基于YOLOv7-CA-BiFPN的路面缺陷检测

Road surface pothole detection based on YOLOv7-CA-BiFPN

摘要:

路面坑洼是主要道路缺陷,会损坏车辆,影响驾驶员的安全驾驶,严重时还会导致交通事故,针对这个问题,本文提出了改进YOLOv7的道路坑洼检测算法。首先,使用Mosaic+Mixup进行内置数据增强,扩充小样本数据集,增强模型泛化能力；其次,引入CA注意力机制,将横纵位置信息编码,保证计算量的同时又能关注大范围位置信息；然后,引入BIFPN双向特征金字塔网络,通过特征融合多尺度语义特征提高检测效率；最后,将损失函数SIoU替换CIoU,有效解决回归中的样本不平衡问题。实验结果表明,改进之后的算法在坑洼数据集的平精度均值和精确率达到了89.42%和90.12%,相比于原本的YOLOv7版本提高了6.18%和1.96%,更准确更快速的应用于道路维修。

Abstract:

Road potholes are the main road defects that can damage vehicles, affect driver safety, and even lead to traffic accidents in severe cases. To address this issue, this paper proposes an improved YOLOv7 road pothole detection algorithm. Firstly, use Mosaic+Mixup for built-in data augmentation, expand small sample datasets, and enhance model generalization ability; Secondly, the CA attention mechanism is introduced to encode the horizontal and vertical position information, ensuring both computational complexity and attention to a wide range of position information; Then, the BIFPN bidirectional feature pyramid network is introduced to improve detection efficiency through feature fusion of multi-scale semantic features; Finally, CIoU is replaced by the loss function SIoU, which effectively solves the sample imbalance problem in regression. The experimental results show that the improved algorithm achieves a mean and accuracy of 89.42% and 90.12% in flat and uneven datasets, which is 6.18% and 1.96% higher than the original YOLOv7 version. It is more accurate and fast in road maintenance applications.