欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 计算机视觉 • 上一篇    下一篇

权值优化集成卷积神经网络及其 在三维模型识别中的应用

  

  1. (长春工业大学计算机科学与工程学院，吉林 长春 130012)
  • 出版日期:2019-12-31 发布日期:2020-01-20
  • 基金资助:
    国家自然科学基金项目(61303132，61806024)；吉林省教育厅“十三五”科学技术项目(JJKH20170574KJ)

Weighted Optimization Integrated Convolutional Neural Network and  Its Application in 3D Model Recognition

  1. (College of Computer Science and Engineering, Changchun University of Technology, Changchun Jilin 130012, China)
  • Online:2019-12-31 Published:2020-01-20

PDF (PC)

84

摘要： 三维模型应用广泛，如何有效地管理和分类这些数据库中的三维模型一直是人们 关注的问题。然而，由于不同三维模型之间的相似性难以测量，因而很难获得一种稳健且广泛 适用的三维模型分类算法。为此，提出了一种权值优化集成卷积神经网络(WOTCNN)模型，并 将其应用到三维模型的分类识别中。首先，获取三维模型的深度投影视图来最大限度地保留三维 模型的空间信息。然后，采用调整的 VGG 网络对各角度的深度投影图像进行训练并提取预测概 率值。最后，通过加权集成算法获得完整三维模型的最终分类结果。对 ModelNet10 及 ModelNet40 数据库的实验表明：三维模型的平均分类准确率达到 92.84%和 86.51%。在预测性能方面，该网 络优于普通的单卷积神经网络；在三维模型识别方面，其分类准确率能够得到显著提升。

关键词: 三维模型分类, 体素化, 卷积神经网络, 集成学习, 权值优化

Abstract: 3D models enjoy a popularity. It has always been our concern as to how to effectively manage and classify the 3D models in these databases. However, due to the similarity between different 3D models is difficult to calculate, it is difficult to obtain a robust and widely applicable 3D model classification algorithm. Thus a weighted optimization integrated convolutional neural network model is proposed and applied to the classification and recognition of 3D models. Firstly, the depth projection view of the 3D model is obtained to maximize the reserve of spatial information of the 3D model. Then, the adjusted VGG network is used to train the depth projection images from different angles and extract the predictive probability values. Finally, the final classification results of the complete 3D model are obtained by weighted ensemble algorithm. The experiments on ModelNet10 and ModelNet40 databases show that the average classification accuracy of the 3D model is 92.84% and 86.51% respectively. In terms of performance prediction, the network is superior to the ordinary single convolution neural network, and its classification accuracy can be significantly improved in 3D model recognition.

Key words:  3D model classification, voxelization, convolutional neural network, ensemble learning, weighted optimization