作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于PMB-DETR的地铁隧道病害检测方法

  • 发布日期:2026-11-18

Subway Tunnel Disease Detection Method Based on PMB-DETR

  • Published:2026-11-18

摘要: 针对地铁隧道复杂环境下多尺度病害特征提取不充分、小目标易漏检、定位精度偏低及模型难以轻量化部署等问题,基于RT-DETR-R18网络提出PMB-DETR轻量化隧道病害检测算法,搭建数据增强预处理、多模块改进网络、多尺度特征融合的全链路检测方案。数据层面,针对隧道病害标注样本稀缺、细微病害纹理生成失真问题,优化CycleGAN生成对抗网络,在原有对抗损失与循环一致性损失基础上引入高层特征感知损失,为模型提供更多数据支撑。网络结构层面设计三项核心改进模块:一是骨干网络嵌入MFPConv多尺度频率概率卷积模块,通过可学习小波滤波器将特征拆分为裂缝高频、脱落中频、渗漏水低频子带,结合频空联合掩码实现稀疏计算,融合部分卷积、结构重参数化与SE通道注意力,在差异化捕捉不同频率病害特征的同时削减模型冗余参数;二是搭建MSMHSA-AIFI多尺度多头稀疏自注意力模块,以各向异性高斯椭圆采样核替代传统固定采样方式,针对不同病害纹理特征自适应采样,提升复杂背景下微小稀疏病害的全局感知能力;三是构建BiEMFPN双向增强特征金字塔,实现语义与细节特征双向回流,通过可学习加权融合,解决浅层小病害细节丢失、深层病害轮廓稀释的问题。在自建地铁隧道病害数据集上开展实验验证,PMB-DETR算法mAP@0.5达88.1%,较基准RT-DETR-R18提升4.5%,模型参数量降低25.6%,各项核心指标优于SSD、Faster R-CNN、YOLO等主流检测模型。实验证明,改进CycleGAN数据增强、MFPConv、MSMHSA-AIFI、BiEMFPN四大模块均可独立提升检测精度,多模块协同优化可实现精度与轻量化最优平衡。本文所提算法精准适配隧道多形态、多尺度衬砌病害检测场景,有效解决微小病害漏检、定位偏差、边缘部署算力受限等核心问题,在检测精度、模型参数量与计算复杂度三者间实现均衡优化,满足工程高精度、轻量化实时检测需求,为地铁隧道智能巡检与安全运维提供高效可行的技术方案。

Abstract: To address insufficient multi-scale defect feature extraction, missed detection of small targets, low localization accuracy, and difficulty in lightweight deployment in complex subway tunnel environments, this paper proposes PMB-DETR, a lightweight tunnel defect detection algorithm based on RT-DETR-R18, and constructs a full-pipeline detection scheme integrating data augmentation preprocessing, multi-module network improvements, and multi-scale feature fusion. At the data level, to tackle the scarcity of annotated tunnel defect samples and texture distortion in generated subtle defects, CycleGAN is optimized by introducing a high-level feature perceptual loss in addition to the original adversarial and cycle-consistency losses, providing additional data support. At the network architecture level, three core improved modules are designed. First, the backbone network embeds the MFPConv multi-scale frequency probability convolution module, which decomposes features into high-frequency crack subbands, mid-frequency spalling subbands, and low-frequency leakage subbands via learnable wavelet filters, achieves sparse computation through frequency-spatial joint masking, and integrates partial convolution, structural reparameterization, and SE channel attention to capture frequency-specific defect characteristics while reducing redundant parameters. Second, the MSMHSA-AIFI multi-scale multi-head sparse self-attention module is constructed, replacing traditional fixed sampling with anisotropic Gaussian elliptical sampling kernels to enable adaptive sampling for diverse defect textures and enhance global perception of tiny sparse defects in complex backgrounds. Third, the BiEMFPN bidirectional enhanced feature pyramid is constructed to realize bidirectional flow of semantic and detail features; through learnable weighted fusion, it resolves the loss of fine details for small defects at shallow levels and contour dilution at deep levels. Experimental validation on a custom subway tunnel defect dataset shows that PMB-DETR achieves an mAP@0.5 of 88.1%, representing a 4.5% improvement over the RT-DETR-R18 baseline with 25.6% fewer parameters, and outperforms mainstream detectors including SSD, Faster R-CNN, and YOLO in core metrics. Experimental results demonstrate that the improved CycleGAN data augmentation, MFPConv, MSMHSA-AIFI, and BiEMFPN can all independently improve detection accuracy, and their collaborative optimization achieves the optimal balance between accuracy and lightweight design. The proposed algorithm is well suited for multi-form, multi-scale lining defect detection scenarios in tunnels, effectively resolving core issues including missed detection of micro-defects, localization deviation, and limited edge computing resources; it achieves balanced optimization among detection accuracy, model parameters, and computational complexity, meets engineering demands for high-precision lightweight real-time detection, and provides an efficient and practical technical solution for intelligent tunnel inspection and safe operation and maintenance.