用于三维物体识别的分组卷积神经网络  

Grouping Convolutional Neural Network for 3D Object Recognition

作　　者：祁少华 杨国为[1] QI Shaohua;YANG Guowei(College of Electronics and Information,Qingdao University,Qingdao 266071,China)

机构地区：[1]青岛大学电子信息学院,山东青岛266071

出　　处：《青岛大学学报（工程技术版）》2022年第3期10-14,22,共6页Journal of Qingdao University(Engineering & Technology Edition)

基　　金：国家自然科学基金资助项目(62172229);江苏省自然科学基金资助项目(BK20211295)。

摘　　要：为寻找一些适合对视图进行分组的函数以对视图进行更合理的分组,本文主要对用于多视角三维物体识别的分组双池化卷积神经网络进行研究。构建了包括以卷积神经网络(convolutional neural network,CNN)为主的骨干网络、L2范数ReLU激活函数(L2-ReLU,LR)分组机制和双池化融合模块的DPCNN网络,LR分组机制能够利用L2范数和ReLU激活函数性质,计算视图的区分度得分,增强分组模块的可解释性,更合理的对多视图进行分组,在组内和组间均使用池化加权进行特征融合,并在主流的ModelNet40数据集上进行实验。实验结果表明,与其他最先进和最有代表性的方法相比,在分类任务上,实例精度提升了0.1%~5.87%,分类精度提升了1.57%~7.57%;在检索任务上,平均精度均值提升了3.31%~13.19%,同时证明LR分组机制对检索任务的性能具有明显的提升,说明本文方法能够达到先进的性能。该研究为多视角3D物体识别领域提供了新的方法思路。In order to find some functions suitable for grouping views in order to group views more reasonably,this paper mainly studies the grouping double pooling convolutional neural network for multi-view 3D object recognition.A DPCNN network including a backbone network dominated by convolutional neural network(CNN),a L2 norm-ReLU activation function(LR)grouping mechanism and a double pooling fusion module is constructed.The LR grouping mechanism can use the properties of L2 norm and ReLU activation function to calculate the discrimination score of views,enhance the interpretability of grouping modules and group multiple views more reasonably,pooling weighting is used for feature fusion within and between groups,and experiments are carried out on the mainstream ModelNet40 dataset.The experimental results show that compared with other most advanced and representative methods,the instance accuracy is improved by 0.1%~5.87%,and the classification accuracy is improved by 1.57%~7.57%;In terms of retrieval tasks,the mean average precision(MAP)is improved by 3.31%~13.19%.At the same time,it is proved that LR grouping mechanism can significantly improve the performance of retrieval tasks,which shows that the method in this paper can achieve advanced performance.This research provides a new method and idea for the field of multi-view 3D object recognition.

关 键 词：多视角 三维物体识别 分组机制 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]