基于深度学习的三维模型检索研究
|
摘要
基于卷积神经网络在图像领域的发展,本文研究了将卷积神经网络应用到三维模型检索领域。对三维模型预处理后,选择6个投影角度把模型投影成6幅二维图像,利用提取的视图作为神经网络的输入,利用深度学习框架提取图像特征作为最终的模型描述符。之后通过比较2个模型多个视角的二维投影的相似度,如果两者间相似,则三维模型也是相似的,再取多维视图的相似度平均值得到2个三维模型的最终相似度,选择最终相似度最大的10个模型作为结果输出。充分利用二维图像领域性能优越的网络架构,并且存在海量图像数据供深度学习模型进行预训练。从检索结果看出,该方法能够得到较好的效果。
Based on the development of Convolution Neural Network in the field of image,this paper studies the application of Convolution Neural Network to 3 D model retrieval. After pre-processing the three-dimensional model,six projection angles are selected to project the model into six two-dimensional images. The extracted views are used as the input of neural network,and the image features are extracted by depth learning framework as the final model descriptors. Then the paper compares the similarity of the two models' two-dimensional projections from multiple perspectives. If the corresponding models are similar,then the threedimensional models are similar. The paper takes the average similarity of the multi-dimensional views to get the final similarity of the two three-dimensional models,and select 10 models with the maximum similarity as the output of the results. It makes full use of the network architecture with superior performance in the field of two-dimensional images,and there is a huge amount of image data for depth learning model pre-training. From the retrieval results,it is demonstrated that this method can get better retrieval.
Based on the development of Convolution Neural Network in the field of image,this paper studies the application of Convolution Neural Network to 3 D model retrieval. After pre-processing the three-dimensional model,six projection angles are selected to project the model into six two-dimensional images. The extracted views are used as the input of neural network,and the image features are extracted by depth learning framework as the final model descriptors. Then the paper compares the similarity of the two models' two-dimensional projections from multiple perspectives. If the corresponding models are similar,then the threedimensional models are similar. The paper takes the average similarity of the multi-dimensional views to get the final similarity of the two three-dimensional models,and select 10 models with the maximum similarity as the output of the results. It makes full use of the network architecture with superior performance in the field of two-dimensional images,and there is a huge amount of image data for depth learning model pre-training. From the retrieval results,it is demonstrated that this method can get better retrieval.
引文
[1]张晓东.三维模型的形状特征提取方法研究[D].北京:中国石油大学,2010.
[2]TANGELDER J W H,VELTKAMP R C.A survey of content based 3D shape retrieval methods[J].M ultimedia Tools and Applications,2008,39(3):441-471.
[3]GAO Yue,ZHANG Hanwang,ZHAO Xibin,et al.Event classification in microblogs via social tracking[J].ACMTransactions on Intelligent System Technology(TIST),2017,8(3):1-14.
[4]ZHAO Sicheng,CHEN Lujun,YAO Hongxun,et al.Strategy for dynamic 3d depth data matching tow ards robust action retrieval[J].Neurocomputing,2015,151(Part II):533-543.
[5]李朋杰.面向三维模型检索的特征提取算法研究[D].北京:北京邮电大学,2012.
[6]BIASOTTI S,CERRI A,BRONSTEIN A,et al.Recent trends,applications,and perspectives in 3D shape similarity assessment[J].Computer Graphics Forum,2016,35(6):87-119.
[7]GAO Yue,WANG Meng,TAO Dacheng,et al.3D object retrieval and recognition w ith hypergraph analysis[J].IEEETransactions on Image Processing,2012,21(9):4290-4303.
[8]GAO Yue,ZHEN Yi,LI Haojie,et al.Filtering of brand-related microblogs using social-smooth multiview embedding[J].IEEETransactions on M ultimedia,2016 18(10):2115-2126.
[9]ZHAO X,WANG N,ZHANG Y,et al.Beyond pairwise matching:Person reidentification via high-order relevance learning[J].IEEE Transactions on Neural Networks and Learning Systems,2018,29(8):3701-3714.
