基于RP-ResNet网络的抓取检测方法  

GRABING DETECTION METHOD BASED ON RP-RESNET NETWORK

作　　者：赵景波[1] 邱腾飞 朱敬旭辉 刘信潮 Zhao Jingbo;Qiu Tengfei;Zhu Jingxühui;Liu Xinchao(Qingdao University of Technology,Qingdao 266000,Shandong,China)

机构地区：[1]青岛理工大学,山东青岛266000

出　　处：《计算机应用与软件》2023年第3期210-216,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(51475251);山东省自然科学基金项目(ZR2013FM014);山东省高等学校科技计划项目(J12LN37);青岛市科技计划项目(15-9-2-109-nsh)。

摘　　要：针对目前机器人抓取中实时性差和小目标物体抓取效果不理想的问题,提出基于RP-ResNet网络的抓取检测方法。该方法以模型ResNet-50为基础,引入Region Proposal Network(RPN)网络与改进型空间金字塔结构,使用SENet机构增强通道注意力,充分提取不同深度语义信息,加强对高低层信息的利用,增加了抓取点检测的准确性。在Cornell数据集上进行实验验证,结果表明,RP-ResNet网络结构达到实验要求,能够提高物体检测的准确度与检测效率。Aimed at the problems of poor real-time performance and poor grasping effect of small target objects in robot grasping,a grasping detection method based on RP-ResNet network is proposed.Based on ResNet-50 model,this method introduced regional proposal network(RPN)and an improved spatial pyramid structure.The SENet mechanism was used to enhance channel attention.It fully extracted semantic information of different depths,enhanced the use of high-level and low-level information,and increased the accuracy of capture point detection.The experimental results on Cornell dataset show that the RP-ResNet network structure meets the experimental requirements and improves the accuracy and efficiency of object detection.

关 键 词：深度学习 区域建议网络 SENet结构 空间金字塔池化(SPP) 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术] TP242