双层卷积和多特征融合的路面裂缝分割方法

Pavement Crack Segmentation Method Based on Bilevel Convolution and Multi-feature Fusion

针对复杂背景下道路裂缝分割方法存在的边缘细节易缺失、目标误判问题,提出双层卷积和多特征融合的路面裂缝分割网络。首先,采用U-Net网络为基础架构,设计双层卷积网络改进编码部分,增大感受野,提取丰富的上下文信息.其次,引入坐标注意力模块优化解码部分,进一步加强网络对裂缝边缘细节的学习.最后,将产生的多级特征反馈至特征融合模块,堆叠通道有效融合深层和浅层特征。此外,损失函数采用二值交叉熵损失和Dice损失函数相结合的方式,有效解决了背景大于裂缝像素点导致的样本不均衡问题。通过在CRACK500、CFD、Cracktree200公开数据集上的实验结果表明:该方法实现MIoU分别为76.8%、65.3%、53.0%,比现有方法MIoU平均提高3.5%,可以实现优异的道路裂缝自动分割效果.

Aiming at the problems of missing edge details and misjudgment of target in road crack segmentation method under complex background,this article proposed a pavement crack segmentation network based on bilevel convolution and multi-feature fusion.Firstly,the U-Net network was used as the infrastructure to design a bilevel convolutional network to improve the coding part,increase the receptive field,and extract rich context information.Secondly,the coordinate attention module was introduced to optimize the decoding part to further enhance the network′s learning of crack edge details.Finally,the generated multilevel features were fed back to the feature fusion module,and the deep and shallow features were effectively fused by stacking channels.In addition,the loss function combined binary cross-entropy loss and Dice loss function,which effectively solved the sample imbalance problem caused by a background larger than crack pixels.The experimental results on CRACK500,CFD,and Cracktree200 show that the MIoU achieved by this method is 76.8%,65.3%,and 53%,respectively,which is 3.5%higher than the result by the existing methods,and it can achieve excellent automatic segmentation effect on road cracks.