The Research Of Gan-based Scheme For Radar Spectrogram Augmentation

The classification of human motion is a research hotspot in the field of computer vision.In recent years,it has received extensive attention from academia and engineering.With the continuous development of deep learning,the use of deep learn ing models has made great progress in classifying human activities.However,the current classification of human activities is based on natural im-ages,and is easily affected by factors such as illumination and air disturbance,affecting the classification effect.The use of micro-Doppler radar spectrograms to classify human activities can effectively r-educe the influence of external fac-tors and improve the accuracy and robustness of the model.However,due to the high cost of micro-Doppler radar spectrograms acquisition,the data set samples are too small,which greatly limits the application of deep learning in micro-Doppler radar spectrograms.In order to solve this problem,the paper proposes data enhancement of micro-Doppler radar images from both quantity and qual-ity.First,spectrograms generation using DCGAN.In this paper,we use DC-GAN to combat the distribution of micro-Doppler radar spectrograms,and gen-erate new spectrograms based on the learned distribution.The generated spec-trograms are mixed with the original training set as a new training set to train the classification model.Then,the training model is trained by adding the training set after the spectrograms is generated,and the test set is used for testing,and the test result is compared with the model trained by the original training set.The experimental results show that the test accuracy of the classification model increased from 92.1%to 96.1%after adding the depth convolution to generate spectrograms against the network.Second,generated spectrograms refined by refiner.To start with the qual-ity of the generated spectrograms.In order to improve the quality of the gen-erated spectrograms,the paper proposes to use the refined network(SimGAN)to refine the generated spectrograms,so that the generated spectrograms would be more similar to the real spectrograms.Then,the refined generated spectro-grams is added to the original training set to train the classification model,and the same test set is used for testing,and the test results are compared.The ex-perimental results show that when the generated spectrograms in the training set are refined by the refining network,the test accuracy of the classification model increased fiom 96.1%to 97.5%...