A publicly available EdmCrack600 dataset and an experimental BJCrack600 dataset (600 asphalt pavement images collected in an experiment) are used to evaluate the performance of the SAU network. Ablation experiments are conducted and the SAU network is compared with state-of-the-art networks (FCN, PSPNet, DeepLabv3, U-net, Attention U-net, and U-net++). For EdmCrack600 dataset, the proposed SAU network outperforms the state-of-the-art networks, with intersection over union (IoU) and F1 score of 50.89% and 83.59%, respectively. Regarding the BJCrack600 dataset, the SAU network demonstrates the best performance among the state-of-the-art networks, achieving IoU and F1 score of 69.69% and 90.90%, respectively. The study findings could provide more intelligent and efficient supports in making advanced decisions of road maintenance.
- road engineering /
- crack segmentation /
- encode-decode network /
- deep learning /
- semantic segmentation

HTML全文

图 1 SAU网络结构示意图

Figure 1. The structure of SAU network

下载: 全尺寸图片幻灯片

图 2 CESP模块

Figure 2. CESP module

下载: 全尺寸图片幻灯片

图 3 CBA结构图

Figure 3. The structure of CBA

下载: 全尺寸图片幻灯片

图 4 不同网络在EdmCrack600数据集上的预测结果

Figure 4. Comparisons of segmentation results on EdmCrack600 dataset among several networks

下载: 全尺寸图片幻灯片

图 5 SAU网络损失函数、F1与MIoU变化曲线

Figure 5. Curve of loss function, F1 and MIoU of SAU

下载: 全尺寸图片幻灯片

图 6 不同网络在自制数据集BJCrack600的预测结果

Figure 6. Comparisons of segmentation results on BJCrack600 among different networks

下载: 全尺寸图片幻灯片

表 1 基于EdmCrack600数据集SAU网络不同损失函数下的精度对比实验

Table 1. comparison of F1 score under different loss functions for SAU network on EdmCrack600 dataset 单位: %

损失函数	MIoU	IoU	F1
loss_focal	62.00	24.72	69.64
loss_dice
loss_focal + loss_dice	75.17	50.89	83.59

下载: 导出CSV

表 2 基于EdmCrack600数据集的SAU网络消融实验

Table 2. An ablation analysis of SAU network on EdmCrac600 dataset 单位: %

模型	MIoU	IoU	F1
U-net*	72.97	46.52	81.60
U-net* + CESP	73.69	47.95	82.27
U-net* + CBA	74.22	49.02	82.75
SAU（U-net* + CESP + CBA）	75.17	50.89	83.59

下载: 导出CSV

表 3 基于BJCrack600数据集的SAU网络消融实验

Table 3. An ablation analysis of SAU network on BJCrack600 dataset 单位: %

模型	MIoU	IoU	F1
U-net*	82.21	65.14	89.26
U-net* + CESP	82.97	66.69	89.82
U-net* + CBA	84.33	69.32	90.78
SAU（U-net* + CESP + CBA）	84.52	69.69	90.90

下载: 导出CSV

表 4 典型图像分割模型在EdmCrack600数据集中的性能对比

Table 4. Comparisons of different image segmentation networks on EdmCrack600 dataset

模型	输入尺寸/（px×px）	IoU/%	MIoU/%	F1/%
FCN	512×512	32.33	65.83	69.93
PSPNet	512×512	27.35	63.33	67.83
U-net*	512×512	46.52	72.97	81.60
Attention U-net	512×512	49.38	74.40	82.91
U-net++	512×512	48.55	74.00	82.54
DeepLabv3	512×512	35.95	67.54	73.10
SAU	512×512	50.89	75.17	83.59

下载: 导出CSV

表 5 典型图像分割模型在BJCrack600数据集的性能对比

Table 5. Comparisons of different image segmentation models on BJCrack600 dataset

模型	输入尺寸/（px×px）	IoU/%	MIoU/%	F1/%
FCN	512×512	28.82	63.74	71.18
PSPNet	512×512	3.37	50.81	52.87
U-net*	512×512	66.86	83.09	89.90
Attention U-net	512×512	67.92	83.62	90.28
U-net++	512×512	67.40	83.35	90.09
DeepLabv3	512×512	39.95	69.37	75.93
SAU	512×512	69.69	84.52	90.90

下载: 导出CSV

参考文献(33)

