关键词:
目标检测
道路损伤
YOLOv10
可逆列网络
摘要:
道路损伤是路面常见的病害之一,及时且有效的检测对于路面的维护与管理至关重要。YOLOv10算法作为一种新提出的YOLO系列算法,其具有的无NMS训练机制、双重分配策略以及轻量化模型设计等特点使得YOLOv10在保持高准确性的同时,实现了更快的推理速度和更高的效率,因此YOLOv10算法特别适用于实时的道路损伤检测任务。在实际应用中,由于移动检测终端的硬件资源有限,对于大参数量模型和大的运算量难以承受,因为本文对YOLOv10算法进行改进,在基本保持较高检测精度的条件下,降低算法的参数量和计算量,以便于其在移动终端的部署。本文采用可逆列网络结构的思想对YOLOv10的主干网络进行重构,重构后的网络使得骨干网络具有了横向的特征信息融合能力,重构后的网络虽然增加了复杂度,但是却大大增强了特征信息融合能力,为了平衡精度和复杂度,本文对YOLOv10骨干网络纵向结构的参数量进行了降维处理,实验结果表明,本文所提的改进YOLOv10算法在检测精度上虽然有略微的下降,但是却能够大幅度地降低算法的参数量和计算量,使得其在移动终端道路损伤检测的应用上更有竞争力。Road damage is one of the common diseases of road surface. Timely and effective detection is crucial for the maintenance and management of road surface. As a newly proposed YOLO series algorithm, YOLOv10 algorithm has no NMS training mechanism, dual allocation strategy and lightweight model design, so that YOLOv10 maintains high accuracy while achieving faster reasoning speed and higher efficiency. Therefore, YOLOv10 algorithm is especially suitable for real-time road damage detection tasks. In practical application, due to the limited hardware resources of the mobile detection terminal, it is difficult to bear the large number of parameters and the large amount of operation. Because this paper improves the YOLOv10 algorithm and reduces the number of parameters and calculation amount of the algorithm under the condition of basically maintaining high detection accuracy, so as to facilitate its deployment in the mobile terminal. In this paper, we use the idea of reversible column network structure to reconstruct the backbone network of YOLOv10. The reconstructed network makes the backbone network with horizontal feature information fusion ability. Although the reconstructed network has increased the complexity, it greatly enhances the ability of feature information fusion to balance the precision and the complexity. In this paper, the number of parameters in the longitudinal structure of the YOLOv10 backbone network is reduced in dimension. The results of the experiments showed that although the pr