Radar Target Recognition In Complex Electromagnetic Environment Based On Artificial Intelligence

Nowadays,with the development and breakthrough of radar technology,the resolution of radar imaging is getting higher and higher,and the recognition requirements for radar targets are getting higher.The image to be processed in this thesis is a synthetic aperture radar(SAR)image,which identifies the target object in the SAR image.How to improve the accuracy and recognition speed of SAR image recognition?The rapid development of artificial intelligence in recent years,especially the rise of deep learning,provides a new path for radar target recognition.The research content of this thesis is based on artificial intelligence radar target recognition technology.The research data are derived from the MSTAR data set,which is a publicly available synthetic aperture radar image data set.This thesis first studies the principle of SAR imaging and its image features,the algorithm process of SAR image target recognition,and the target characteristics of semi?physical simulation scattering points in complex electromagnetic environment.Then the deep neural network algorithm,convolutional network model,restricted Boltzmann machine and autoencoder are studied.The SAR image noise is processed based on these algorithm models.Then the autoencoder and support vector machine combination model is proposed for target recognition.The model uses the autoencoder to implement feature selection and data dimensionality reduction and uses the support vector machine to perform the final target classification.This thesis analyzes the strengths and weaknesses of the model and proposes ideas for improvement.Then,the radar target recognition based on convolutional neural network is studied in this thesis.By analyzing various current deep neural network framework models,their advantages and disadvantages are analyzed.A new deep convolutional network framework is proposed based on AlexNet model and LeNet model.The advantage of this model is that it has enough depth to extract high-level features in the original image,and avoids the disadvantages of slow training caused by too deep network.Before the training,the original SAR image is preprocessed by many preprocessing methods,such as noise processing,image normalization,and data enhancement.Then,the selection of the convolution kernel size of the convolutional neural network,the determination of the pooling method,the use of batch normalization and dropout,the advantages and disadvantages of the activation function,the use of the optimization function,and the necessity of regularization were tested.Select the best network mechanism and training skills to get the highest recognition accuracy;finally,based on the model to identify the radar target,and analyze the parameter factors that identify the impact.