针对已有目标检测方法在混凝土桥梁表观病害检测的应用上识别精度低且伴随较多误检和漏检的问题，提出一种改进的YOLOv5s桥梁表观病害检测方法。首先，针对目前桥梁表观病害特征成分较复杂的问题，为了更有效的利用不同尺度的缺陷特征，在主干网中添加修改后的空间金字塔池化模块，从而提高了整体网络对缺陷特征信息的获取能力，同时减少了运算工作量；其次，针对由病害图像中不同缺陷特征交叉分布导致的误检率、漏检率高的问题，在YOLOv5s网络中加入轻量化注意力模块；最后，针对桥梁缺陷尺寸差异大、分类困难以及数据集小而导致的边界回归不匹配的问题，采用考虑了向量角度的损失函数。实验证明，改进后的YOLOv5s检测器在桥梁表观病害目标检测识别任务中能够有效提高精度、降低误检率和漏检率。

To solve the problems of low accuracy, high false detection rate and missed detection rate of current target detection methods in concrete bridge apparent disease detection, an improved YOLOv5s method on bridge apparent disease detection is proposed. Firstly, to achieve more effective fusion of features at different scales and increase receptive fields, an improved spatial pyramid pooling module is added to YOLOv5s network to enhance feature extraction capability and reduce computational cost. In addition, aiming at the problem of high false detection rate and missed detection rate caused by cross distribution of different defect features in disease images, a light attention module is added in YOLOv5s network. At last, the loss function considering vector angle is adopted to solve the problems caused by the different size of bridge defect, classification difficulty and small data set leading to the boundary box regression mismatch. Experimental results show that the improved YOLOv5s detector can effectively improve the accuracy, reduce the false detection rate and loss detection rate in bridge apparent disease detection task.

TP391.4

国家自然科学基金资助项目(51775072)；重庆市科技创新领军人才支持计划项目(CSTCCCXLJRC201920)；重庆市高校创新研究群体(CXQT20019)。