球面深度全景图表示下的三维形状识别-计算机辅助设计与图形学学报.PDFVIP

29 9 Vol.29 No.9 2017 9 Journal of Computer-Aided Design Computer Graphics Sep. 2017 球面深度全景图表示下的三维形状识别 1) 1) 1) 1) 2) 1)* 冯元力 , 夏 梦 , 季鹏磊 , 周 潇 , 曾 鸣 , 刘新国 1) (浙江大学CADCG 国家重点实验室 杭州 310058) 2) (厦门大学软件学院 厦门 361005) (xgliu@cad.zju.edu.cn) : 三维形状识别是近年来较为热门的研究方向, 针对其中的三维模型形状的表达方法和识别问题, 提出一种 多分支卷积神经网络下的三维模型识别方法. 该方法通过对三维模型进行球面深度投影得到球面全景图; 为了提高 识别精度, 将每个模型的球面全景图从多个角度展开, 创建多幅平面图像作为识别系统的输入; 识别系统使用多分 支的卷积神经网络, 并将多幅全景图进行整合分析, 最终得到一个三维模型的识别结果. 对三维模型进行分类和检 索的实验结果表明, 文中方法的识别效果优于近年来的前沿方法, 对三维模型进行检索的准确度甚至超过了多视图 识别方法. : 卷积神经网络; 形状识别; 球面投影; 全景图 : TP391.41 Deep Spherical Panoramic Representation for 3D Shape Recognition Feng Yuanli1), Xia Meng1), Ji Penglei1), Zhou Xiao1), Zeng Ming2), and Liu Xinguo1)* 1) (State Key Laboratory of CADCG, Zhejiang University, Hangzhou 310058) 2) (Software School of Xiamen University, Xiamen 361005) Abstract: 3D shape recognition is a hot topic in recent years. This paper proposed a 3D model recognition method with multi-branch convolutional neural network (CNN) to address the problems of 3D shape representa- tion and recognition. The inputs of the proposed method are spherical panoramas by deep spherical projection of 3D models; to improve recognition accuracy, the spherical panorama of the shape first unfolded on various orien- tations to produce multiple rectified images as input of recognition frame; the recognition system consists of a multi-branch CNN, which analyzes the panoramas as a whole to produce the final recognition result. The experi- ment results of retrieval and classification on various of 3D dataset showed t