基于检测头结构优化的沥青路面病害检测模型轻量化研究

Lightweight distress detection models for asphalt pavement based on optimized detection head structure

关键词:  沥青路面  病害检测  轻量化算法  检测头  结构优化

Key words:  asphalt pavement  surface defect detection  lightweight algorithm  detection head  structural optimization

摘要:

道路病害目标普遍具有形态不规则、细节微弱、小目标密集和尺度跨度大的特征,传统检测头在空间建模、尺度感知与语义共享方面存在不足,难以兼顾检测精度与计算效率。为此,以YOLOv11为基础,提出3种轻量化检测头结构优化方案,构建了兼顾高效性与适应性的道路病害智能检测框架。其中,DyHead模块通过引入多维注意力机制实现动态特征建模,显著提升了尺度与空间感知能力；LSCD结构采用轻量共享卷积设计,在保持特征表达能力的同时实现参数量减少30%、FLOPs降低22%,有效改善模型的部署性能；GLSA模块融合全局局部空间注意力机制,提升了模型对复杂结构特征的判别能力与多尺度适应性。研究结果表明:3种改进模型在精度与效率上均取得显著优化,其中DyHead在mAP@0.5上提升至64.95%,LSCD在资源受限条件下表现最优,GLSA在综合性能方面最为均衡。研究成果为道路病害检测模型的轻量化设计与嵌入式部署提供了新思路,对智慧交通基础设施的智能运维具有重要的工程应用价值。

Abstract:

Pavement distresses are typically characterized by irregular morphology, weak texture features, dense small targets, and large-scale variations. Traditional detection heads suffer from limited spatial modeling, poor scale perception, and insufficient semantic sharing, making it difficult to balance detection accuracy and computational efficiency. To address these issues, this study proposed three lightweight detection head optimization schemes based on the YOLOv11 framework, establishing an intelligent and efficient detection architecture for pavement distresses. Specifically, the DyHead module introduces a multi-dimensional attention mechanism to achieve dynamic feature modeling, significantly enhancing scale and spatial perception; the LSCD structure adopts a lightweight shared convolution design, reducing the number of parameters by 30% and FLOPs by 22% while maintaining feature representation capacity, thereby improving deployment efficiency; and the GLSA module integrates a global-local spatial attention mechanism to strengthen feature discrimination and multi-scale adaptability under complex road conditions. Experimental results demonstrate that all three improved models achieve notable gains in both accuracy and efficiency: DyHead achieves an mAP@0.5 of 64.95%, LSCD performs best under resource-constrained conditions, and GLSA exhibits the most balanced overall performance. The proposed methods provide new insights for lightweight model design and embedded deployment of pavement distress detection, offering valuable technical support for the intelligent maintenance of smart transportation infrastructure.