【摘要】 针对传统装配机器人工件自动识别速度慢、精度低等问题，提出一种基于改进YOLOv4的工件识别方法。首先，采用简化Moblienetv3替换YOLOv4框架的特征提取网络；其次，将YOLOv4深层网络的标准卷积层替换为深度可分离卷积层；最后，增加104×104的特征检测尺度，并使其与另外三个尺度(52×52,26×26,13×13)进行融合。搭建了螺栓、螺母、轴等典型工件的图像数据采集平台，构建训练样本集并进行网络训练。测试结果表明，在工件识别任务中，与传统YOLOv4算法相比，网络参数规模减小87.10%,检测速度提高39.86%,初步满足实际生产中工件自动检测需求。更多还原

【Abstract】 In this paper, a novel workpiece recognition method was proposed based on improved YOLOv4.The proposed approach aims to solve the problem of slow speed and low detection accuracy of traditional assembly robots.Firstly, the feature extraction network of the YOLOv4 framework was replaced by an simplify Moblienetv3;Secondly, the standard convolutional layer of the YOLOv4 deep network was replaced by the depthwise separable convolutional layer; Finally, the feature detection scale of 104×104 was added and fused with the other three scales(52×52,26×26,13×13).AN image data collection platform for typical workpieces such as bolts, nuts and shafts was built, a training sample set was constructed and network training was carried out.The comparison test results show that compared to the traditional YOLOv4 algorithm, the network parameters scale is reduced by 87.10%,the detection speed is improved by 39.86%.The results can meet the requirement of automatic detection of the workpiece in actual production.更多还原