Real Beam Scanning Radar Superresolution Algorithm Based On Deep Learning

Real beam scanning radar is capable of obtaining the actual beam image of the forwardlooking area of a flight platform,and has significant applications in terrain avoidance,autonomous landing,maritime search,and ground attack.However,the azimuth resolution of the beam image is often poor due to antenna size and radar range limitations.Although traditional convolution inversion technology can enhance the azimuth resolution to some extent,it is characterized by slow convergence and difficult parameter selection.Deep learning method has been widely studied in various fields in recent years.It is of great significance to explore its feasibility in azimuth super resolution of real beam scanning radar.This paper focuses on the application of deep learning methods in real beam scanning radar and studies both model-driven and data-driven deep learning superresolution methods.The main objectives of this work are as follows:1、The convolution model of azimuth-echo is derived,and the applicable conditions of the model are discussed.The deep learning theory and its application potential in the field of radar superresolution are explained,which lays a foundation for the subsequent research of theoretical methods.2、Aiming at the problems of slow convergence and difficult tuning of traditional superresolution algorithms,a real beam scanning radar superresolution method based on threshold iterative shrinkage network is studied,and the azimuth superresolution of real beam scanning radar is effectively realized.At the same time,the attention module is introduced to enhance the representation of key information,and further improve the superresolution performance and processing efficiency under complex scene conditions.3、A data-driven superresolution method based on adaptive multi-scale cyclic convolutional neural network is proposed to solve the problem of azimuth convolution model mismatch in high-speed platform real beam scanning radar.It uses the coding and decoding network to realize feature extraction and image reconstruction,and introduces the multi-scale cyclic network to avoid the information loss of single feature.Meanwhile,it improves the general residual module and sets the weight adaptive learning module to realize the adaptive fusion of different levels of features,and effectively realizes the bearing super resolution of the real beam scanning radar under the condition of high-speed platform.The above-mentioned methods have been verified using simulation data and can effectively realize azimuth superresolution imaging of real beam scanning radar.