计算机与现代化 ›› 2022, Vol. 0 ›› Issue (07): 33-39.

• 图像处理 • 上一篇    下一篇

ERCUnet:一种基于U-Net改进的道路裂缝检测模型

  

  1. (1.郑州大学软件学院,河南郑州450001;2.中国科学院空天信息创新研究院,北京100101)
  • 出版日期:2022-07-25 发布日期:2022-07-25
  • 作者简介:刘宇翔(2000—),男,河南焦作人,本科生,研究方向:人工智能,大数据分析,图像识别,E-mail: lyx2019@stu.zzu.edu.cn; 佘维(1977—),男,湖南常德人,教授,博士,研究方向:人工智能,信息安全,能源区块链,E-mail: wshe@zzu.edu.cn; 通信作者:沈占峰(1977—),男,黑龙江大庆人,研究员,博士,研究方向:高分辨率遥感影像信息提取,深度学习,E-mail: Shenzf@radi.ac.cn; 谭帅(2003—),男,河南信阳人,本科生,研究方向:物联网,人工智能,E-mail: 2826598617@qq.com。
  • 基金资助:
    郑州大学大学生创新创业训练计划资助项目(202110459167); 国家自然科学基金资助项目(41971375); 国家重点研发计划项目(2018YFB0505000)

ERCUnet: An Improved Road Crack Detection Model Based on U-Net

  1. (1. School of Software Technology, Zhengzhou University, Zhengzhou 450001, China;
    2. Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100101, China)
  • Online:2022-07-25 Published:2022-07-25

摘要: 针对传统的道路裂缝检测方法存在灵活度不高、普适性不强等问题,本文参考ResNet中的残差设计和U-Net模型的U形编码解码结构,设计一种基于U-Net改进的道路裂缝检测模型——ERCUnet。该模型以残差块为主体,针对裂缝检测优化不同深度卷积层的卷积核数量,模型中所有的残差块结构相同,模型整体结构更加整齐和简单,具有可塑性好、结构性强,残差结构不仅使特征融合更加充分,也避免了深层卷积神经网络梯度消失的问题。实验在CrackForest数据集上进行,将CrackForest的118张含标注图片按照5〖DK〗∶1的比例划分训练集和测试集,通过一系列数据增广方法,有效缓解了训练数据过少的问题。损失函数融合了交叉熵和F1分数,缓解了正负样本不均衡的问题,最终的实验结果显示ERCUnet模型参数量仅为U-Net(BN)模型的13.30%,在测试集上的查全率、查准率、F1值均达70%以上,噪声率、准确率分别为29.05%、99.01%。为证实ERCUnet的可塑性,通过修改模型参数得到ERCUnet-tiny模型,其参数量仅为U-Net(BN)模型的2.39%,在测试集上取得了与U-Net(BN)相近的效果。

关键词: 道路裂缝检测, U-Net, 残差结构, 数据增广

Abstract: Aiming at the problems of traditional road  crack  detection methods,  such as low flexibility and poor universality, refering to the residual design in ResNet and the U-shaped encoding and decoding structure of U-Net model, an improved road crack detection model based on U-Net, named ERCUnet,  is designed. The model takes residual blocks as the main body, and optimizes the number of convolution cores of convolution layers at different depths for crack detection. All residual blocks in the model have the same structure, the overall structure of the model is more neat and simple, with good elasticity and strong structure. The residual structure not only makes the feature fusion more sufficient but also avoids the problem of gradient disappearance of deep convolution neural network. The experiment is conducted on the CrackForest dataset. The 118 labeled pictures of CrackForest are divided into training set and testing set according to the ratio of 5〖DK(〗∶〖DK)〗1. Through a series of data expansion methods, the problem of too little training data is effectively alleviated. The loss function combines cross entropy and F1 score to alleviate the imbalance between positive and negative samples. The final experimental results show that the number of parameters of ERCUnet model is only 13.30% of that of U-Net model, the recall, precision, and F1 are all greater than 70%, and noise rate and accuracy are 29.05% and 99.01% on testing set. ERCUnet-tiny model is obtained by modifying model parameters to confirm the plasticity of ERCUnet, and the number of its parameters is only 2.39% of that of U-Net model, similar effect to U-Net is achieved  on testing set.

Key words: road crack detection, U-Net, residual structure, data expansion