基于RCGELAN-YOLOv11的路面损伤检测算法

Road Surface Damage Detection Algorithm Based on RCGELAN-YOLOv11

摘要: 本文针对路面损伤检测任务中传统方法效率低、易受环境干扰的痛点，提出了一种基于RCGELAN-YOLOv11的改进算法。在YOLOv11算法的基础上改进网络结构，通过设计RC-G-ELAN模块替代YOLOv11中的C3k2模块，实现了检测精度与计算效率的双重优化。具体而言，RC-G-ELAN模块引入重参数化卷积RepConv增强特征提取能力，采用Conv 3 × 3模块替代C3k模块简化网络结构，将串行结构改为并行结构减少冗余计算，并在网络后端添加Conv 1 × 1模块促进通道间信息交互。实验采用GRDDC2020数据集(含14,569张标注图像，覆盖10类损伤类型)，以7:1:2比例划分训练集、验证集和测试集。实验结果表明，改进后的算法与YOLOv11算法的检测精度相当，但是改进算法的网络层数降低19.33%，参数量减少15.84%，梯度计算量降低22.81%，GFLOPs降低6.15%，大幅地降低了计算复杂度，为复杂场景下的路面损伤检测提供了高效可靠的技术支撑。

Abstract: This paper addresses the pain points of low efficiency and environmental interference in traditional road surface damage detection methods by proposing an improved algorithm based on RCGELAN-YOLOv11. Building upon the YOLOv11 framework, we optimize the network architecture by replacing the C3k2 module with the RC-G-ELAN module, achieving dual improvements in detection accuracy and computational efficiency. Specifically, the RC-G-ELAN module introduces RepConv (reparameterized convolution) to enhance feature extraction capabilities, adopts Conv 3 × 3 modules to simplify the network structure, converts serial architecture to parallel architecture to reduce redundant computations, and adds Conv 1 × 1 modules in the backend to facilitate channel interaction. Experiments were conducted using the GRDDC2020 dataset (containing 14,569 labeled images covering 10 damage types), with training, validation, and test sets divided in a 7:1:2 ratio. Results demonstrate that the improved algorithm achieves comparable detection accuracy to YOLOv11 while reducing network layers by 19.33%, parameter volume by 15.84%, gradient computation by 22.81%, and GFLOPs by 6.15%. This significant reduction in computational complexity provides efficient and reliable technical support for road surface damage detection in complex scenarios.