Synthetic Aperture Radar Image Target Recognition Research Based On Deep Learning

Synthetic aperture radar(SAR)is a kind of high-resolution imaging radar,which can obtain high-resolution radar images regardless of any weather conditions and illumination conditions,and is widely used in the fields of military investigation and geographical exploration.In the field of military investigation,after the space borne or airborne SAR imaging to ground,because the process of radar image generation is affected by the strong electromagnetic scattering and background noise,it is time-consuming,difficult and inefficient to analyze the target in the image if using human's eyes,so the research of automatic target recognition of synthetic aperture radar(SAR-ATR)becomes quite important.In recent years,deep learning has become the most popular method in the field of target recognition by automatic learning high level features in deep architectures.In the field of image recognition,convolution neural network(CNN)with its special methods of feature mapping such as weighted value sharing and feature aggregation,obtains the extremely good effect in many kinds of image recognition tasks,and is favored by the many research domains such as pattern recognition,artificial intelligence and computer vision.This paper first analyzes the present situation of the SAR automatic target recognition methods based on deep learning,and expounds the principle of forward and back propagation and special structure in deep learning.Then,aiming at the real radar target image dataset,a convolution neural network which can map and aggregate multi-scale features is constructed,and the model recognition experiment and the posture estimation experiment of targets are carried out.High recognition accuracies are achieved on several datasets.After that,in order to verify the latent disadvantage of the real radar image set in training,the radar image simulation platform is established.Combined with the real radar image background data,different background superposition strategies are used to form the contrast datasets.A deep network with strong ability of preventing over-fitting is constructed and these datasets are trained respectively on this deep network.The disadvantage of background similarity of real radar image sets is proved by the results of each model's own test and cross verification between different models and datasets.Then,in view of the problem of target recognition in large-scene multi-target radar images,the three-level series system of detection,identification and recognition and the detection-recognition system based on candidate box optimization are established respectively.The three-level series system is consist of the detection by simple image processing method,the major class identification network and the model recognition networks.Through the target position multiplicity and the self-rotation amplification,unifying the later network parameter tuning,the recognition model which has good generalization ability is trained,finally the results in the constructed end to end system verify the validity of the three-level series system.The recognition system based on the candidate box optimization adopts the joint training technique of detection and recognition,and the confidence solving and the non-maximum suppression are done on the candidate boxes selected according to certain rules.It obtains the high overlap rate and the recognition precision in the detection and recognition of the small radar images.Then the end to end system which is suit for large-scene is constructed.The experimental results show the good recognition performance of the system based on candidate box optimization.Finally,the target tracking technology after recognition and location is explored,and radar video target tracking based on correlation filter tracking algorithm is carried out.