改进YOLOv5的路面裂缝检测模型研究

Research on improved YOLOv5 pavement crack detection model

针对传统的路面裂缝检测方式耗时耗力、成本高、主观性强等问题,提出了一种基于YOLOv5的路面裂缝检测模型YOLOv5-Crack。首先在主干部分处引入坐标注意力机制并优化成CA-plus结构以提高裂缝特征关注度;其次提出一种全新的特征融合网络ESPP,降低部分计算量的同时提升特征融合能力;然后,在颈部网络中使用重影混洗卷积替代传统卷积,尽可能保持通道语义信息的同时降低计算成本;最后,整体引入SIoU损失函数提升回归精度。为验证改进模型YOLOv5-Crack的有效性,在数据集GRDDC 2020上进行对比实验,结果表明其F1分数分别为58.43%和58.21%,与原YOLOv5模型相比分别提升了4.05%和3.93%,并且降低了7.8 GFLOPs的计算量,FPS达到37.9,有效解决了路面裂缝检测的弊端;同时与主流目标检测算法相比,所提出的YOLOv5-Crack模型在路面裂缝检测方面更具有优越性。

In response to the problems of traditional road crack detection methods,such as time-consuming,labor-intensive,high cost,and subjectivity,a YOLOv5-based road crack detection model,named YOLOv5-Crack,is proposed.Firstly,a coordinate attention mechanism is introduced in the backbone and optimized as a CA-plus structure to enhance the crack feature focus.Secondly,a novel feature fusion network ESPP is proposed to reduce some computational costs while improving the feature fusion capability.Then,the heavy Ghost-Shuffle convolution is used in the neck network to replace the traditional convolution,which can keep the channel semantic information as much as possible while reducing computational costs.Finally,the SIoU loss function is introduced to improve the regression accuracy.To validate the effectiveness of the improved YOLOv5-Crack model,comparative experiments are conducted on the GRDDC 2020dataset,and the results show that the F1scores are 58.43%and 58.21%,respectively,which are 4.05%and 3.93%higher than those of the original YOLOv5model,and the computational cost is reduced by 7.8 GFLOPs,with an FPS of 37.9,effectively addressing the shortcomings of road crack detection.Furthermore,compared with mainstream object detection algorithms,the proposed YOLOv5-Crack model has superior performance in road crack detection.