Research And Implementation Of Point Cloud Multi-Scale Classification And Segmentation Network Model Based On Deep Learning

At present,deep learning networks are widely used in shape classification and component segmentation of point cloud models.However,although these network models have excellent performance,they often only focus on point cloud processing within a certain scale range,lack of versatility,and perform poorly when encountering some point cloud data with large scale spans.There are mainly two reasons for this problem.First of all,the classic deep learning point cloud processing model has a single sampling method and lacks flexibility in the face of multi-scale point cloud raw data,which leads to deviations in data collection and utilization,which fundamentally limits the performance of the network model.Secondly,the mainstream point cloud processing deep learning model lacks the ability to comprehensively utilize contextual information from both global and local perspectives,resulting in the model being prone to problems such as suspended animation and smooth transition when processing multi-scale point cloud data,which further limits the network model in Performance in whole and part detail.Therefore,in order to solve this problem,new and more flexible point cloud sampling methods need to be explored to improve the sampling ability of raw point cloud data.At the same time,it is necessary to design a deep learning network model that can more accurately process multiscale point cloud data,and comprehensively utilize contextual information from both global and local perspectives to better cope with the diversity brought by more multi-scale point cloud data in the future.and complexity challenges.In this paper,through the analysis of the existing deep learning network model in the field of point cloud processing,the current mainstream network model has a single sampling method,low efficiency,and difficulty in adapting to the needs of multi-scale point cloud processing.An efficient point cloud multi-scale classification and segmentation network model PMNN(Pointcloud multi-scale classification and segmentation network model)is proposed.It is mainly composed of adaptive point cloud sampling,triple hybrid attention mechanism,and adopts feature pyramid network structure.Among them,the triple hybrid attention mechanism includes spatial,channel and local area attention mechanisms,and is constructed based on dynamic graph convolution to optimize the processing process of side convolution and improve the feature extraction effect and the ability of the network.With this approach,we are able to more comprehensively utilize the contextual information in point cloud data and improve the accuracy performance of point cloud objects in detail parts.And it is proved that PMNN can effectively improve the accuracy and accuracy of point cloud processing on multiple experimental data sets of different scales.Finally,the impact of each module under the model on the overall performance is further verified by ablation experiments.