摘要： 道路裂缝是道路安全检测的重要部分，随着深度学习和计算机视觉的发展，利用深度学习对道路图像中裂缝信息提取的方法趋于成熟。现有深度学习道路裂缝检测方法对细小裂缝提取不完整以及受背景因素干扰，导致检测精度降低。基于CBAM注意力机制和残差网络，改进U-Net神经网络模型，构建一种融合残差和注意力机制的道路裂缝检测深度学习网络模型。该模型在U-Net网络的上采样和下采样过程中分别嵌入通道注意力机制和空间注意力机制。CBAM注意力机制在通道和空间维度上同时进行全局平均和全局最大混合池化，以提取更多有效的全局和局部细节信息。同时，在U-Net网络中融合残差模块，有效解决网络梯度消失、梯度爆炸以及网络退化的问题，进一步提高道路裂缝的检测能力。实验结果表明，在上采样和下采样过程中嵌入CBAM注意力机制网络的F1值提升到81.02%，相比U-Net原始网络，提升13.76个百分点。融合残差模块并在下采样过程中嵌入CBAM注意力机制网络的F1值达到85.82%，相比只嵌入CBAM注意力机制的网络，提升了4.8个百分点。

关键词: 裂缝检测, 深度学习, U-Net神经网络, 注意力机制, 残差结构

Abstract: Road cracks are an important part of road safety detection，and with the development of deep learning and computer vision，methods for extracting crack information in road images using deep learning methods are maturing.Existing deep learning road crack detection methods cannot extract small cracks and are affected by background factors，resulting in a decrease in detection accuracy.Based on the Convolutional Block Attention Module（CBAM） attention mechanism and the residual network，a deep learning network model for road crack detection incorporating the residual and attention mechanisms is established by improving the U-Net neural network model.The model embeds the channel attention mechanism and spatial attention mechanism in the up-sampling and down-sampling processes of the U-Net network，respectively.The CBAM attention mechanism performs both global average and global maximum mixed pooling on both channel and spatial dimensions，producing more effective global and local detail information.Meanwhile，integrating residual modules in the U-Net network effectively solves the problems of network gradient disappearance，gradient explosion，and network degradation，further improving the detection ability of road cracks.The experimental results show that compared with the U-Net original network，the F1 value of the U-Net network embedded with CBAM attention mechanism in the up-sampling and down-sampling processes to 81.02%，an increase of 13.76 percentage points.Further，compared with the network that only embeds the CBAM attention mechanism，the F1 value of the network that integrates residual modules and embeds the CBAM attention mechanism in the down-sampling processes reaches 85.82%，an increase of 4.8 percentage points.

Key words: crack detection, deep learning, U-Net neural network, attention mechanism, residual structure

中图分类号: 

• P237