基于改进YOLOv5s模型的金属表面缺陷检测方法

Metal surface defect dectection method based on an improved YOLOv5s model

精密零件广泛应用于国防、医疗器械、航空航天、电子等产业,是实现“中国制造2025”宏伟蓝图的基础部件之一。然而,在精密零件的加工过程中,零件表面不可避免地会出现裂痕、点蚀表面、划痕等缺陷,导致零件质量和精度下降,进一步地影响设备性能和使用寿命。因此,零件表面缺陷的有效检测至关重要。为解决金属表面小目标缺陷漏检误检、精确度低的问题,文章设计了一种用于金属零件缺陷检测的改进YOLOv5s模型。首先,针对零件表面缺陷成像多为条纹状或点状特征、类型相对简单的特点,设计了一种轻量化的GhostC3模块,使用更少的参数,通过减少计算量生成特征图;其次,结合点蚀表面、轧制氧化皮等类型缺陷尺寸普遍偏小且分布不均匀的特点,设计一种基于双向特征金字塔网络BiFPN的Concat_BiFPN模块,充分利用BiFPN对不同尺度特征的融合能力,提高小目标检测的精度和稳定性;最后,充分利用SIoU(Spatial Intersection over Union loss)惩罚函数同时考虑目标形状、空间关系的特点,更好地捕捉不同尺寸的目标在图像中的位置关系,从而提高目标位置的精确性。在NEU-DET数据集上的消融实验和对比实验表明,该方法在参数量大幅降低的情况下获得了更高的多类别平均精度。

Precision components are widely used in various industries,including national defense,medical devices,aerospace,and electronics.They are among the fundamental elements for realizing the grand vision of“Made in China 2025.”However,during the processing of precision components,surface defects such as cracks,pitted-surfaces,and scratches are inevitable,which leads to a decline in the quality and precision of the components.Furthermore,these surface defects will affect the per-formance and lifespan of the equipment.Therefore,effective detection of surface defects is crucial.To address the issues of missed and false detections of metal surface defects and to improve detection accuracy,this paper proposes an improved YOLOv5s model for surface defect detection in metal com-ponents.First,considering that metal surface defects often present as streaks or spots which express as relatively simple types,a lightweight GhostC3 module is designed by generating feature maps using fewer parameters and resulting in reduced computational complexity.Second,given the characteristics of defects such as pitted-surfaces and rolled-in scale which are generally small and unevenly distribu-ted,a Concat_BiFPN module based on a bidirectional feature pyramid network was designed.This module makes full use of the ability of BiFPN to fuse features at different scales,and improves the ac-curacy and stability of small target detection.Finally,the SIoU loss function is utilized,which consid-ers the shape and spatial relationships of targets to better capture the positional relationships of differ-ent targets among the image,thus resulting in enhancing the precision of target localization.Ablation and comparative experiments on the NEU-DET dataset demonstrate that the proposed method achieves higher multi-class average precision with a significantly reduced number of parameters.