Study On ISAR Taget Recognition Based On Deep Learning

Synthetic aperture radar(SAR)image technique has developed in recent years.Radar image classification and target recognition have become new hot topics.In target recognition tasks,deep neural networks mainly focus on optical images.But SAR images are complex images and the conventional real-valued neural networks cannot make full use of the phase information in the complex image.Although deep neural networks achieve a good performance in target recognition,the principle is not clear and the model itself is like a ’black box ’.The fine-tuning for the model is only dependent on trial and error,lacking necessary theoretical support which has become an obstacle of the development of the theory.In this paper,we analyze the status of the research on deep neural networks at home and abroad.We make further research on complex-valued convolutional neural networks and network visualization theory.Besides,the electromagnetic characteristics of the inverse synthetic aperture radar(ISAR)target is combined with deep neural networks.The framework of complex neural networks and the visualization theory are proposed in this paper.The classification accuracy of the deep network and the interpretability of the model are improved.The main content of the dissertation is summarized as follows:1.The framework of complex-valued convolutional neural networks.The real-valued deep neural networks mainly extract the amplitude features of input samples,while SAR images are complex images,which not only have amplitude information but also contain phase information.Real-valued neural networks are difficult to make full use of the phase information of SAR complex images.Therefore,based on the detailed analysis of the realvalued deep network structure in the second chapter,we further extend the network structure to the complex domain,and redesigns the main structures of the complex domain network,such as convolutional layer,maxpooling layer and activation function,so that it can extract the phase information of the input complex image samples.2.Weight initialization by electromagnetic feature transfer learning.Since the parameters of the complex neural network are nearly doubled compared to the real network,it is more difficult for complex convolutional neural networks to converge.The times of iterations required for training will increase.Inspired by the idea of transfer learning in deep neural networks,a parameter initialization method of feature transfer learning is proposed.This method combines the electromagnetic feature of the target in ISAR image with the transfer learning theory of deep neural networks.The attribute scattering center model is applied to initialize the weight parameters in the convolution layer of complex convolutional neural networks,which improves the ability of feature extraction from the input SAR complex image samples and accelerates the training process of the network.The method can reduce the number of iterations and improve training efficiency of the complex neural networks.3.Visualization of weight parameters in deep neural networks for ISAR target recognition.Deep neural networks have a good fitting capacity of the input data.Although the performances of the visual tasks are good,the principle is not clear.It is difficult to understand how the model realizes image classification.The main reason for the problem is that there is a hidden layer in the deep network,and the transform relations between the input sample and the classification output cannot be obtained directly.The traditional deconvolution method cannot accurately explain the meaning of the results in reality and is difficult to achieve quantitative analysis.In order to solve the problems above,this paper proposes a weight visualization method based on model reconstruction for deep neural networks in the fourth chapter.The prior information of the parameters in the original deep neural network model is obtained by using the input samples.The internal structures of the model are adjusted by using the prior information to keep the transform relations unchanged for further process.After the model structures are adjusted,a group of orthogonal vectors is used to calculate the response of the network input to obtain the transform relations of the model.Compared with the method based on back propagation,this new approach has high efficiency in analyzing the transform relations of multi-layers in deep neural networks and can analyze the transform relations of hidden layer in deep network as well.4.The analysis of the principle of deep neural networks for target recognition.On the basis of weight visualization of deep neural networks,to understand the principle of deep neural networks for target classification,it is necessary to analyze the weight visualization results of deep neural networks.In the fourth chapter,the results of weight visualization of deep neural networks are analyzed from two aspects.The ISAR aircraft target data is used as the input sample and the optimal image sample fit for the network model is analyzed when the weight of the deep neural network is given.The training method,back propagation algorithm,is further extended.The gradient of the loss function is extended to the input sample,and the input samples of the model are iterated without changing the weight.Finally,the optimal sample suitable for the model is analyzed.On the other hand,based on the analysis of the principle of deep neural networks for recognition of single input sample,a recognition analysis method for ISAR aircraft samples is proposed.The weight parameters after the training process are not changed in the process of recognition task with different input samples,but the state of some structures in the model is changed.According to the visualization results of deep network weights obtained in the fourth chapter,the method by replacing the state of the single sample with the average of the internal state of the network is proposed to obtain the visualization results of the weights for the whole data samples to analyze the principle of the network for the dataset.At the same time,for ISAR aircraft data,the principle of deep neural networks is analyzed by using the time series information of samples under different attitudes.