摘要
为了进一步提高三维模型的识别精度,提出了一种基于深度卷积神经网络的三维模型识别方法。将点云数据通过占用网格规范化计算转化为二值3D体素矩阵,通过附加正则化项的随机梯度下降算法提取体素矩阵的特征,再通过共享权重的旋转增强对训练集进行数据增广并以此对模型标签进行预测。实验结果表明,该算法在公开数据集ModelNet40及悉尼城市模型数据集上的识别精度均达到85%左右。与基于同类机器学习的三维模型识别算法相比,在相同训练数据集上该方法网络训练时间短,在相同测试数据集上模型识别准确率高,检索速度快。提出的体素占用网格模型的深度卷积神经网络,可以实现三维点云模型数据集及规范化体素模型数据集的识别和分类工作。
A 3 D model recognition algorithm based on deep convolution neural network is proposed to improve the recognition accuracy for 3 D models.Firstly,the point cloud data is transformed into binary 3 D voxel matrix by grid-based normalized computation.Secondly,the feature of voxel matrix is extracted by random gradient descent algorithm with regularization term.Finally,the training set is augmented by rotation enhancement with shared weights,and the model labels are predicted.The recognition accuracy of this algorithm on the public dataset ModelNet40 and the Sydney City Model dataset is about85%.Compared with the similar model recognition algorithm based on machine learning,our network has shorter training time on the same training data set and higher model recognition accuracy on the same test data set,and the retrieval speed is fast.The deep convolution neural network based on voxel occupancy grid model proposed in this paper can effectively realize recognition and classification of 3 D models,including point cloud dataset and normalized voxel model dataset.
作者
杨军
王亦民
YANG Jun;WANG Yimin(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,P.R.China)
出处
《重庆邮电大学学报（自然科学版）》
CSCD
北大核心
2019年第2期253-260,共8页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金(61462059)~~
关键词
卷积神经网络
占用网格
体素化
随机梯度下降
旋转增强
convolutional neural network
occupancy grid
voxelization
stochastic gradient descent
rotation enhancement