[1]	沙爱民, 童峥, 高杰. 基于卷积神经网络的路表病害识别与测量[J]. 中国公路学报, 2018, 31(1): 1-10. https://www.cnki.com.cn/Article/CJFDTOTAL-ZGGL201801002.htm SHA A M, TONG Z, GAO J. Recognition and measurement of road surface diseases based on convolutional neural networks[J]. China Journal of Highway and Transport, 2018, 31 (1): 1-10. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-ZGGL201801002.htm
[2]	杭伯安. 基于改进k-means算法的沥青路面裂缝分割方法[J]. 公路交通科技, 2023, 40(4): 1-8. https://www.cnki.com.cn/Article/CJFDTOTAL-GLJK202304001.htm HANG B A. Asphalt pavement crack segmentation method based on improved K-means algorithm[J]. Journal of Highway and Transportation Research and Development, 2023, 40 (4): 1-8. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-GLJK202304001.htm
[3]	沈照庆, 彭余华, 舒宁. 一种基于SVM的路面影像损伤跨尺度识别方法[J]. 武汉大学学报(信息科学版), 2013, 38 (8): 993-997. https://www.cnki.com.cn/Article/CJFDTOTAL-WHCH201308025.htm SHEN Z Q, PENG Y H, SHU N. A SVM-based cross-scale damage recognition method for pavement image[J]. Geomatics and Information Science of Wuhan University, 2013, 38 (8): 993-997. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-WHCH201308025.htm
[4]	张娟, 沙爱民, 孙朝云, 等. 基于相位编组法的路面裂缝自动识别[J]. 中国公路学报, 2008(2): 39-42. https://www.cnki.com.cn/Article/CJFDTOTAL-ZGGL200802007.htm ZHANG J, SHA A M, SUN C Y, et al. Automatic pavement crack recognition based on phase grouping method[J]. China Journal of Highway and Transport, 2008(2): 39-42. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-ZGGL200802007.htm
[5]	OLIVEIRA H, CORREIA P L. CrackIT: An image processing toolbox for crack detection and characterization[C]. 2014 IEEE International Conference on Image Processing(ICIP), Paris, France: IEEE, 2014.
[6]	ZHANG A, WANG K C, LI B, et al. Automated pixel-level pavement crack detection on 3D asphalt surfaces using a deep-learning network[J]. Computer-Aided Civil and Infrastructure Engineering, 2017, 32(10): 805-819. doi: 10.1111/mice.12297
[7]	ZHANG A, WANG K C, FEI Y, et al. Deep Learning-based fully automated pavement crack detection on 3D asphalt surfaces with an improved CrackNet[J]. Journal of Computing in Civil Engineering, 2018, 32(5): 04018041. doi: 10.1061/(ASCE)CP.1943-5487.0000775
[8]	FEI Y, WANG K C, ZHANG A, et al. Pixel-level cracking detection on 3D asphalt pavement images through deep-learning-based CrackNet-Ⅴ[J]. IEEE Transactions on Intelligent Transportation Systems, 2020, 21(1): 273-284. doi: 10.1109/TITS.2019.2891167
[9]	LONG J, SHELHAMER E, DARRELL T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(4): 640-651. doi: 10.1109/TPAMI.2016.2572683
[10]	CHEN L C, PAPANDREOU G, KOKKINOS I, et al. Semantic image segmentation with deep convolutional nets and fully connected CRFs[OL]. (2016-06-07)[2022-12-11]. http://arxiv.org/abs/1412.7062.
[11]	CHEN L C, PAPANDREOU G, KOKKINOS I, et al. DeepLab: semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 40(4): 834-848
[12]	CHEN L C, PAPANDREOU G, SCHROFF F, et al. Rethinking atrous convolution for semantic image segmentation[OL]. (2017-12-05)[2022-12-11]. http://arxiv.org/abs/1706.05587.
[13]	CHEN L C, ZHU Y, PAPANDREOU G, et al. Encoder-Decoder with atrous separable convolution for semantic image segmentation[OL]. (2018-08-22)[2022-12-11]. http://arxiv.org/abs/1802.02611.
[14]	RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation[OL]. (2015-05-18)[2022-12-11]. http://arxiv.org/abs/1505.04597.
[15]	OKTAY O, SCHLEMPER J, FOLGOC LL, et al. Attention u-net: Learning where to look for the pancreas[OL]. (2018-05-20)[2023-08-27]. https://arxiv.org/abs/1804.03999.
[16]	ZHOU Z, RAHMAN SIDDIQUEE M M, TAJBAKHSH N, et al. Unet++: A nested u-net architecture for medical image segmentation[C]. Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4^th International Workshop, Granada, Spain: Springer, 2018.
