基于注意力机制的单阶段抓取检测研究

Research on single-stage grab detection based onattention mechanism

为了提高机器人抓取算法的检测性能,结合卷积神经网络(CNN)和深度学习理论基础,设计了一种基于Inception模块和注意力机制SENet网络的轻量级卷积模块SE-Inception-Lite,并对SE-Inception-Lite模块进行了残差机制算法融合。在此基础上,设计了基于SE-Inception GraspNet抓取网络的机器人多模态抓取检测算法,并在康奈尔抓取数据集上进行验证。实验结果表明:本文算法在图像分割数据和对象分割数据中的准确率分别为96.5%和96.3%,检测速度为22 ms。本文算法的抓取检测精度比同类算法更高更快。本文网络的网络参数也较少,算法运行需要的计算机资源更少,对机器人抓取检测的推广起了重要作用。

In order to improve the detection performance of robot grasping algorithm,a lightweight convolution module,SE-Inception Lite,based on the Inception module and the attention mechanism SENet network,is designed by combining the theoretical basis of convolutional neural network(CNN)and deep learning.The residual mechanism algorithm fusion of SE-Inception-Lite module is implemented.On this basis,a robot multimodal grasp detection algorithm based on the SE Inception Graspnet grasp network is designed and verified on the Cornell grasp data set.The experimental results show that the accuracy of the proposed algorithm in image segmentation data and object segmentation data is 96.5%and 96.3%,respectively,and the detection speed is 22 ms.The grasping detection precision of this algorithm is higher and faster than similar algorithms.The network parameters are also less,and the algorithm operation requires less computer resources,which plays an important role in the popularization of robot grasp detection.