基于改进YOLOv7-tiny的工业金属表面缺陷检测方法

Industrial metal surface defect detection method on improved YOLOv7-tiny

针对传统检测方法对于工业金属表面缺陷检测存在效率低,难以适应工业生产的缺点,以及目前YOLOv7-tiny算法中容易造成漏检等问题,该文提出了一种改进YOLOv7-tiny的工业金属表面缺陷检测方法。通过在主干网络中引入注意力机制相关的CBAM模块,并且使用SiLU激活函数对原有的LeakyReLU激活函数进行替换,即替换CBL模块为CBS模块。在多尺度特征融合网络中增加了一个更大尺度的特征图进行特征融合,并增加一个相应尺度的检测头。使用EIoU作为损失函数来提高检测精度。GC10-DET工业金属表面缺陷数据集上的实验结果表明,改进后算法的平均检测精度相比原算法提高了4.29%,有效改善了漏检等情况,提高了检测精度。

In view of the shortcomings of traditional detection methods for industrial metal surface defect detection,such as low efficiency and difficulty in adapting to industrial production,and the current YOLOv7-tiny algorithm is prone to missing detection and other problems.This paper presents an improved YOLOv7-tiny method for detecting defects on industrial metal surfaces.By introducing the attention-mechanism related CBAM module into the backbone network and replacing the original LeakyReLU activation function with the SiLU activation function,the CBL module is replaced with the CBS module.In the multi⁃scale feature fusion network,a larger scale feature map is added for feature fusion,and a corresponding scale detection head is added.The EIoU function is used as a loss function to improve the detection accuracy.The experimental results on GC10-DET data set of industrial metal surface defects show that the average detection accuracy of the improved algorithm is 4.29%higher than that of the original algorithm,which effectively improves the cases of missing detection and improves the detection accuracy.