Radar Image Fusion Recognition Based On Deep Learning

Radar target recognition is a kind of technology which can automatically judge the target category by radar echo signal,which is widely used in civil and military fields.Broad band Radar target recognition includes recognition based on High Resolution Range Profile(HRRP)and recognition based on Inverse Synthetic Aperture Radar(ISAR)image.It can obtain the structure information of the target,and then achieve more accurate classification of the target.Because the observation data of single sensor are imperfect and uncertain,the performance of single mode recognition is not ideal sometimes.Based on the sharing and comprehensive utilization of observation information,multi-mode fusion recognition achieves more accurate and reliable target recognition results.Traditional fusion methods rely on expert experience in feature extraction and feature selection,and each stage is independent from each other,so endto-end automatic recognition cannot be achieved.Given deep learning has lots of advantages in the information processing such as the powerful ability of data fitting,the characteristics of automatic extraction and strong portability,in order to realize the goal of fusing multiple radar information in an end-to-end identification process,multi-mode fusion and deep learning are combined to improve the performance of radar target recognition.In this paper,the HRRP fusion recognition and ISAR image fusion recognition are studied respectively,and the ISAR image fusion recognition under modal loss is discussed.The main research contents and achievements of this paper are summarized as follows:1.Aiming at the problem of high feature dimension in the fusion process of multiple HRRP,a HRRP fusion recognition method based on attention mechanism dimension reduction is proposed.In this paper,the preprocessed multi-source HRRP was firstly extracted and normalized by Long short-term Memory(LSTM)network respectively,and then the features were cross-product to obtain the fusion features.In order to achieve effective utilization of high-dimensional features and reduce network parameters,This paper introduces attention mechanism to achieve matrix and vector dimensionality reduction respectively.Finally,the feature vector is input into the fully connected classifier to achieve target classification.The experimental results show that the recognition rate is better than that of single mode recognition and general multi-mode recognition methods.2.Aiming at the problem of information redundancy in the fusion of multiple ISAR images,a fusion recognition method of ISAR images based on shared hidden space is proposed.In this paper,a shared hidden space with multiple inputs is constructed through representational learning to represent and reduce redundant information with the same feature.At the same time,an attention module is proposed to enhance the information interaction between network branches,which can effectively improve the speed of network convergence.Simulation results show that this method is superior to other multiview fusion methods in recognition rate and number of network parameters.3.In view of the lack of modes in the complex environment,the ISAR image fusion identification method based on Bot Net and Embrace Net was proposed.Due to the complex environment and the motion characteristics of non-cooperative targets,ISAR imaging conditions are sometimes difficult to guarantee,that is,there may be a lack of input modes.Therefore,a fusion recognition model named “Embrace Net” is adopted in this paper,which not only achieves improved recognition rates with reliable inputs from multiple sources,but also ensures a higher recognition rate than with single inputs in the absence of modes.