Research On 3D Point Cloud Shape Classification And Semantic Segmentation Based On Attention And Deep Networks

Technologies processing 3D point cloud,such as shape classification,sematic segmentation and so on,have important research value and broad application prospects in the field of computer vision.The 3D point cloud classification and segmentation network based on deep learning is one of the popular research directions of 3D point cloud processing.Because the disordered point cloud data is unstructured and numerous along with noise interference,it’s so difficult to extract features that there are still challenges to improve the performance of 3D point cloud deep networks.As a feature selecting method,attention mechanism is widely used in natural language processing,2D image processing and other fields,and can greatly improve the performance of the model.Therefore,it has become a common approach in model construction to design modules based on attention mechanism in recent years.Based on the deep dive into the attention mechanism and deep neural network,this paper conducts research on 3D point cloud shape classification and semantic segmentation methods based on attention mechanism and deep neural network.The main work and innovations of this paper are as follows:1)In order to improve the robustness of the point cloud shape classification network against noise interference,an independent block called competitive attention fusion block is designed,which can be flexibly transferred to different point cloud network architectures.This block enhances the representation of the intermediate features from the point cloud network by competitively fusing two different attention weights.This block contains two independent sub-blocks to obtain different attention weights: Firstly,the multi-layer feature squeeze and excitation block,derived from the attention mechanism in 2D image processing,is used to focus on the intermediate features of different levels.Secondly,the feature inner connection self-attention block,introduced from the self-attention mechanism in natural language processing,is meant to measure the internal similarity of intermediate features.The two sub-blocks demonstrate the importance of each feature channel from different perspectives and implement the selection in feature channels through a competitive fusion mechanism.Finally,the structure of the point cloud classification network embedded with the competitive attention fusion block is described.Experiments of shape classification,robustness,comparison,and ablation were performed on the Model Net40 dataset.The results show that the competitive attention fusion block can be embedded in the classic network Point Net++ and the advanced network Point ASNL.While achieving higher classification accuracy,the robustness of the network is significantly improved against a variety of different noises.Compared with traditional anti-noise methods,this method has better anti-noise performance and less time complexity.2)For the sake of improving the feature extraction capability of the point cloud semantic segmentation network,a point cloud Transformer cell is designed.This cell is based on the attention mechanism used in Transformer in the field of machine translation,and aims to learn the association between the coordinate space and the feature space of the sampled points,aiming to achieve the fusion of attention features from different feature subspaces.The point cloud Transformer cell,which is embedded into the set abstraction layer and feature propagation layer of the semantic segmentation network,is meant to extract global and local features within the layer and then fuse them with other features from the same layer,resulting benefits like enhancing the semantic context,augmenting representation capability of feature and improving the performance of point cloud semantic segmentation network.Finally,a point cloud semantic segmentation network is constructed which integrates a point cloud Transformer cell and a competitive attention fusion block.The part semantic segmentation experiment was carried out on the Shape Net Part datasets,and the large-scale indoor scene semantic segmentation experiment was performed on the S3 DIS datasets.The results showed that the point cloud semantic segmentation network,fused with the point cloud Transformer cell and the competitive attention fusion module,achieves higher accuracy in semantic segmentation tasks and exceeds the benchmark network in the segmentation performance on multiple semantic objects.By conducting two ablation experiments,it is proved that the proposed method can increase the feature extraction ability of the point cloud network and improve the segmentation accuracy of the network for multiple semantic targets.Finally,the visualization results have demonstrated superiority of the proposed method on semantic segmentation tasks.