Abstract: In industry,equipments for surface defect detection currently face performance bottlenecks due to limited computational resources of hardware.Orienting the trade-off between high efficiency and high accuracy,a surface-defect-detection model based on scaling Cross Stage Partial Network(CSPNet) was established.A YOLOv5s-P model that be scaled up and down for networks of different sizes was established based on the scaling algorithm of cross-stage partial network and YOLOv5s model.Specially,the neck network structure was CSP-ized to improve the feature extraction capability of the model.The SoftPool downsampling method was used to optimize network structure and parameters of the Spatial Pyramid Pooling(SPP) module,and the depthwise separable convolution was introduced to make the model lightweight while avoiding accuracy loss.Experiments showed that 96.1% mAP was obtained on DAGM 2007 surface defect data set,which improved the detection accuracy by 5% and reduced the parameter amount by 1.7% compared with the original model.At the same time,the detection speed was 4fps when deployed on the edge device Raspberry Pi4B.

Key words: deep learning, convolutional neural network, cross stage partial network, defect detection

摘要： 工业生产中,表面缺陷检测设备常因硬件计算资源有限而存在性能瓶颈。针对检测中高实时性与高精确度的权衡问题,建立了一种基于扩展跨阶段局部网络的表面缺陷检测模型。首先,基于扩展跨阶段局部网络算法和YOLOv5s模型,构建可上下缩放且适用于不同规模网络的YOLOv5s-P系列模型,将网络颈部CSP化提高模型特征提取能力;其次,采用SoftPool下采样方法对颈部空间金字塔池化(SPP)模块进行结构和参数优化,并引入深度可分离卷积使模型轻量化的同时避免精度损失。实验表明,在德国DAGM 2007表面缺陷数据集上获得96.1%的mAP,较原模型在检测精度上提升了5%、参数量降低了1.7%,同时部署到边缘设备Raspberry Pi4B上时检测速度为4fps。

关键词: 深度学习, 卷积神经网络, 跨阶段局部网络, 缺陷检测