Research On Lunar Terrain Classification Algorithm

With the development of lunar exploration projects,the classification of lunar terrain is particularly important.The purpose of lunar terrain classification is to measure whether the lunar rover can pass the current terrain smoothly.This article mainly starts from the perspective of non-geometric traversability,constructing an automated lunar terrain classification system,so that the lunar rover can automatically detect the traversability of the current terrain,and provide a guarantee for the cruise mission.First,for terrain classification tasks,previous methods used a single scale or a single model to extract image features,high-and low-resolution networks to extract image features,and networks with no relationship between channels.These methods will lead to insufficient feature extraction.Therefore,the classification accuracy will be reduced.The samples in the terrain classification task are different from the samples in other image classification tasks.The difference between samples in the terrain classification task is smaller than other imagelevel classification tasks.And in the terrain classification,the color of each sample is similar.Therefore,we need to maintain the high resolution of the features and establish the interdependence between channels to highlight the image features.This kind of network can improve the accuracy of classification.Secondly,in order to overcome these challenges,this paper proposes a lunar terrain classification algorithm using a deep integrated network.We optimized the activation function and structure of the convolutional neural network to better extract the fine features of the image and infer the terrain category of the image.In particular,this article has made some contributions: establishing interdependencies between channels to highlight features,and maintaining high-resolution representations throughout the process to ensure the extraction of fine features.Multi-model collaborative judgment can make up for the shortcomings of a single model structure design,make the models form a competitive relationship,and improve accuracy.On our data set,the overall classification accuracy of this method reaches 91.57%,and the accuracy is higher on some terrains.