Research On Low-resolution Radar Target Recognition Based On Deep Learning

Radar target recognition plays an important role in modern warfare.Since most active radars are low-resolution,it is of great military significance to study the low-resolution radar target recognition.Traditional radar target recognition has several drawbacks(e.g.,low recognition rate and poor generalization)in the conditions of limited training data,sample imbalance and other complex electromagnetic environment.The low-resolution radar target recognition technology based on deep learning is studied in this dissertation.The main research contents are as follows:Traditional radar target recognition adopts two-step method,which extracting features first and then recognizing based on features.Firstly,a new method of one-step radar target recognition based on Convolutional Neural Networks(CNN)is proposed.Sample data is directly taken as input in this method.CNN extracts the deep features of the sample data automatically and recognizes radar target in one-step.Secondly,in view of the low recognition rate of radar target with a limited training data,a radar target recognition algorithm based on Weighted Auxiliary Classifier Generative Adversarial Networks(WACGAN)is proposed.By using WACGAN to produce and automatically select the high-quality generated samples,the discriminator can extract deeper features of data.This method improves the recognition rate of the one-step radar target recognition under the condition of limited training data.Then,a radar target recognition algorithm based on piecewise loss function is proposed,which resolves the low recognition rate under the condition of unbalanced samples.CNN trains the network with appropriate loss function in different training periods.Several numerical experiments have been carried out to demonstrate the effectiveness of this algorithm.Finally,aimed at the utilization of unlabeled sample in radar target recognition,a semi-supervised radar target recognition algorithm based on GAN is proposed.Based on semi-supervised GAN model,the algorithm uses CNN as discriminator and makes adversarial learning between unlabeled samples and generated samples.The method improves the recognition rate of CNN on the basis of supervised learning of labeled samples.