基于U2net神经网络的混凝土表面裂缝检测研究

Research on the detection of concrete surface cracks based on U2net neural network

本研究提出了一种基于U2net神经网络的改进网络AM-U2net,以实现对混凝土结构裂缝更为准确和快速地检测。首先,在原U2net网络的跳跃连接部分施加注意力机制,在网络结构的上层部分添加CBAM注意力模块,下层部分添加ECA注意力模块。然后,引入Mish激活函数,提出一种AM-U2net网络。结果显示:改进网络AM-U2net的性能要优于对比网络Unet、Unet++和原U2net,准确率、精确率、召回率和f1score分别达到99.1%、64.2%、65.4%和64.8%。最后,采用骨架线的方法对模型输出的混凝土裂缝预测图像,在像素级别上标注计算裂缝的长宽属性,使得模型预测裂缝的结果能够反映出混凝土裂缝的一些基本属性。本研究提出的改进算法在混凝土裂缝识别任务中具有较好的表现,为裂缝检测的后续研究提供思路。

An improved network AM-U2net based on U2net neural network is proposed to achieve more accurate and rapid detection of concrete cracks.Firstly,the attention mechanism is applied to the skip connection of the original U2net network.CBAM attention module and ECA attention module are added to the upper and lower layers of the network structure respectively.Then,an AM-U2net network is proposed after introducing Mish activation function.The results show that the performance of AM-U2net is superior to the comparison networks Unet,Unet++and the original U2net,with accuracy,precision,recall and f1score reaching 99.1%,64.2%,65.4%and 64.8%respectively.The concrete crack prediction image output by the model is based on the skeleton line method.The length and width attributes of the cracks are marked and calculated at the pixel level,so that the model′s predicted crack results can reflect some basic properties of concrete cracks.The improved algorithm proposed in this study can effectively identify concrete cracks,providing ideas for subsequent research.