基于多尺度特征融合网络的路面裂缝分割方法研究

Research on road crack segmentation method based on multi-scale feature fusion network

针对路面缺陷检测设备采集到的路面影像中裂缝像素数占比较小,而现有深度学习的语义分割网络难以分割图像中小目标的问题,本文提出了一种融合了多尺度特征信息的编解码结构语义分割网络模型。首先,在经典编解码结构分割网络U-Net(u-shaped network)的基础上,将编码结构中不同层级的特征图分别与解码结构中的特征图越级融合;其次,在网络的误差反向传播过程中选择适用于小目标物体分割的损失函数Focal Loss;最后,在CrackForest数据集上对本文方法和常用四种模型进行了对比分析。结果表明,改进的U-Net模型在交并比、F1分数两种综合评价指标上的表现均要优于其他模型,且在细小裂缝的分割结果上提升更为明显。这说明本研究能够为路面影像中裂缝分割任务提供技术参考。

A method to solve the low segmentation accuracy in identifying small cracks in pavement images is presented.First,the structure and composition of the classic encoder-decoder segmentation network,U-Net,and analyze the challenges this model faces when it comes to segmentatiing small objects.From a structural perspective,the decoding structure of the network lacks low-level features,such as boundary information,which results in lower segmentation accuracy when dealing with objects like cracks.Secondly,based on U-Net,a classical codec structure segmentation network,we merge feature maps from different levels in the encoding structure with those in the decoding structure.This approach allows us to recover multi-scale crack information within the decoding structure,reducing the loss of crack information during the training process.Then,we employ a specialized loss function called Focal Loss,which is suitable for small target segmentation,during the error backpropagation stage to further improve the segmentation accuracy of cracks.Then,the segmentation performance of the network is compared with that of FCN-8s,U-Net,SegNet,and CrackU-net using CrackForest dataset,a widely used dataset for testing road crack segmentation.Finally,we compare the performance of our method when removing the multi-scale feature fusion layer,removing Focal Loss,and removing both.The results show that our method exhibits higher segmentation accuracy compared to the traditional four methods and is more suitable for crack segmentation in road images.The proposed multi-scale feature fusion layer and Focal Loss both improve the segmentation performance.In summary,our method effectively improve the segmentation accuracy of small and medium targets in images,with significant improvements in the task of road crack segmentation.As a result,it holds practical value in road crack segmentation tasks.