Research For 3D Object Recognition Based On Convolutional Neural Network

Nowadays,with the blossom of Artificial Intelligence,deep learning methods are widely focused and applied,especially Convolutional Neural Network.As the priority research areas of Artificial Intelligence,computer vision has been developing fast in recent years.Missions using CNN with organism structure imitation;recognition,detection and segmentation to 2-D image domain;have found mature solutions.However,deep learning method research for 3-D data is still in a fledgling stage.Different from 2-D data,3-D data have a series of problems including limited collection means,deficient data size,complex organizational form,large space overhead,etc.These problems greatly challenge CNN methods which use 3-D data.Aiming at problems of data size and space overhead,networks need to not only enhance strong catching ability to structure features,but also control parameter scale and extract validate features adequately.Moreover,complex data organizational form makes research of 3-D data cannot accommodate all forms with only one method.Therefore,correlational researches are divided into multiple directions based on different data organizational forms.Currently,usual 3-D data forms include voxel,point cloud and multi-view.Focusing on voxel and point cloud,this thesis summarizes current deep learning methods based on these two forms in the first place.It afterwards designs feature-reorganization-based CNN with the peculiarity of voxel data and use short connection structure to improve feature availability for the better recognition result of voxel data.Finally,aiming at point cloud data,this thesis proposes an graph convolutional network structure with an amalgamation of multi-scale neighborhood characteristics for effectively catching structure features in irregular point cloud,and this model are amply compared with current method in experiment.Main researches in this thesis are summarized as follows:(1)Summarizing and analyzing CNN method pointing at voxel and point cloud: this thesis summarizes CNN method based on voxel and point cloud data.Researches,on one hand,analyze and summarize problems in networks which use voxel data,and analyze advantages and shortages of current methods;on the other hand,analyze main stream in point cloud data researches and compare with advantages and shortages of current methods.(2)Proposing voxel feature reorganization network and comparing it with main performance index of current methods: To solve problems of limited resolution ratio and texture deletion of voxel data,researches use short connection structure to improve characteristic multiplexing rate,simplify network parameters and prove networks can fully extract structure features of voxel data.Meanwhile,researches combine global average pooling to further restrain the over fitting phenomenon.By comparing with current voxel-based method,researches gain more excellent recognition result.(3)Proposing an graph convolutional network with an amalgamation of multiscale point cloud characteristics and compare it with mainstream methods: With the considering about disorder and irregularity,researches use graph convolution to unite local point set to extract structure features.In the meantime,networks with multibranch design fuse multiscale neighborhood characteristics and lead networks to focus on key points in point cloud structure by calibration technology,causing further improvement of networks effect.Finally,researches analyze effectiveness by using visualization means to make comparison with current methods.