Cloud Classification Of Large Scale Dataset Based On Convolution Neural Network

With the continuous development of 3D data acquisition technology in recent years,it is more and more convenient to obtain the cloud data of 3D scenic spots.Nowadays,in autopilot,unmanned ship cruise and urban three-dimensional modeling scenes,large field scenic spot cloud is used to solve the corresponding problems.Large field scenic spot cloud data contains a variety of rich information.How to effectively and accurately classify,segment and identify large field scenic spot cloud data has become a hot topic in the field of computer vision.At present,the deep learning network model framework is more and more mature in the field of image processing.Because the point cloud is a kind of three-dimensional irregular data,the three-dimensional convolutional neural network in deep learning directly processes the three-dimensional data of large scene,which has some problems such as low classification accuracy and complex calculation.In order to effectively solve the above problems.In this paper,two different methods are proposed for cloud classification of large scenic spots.The first method is based on blueprint separation convolution neural network,and the second method is based on binary neural network:(1)Aiming at the problem of low efficiency of cloud classification for large scenic spots,this paper proposes a cloud classification method for large scenic spots based on blueprint separation convolution neural network,which uses the blueprint separation convolution layer to replace the traditional convolution layer,so as to improve the accuracy of classification and optimize the time efficiency.The experimental results on Oakland dataset show that the proposed method is effective and the total classification accuracy is 98.0%.(2)In order to solve the problem of high computational complexity in large scale scenic spot cloud data set classification,this paper proposes a large scale scenic spot cloud classification method based on binary neural network.It adopts the binarization method of Libra-PB parameters,and processes the input two-dimensional feature images through the convolution neural network framework of Res Net18.Finally,the classification results are obtained.This paper compares three different binary neural network methods,The experimental results show that the classification accuracy of this method is 97.6% on Oakland dataset,and the training time is about 6 hours,which can effectively reduce the computational complexity and improve the time efficiency.