基于改进YOLOv5s的动车配件裂纹荧光磁粉检测研究

Research on Fluorescent Magnetic Particle Crack Detection of Bullet Train Accessories Based on Improved YOLOv5s

摘要: 针对目前动车配件裂纹缺陷磁粉检测存在数据样本少，没有公开数据集，网络模型体积大，识别精度低等问题，本文建立数据集并提出一种改进YOLOv5s目标检测算法。首先，针对数据样本少的问题，为了提高模型泛化性，同时针对荧光磁粉图像的特点，对小样本数据集使用HSV (Augment HSV (Hue, Saturation, Value))随机增强图像进行数据增强。然后，将YOLOv5s中backbone层的主干网络CSPDarknet-53更换为轻量化ShuffleNet-v2，减少模型体积。其次，在此基础上添加A2-Nets注意力模块，并结合Focal Loss损失函数使用。以确保网络更好地关注细小裂纹，改善裂纹检测任务精度。配合荧光磁粉技术，使动车配件裂纹增强显示，有利于算法更好地识别到裂纹。实验结果表明：本文提出的改进YOLOv5s目标检测算法提高了检测精度，具有较强的合理性。在动车裂纹荧光磁粉数据集上，检测精度mAP_0.5由86.00%提高到93.74%。为YOLOv5s目标检测算法在荧光磁粉检测项目上应用提供借鉴。

Abstract: Aiming at the problems of limited data samples, no public data set, the network model is large, and low recognition accuracy in magnetic particle detection of crack defects of bullet train parts, this paper establishes a data set and proposes an improved YOLOv5s target detection algorithm. Firstly, for the problem of small data samples, in order to improve the generalization of the model and the characteristics of fluorescent magnetic powder images, HSV (Hue, Saturation, Value) randomly en-hanced images were used to Augment the small sample data set. Then, the backbone network CSPDarknet-53 of the backbone layer in YOLOv5s is replaced with lightweight ShuffleNet-v2 to re-duce model size. Secondly, the A2-Nets attention module is added on this basis and used in combi-nation with Focal Loss function. To ensure that the network pays more attention to small cracks and improve the precision of crack detection tasks. With the help of fluorescent magnetic particle tech-nology, the crack display of bullet train parts is enhanced, which is helpful for the algorithm to identify the crack better. The experimental results show that the improved YOLOv5s target detec-tion algorithm proposed in this paper can improve the detection accuracy and has strong rationali-ty. The detection accuracy of mAP_0.5 is increased from 86.00% to 93.74% on the fluorescent mag-netic particle data set of the crack of the motor vehicle. This paper provides reference for the appli-cation of YOLOv5s target detection algorithm in fluorescent magnetic particle detection projects.