基于改进特征融合和区域生成网络的Mask R-CNN的管件分拣研究  

Research on Pipe Sorting Based on Improved Mask R-CNN Using Feature Fusion and Region Generation Network

作　　者：韩慧妍 吴伟州 王文俊 韩燮[1,2,3] HAN Huiyan;WU Weizhou;WANG Wenjun;HAN Xie(School of Computer Science and Technology,North University of China,Taiyuan 030051,Shanxi,China;Shanri Key Laboratory of Machine Vision and Virtual Reality,Taiyuan 030051,Shanxi,China;Shanci Province's Vision Information Processing and Intelligent Robot Engineering Research Center,Taiyuan 030051,Shanxi,China)

机构地区：[1]中北大学计算机科学与技术学院,山西太原030051 [2]机器视觉与虚拟现实山西省重点实验室,山西太原030051 [3]山西省视觉信息处理及智能机器人工程研究中心,山西太原030051

出　　处：《应用科学学报》2023年第5期840-854,共15页Journal of Applied Sciences

基　　金：山西省自然科学基金(No.202303021211153);国家自然科学基金(No.62106238);山西省科技成果转化引导专项(No.202104021301055);山西省研究生创新项目(No.2021Y626)资助。

摘　　要：针对管件分割任务中各类管件区分难度大,光线和阴影对分割精度存在干扰等问题,提出了一种改进的掩膜区域卷积神经网络(mask region-convolutional neural network,Mask R-CNN)的管件分拣算法。通过增加低层特征图以改进特征融合网络,提高小型管件的识别率;根据管件尺寸比例改进区域生长网络的生成框,以加快模型收敛速度;增加通道和空间注意力模块,提升管件识别精度及掩膜效果。将改进后的Mask R-CNN用于四类管件的分拣任务,实验结果表明,改进后Mask R-CNN的掩膜检测平均精度均值(mean average precision, mAP)和平均召回率(mean recall, mRecall)值分别提高了1.5%和1.7%,对管件位置、类型和尺寸的判别能力更强,能够满足实际生产中机器人分拣管件的精度要求。In order to solve the problems of difficulty in distinguishing various pipe fittings and interference of light and shadow on segmentation accuracy,a pipe fittings sorting algorithm based on improved Mask R-CNN is proposed.The feature fusion network is improved by incorporating low-level feature map,improving the recognition rate of small pipe fittings.The generation box of area growth network is modified according to the size ratio of pipe fitting,so to accelerate the convergence rate of the model.Introduction of the channel and space attention module enhances the identification accuracy of pipe fitting and mask effect.The improved Mask R-CNN is applied to the sorting task of four types of pipe fittings,demonstrating increased mAP and mRecall values for mask detection(1.5% and 1.7% improvement,respectively).The robot exhibits enhanced capabilities in discriminating the location,type and size of pipe fittings,thereby meeting the accuracy requirements of sorting pipe fittings in actual production.

关 键 词：管件分拣 低层特征 区域生成网络 混合注意力机制 实例分割 

分 类 号：P315.69[天文地球—地震学]