A Target Recognition Method of Small Sample Based on RCS Data

Abstract

Abstract During the training of target recognition models based on Radar Cross Section (RCS) data, a persistent challenge arises in sampling due to the inherent difficulty in acquiring a sufficient number of samples. This scarcity of data poses a significant impediment to the effective training of models, resulting in diminished accuracy in target recognition. To address this issue, this article proposes a target classification method based on RCS data under small sample conditions. The approach adopts the fundamental concept of Model-Agnostic Meta-Learning (MAML) to train the target recognition model, enhancing the structure of MAML model. An hourglass-shaped convolution layer is introduced to the input layer, with an additional convolution layer preceding the output layer, and a switch to a central loss function. To substantiate the efficacy of the improved MAML model, comprehensive comparative analyses are conducted with benchmark models, including MAML, ResNet 18-layers, Long Short-Term Memory (LSTM), among others. Experimental results conclusively demonstrate the superior performance of the refined MAML model in target recognition under conditions of limited samples, attaining an average prediction accuracy of 85.62%. This signifies a noteworthy 5-percentage-point improvement compared to the baseline model prior to the introduced enhancements.