满轲, 孙东兴, 刘晓丽,等.融合多尺度注意力机制的轻量化地下空间裂缝分割方法[J].河南理工大学学报（自然科学版）,2026,45(4):105-113.

MAN K, SUN D X, LIU X L,et al.Lightweight crack segmentation method for underground spaces integrating a multi-scale attention mechanism[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(4):105-113.

融合多尺度注意力机制的轻量化地下空间裂缝分割方法

满轲1, 孙东兴1, 刘晓丽2, 宋志飞1, 邓稀肥3, 鞠兴军4

1.北方工业大学 土木工程学院，北京  100144;2.清华大学 水沙科学与水利水电工程国家重点实验室，北京  100084;3.中铁四局集团有限公司，安徽 合肥  230023;4.国家能源集团宝日希勒能源有限公司，内蒙古 呼伦贝尔  021008

摘要: 目的 针对现有深度学习裂缝检测模型参数量大、计算复杂、难以满足现场实时检测需求等问题，研发一种可在移动端和边缘设备实时部署的高精度轻量化检测方法。  方法提出一种改进的DeepLabV3+模型算法，用于检测结构受损裂缝。改进后的模型结合混合感受野设计，采用递进式膨胀率的空洞卷积，有效扩大感受野并避免网格效应；以轻量级MobileNetV2替换原始主干网络，降低特征通道维度；在解码器中引入GhostNet卷积操作，通过线性变换生成冗余特征图，减少参数规模和计算成本；结合多尺度卷积融合与通道注意力机制，设计基于方向特征提取的增强型裂缝注意力模块，强化对裂缝边缘和复杂走向的感知；优化损失函数，引入Dice与Focal损失权重，提升模型对细小裂缝的检测能力，缓解类别划分不均衡问题；融合高低级语义特征，增强裂缝的表征能力。 结果结果表明，改进后的模型参数量降至原来的25.6%，测试集的F₁分数、平均交并比、平均像素准确率分别提升了3.37%，5.53%和6.20%。 结论改进后的模型减少了特征信息的丢失，增强了裂缝边缘细节与复杂走向的识别能力，表现出良好的泛化性，新的改进算法为隧道、深基坑和其他地下空间结构病害的自动化监测提供了一种新思路。

关键词:深度学习;裂缝检测;语义分割;注意力机制;空洞卷积

doi:10.16186/j.cnki.1673-9787.2025100033

基金项目:国家重点研发计划项目（2023YFB4005505）；国家自然科学基金资助项目（51522903，51774184）

收稿日期:2025/10/24

修回日期:2026/03/23

出版日期:2026/06/17

Lightweight crack segmentation method for underground spaces integrating a multi-scale attention mechanism

Man Ke1, Sun Dongxing1, Liu Xiaoli2, Song Zhifei1, Deng Xifei3, Ju Xingjun4

1.School of Civil Engineering， North China University of Technology， Beijing  100144， China;2.State Key Laboratory of Hydroscience and Engineering， Tsinghua University， Beijing  100084， China;3.China Railway No.4 Engineering Group Co.， Ltd.， Hefei  230023， Anhui， China;4.Baorixile Energy Co.， CHN Energy， Hulunbuir  021008， Inner Mongolia， China

Abstract: Objectives With the rapid development of urban rail transit systems in China， the safety inspection demands for underground space structures have increased significantly. Existing deep learning-based crack detection models suffer from large parameter sizes and high computational complexity， making it difficult to satisfy the requirements of real-time on-site detection. Therefore， a high-precision lightweight detection method suitable for real-time deployment on mobile terminals and edge devices was developed. Methods An improved DeepLabV3+ model was proposed for structural crack detection. The improved model integrates a hybrid receptive field design and employs dilated convolutions with progressive dilation rates to effectively enlarge the receptive field while avoiding grid artifacts. A lightweight MobileNetV2 was adopted to replace the original backbone network， thereby reducing feature channel dimensions. GhostNet convolution operations were introduced into the decoder to generate redundant feature maps through linear transformations， reducing the number of parameters and computational cost. In addition， an enhanced crack attention module based on directional feature extraction was designed by combining multi-scale convolution fusion and channel attention mechanisms， thereby improving the perception of crack edges and complex propagation patterns. The loss function was optimized by integrating weighted Dice loss and Focal loss to improve sensitivity to fine cracks and alleviate class imbalance. Furthermore， high-level and low-level semantic features were fused to enhance crack representation capability. Results The results show that the number of model parameters is reduced to 25% of the original model， while the F₁-score， mean intersection over union， and mean pixel accuracy on the test set are increased by 3.37%， 5.53%， and 6.2%， respectively.  Conclusions The improved model effectively reduces feature information loss and enhances the recognition capability for crack edge details and complex propagation patterns， demonstrating good generalization performance. The proposed algorithm provides a new approach for the automated monitoring of structural defects in tunnels， deep foundation pits， and other underground spaces.

Key words:deep learning;crack detection;semantic segmentation;attention mechanism;dilated convolution