Few-Shot Transfer Learning for SAR Image Classification Without Extra SAR Samples

Abstract

Deep learning-based synthetic aperture radar (SAR) image classification is an open problem when training samples are scarce. Transfer learning-based few-shot methods are effective to deal with this problem by transferring knowledge from the electro-optical (EO) to the SAR domain. The performance of such methods relies on extra SAR samples, such as unlabeled novel class's samples or labeled similar classes samples. However, it is unrealistic to collect sufficient extra SAR samples in some application scenarios, namely the extreme few-shot case. In this case, the performance of such methods degrades seriously. Therefore, few-shot methods that reduce the dependence on extra SAR samples are critical. Motivated by this, a novel few-shot transfer learning method for SAR image classification in the extreme few-shot case is proposed. We propose the connection-free attention module to selectively transfer features shared between EO and SAR samples from a source network to a target network to supplement the loss of information brought by extra SAR samples. Based on the Bayesian convolutional neural network, we propose a training strategy for the extreme few-shot case, which focuses on updating important parameters, namely the accurately updating important parameters. The experimental results on the three real-SAR datasets demonstrate the superiority of our method.