[10]孙志军,薛磊,许阳明,等.深度学习研究综述[J].计算机应用研究,2012,29(8):2806-2810.
[11]席晓聪.图像分类方法研究[D].济南:山东大学,2013.
[12]郭裕兰.三维形状数据的深度特征表示[J].CCF计算机视觉专委简报,2017(2):8-11.
[13]潘翔,张三元,张引,等.一种基于拓扑连接图的三维模型检索方法[J].计算机学报,2009,27(9):1250-1255.
[14]WANG Dong,WANG Bin,ZHAO Sicheng,et al.View-based3D object retrieval w ith discriminative view s[J].Neurocomputing,2017,252:58-66.
[15]HU Fan,XIA Guisong,HU Jingwen,et al.Transferring deep convolutional neural netw orks for the scene classification of highresolution remote sensing imagery[J].Remote Sensing,2015,7(11):14680-14707.
[16]YANG Songfan,RAMANAN D.Multi-scale recognition with DAG-CNNS[C]//2015 IEEE International Conference on Computer Vision.Santiago,Chile:IEEE,2015:1215-1223.
[17]杨荣,冯有前,袁修久.利用现有模型修复不完整三维模型[J].计算机辅助设计与图形学学报,2015,27(1):98-105.
[2]TANGELDER J W H,VELTKAMP R C.A survey of content based 3D shape retrieval methods[J].M ultimedia Tools and Applications,2008,39(3):441-471.
[3]GAO Yue,ZHANG Hanwang,ZHAO Xibin,et al.Event classification in microblogs via social tracking[J].ACMTransactions on Intelligent System Technology(TIST),2017,8(3):1-14.
[4]ZHAO Sicheng,CHEN Lujun,YAO Hongxun,et al.Strategy for dynamic 3d depth data matching tow ards robust action retrieval[J].Neurocomputing,2015,151(Part II):533-543.
[5]李朋杰.面向三维模型检索的特征提取算法研究[D].北京:北京邮电大学,2012.
[6]BIASOTTI S,CERRI A,BRONSTEIN A,et al.Recent trends,applications,and perspectives in 3D shape similarity assessment[J].Computer Graphics Forum,2016,35(6):87-119.
[7]GAO Yue,WANG Meng,TAO Dacheng,et al.3D object retrieval and recognition w ith hypergraph analysis[J].IEEETransactions on Image Processing,2012,21(9):4290-4303.
[8]GAO Yue,ZHEN Yi,LI Haojie,et al.Filtering of brand-related microblogs using social-smooth multiview embedding[J].IEEETransactions on M ultimedia,2016 18(10):2115-2126.
[9]ZHAO X,WANG N,ZHANG Y,et al.Beyond pairwise matching:Person reidentification via high-order relevance learning[J].IEEE Transactions on Neural Networks and Learning Systems,2018,29(8):3701-3714.
[10]孙志军,薛磊,许阳明,等.深度学习研究综述[J].计算机应用研究,2012,29(8):2806-2810.
[11]席晓聪.图像分类方法研究[D].济南:山东大学,2013.
[12]郭裕兰.三维形状数据的深度特征表示[J].CCF计算机视觉专委简报,2017(2):8-11.
[13]潘翔,张三元,张引,等.一种基于拓扑连接图的三维模型检索方法[J].计算机学报,2009,27(9):1250-1255.
[14]WANG Dong,WANG Bin,ZHAO Sicheng,et al.View-based3D object retrieval w ith discriminative view s[J].Neurocomputing,2017,252:58-66.
[15]HU Fan,XIA Guisong,HU Jingwen,et al.Transferring deep convolutional neural netw orks for the scene classification of highresolution remote sensing imagery[J].Remote Sensing,2015,7(11):14680-14707.
[16]YANG Songfan,RAMANAN D.Multi-scale recognition with DAG-CNNS[C]//2015 IEEE International Conference on Computer Vision.Santiago,Chile:IEEE,2015:1215-1223.
[17]杨荣,冯有前,袁修久.利用现有模型修复不完整三维模型[J].计算机辅助设计与图形学学报,2015,27(1):98-105.