基于注意力机制的实时性抓取检测算法  

Real-time grab detection algorithm based onattention mechanism

作　　者：郝宸 田瑾[1] 韩华[1] 吴飞[1] 李洪芹[1] HAO Chen;TIAN Jin;HAN Hua;WU Fei;LI Hongqin(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《传感器与微系统》2022年第1期131-134,共4页Transducer and Microsystem Technologies

基　　金：上海市科委重点科技项目(18511101600);上海工程技术大学检测技术与自动化装置学科学位点建设项目(19XXK003)。

摘　　要：针对在工业环境中,机器人的抓取精准度受到物体不同的姿态和角度等影响的问题,提出一种基于注意力机制的实时性逐像素网络抓取检测方法。首先,以轻量级网络—全卷积网络(FCN)模型为基础进行逐像素提取抓取框位置及抓取角度;其次,提出了一种新型的注意力机制—DAM,将其整合到轻量级网络中,自动学习输入图像的显著特征以减少局部信息的丢失。与常用的滑动窗口方法相比,本文模型训练时损失率下降更快以及不容易过拟合。最后,在Cornell数据集上进行实验,结果表明:所提方法在物体抓取的准确性和实时性方面有很大提高,达到实时性抓取检测要求。Aiming at the problem that in industrial environment,grasping accuracy of robot is affected by different postures and angles of objects,a real-time pixel-by-pixel network grasping detection method based on attention mechanism is proposed.Firstly,the method uses lightweight network fully convolutional networks(FCN)model as a basis to extract the position of grasping box and angle of grasping.Secondly,a method uses a new type of attention mechanism double attention model(DAM)is proposed and integrates it into the lightweight network.Automatically learn the salient features of the input image to reduce the loss of local information.Compared with the commonly used sliding window method,the proposed model drops faster in training loss and is not easy to overfit.Finally,experiments on the Cornell dataset show that the accuracy and real-time performance of the method is greatly improved in object grasping to meet the requirements of real-time grasping detection.

关 键 词：像素检测 实时性 注意力机制 轻量级卷积神经网络 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]