Robust Target Recognition Based On High-resolution ISAR Images

Containing essential information such as shape,structure,and size,high-resolution ISAR images of aerospace targets play a vital role in radar automatic target recognition.ISAR targets are usually non-cooperative,whose motion attitude,radar line of sight,imaging accumulation angle,etc.are unknown,resulting in unknown image domain deformations such as stretching,compression,and rotation and add great difficulties to effective feature extraction and classifier designing.At present,how to extract deformation robust features from high-resolution ISAR images and design effective classification methods have become key issues in the radar automatic target recognition community.In response to the above problems,this thesis utilizes advanced machine learning theories such as deep neural networks and conducts in-depth research on deformation robust recognition methods for high-resolution ISAR images.Specific contents include:deformation robust target recognition based on Trace features;deformation robust target recognition based on Deep Convolutional Neural Network(DCNN);deformation robust target recognition based on Spatial Transformer Network(STN).The related research provides theoretical and technical support for solving the difficult problem of ISAR target recognition with complex unknown deformations.The first part studies deformation robust target recognition based on Trace transformation.This method first calculates the target principal axis of the preprocessed ISAR image,and then extracts trace features in the local area near the principal axis,which are not sensitive to image domain deformation such as zooming,translation,and rotation,and finally utilizes the nearest neighbor classifier to achieve target classification.Finally,the ISAR image data set I containing seven types of aircraft is constructed,and the effectiveness of the method is proved through a series of deformation experiments.The second part studies deformation robust target recognition based on DCNN.Aiming at the problems that the deformation robust recognition method based on Trace transform cannot achieve end-to-end training and the recognition performance is poor in complex imaging environments,deformation robust target recognition based on DCNN is discussed.This method first builds a DCNN model,and then applies ISAR images with various deformation to perform end-to-end training and obtain the optimal network parameters.Compared with the Trace transformation,DCNN has achieved better recognition accuracy on aircraft data set I with seven types of airplanes.The third part studies robust target recognition based on STN.Aiming at the limited adaptation of DCNN to image deformation,a robust target recognition method based on STN is proposed.This method introduces STN into DCNN,estimates data-based affine parameters and adjusts the deformed image.For aircraft data set I with seven types of airplanes and aircraft data set II with five types of airplanes,the proposed method can achieve better recognition performance than Trace transform and DCNN.