Cross-task and cross-domain SAR target recognition: A meta-transfer learning approach

Abstract

Meta learning and transfer learning offer promising solutions to the problem of requiring large amounts of data in deep learning approaches for synthetic aperture radar (SAR) target recognition. To improve their performance further, we propose a novel Meta-transfer learning approach for cross-task and cross-domain SAR target recognition (MetraSAR). In the meta training phase, we train a robust meta learner with the human-like ability to master new knowledge quickly across tasks and domains. By designing the weighted classification loss with class weights, we conduct hard class mining that forces the meta learner to grow stronger. In addition to the external knowledge transfer across different tasks, we achieve the internal transfer across domains by using the domain confusion loss with a domain discriminator. To balance the two designed loss terms, we adopt the multi-gradient descent algorithm to optimize the meta learner adaptively. In the meta testing phase, the trained robust meta learner is transferred to solve the new task with few shot samples and a quick generalization. Extensive experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset validate that MetraSAR has better performance than conventional SAR target recognition methods.