Arbitrary size ratio canonical object classification based on Convolutional Neural Network

Abstract

Based on the research of radar target characteristics and the need for classification and recognition of targets, this paper presents the application of convolutional neural network(CNN) to the classification of canonical objects with non-fixed size ratio. First, the whole-angle RCS data of the canonical objects is preprocessed, then it is used for training CNN network, thus, the classification and recognition of five kinds of canonical objects with arbitrary size ratio(i.e., plate, dihedral, trihedral, cylinder, top-hat) are realized. The training accuracy is 99% and the validation accuracy is 97% on the data set simulated on canonical objects with different size ratios. This study can provide a research foundation and reference for the classification and recognition of complex objects.