基于多尺度空洞卷积结构的路面裂缝分割方法  

A Method for Pavement Crack Segmentation Based on Multi-scale Cavity Convolution Structure

作　　者：何宇超 段中兴[1] 高静 HE Yu-chao;DUAN Zhong-xing;GAO Jing(College of Information and Control Engineering,Xi’an University of Architecture and Technology,Xi’an Shaanxi 710055,China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,陕西西安710055

出　　处：《公路交通科技》2024年第1期1-9,17,共10页Journal of Highway and Transportation Research and Development

基　　金：国家自然科学基金项目(51678470)。

摘　　要：为了实现道路裂缝的自动化检测,改善现有裂缝分割模型存在分割不连续与嘈杂背景误分割等问题,提出了一种基于多尺度空洞卷积结构的裂缝分割模型MAC-UNet。以UNet作为基础网络,首先提出了多尺度空洞卷积结构,替换编码器与解码器中的双卷积结构,提升了网络对复杂拓扑结构的分割性能。然后,构建了交叉注意力机制,使用金字塔注意力模块代替编、解码器之间的跳跃连接,保留因池化丢失的空间特征。增加通道注意力引导多尺度信息,有效地融合到解码器特征中,使得恢复裂缝时,细节更加丰富,定位更准确。最后,在道路裂缝数据集CFD和GAPS384上与FCN、PSPNet等5种方法进行试验对比,相较于UNet,在CFD数据集上,MIOU和Kappa系数分别提升了8.4%和8.52%。在GAPS384数据集上,分别提升了6.84%和8.23%,对于道路裂缝的分割更加清晰与完整。结果表明:与主流的分割算法相比,所提出算法的识别精度方面具有较明显的优势,在光照不均匀、各种噪音干扰、背景灰度水平不同的情况下,所提模型仍然能够获取稳定的检测结果,能够应对复杂裂缝分割问题,并且可视化裂缝检测误差较小,符合实际工程需求,且模型体积较小,具有一定的工程应用价值。In order to realize the automatic detection of road cracks and improve the problems of segmentation discontinuity and false segmentation of noisy background in the existing crack segmentation model,a crack segmentation model MAC-UNet based on multi-scale cavity convolution structure is proposed.Taking UNet as a basic network,the multi-scale cavity convolution structure is firstly proposed to replace the double convolution structure in the encoder and decoder,that improves the segmentation performance of the network for complex topology structure.Then,the cross-attention mechanism is constructed,and the pyramidal attention module is used to replace the jump connection between the encoder and decoder.The spatial characteristics lost due to pooling are retained,and the channel attention is increased to guide the multi-scale information to be effectively integrated into the decoder characteristics,so that when the crack is restored,the details are more abundant,and the positioning is more accurate.Finally,the experimental comparison with FCN,PSPNet and other 5 methods is carried out on the road crack data set CFD and GAPS384.Compared with FCN,PSPNet and other 5 methods on the road crack dataset CFD and GAPS384,MIOU and Kappa coefficient increase by 8.4%and 8.52%respectively on the CFD dataset compared with UNet.On the GAPS384 dataset,it is improved by 6.84%and 8.23%respectively,and the segmentation of road cracks is clearer and more complete.The result shows that compared with the mainstream segmentation algorithm,the proposed algorithm has obvious advantages in recognition accuracy.Under the conditions of uneven illumination,various noise interference and different background gray levels,the proposed model can still obtain stable detection results and can deal with complex crack segmentation problems.The visual crack detection error is small,which meets the actual engineering requirements,and the model volume is small,which has certain engineering application value.

关 键 词：道路工程 裂缝识别 深度学习 道路裂缝 空洞卷积 多尺度特征 

分 类 号：U418.6[交通运输工程—道路与铁道工程] TP391.41[自动化与计算机技术—计算机应用技术]