铁道建筑

针对无人机铁路沿线巡检图像中背景干扰复杂、路基及附属设施尺度差异显著及部分目标具有极端长宽比，导致现有目标检测算法识别精度不足与易漏检等问题，本文提出一种基于YOLO26改进的铁路路基及附属设施检测模型。该模型以华北某重载干线铁路某工务段无人机巡检图像为数据源，通过引入卷积块注意力模块(Convolutional Block Attention Module， CBAM)、空洞残差模块(Dilation-wise Residual， DWR)及条形池化模块(Strip Pooling Module， SPM)对YOLO26网络结构进行深度优化。其中，CBAM模块通过串联通道与空间注意力并利用膨胀卷积扩大特征感知范围，实现了对全局上下文信息的精准捕捉与背景噪声的有效抑制；DWR模块凭借多分支并行的空洞卷积结构动态调整感受野，显著强化了网络对不同尺度目标的特征建模能力；SPM模块通过垂直与水平方向的条形感知算子，针对性优化了铁轨、护栏等具有极端长宽比目标的特征提取效果与几何形状适配度。实验结果表明，在包含铁轨、栅栏、排水沟及护坡等四类目标的铁路数据集上，改进后的模型性能显著优于YOLOv8、YOLO11等主流算法。相比基线模型，本文方法的mAP@50和mAP@50:95分别提升了2.3%和3.7%。在针对特定类别的检测中，排水沟和铁轨等核心类别的精度提升尤为显著，其中排水沟的mAP@50和mAP@50:95分别增长了4.5%和9.8%。通过Grad-CAM热力图分析证实，本研究能更准确地聚焦于目标特征区域，有效满足了铁路安全巡检对路基及附属设施检测的高精度与高效率需求。

To address the issues of complex background interference， significant scale differences of subgrade and ancillary facilities， and extreme aspect ratios of certain targets in UAV-based railway inspection images， which lead to insufficient recognition accuracy and high missed detection rates in existing object detection algorithms， this paper proposed an improved railway subgrade and ancillary facility detection model based on YOLO26. The model utilized UAV inspection images from a maintenance depot of a heavy-haul trunk railway in North China as the data source， and deeply optimized the YOLO26 network architecture by introducing the Convolutional Block Attention Module (CBAM)， the Dilation-wise Residual (DWR) module， and the Strip Pooling Module (SPM). Specifically， the CBAM module cascaded channel and spatial attention and employed dilated convolutions to expand the feature perception range， achieving accurate capture of global contextual information and effective suppression of background noise. The DWR module dynamically adjusted the receptive field through a multi-branch parallel dilated convolution structure， significantly enhancing the network's feature modeling capability for targets at different scales. The SPM module specifically optimized the feature extraction and geometric shape adaptation for targets with extreme aspect ratios， such as rails and fences， through vertical and horizontal strip perception operators. Experimental results demonstrate that on a railway dataset containing four target categories (rails， fences， drainage ditches， and slope protection)， the performance of the improved model is significantly superior to mainstream algorithms such as YOLOv8 and YOLO11. Compared with the baseline model， the mAP@50 and mAP@50:95 of the proposed method improve by 2.3% and 3.7%， respectively. In the detection of specific categories， the accuracy improvements of core categories such as drainage ditches and rails are particularly notable， with the mAP@50 and mAP@50:95 of drainage ditches increasing by 4.5% and 9.8%， respectively. Grad-CAM heatmap analysis confirms that this study can more accurately focus on target feature regions， effectively meeting the high-precision and high-efficiency requirements of railway safety inspection for the detection of subgrade and ancillary facilities.