基于改进YOLOv3算法的刀具表面缺陷检测

Tool Surface Defect Detection Based on Improved YOLOv3 Algorithm

目前,刀具的缺陷检测大都依赖人工完成,效率低、误检率高而且对工人视力有一定的损害。为此,开发自动化视觉检测平台和算法。针对刀具表面的缺陷检测任务,对相机、镜头和光源进行选型,设计了视觉检测实验平台。同时对图像降噪、增强等预处理算法进行研究。使用标准YOLOv3算法存在一定的漏检和检测框定位不准问题,从特征层、聚类算法和损失函数三方面对YOLOv3算法进行改进。实验表明,改进后的算法将AvgIoU从0.76提升到0.85。mAP从79.47%提升到84.02%,检测速度达到22FPS。搭建视觉检测平台、开发视觉检测算法,能够代替现有的人工来进行刀具的缺陷检测。

At present,most of the tool defect detection depends on manual completion,low efficiency,high error detection rate and a certain damage to the vision of workers.Therefore,the automatic visual detection platform and algorithm are developed.In this paper,the camera,lens and light source were selected for the defect detection task of the cutter surface,and a visual detection experimental platform was designed.At the same time,the image denoising,enhancement and other pre-processing algorithms are studied.The standard YOLOV3 algorithm has some problems of missed detection and inaccurate positioning of the detection frame from the three aspects of feature layer,clustering algorithm and loss function to improve YOLOV3 algorithm.Experimental results show that the improved algorithm increases AvgIoU from 0.76 to 0.85.mAP increased from 79.47%to 84.02%,and detection speed reached 22FPS.To build a visual inspection platform and develop a visual inspection algorithm,which can replace the existing manual tool defect detection.