Few-Shot Radar Target Recognition based on Transferring Meta Knowledge

Abstract

Aiming at new types of few shot enemy targets in the combat scenario, a few shot radar target recognition technology based on transferring meta knowledge is proposed. This technology simulates the human learning process. First, a learning mechanism for multiple recognition tasks is built. Secondly, By learning recognition process of different tasks, the ability such as quickly adaption, strengthen generalization is gained, meta knowledge is accumulated gradually. Finally, the meta knowledge is transferred to support the model to achieve fast and accurate learning in the new few-shot recognition scenario. The proposed algorithm can realize the fast and accurate recognition of new types of radar targets in few shot scenarios (5 samples or less). Good results have been achieved in the full-angle field measured data set of a HRRP dataset and the public MSTAR dataset. It has a recognition rate of 97.4% when only 5 samples of new types of targets are learned in the MSTAR dataset. The optimal recognition rate based on HRRP dataset is 79.1%. Under the condition that only one sample is learned for each new type of target, the average recognition accuracy of three classifications can reach 64.1%. At the same time, the algorithm can overcome the angle sensitivity of radar data to a certain extent, which is very utility in practical scenarios.