基于改进YOLOv5的桥梁裂缝模型研究

Research on bridge crack model based on improved YOLOv5

摘要:

桥梁裂缝人工检测耗时费力、安全性不高，为了高效、准确、无接触的对桥梁裂缝进行识别检测，提出一种基于改进YOLOv5的桥梁裂缝检测模型YOLOv5-SA；该方法在YOLOv5s模型的基础上，首先对收集的数据集利用几何变换、光学变换等操作进行数据增强；其次将融合视觉注意力机制 (SKNet，Selective Kernel Networks)添加到Head部分来提高模型对裂缝特征的表示能力；最后在金字塔特征表示法(FPN)的基础上利用自适应空间特征融合(ASFF, Adaptively Spatial Feature Fusion)模块加强网络特征融合能力，增加对桥梁裂缝小目标的检测；结果表明：改进后的模型相对于YOLOv5s模型能更好的抑制非关键信息，减少背景中的无效信息干扰，提高桥梁裂缝目标检测精准度；改进后的YOLOv5-SA模型准确率达到88.1%，与原YOLOv5s模型相比提高了1.6%；平均精度均值mAP0.5和mAP0.5-0.95分别达到90.0%、62.1%，相比而言分别提高了2.2%、2.4%；与其他桥梁裂缝检测相关方法(Faster-RCNN、YOLOv4tiny)相比，提出的YOLOv5-SA模型也具有相当或更好的检测性能；由此可见改进后的模型能更高效的检测复杂环境下的桥梁裂缝，可以为工业检测提供一部分思路。

Abstract:

Manual detection of bridge cracks is time-consuming and laborious, and the safety is not high. In order to identify and detect bridge cracks efficiently, accurately and without contact, a bridge crack detection model YOLOV5-SA based on improved YOLOv5 is proposed. Based on the YOLOv5s model, firstly, the collected data set is enhanced by geometric transformation and optical transformation. Secondly, Selective Kernel Networks (SKNet, Selective Kernel Networks) were added to the Head part to improve the representation ability of crack features. Finally, on the basis of pyramid Feature notation (FPN), Adaptively Spatial Feature Fusion module was used to strengthen the network feature fusion ability, and to increase the detection of small targets for bridge cracks. The results show that compared with the YOLOv5s model, the improved model can suppress non-critical information better, reduce the interference of invalid information in the background, and improve the accuracy of bridge crack target detection. The accuracy of the improved YOLOv5-SA model reaches 88.1%, which is 1.6% higher than that of the original YOLOv5s model. The average accuracy of mAP0.5 and MAP0.5-0.95 reached 90.0% and 62.1%, respectively, which increased by 2.2% and 2.4%. Compared with other methods related to bridge crack detection (Faster-RCNN, YOLOv4tiny), the proposed YOLOv5-SA model also has comparable or better detection performance. It can be seen that the improved model can detect bridge cracks in complex environments more efficiently, which can provide some ideas for industrial detection.