改进 YOLOv8 的轨道扣件状态检测方法

针对现有轨道扣件状态检测算法对小目标物体与复杂形状物体的检测能力差而导致检测结果异常，以及小目标层特 征冗余等问题，提出了一种改进 YOLOv8 的轨道扣件状态检测方法。在 YOLOv8 网络中增加可变形空间金字塔扩张卷积模 块，以提高模型对小目标物体以及形变复杂物体的检测精度。同时增加小目标空间重构单元以减少小目标特征冗余，促进小 目标特征的学习。根据采集到的轨道扣件数据集进行模型的训练和测试，并与多组轨道扣件状态检测算法进行对比，实验结 果表明，相较于对比算法，所提算法精确度平均提升3.20%,召回率平均提升3.34%,平均精度平均提升3.96%。实验证明 所提算法能够有效进行轨道轨道扣件状态检测，并且具有较强的泛化能力，可以部署于复杂交通场景。

Improved YOLOv8 track fastener state detection method

Aiming at the problems of poor detection ability of existing track fastenings state detection algorithms on small target objects and complex shapes,which lead to abnormal detection results,as well as feature redundancy of small target layer,an improved YOLOv8 track fastenings state detection method was proposed.A deformable space pyramid expansion convolutional module is added in YOLOv8 network to improve the detection accuracy of small target objects and deformable complex objects.At the same time,the reconstruction unit of small target space is added to reduce the redundancy of small target features and promote the learning of small target features.The model is trained and tested according to the collected data set of track fasteners,and compared with multiple groups of track fasteners state detection algorithms.The experimental results show that compared with the comparison algorithm,the accuracy of the proposed algorithm is increased by 3.20%on average,the recall rate is increased by 3.34%on average,and the average accuracy is increased by 3.96%on average.The experiments proves that the proposed algorithm can detect the state of rail fastenings effectively,and has strong generalization ability,and can be deployed in complex traffic scenarios.