基于注意力机制的三维模型多视图分类算法研究

Research on Multi-view 3D Model Classification Algorithm Based on Convolutional Block Attention Module

对于三维模型的分类,根据其不同的模型类型有不同的分类方法:利用三维模型多视角的二维视图进行分类,利用点云格式的三维模型分类以及利用体素模型进行分类。由于三维模型结构的复杂性和不规律性,且因为二维图像的标记数据量远远大于三维模型,所以一般使用三维模型的多视图图像来对三维模型进行分类研究。其中选择了多视图卷积神经网络(Multi-View Convolutional Neural Networks,MVCNN),因其网络结构不复杂,耗时适中且具有不错的分类结果。并通过向MVCNN中引入注意力机制,使得图像的主要特征位置以及特点显著,能够更好地对三维模型的图像进行特征提取以及分类。实验结果表明,引入注意力机制的多视图三维模型分类方法较原方法,能够将三维模型多视图的分类准确率提高约3%。

For the classification of 3D model,there are different classification methods according to different model types:using multi-view of 3D model for classification,using point cloud of 3D model for classification and using voxel model.However,due to the complexity and irregularity of the structure of 3D model,and because the amount of label data of 2D image is much larger than 3D model,multi view image of 3D model is generally used to classify 3D model.Among them,mvcnn(multi view convolutional neural networks)is selected.Because of its uncomplicated network structure,moderate time consumption and good classification results.By introducing attention mechanism into mvcnn,the main feature positions and features of the image are significant,which can better extract and classify the 3D model images.The experimental results show that the attention module can improve the classification accuracy by about 3%compared with the original method.