[17]	YANG X, LI H, YU Y, et al. Automatic pixel-level crack detection and measurement using fully convolutional network[J]. Computer-Aided Civil and Infrastructure Engineering, 2018, 33(12): 1090-1109. doi: 10.1111/mice.12412
[18]	张志华, 邓砚学, 张新秀. 基于改进SegNet的沥青路面病害提取与分类方法[J]. 交通信息与安全, 2022, 40(3): 127-135. doi: 10.3963/j.jssn.1674-4861.2022.03.013 ZHANG Z H, DENG Y X, ZHANG X X. Asphalt pavement disease extraction and classification method based on improved SegNet[J]. Journal of Transport Information and Safety, 2022, 40(3): 127-135. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2022.03.013
[19]	史梦圆, 高俊钗. 改进U-Net算法的路面裂缝检测研究[J]. 自动化与仪表, 2022, 37(10): 52-55, 67. https://www.cnki.com.cn/Article/CJFDTOTAL-ZDHY202210011.htm SHI M Y, GAO J C. Research on pavement crack detection based on improved U-Net algorithm[J]. Automation & Instrumentation, 2022, 37(10): 52-55, 67. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-ZDHY202210011.htm
[20]	GAO X, TONG B. MRA-UNet: balancing speed and accuracy in road crack segmentation network[J]. Signal, Image and Video Processing, 2023, 17(5): 2093-2100. doi: 10.1007/s11760-022-02423-9
[21]	惠冰, 李远见. 基于改进U型神经网络的路面裂缝检测方法[J]. 交通信息与安全, 2023, 41(1): 105-114, 131. doi: 10.3963/j.jssn.1674-4861.2023.01.011 HUI B, LI Y J. Pavement crack detection method based on improved U-type neural network[J]. Journal of Transport Information and Safety, 2023, 41(1): 105-114, 131. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2023.01.011
[22]	LIU C, ZHU C, XIA X, et al. FFEDN: feature fusion encoder decoder network for crack detection[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23 (9) : 15546-15557. doi: 10.1109/TITS.2022.3141827
[23]	MEI Q, GÜL M. A cost effective solution for pavement crack inspection using cameras and deep neural networks[J]. Construction and Building Materials, 2020, 256: 119397. doi: 10.1016/j.conbuildmat.2020.119397
[24]	ZHONG J, ZHU J, HUYAN J, et al. Multi-scale feature fusion network for pixel-level pavement distress detection[J]. Automation in Construction, 2022, 141: 104436. doi: 10.1016/j.autcon.2022.104436
[25]	唐港庭, 尹超, 王绍平, 等. 基于改进GoogLeNet的沥青路面裂缝识别算法[J]. 智能计算机与应用, 2023, 13(3): 202-206. https://www.cnki.com.cn/Article/CJFDTOTAL-DLXZ202303033.htm TANG G T, YIN C, WANG S P, et al. Asphalt pavement crack recognition algorithm based on improved GoogLeNet[J]. Intelligent Computer and Applications, 2023, 13(3): 202-206. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-DLXZ202303033.htm
[26]	HOU Q, ZHANG L, CHENG M M, et al. Strip pooling: rethinking spatial pooling for scene parsing[OL]. (2020-03-20)[2023-02-23]. http://arxiv.org/abs/2003.13328.
[27]	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[OL]. (2018-05-16)[2023-03-26]. https://openaccess.thecvf.com/content_cvpr_2018/html/Hu_Squeeze-and-Excitation_Networks_CVPR_2018_paper.html.
[28]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]. The IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, America: IEEE, 2016.
[29]	WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]. The European Conference on Computer Vision, Munich, Germany: Springer, 2018.
[30]	LI X, SUN X, MENG Y, et al. Dice loss for data-imbalanced nlp tasks[OL]. (2020-08-29)[2023-06-21]. https://arxiv.org/abs/1911.02855v3.
[31]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[OL]. (2018-02-07)[2023-06-28]. http://arxiv.org/abs/1708.02002.
[32]	RUSSELL B C, TORRALBA A, MURPHY K P, et al. LabelMe: a database and web-based tool for image annotation[J]. International Journal of Computer Vision, 2008, 77 (1): 157-173.
[33]	WU Y, CAO W, SAHIN S, et al. Experimental characterizations and analysis of deep learning frameworks[C]. 2018 IEEE International Conference on Big Data, Seattle, USA: IEEE, 2018. International Workshop, Granada, Spain: Springer, 2018.