一种轻量化网络的道路病害检测方法研究

A Study on Road Defect Detection Method for Lightweight Networks

摘要:

针对道路病害检测中道路情况复杂、实时检测较为困难，缺检漏检等问题，采集并制作了多类型道路病害数据集R-CRACK，在YOLOv5s模型基础上，在Neck模块上使用轻量化卷积GSConv模块替换部分标准卷积构建轻量化网络颈部；在Head模块上利用SimSPPF对空间金字塔池化方式进行改进并应用轻量级上采样算子CARAFE，而后得到GSC-YOLO模型。将GSC-YOLO模型利用矩形推理、图像加权，标签平滑处理方式对数据集R-CRACK中训练集部分进行训练。模型训练后的结果表示，与YOLOv5s基础模型相比，GSC-YOLO参数量减少6.8%、计算量减少4.8%、mAP(@.5)上涨了9.2%。利用改进前后的网络模型分别对单一及复杂环境下的道路病害进行检测，通过对比不同模型的检测效果，证明了GSC-YOLO模型针对YOLOv5s缺检漏检等问题有所改进，此类轻量化检测网络对解决道路病害检测有着重要意义。

Abstract:

In response to the challenges of complex road conditions, difficult real-time detection, and issues such as missing detections and false negatives in road distress detection, a multi-type road distress dataset named R-CRACK was collected and created. Based on the YOLOv5s model, lightweight GSConv modules were used to replace some standard convolutions in the Neck module to construct a lightweight network neck. In the Head module, SimSPPF was applied to improve the spatial pyramid pooling method, and a lightweight upsampling operator CARAFE was utilized, resulting in the GSC-YOLO model. The GSC-YOLO model was trained on a portion of the training set of the R-CRACK dataset using rectangle inference, image weighting, and label smoothing techniques. The results of the trained model show that compared to the base YOLOv5s model, the GSC-YOLO model reduces parameters by 6.8%, computation by 4.8%, and improves mAP(@.5) by 9.2%. The improved network models were used to detect road distress in both single and complex environments. By comparing the detection results of different models, it was demonstrated that the GSC-YOLO model improves upon the shortcomings of YOLOv5s in terms of missing detections and false negatives. This type of lightweight detection network is of significant importance for solving road distress detection.