改进YOLOv5s的小样本3D打印点阵结构表面缺陷检测

Surface defect detection of small sample for 3D printed latticestructures by improved YOLOv5s model

3D打印点阵结构已经广泛应用于航空航天、机械和建筑等行业,其表面缺陷分布不均匀且特征微弱,常常造成漏检和误检。针对这一问题,提出了一种YOLOv5s-PD模型。在该模型中,添加了XSPPF模块和空洞金字塔池化模块,提高了模型对不同缺陷特征的获取能力;针对3D打印点阵结构表面缺陷分布杂乱导致误检率高的问题,在YOLOv5s模型中加入了ECA模块;考虑到3D打印点阵结构表面缺陷尺寸信息无规律并且差异较大而导致的预测框与真实框间方向不一致,采用了SIoU损失函数。采用改进模型对制作的3D打印点阵结构表面缺陷数据集进行检测,结果表明:缺陷检测的召回率达到94.0%,平均精度mAP@0.5达到96.2%,所提出的改进算法可以实现对3D打印点阵结构表面缺陷自动检测。

3D printed lattice structures are widely used in aerospace,machinery,and construction industries,but their surface defect distribution is uneven and weak,often resulting in missed and false detection.To solve the problem,this paper proposes a YOLOv5s-PD model which adds the XSPPF module and the Atrous Spatial Pyramid Pooling module to improve the ability of the model to acquire different defect features.To address the high false detection rate caused by disordered defect distribution on the surface of the 3D printed lattice structure,the ECA module is added to the YOLOv5s model.The SIoU loss function is adopted to consider the inconsistency between the predicted frame and the real frame due to the irregular and large difference in the surface defect size information of the 3D-printed lattice structure.The improved model is employed to detect the surface defect dataset of lattice structures.Our results show that recall of defect detection is 94.0%,and the mAP@0.5 stands at 96.2%.Our proposed improved network achieves automatic identification of surface defects of 3D printed lattice structures.