Deep Learning Classification Of 3D Point Sets Based On Siamese Network

As the foundation and the core of the computer vision and image processing,the classification of 2D images has achieved considerable development with the continuous progress of deep learning in recent years.We live in a 3D space,feeling and touching the surrounding 3D objects at all times.Compared with a 2D image,a 3D object adds a dimension,which brings more visual information,geometric information and position information.However,it is this extra dimension that makes the volume of the network grow exponentially when the convolutional neural network is applied in the 3D field.And neither the storage nor the amount of calculation can be borne by the existing hardware equipment.Moreover,in real life,unlike mobile phones that can take 2D pictures at any time,the equipment for acquiring 3D data has not been popularized.And the obtained 3D data have the characteristics of small quantity and poor quality,which obviously cannot use convolutional neural network for classification directly.In order to realize the recognition of 3D point cloud objects by deep learning,this paper proposes a network using Siamese network to solve the problem of point cloud recognition.It can solve the overfitting problem of point cloud data obtained in real scenes and the poor performance of the stacked convolutional neural network under small-scale data sets.The main work of this paper is as follows:(1)This paper compares the effects of the stacked convolutional neural networks and Siamese networks in different data sets in the field of image classification.And the advantages of Siamese network are demonstrated in many cases,such as fewer data sizes,more sample categories but fewer samples per category,and the difference between different types of samples is small,etc.Siamese network is a kind of network to measure the similarity of two samples.Different from the stacked convolutional neural networks,which require large-scale training sets and accumulated depths,Siamese network is still possible to make a more accurate match between samples of the same category in the case of small-scale data sets with small inter-class differences.(2)Data normalization and data augmentation are used to improve the accuracy of point cloud data classification of Point Net under real scenes.Point Net directly applies the convolutional neural network on the point cloud to realize the classification of 3D objects,and solves the problem that the point cloud is not fixed in number and point order,as well as the loss and noise of some points.However,Point Net does not recognize it well with point cloud samples taken in real scenes.In this paper,the reasons are analyzed through comparative experiments.And the classification performance of Point Net in real scenes is improved by normalizing the input point cloud and using the simulated real-world point cloud sample to augment the data set.(3)This paper proposes a 3D point cloud classification and recognition algorithm based on Siamese networks.Since the hardware equipment for acquiring 3D data is not mature enough,the amount of point cloud samples that can be obtained in reality is very small,and even cannot satisfy the training of normal convolutional neural networks.This paper designs a network that combines Siamese network that can solve small-scale sample training with Point Net for 3D point cloud classification,namely Point Net based on Siamese Net,to achieve complementary advantages.As a solution to the recognition problem of small-scale point cloud samples,it further improves the accuracy of point cloud recognition in real scenes.