Research On Plant Point Cloud Data Patching Algorithm Based On Deep Learning

As information technology pushes the world from traditional industries to the information society,agriculture has also entered a new era of digitalization and intelligence.The study of plant shape expression and phenotype analysis based on 3D point cloud technology is of great significance for promoting high yield and high quality production in agriculture.However,in the 3D reconstruction of plant seedlings,there are often factors such as unstable outdoor light,occlusion between leaves,limited camera field of view,and low scanning accuracy,which lead to noise,sparseness and lack of point clouds of plant seedlings,which affect the accuracy of plant phenotypic analysis.Therefore,studying the patching algorithm of plant point cloud data can assist the complete storage and non-destructive analysis of the three-dimensional shape of plants,which is of great significance to the study of plant phenotypes.This thesis took the point cloud of plant seedlings as the research object,and focused on the point cloud inpainting algorithms based on deep learning,which were two tasks of point cloud denoising and point cloud completion.Due to the unique shape characteristics of plant point clouds,the existing point cloud processing algorithms cannot well capture the characteristics of different plant organs,so that it is difficult to patch plant point clouds.To this end,this thesis focused on in-depth research on the attention mechanism and point cloud feature extraction network,and designed a denoising network and a completion network for plant point clouds.The main research and contributions are as follows:（1）Plant multi-angle image acquisition platform was designed and built.It used low-cost equipment to collect plant 3D point cloud data,and obtained a clean plant point cloud data through 3D reconstruction based on multi view and a series of point cloud preprocessing.Through experimental verification,the optimal number of image acquisition and camera installation scheme of the system were determined to reduce the time cost of point cloud reconstruction.（2）Plant point cloud denoising network based on point cloud attention feature extraction module（PAFM）was proposed.Firstly,it used a point cloud attention feature extraction module,which consists of a point cloud channel attention module and a point cloud space attention module to enhance the model’s ability to learn the local detailed features of point clouds.Besides,it integrated multi-scale feature vectors by aggregating feature representation vectors in different feature dimensions,which can prevent the loss of shallow features and enhance the ability to represent global shape and local geometric details.This experiments were carried out on the plant dataset,and the results shown that the network can effectively remove noise points near plants while maintaining the geometric features of plants.（3）Plant point cloud completion network based on multi-scale geometry-aware point Transformer（MGA-PT）was proposed.Firstly,it used the down-sampling feature extraction module to solve the problem of excessive input data size and single features.Then,the Multi-Scale Geometry-Aware Point Transformer is used to perform feature extraction and point cloud generation.In the process of integrating geometric information,it solved the problem of lack of inductive preference of the native Transformer,so that the ability of feature extraction was enhanced.Finally,the dual-path dense point cloud generation module was used to process the input part and the missing part separately.It can reduce the loss of input point cloud features which ensures the actual distribution differences of different plant organs.Compared with other models,this method performed well on the plant seedings data.The experimental results shown that it can effectively predict the point cloud of the missing parts of the plant.And the generated dense point cloud can effectively maintain the shape details of the input point cloud.