基于改进YOLOv5s网络的机械零件目标检测

Target Detection of Mechanical Parts Based on Improved YOLOv5s Network

针对传统机械零件目标检测网络受到外界光照、小目标物体误检漏检的影响,检测精度不高实时性较差的问题,提出一种基于改进YOLOv5s网络的机械零件检测算法。首先,使用幽灵模块和幽灵瓶颈层替换原YOLOv5s的部分卷积结构和C3结构,降低浮点运算量(GFLOPs);其次,使用一个新的、高效的特征金字塔网络BIFPN,它在训练过程中不断更新分支的权重,从而融合不同分辨率的输入特征;最后,在主干网络后面加入一种高效注意力机制(CA),实现了高效地提取机械零件的深层特征。结果表明,改进后的YOLOv5s深度学习算法能够更加精准快速的识别机械零件的类别,并且具有较小的模型体积,方便在运算量低的小型设备上使用。

Aiming at the problem that the traditional mechanical parts target detection network is affected by external illumination and false detection and missed detection of small target objects,and the detection accuracy is not high and the real-time performance is poor,a mechanical parts detection algorithm based on the improved YOLOv5s model is proposed.First of all,the partial convolutional structure and C3 structure of the original YOLOv5s were replaced with ghost module and ghost bottleneck layers to reduce the amount of floating-point operations(GFLOPs).Second,a new,efficient feature pyramid network BIFPN is used,[JP]which constantly updates the weights of the branches during training to fuse input features of different resolutions.Finally,an Efficient Attention Mechanism,Coordinate attention(CA)is added to the backbone network to efficiently extract the deep features of mechanical parts.The results show that the improved YOLOv5s deep learning algorithm can identify the categories of mechanical parts more accurately and quickly,and has a small model volume,which is convenient for use on small devices with low computing volume.