基于卷积神经网络的混凝土桥梁表观病害识别模型

Concrete Bridge Surface Disease Identification Model Based on Convolutional Neural Network

为响应交通强国公路数字化的转型号召,针对公路桥梁数字化技术研发与应用需求,从三维实景基础数据采集和病害识别分析模型方面研制量化快速分析评估软件。针对病害识别,基于卷积神经网络模型构建桥梁表观病害智能识别模型架构,并对桥梁表面典型病害数据集进行多种病害识别模型构建。为贴合实际应用环境,选取裂缝、脱落、渗水病害,提出了一种低样本训练图像识别算法。3种病害各选取100张图像以8∶2的比例进行训练和验证。最终以VGG16V2为基础构建的低样本优化训练模型的病害识别率相比原始VGG16V2算法模型提升了3.3%,针对3种病害的识别准确率为93.3%。为低样本下的病害识别模型训练提出的优化架构,有助于在低样本的情况下实现全自动化桥梁表观病害分类技术的研究,同时可以自动化的为目标检测等高度数据依赖的任务提供高质量的样本数据,节省人工,提高效率。

In response to the call for highway digital transformation of transportation power,and to the study and application needs of digital technology for highway bridges,the quantitative and rapid analysis and evaluation software were planned to develop based on the three-dimensional real scene basic data collection models and disease identification analysis models.The intelligent identification model architecture for bridge surface diseases based on CNN model was built.The constructing multiple disease identification models were constructed based on the typical bridge surface diseases dataset.To fit the practical application environment,the training image recognition algorithm with samples was proposed by selecting diseases included cracks,detachment,and water seepage.100 images were selected for each of the three diseases to train and validate in the ratio of 8∶2.The disease recognition rate of few samples optimized training model based on VGG16V2 was improved by 3.3%,compared to the original VGG16V2 algorithm model.The recognition accuracy for three types of diseases was 93.3%.The optimized architecture proposed for disease recognition model training with samples can help to achieve fully automated study on bridge surface disease classification technology with samples.Simultaneously,the high-quality sample data can be automatically provided for highly data dependent tasks such as target detection,saving manpower and improving efficiency.