基于改进的YOLO v3的工件识别方法研究

Research on Identification Method of Workpiece Based on Improved YOLO v3 and Deep Separable Convolutional Networks

针对传统识别方法受工件的摆放角度、位置影响较大,鲁棒性差的问题,提出了一种基于改进的YOLO v3的工件识别方法。首先,采用深度可分离卷积网络对Darknet网络进行结构改进,减少了计算复杂度,提高了检测速率;其次通过K-means聚类方法对数据集参数进行聚类,得到了更适合工件识别的预测框,确定了预测框的参数;然后运用了图像增强的方法对采集的图像集进行处理,扩充了训练样本;最后在训练迭代后,采用多种评价指标联合评价的方法,将所设计的算法与多种检测算法进行对比。实验结果表明,基于改进的YOLO v3的工件识别方法在测试集的准确率达96.5%,召回率达93.4%,识别速率达63fps,更能满足工业生产中工件识别的需求。

Aiming at the problem that the traditional identification method is greatly affected by the placement Angle and position of the workpiece,and the robustness is poor,a workpiece identification method based on the improved YOLO v3 is proposed.Firstly,a deep separable convolutional network is used to improve the structure of Darknet network,which reduces the computational complexity and improves the detection rate.Secondly,k-means clustering method is used to cluster the data set parameters,and the prediction box more suitable for workpiece identification is obtained,and the parameters of the prediction box are determined.Then the image set is processed by image enhancement method to expand the training sample.Finally,after training iteration,the method of joint evaluation of multiple evaluation indexes is adopted to compare the designed algorithm with various detection algorithms.Experimental results show that the workpiece identification method based on the improved YOLO v3 has an accuracy rate of 96.5%,recall rate of 93.4%and identification rate of 63fps in the test set,which can better meet the requirements of workpiece identification in industrial production.