基于改进YOLOv3深度卷积网络的竹垫智能装配方法

针对我国竹垫人工组装劳动强度大、效率低的问题,提出一种基于改进YOLOv3深度卷积网络的竹垫智能装配方法,通过智能识别、定位、摆放和组装实现卡扣式竹垫的装配生产模式。YOLOv3深度卷积网络的改进包括:首先通过在原YOLOv3的特征提取网络Darknet-53中加入DenseNet密集型连接网络结构,提高特征提取能力;再根据实际检测需求减少网络预测尺度,提升检测精度;最后采用K-means聚类算法对YOLOv3中的初始锚点框参数进行优化,以加快识别速度。笔者还设计了试验机进行装配生产试验。该试验机使用CCD相机采集不同图案的竹垫样品图像,对竹垫样品图像进行识别定位处理,获得目标竹垫零件的位置和颜色信息,然后控制摆放系统对目标竹垫零件进行吸取摆放,再启动传送机构输送竹垫零件,最后通过组装系统实现对目标竹垫零件的组装,实现了竹垫检测、摆放、传送、组装自动化。试验结果表明,视觉系统在GPU和CPU下识别定位竹垫零件的平均时间为16.7和105.3 ms,识别均值平均精度M_AP为99.86%,平均组装一行竹垫零件的时间为24.63 s,验证了本方法的可行性。

Bamboo mat intelligent assembly method using improved YOLOv3 deep convolutional network

Aiming at solving problems of high labor intensity and low efficiency of manual assembly of bamboo mats in China,an intelligent assembly method of bamboo mats using the improved YOLOv3 deep convolutional network was proposed to realize the intelligent assembly of bamboo mats,reduce the labor intensity in the process of bamboo mats assembly,and lower the production cost.Through intelligent identification,positioning,placement,and assembly,the production model of the snap-on bamboo mat was realized.The improvements of the YOLOv3 deep convolutional network include:firstly,the DenseNet dense connection network structure was added to the original YOLOv3 feature extraction network Darknet-53 to improve the feature extraction ability.Then the network prediction scale was reduced according to the actual detection requirements to improve the detection accuracy.Finally,the K-means clustering algorithm was used to optimize the initial anchor frame parameters in YOLOv3 to speed up the recognition speed.This study also designed a test machine to assembly the production test.Firstly,the CCD camera was used to collect the bamboo mat sample images of different drawings,identify and locate the bamboo mat sample images,and obtain the position and color information of the target bamboo mat parts.Then,the placement system was controlled to absorb and place the target bamboo mat parts.The system included a rectangular coordinate machine arm and a vacuum adsorption end effector.It was ensured the accurate positioning of the mechanical arm and the rapid absorption and placement of the bamboo mat parts by the vacuum adsorption end effector,then started the transmission mechanism to transport the bamboo mat parts,and finally realized the assembly of the target bamboo mat parts through the assembly system.The device visual recognition adopted the improved YOLOv3 depth convolution network to realize the speed and accuracy of recognition and positioning.The test results showed that the average recognition time of vision system on GPU and CPU was 16.7 and 105.3 ms,respectively.The average recognition accuracy M_AP was 99.86%and the average assembly time was 24.63 s,from which the feasibility of the proposed method was verified.