基于改进YOLOv3_tiny算法的工件表面缺陷检测

Surface Defect Detection of Workpiece Based on Improved YOLOv3_tiny Algorithm

针对当前工件表面缺陷检测算法模型大、实时性差、难以在性能受限的嵌入式系统中运行的问题,以轻量级算法YOLOv3_tiny为基础,提出一种改进算法Defect_YOLOv3_tiny。首先,应用K-means算法生成适用于缺陷特征的先验框;其次,将空间金字塔池化(SPP)模块添加到网络,同时引入注意力机制优化缺陷检测精度;最后,增加算法检测分支以遏制微小缺陷漏检。实验结果显示,改进算法检测速度为81.41fps,平均检测精度mAP为94.7%。与原有YOLOv3_tiny相比,改进算法检测速度仅降低11.6%,mAP由89.8%提升至94.7%,表明改进后的检测算法满足嵌入式系统对缺陷检测的轻量化与准确性需求。

To address the problems of current workpiece surface defect detection algorithms with large models,poor real-time performance,and difficulty in running in performance-constrained embedded systems,an improved algorithm Defect_YOLOv3_ti⁃ny is proposed based on the lightweight algorithm YOLOv3_tiny.Firstly,the K-means algorithm is applied to generate a priori box suitable for defect features.Secondly,the spatial pyramid pooling(SPP)module is added to the network,and the attention mecha⁃nism is introduced to optimize the defect detection accuracy.Finally,the algorithm detection branch is added to contain minor de⁃fects missed.The experimental results show that the detection speed of the improved algorithm is 81.41fps,and the average detec⁃tion accuracy mAP is 94.7%.Compared with the original YOLOv3_tiny,the detection speed of the improved algorithm is only re⁃duced by 11.6%,and the mAP is increased from 89.8%to 94.7%,indicating that the improved detection algorithm meets the re⁃quirements of the embedded system for lightweight and accuracy of defect detection.