基于GS-MobileNet的轻量级道路异常检测算法

A Lightweight Road Anomaly Detection Algorithm Based on GS-MobileNet

摘要: 针对传统路面缺陷检测方法在复杂场景下实时性差、误检率高，以及现有深度学习模型难以兼顾精度与边缘部署效率的问题，本文构建了轻量化的路面缺陷检测算法GS-MobileNet。基于MobileNetV4-Ghost的主干网络架构。通过引入Ghost模块的廉价线性变换操作，在保证特征表达能力的同时显著降低模型参数量。创新设计的局部–全局协同注意力机制，在浅层和中层网络嵌入BoT模块强化裂纹边缘连续性建模，深层采用SimAM无参注意力动态抑制背景噪声。进一步通过BiFPN动态加权多尺度融合策略，有效改善微小裂纹的检出能力，并结合分组查询解码器实现分类–定位任务的高效对齐。最终构建GS-MobileNet模型以3.12 M参数量取得91.3%的精确率与92.1%的mAP。

Abstract: Addressing the issues of poor real-time performance and high false positive rates of traditional road defect detection methods in complex scenarios, as well as the difficulty of existing deep learning models in balancing accuracy and edge deployment efficiency, this paper constructs a lightweight road defect detection algorithm named GS-MobileNet. The algorithm is based on the MobileNetV4-Ghost backbone network architecture. By introducing the Ghost module’s cheap linear transformation operations, it significantly reduces the number of model parameters while ensuring feature representation capability. The innovatively designed local-global cooperative attention mechanism embeds the BoT module in shallow and middle layers to enhance crack edge continuity modeling, while the deep layers adopt the SimAM parameter-free attention to dynamically suppress background noise. Furthermore, the BiFPN dynamic weighted multi-scale fusion strategy effectively improves the detection capability for tiny cracks, and combined with the group query decoder, it achieves efficient alignment of classification and localization tasks. The final GS-MobileNet model, with only 3.12 M parameters, achieves a precision of 91.3% and an mAP of 92.1%.