基于分层特征残差神经网络的桥梁水下结构裂缝分割

Crack Segmentation of Underwater Structures of Bridges Based on Hierarchical Feature Residual Neural Network

为提高桥梁水下结构裂缝检测任务的自动化水平,提出了一种基于分层残差神经网络的裂缝检测方法。该方法利用多层次特征残差连接机构,抑制了建筑表面噪声特征的干扰,提取并融合了不同层次的特征图像,增强了模型对裂缝区域和非裂缝区域的精确划分能力。同时借助迁移学习方法,利用预训练模型的参数初始化模型并借助水下裂缝数据集调整权重,使模型具备分析数据量极少的桥梁水下结构裂缝数据集的能力。该模型在自采集的桥梁水下结构裂缝数据集上进行了试验验证。结果表明:分层残差神经网络具备精确划分裂缝像素与非裂缝像素的能力,预测像素准确率达到87.2%,证明了该方法的可行性。该模型为桥梁水下结构裂缝检测任务的自动化提供了一种有效的解决方案,同时也为其他类似的图像检测任务提供了一种参考思路。

A crack detection method based on hierarchical residual neural network is proposed to improve the automation of the crack detection task for underwater structures of bridges.The method utilizes a multi-level feature residual linkage mechanism,which suppresses the interference of noise features on the building surface,extracts and fuses feature images at different levels,and enhances the model's capacity to accurately delineate cracked and non-cracked regions.Meanwhile,with the help of transfer learning method,the model is initialized with the parameters of the pre-trained model and the weights are adjusted with the underwater crack dataset,so that the model has the capacity to analyze the bridge underwater structure crack dataset with very small amount of data.The model was experimentally validated on a self-collected bridge underwater structural crack dataset.The results showed that the hierarchical residual neural network could accurately classify cracked pixels from non-cracked pixels,and the predicted pixel accuracy reached 87.2%,which proved the feasibility of the method.The model provides an effective solution for automating the bridge underwater structural crack detection task,and also provides a reference idea for other similar image detection tasks.