基于改进YOLOv4算法的零件识别与定位

Recognition and Localization Method of Workpiece Based on Improved YOLOv4

针对多角度无序零件视觉识别难度大,定位精准性差等亟待解决的问题,提出一种基于改进YOLOv4的零件识别与定位方法。首先,采用AdaBelief(Adaptive"belief"stepsize)优化算法代替原有SGDM(Stochastic gradient descent with Momentum)优化算法,提高收敛速度和识别精度;其次,利用Canny边缘检测和Sklansky算法将原有预测边界框改进为凸包和最小外接矩形框,提高定位精度;最后,在制作的零件数据集上进行零件识别与定位实验。实验结果表明,改进YOLOv4在零件数据集上的测试准确率(Precision)达93.37%,提高了3.47%,F 1(参数α=1时Precision和Recall的加权调和平均)由94.32%提升至96.16%,被检测零件的定位结果从原有预测边界框缩小到可表示零件形状的凸包和最小外接矩形框,以及最小外接矩形框的角点坐标,零件识别精度和定位精准性上均优于原有YOLOv4,满足视觉引导下对零件精准识别与定位的要求。

Based on Yolov4 Object Detection,an improved method is proposed to solve the problems of low precision for multi-angle and disordered workpieces.Fristly,AdaBelief Optimizer(Adapt the stepsize according to the"belief"in the current gradient direction)is used to substitute SGDM(Stochastic gradient descent with Momentum)and improve the convergence speed and recognition precision of YOLOv4;secondly,Canny edge detection and Sklansky Algorithm(Finding the convex hull of any simple polygon)are used to calculate convex hull and minimum bounding rectangle box,which optimizes the original Bounding Box prediction and obtains more precise workpiece location;at last,Experi-mental results show that the improved YOLOv4 achieves 93.37%precision for workpiece dataset,improves by 3.47%,F 1(α=1,F 1-Measure)achieves 96.16%from 94.32%,the location results are convex hull and minimum bounding rectangle box that can represent the shape of the workpiece better,and the corner coordinates of the minimum bounding rectangle box,which can meet the precision requirement of workpiece recognition and localization in the field of industrial intelligent manufacturing.