基于残差网络的混凝土结构病害分类识别研究

针对混凝土结构病害识别类型单一、精度较低的现状,提出了基于残差网络和迁移学习的病害分类识别方法,通过构建多属性病害数据集,利用迁移学习优化残差网络模型,提出混凝土结构健康状态识别的多个任务。首先收集混凝土结构的病害状态图像,依次通过数据清洗、尺寸均一化、数据扩增和多人投票标注,最终得到包含6 680张图像的混凝土结构病害多属性数据集,并依据不同标注属性进行了相应训练集、验证集和测试集的划分; 然后利用迁移学习对预训练的ResNet-34网络前3个部分进行参数冻结,后续2个部分的参数进行重新训练,并在模型末端添加新的参数,基于已构建的数据集进行训练; 最后在提出的构件类别检测、剥落检测、病害检测和病害类别检测任务中,分别获得84.88%、98.56%、97.18%和85.34%的F1分数。结果表明:通过构建多属性标注的混凝土结构病害数据集训练深度学习模型,可较好地实现多场景特征下的病害识别效果; 采用迁移学习技术可从开源数据中获取较好的特征提取效果; 改进的ResNet-34网络可克服网络退化问题,并针对混凝土结构病害识别的多个任务获得较好的效果; 相对于单一的混凝土结构病害识别,进行病害部位、程度、多类别的系统性检测,可为结构状态评估提供详细信息,更贴合工程实际需要。

Research on Classification and Recognition of Concrete Structure Diseases Based on Residual Network

In view of the single type and low accuracy of concrete structure disease recognition, a disease classification and recognition method based on residual network and transfer learning was proposed. By establishing a multiattribute disease dataset and using transfer learning to improve the residual network, the multiple tasks for identifying the health status of concrete structures were proposed. First of all, the original images of disease state of concrete structure were collected and sequentially underwent data cleaning, size normalization, data augmentation and multi-person voting to label. Then, a multiattribute dataset of concrete structure diseases containing 6 680 images was obtained, and corresponding training set, validation set and testing set were divided according to different label attributes. Sequentially, the first three parts of parameters of the pre-trained ResNet-34 were frozen by transfer learning. The parameters of subsequent two parts were retrained, and new parameters were added at the end of the model and training was conducted based on the established dataset. Finally, the F1 score of 84.88%, 98.56%, 97.18% and 85.34% were obtained in the proposed tasks of component type detection, spalling detection, disease detection and disease type detection, respectively. The results show that by building a multiattribute labeled dataset of concrete structure diseases to train a deep learning model, the disease recognition under multi-scene conditions can be better achieved. Better feature extraction results can be obtained from open source data by using transfer learning.The improved ResNet-34 can overcome the problem of network degradation, and achieve better results for multiple tasks of concrete structure disease identification. Compared with the single disease identification of concrete structure, systematic detection of disease location, degree, and multiple categories can provide detailed information for structural state assessment, which is more suitable for the actual needs of the project.