| Radar can use electromagnetic wave to detect,track and measure objects which are in its range of applications,so it plays an extremely important role in both military field and civil field.With the development of modern science and technology,radar automatic target recognition technology(RATR)has received more and more attention.This technique extracts information which is useful for identifying targets from target scattering echoes containing target features and uses certain criteria to determine target categories and related attributes.To obtain higher target recognition accuracy,we can use signal echos containing abundant target information to identify targets,that is,use one-dimensional high-resolution range profile(HRRP)data acquired from a wide band radar.A HRRP denotes the coherent summation of projection vectors of complex echoes from target scatters along the radar lineof-sight(LOS).It is a strong function of the target radar aspect angle and contains abundant informative target structure signatures,e.g.target size,scatters distribution,etc.A HRRP is easy to obtain,store and process,which has become one of the hotspots in radar target recognition domain.This paper combines deep learning methods to study the recognition of HRRP data.The main contents of this paper are as follows:1.This paper outlines the research background of HRRP target recognition technology and the development trend of domestic and foreign research,then introduces the main content and working arrangement of the paper.Later,the deep learning models with strong expressive power are introduced and the great success of such models in the field of target recognition such as computer vision,natural language processing and speech recognition is introduced.The advantages of deep learning models are studied in combination with specific frameworks.According to the above passage,the convolutional neural networks(CNNs)and the recurrent neural networks(RNNs)in deep learning are used to recognize HRRP data in this paper.2.This paper studies the HRRP sequence recognition based on deep neural networks.Combining the latest research results of internationally relevant deep neural networks along with target recognition in recent years,this paper aims at HRRP data recognition of three types of aircrafts.We use RNNs to model HRRP sequence to verify the feasibility of RNNs,then two HRRP target recognition methods based on recurrent CNN models are proposed.In the first place,this paper introduces the acquisition method of HRRP data and its basic mathematical model,summarizing its characteristics and giving corresponding preprocessing methods for its attitude sensitivity,range sensitivity and translation sensitivity.Then two different methods based on recurrent CNN models are used to recognize HRRP sequence.3.This paper studies the HRRP recognition using two different methods based on recurrent CNN models.In the first method,the CNN's multi-level structure from concrete to abstract is used to extract the distinguishable features of HRRP,then the RNN is used to identify the feature sequences.In the second method,we replace the matrix multiplication in both inputto-hidden state and hidden-to-next hidden state in every RNN cell,then use this improved RNN to classify HRRP sequence,which enables us to process sequential information and extract spatial information of HRRP at the same time.Compared to machine learning methods,these recurrent CNN models based on deep neural network can achieve target recognition in an end-to-end manner.Compared to existing deep learning methods,the proposed models can aquire higher recognition accuracy rate. |
