3D形状体素的多尺度残差特征编码方法研究

Research of Multilevel Residual Features Coding Method Based on 3D Shape Representation

体素形状表达一直是三维形状表达方法中的重要研究内容。现有的基于深度学习的体素形状表达的特征编码学习是一个重要的研究问题。目前,体素的特征编码采用3DCNN,只利用了体素的层次特征。同时,计算的代价十分昂贵,需要占用大量的显存。更高精度的特征编码方法有待进一步研究。针对体素的特征编码问题,提出了一种基于多尺度残差特征的形状表达编码模型,学习更有效的特征编码。在此基础上在解码器中引入自注意机制,进一步提高解码模型的精度。此外,还探索研究了一种体素残差编码的深度分离卷积方法,在精度下降不大的条件下有效降低了模型的参数。在ShapeNet数据集上的实验表明,所提方法比3DCNN有更高的编码精度,能得到更好的实验结果。

Voxel shape representation has always been an important research content in 3D shape representation methods.The existing feature coding learning of voxel shape representation based on deep learning is an important research problem.At present,the feature coding of voxels adopts 3DCNN,and only the hierarchical features of voxels are used.At the same time,the cost of calculation is very expensive and requires a large amount of memory.More high-precision feature coding methods need to be further studied.Aiming at the feature coding problem of voxels,a shape representation coding model based on multilevel residual features is proposed to learn more effective feature coding.On this basis,a self-attention mechanism is introduced into the decoder to further improve the accuracy of the decoding model.In addition,a depthwise separatable convolution method based on voxel residual coding is explored,which effectively reduces the parameters of the model under the condition that the accuracy is not greatly reduced.Experiments of the ShapeNet dataset show that the proposed method has higher coding accuracy than 3DCNN,and can also get better experience results.