一种基于桥梁横向裂缝的病害识别方法

桥梁安全是交通安全中最重要的部分之一，桥梁定检是对桥梁安全最有效的评价方法，而在桥梁定检中对病害的识别尤其重要。据相关统计，我国混凝土桥梁的病害中有80%以上是由裂缝造成的，因此对桥梁的健康检测主要是对桥梁表观裂缝的检测与测量。为了提高检测的准确性和效率，统计桥梁裂缝病害图片的识别精度与收敛性，本文结合卷积神经网络（Convolutional Neural Networks, CNN）技术，利用TensorFlow提供的双线性插值法预处理裂缝图像。将裂缝图像预处理成分辨率为75 px、150 px和300 px；并在TensorFlow架构中构建CNN的训练模型，用该模型分别训练分辨率为75 px、150 px和300 px的横向裂缝病害图片，分析在不同分辨率情况下横向裂缝病害的训练精度、收敛性和效率。将3种不同分辨率的横向裂缝病害图片训练结果模型保存下来作为横向裂缝病害识别的测试模型，并将这些模型分别用于测试实际工程应用中的横向裂缝图片和测试集中的横向裂缝图片，分析在经过预处理后的测试集图片与实际环境中拍摄的横向裂缝图片的识别精度、收敛性和效率。实验结果表明,与刘洪公等人所用的识别算法相比，本文算法的训练精度、可扩展性和实时性更好，识别率高达99%及以上，为其他裂缝病害的识别奠定了良好的基础，也能够为桥梁裂缝病害的检测提供参考数据。

A Method to Identify Diseases of Bridges Based on Transverse Cracks

Bridge safety is one of the most important parts of traffic safety. Bridge inspection is the most effective evaluation method for bridge safety, and the identification of diseases is especially important in bridge inspection. According to relevant statistics, more than 80% of the diseases of concrete bridges in China are caused by cracks. Therefore, the health inspection of bridges is mainly the detection and measurement of the apparent cracks of bridges. In order to improve the accuracy and efficiency of detection, and to count the accuracy and convergence of bridge crack disease images, this paper combines Convolutional Neural Networks (CNN) technology and uses bilinear interpolation provided by TensorFlow to preprocess crack images. We preprocess crack images into resolutions of 75 px, 150 px and 300 px; and build a CNN training model in the TensorFlow architecture, use this model to train horizontal crack disease pictures with resolutions of 75 px, 150 px and 300 px respectively, and analyze at different resolutions under the circumstances, the training accuracy, convergence and efficiency of the lateral crack disease. The three different resolutions of the transverse crack disease picture training result model are saved as the test model of the lateral crack disease identification, and these models are used to test the lateral crack picture in the actual engineering application and the lateral crack picture in the test set for analyzing the recognition accuracy, convergence and efficiency of the test set pictures after preprocessing and the transverse crack pictures taken in the actual environment. Experimental results show that compared with the recognition algorithm used by Liu Hong-gong et al, the training accuracy, scalability and real-time performance of the algorithm in this paper are better, and the recognition rate is as high as 99% and above, which lays a good foundation for the recognition of other crack diseases. The foundation can also provide reference data for the detection of bridge cracks.