基于多视图和注意力推荐网络的三维物体识别方法  

3D Object Recognition Method Based on Multi-view and Attention Recommendation Network

作　　者：张满囤 权子洋 师子奇 刘川 申冲 吴清 田琪 ZHANG Mandun;QUAN Ziyang;SHI Ziqi;LIU Chuan;SHEN Chong;WU Qing;TIAN Qi(College of Artificial Intelligence and Data Science,Hebei University of Technology,Tianjin 300130,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300130

出　　处：《郑州大学学报（理学版）》2023年第1期57-63,共7页Journal of Zhengzhou University:Natural Science Edition

基　　金：河北省自然科学基金项目(F2019202054)。

摘　　要：传统物体识别方法是从单一图像中通过人工提取图像特征,存在成本高、质量低等问题。针对上述问题,提出一种基于多视图和注意力推荐网络的三维物体识别方法,多视图很好地保留了物体在局部和全局上的特征;注意力模块可以有效地对视图上关键的特征聚焦,忽略无关或干扰特征。该方法利用一组多视图作为输入数据,通过卷积神经网络端到端提取物体特征,在卷积层加入注意力模块,实现视图关键区域的定位和剪裁,将处理后的视图送入另外一个卷积层,两个相同卷积操作提取的特征在池化层聚合,利用稀疏表示分类器对特征描述子进行分类识别。通过两个公开数据集的实验表明,所提算法对物体图像的识别准确度优于传统算法。Traditional object recognition methods of extracting image features manually from a single image had the problems of high cost and low quality. To solve the above problems, a 3 D object recognition method based on multi view and attention recommendation network was proposed. Multi view retained the local and global features of objects well;the attention module could effectively focus on the key features in the view and ignore the irrelevant or interfering features. In this method, a group of multiple views were used as input data;object features were extracted end-to-end through convolution neural network;attention module was added to the convolutional neural network to realize the positioning and clipping of key areas of the view;the processed view was sent to another convolution layer;and the features extracted by two identical convolution operations were aggregated in the pool layer. Finally, the sparse representation classifier was used to classify and recognize the feature descriptors. The experimental results on two public data sets showed that the recognition accuracy of the proposed algorithm was 10% higher than that of the traditional algorithm.

关 键 词：三维物体识别 多视图 注意力模块 卷积神经网络 稀疏表示分类器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]