基于轻量级Transformer的隧道裂缝分割  

Tunnel crack segmentation based on lightweight Transformer

作　　者：邝先验 徐姚明 雷卉 程福军 桓湘澜 KUANG Xianyan;XU Yaoming;LEI Hui;CHENG Fujun;HUAN Xianglan(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000

出　　处：《铁道科学与工程学报》2024年第8期3421-3433,共13页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(51268017,72061016)。

摘　　要：裂缝检测对保证隧道结构安全至关重要,及时发现隧道裂缝缺陷,有利于降低工程维修成本和保障行车安全。然而,传统卷积神经网络在隧道裂缝检测任务中主要侧重提高检测精度和算法复杂度,如何平衡裂缝检测的精度和实时性是当前研究的一个难点。针对这一问题,本文提出一种基于轻量级Transformer的裂缝分割方法 CrackViT。首先,采用卷积神经网络与Transformer混合的MobileViT网络构建裂缝特征提取网络,减少网络模型参数和计算量,并且有效提取裂缝图像全局信息和局部特征信息。然后,提出改进空洞空间金字塔池化解码器实现不同尺度的特征提取和信息融合,实现像素级概率分布。同时,裂缝图像存在细节信息缺失问题,引入高效通道注意力模块,增强对裂缝特征信息的提取能力。此外,针对裂缝与背景类别不平衡问题,设计了在线困难样本挖掘损失函数进行缓解。实验结果表明:在单个3050Ti GPU上,CrackViT算法最终在裂缝数据集上以63 FPS的速度获得了75.62%的IoU,模型参数量仅为2.43 M。CrackViT-L模型精度IoU为76.83%,模型参数量为3.56 M,模型推理速度达到61FPS。算法测试精度优于大多数主流模型,并且需要更少的模型参数。研究结果表明,CrackViT所预测的隧道裂缝分割图像边缘更加清晰和完整,保持推理速度的同时,能够有效检测裂缝,该算法有助于隧道裂缝检测实际应用。Crack detection is crucial to ensuring the safety of the tunnel structure,and the timely detection of tunnel crack defects is conducive to reducing the project maintenance cost and guaranteeing traffic safety.However,the traditional convolutional neural network in tunnel crack detection tasks mainly focuses on improving detection accuracy and algorithm complexity.How to balance accuracy and real-time crack detection is a difficult point in the current research.To address this problem,this paper proposed a crack segmentation method called CrackViT based on a lightweight Transformer.First,the MobileViT network,which is a hybrid of convolutional neural networks and Transformer,was used to construct a crack feature extraction network.It reduced the parameters of the network model and the amount of computation and efficiently extracts the global information and the local feature information of the crack image.Then,an improved atrous spatial pyramid pooling decoder was proposed to realize feature extraction and information fusion at different scales and achieve pixel-level probability distribution.Meanwhile,the crack image suffers from the problem of missing detail information,and an efficient channel attention module was introduced to enhance the extraction ability of the crack feature information.In addition,for the problem of imbalance between crack and background categories,an online difficult sample mining loss function was designed to mitigate it.The experimental results show that the CrackViT algorithm finally achieves 75.62%IoU on the crack dataset with 63 FPS on a single 3050Ti GPU,with a model parameter count of only 2.43 M.The CrackViT-L model accuracy IoU is 76.83%,with a model parameter count of 3.56 M,and the model inference speed reaches 61 FPS.The algorithm’s tested accuracy is better than most mainstream models and requires fewer model parameters.The results show that the edges of the tunnel crack segmentation images predicted by CrackViT are clearer and more complete,and the cracks can be effectivel

关 键 词：裂缝分割 TRANSFORMER MobileViT 空洞空间金字塔池化 轻量级模型 

分 类 号：U457.5[建筑科学—桥梁与隧道工程]