Research Of Point Cloud Segmentation Algorithm Based On Feature Fusion

Point cloud semantic segmentation is the basis for understanding 3D scene.It has broad application in navigation and positioning,medical image analysis,mapping geographic information,and digital cultural relics protection.With the improvement of point cloud semantic segmentation neural network,the model is becoming more and more complex.This kind of models requires large GPU memory and segmentation time.So,how to use lightweight neural networks is of great significance for efficient and high-performance point cloud semantic segmentation.This thesis studies the lightweight point cloud semantic segmentation network and its application in medical images.The main contents are as follows:A lightweight point cloud semantic segmentation network Point ANL is designed.Firstly,the large-scale point cloud data is transformed into small point cloud datasets that can be accepted by the network through sampling and normalization.Next,by improving the non-local module that consumes time and memory,the asymmetric non-local module is designed to reduce the computational cost.At the same time,it is combined with the local module to obtain local and global features in order to enhance the feature representation of point cloud.And then the lightweight network is trained by using Focal Loss function such that the model can focus on hard samples quickly.The simulation experiment is based on the ISPRS remote sensing dataset.Compared with DAPNet network model,the present Point ANL network is 3.1 times faster.The total model parameters are reduced by 7.6 times,and the GPU running memory usage is reduced by 2.3 times while the accuracy is similar.A coarse-to-fine multi-organ segmentation model C2 FMOS is designed.The sparsity of point cloud is used to realize abdominal multi organ segmentation and reduce class imbalance.By converting dense voxels into sparse point clouds,C2 FMOS obtains effective shape representation and spatial information.The richer image information is obtained through target structure enhancement and image enhancement comparison.By adding the conditional task encoding module,C2 FMOS can efficiently segment multiple organs.For resolving the input and output imbalance,a new loss function is designed.Moreover,we proposed a visual network for point cloud semantic segmentation process.This network can show the processing process of network evolution.The experimental results show that C2 FMOS model achieves the highest overall performance on the seven organ segmentation datasets,with an average Dice of 76.96%,which is 1.32% higher than that of Do DNet,an average HD of 16.33 mm.Meanwhile,it is 3.17 mm lower than that of Do DNet.Compared to the state-of-the-art methods,the Dice on the binary datasets colon,lung and spleen shows an increase of 3.24%,5.14% and 2.13%,respectively.