基于融合多尺度多通道特征的深度监督网络实现裂缝检测  

Crack Detection Using Deep Supervised Networks with Multi-scale and Multi-channel Feature Fusion

作　　者：朱俊彬 杜斌 许世敏 ZHU Jun-bin;DU Bin;XU Shi-min(College of Civil Engineering,Guizhou University,Guiyang 550025,China;Guiyang Highway Administration Bureau,Guiyang 550002,China)

机构地区：[1]贵州大学土木工程学院,贵阳550025 [2]贵阳公路管理局,贵阳550002

出　　处：《科学技术与工程》2024年第13期5595-5603,共9页Science Technology and Engineering

基　　金：贵州贵阳国家高新区科技计划(GXCX-2018-016);贵州省科技计划项目(黔科合基础-ZK[2021]一般290)。

摘　　要：针对卷积神经网络在提取宽裂缝时,浅层卷积层在提取一些细节信息的同时也提取到了大量噪声的问题。提出了一个基于VGG-16骨架并融合深层特征的全卷积神经网络(fully convolutional network,FCN)分割网络,并在每层加入侧边输出以直接监督模型学习到更多的有用信息。此外,还采用了一种Focal Loss损失函数来解决数据集本身正负样本分类不平衡的问题。这种多尺度多通道深层特征与独特的损失函数融合应用,使网络具备很强的抗干扰性和较快的收敛速度。在DeepCrack数据集上,所提出的深层特征融合网络(deep feature fusion network,DFFN)与整体嵌套边缘检测(holistically-nested edge detection,HED)、FCN和DeepCrack相比,表现出更好的性能和更快的推理速度。A fully convolutional network(FCN)segmentation network based on VGG-16 skeleton and fused with deep features was proposed to address the problem of shallow convolutional layers extracting a large amount of noise while extracting some detailed information when using convolutional neural networks to extract wide cracks.Side outputs are added to each layer to directly supervise the model's learning of more useful information.Furthermore,a Focal Loss function was adopted to address the issue of imbalanced classification of positive and negative samples in the dataset itself.This fusion application of multi-scale and multi-channel deep features with unique loss functions enables the network to have strong anti-interference ability and fast convergence speed.On the DeepCrack dataset,the proposed deep feature fusion network(DFFN)exhibits better performance and faster inference speed compared to holistically nested edge detection(HED),FCN and DeepCrack.

关 键 词：卷积神经网络 裂缝分割 侧边监督 样本平衡 

分 类 号：U495[交通运输工程—交通运输规划与管理]