Few-shot unsupervised domain adaptation for SAR target recognition based on input alignment and class confusion loss

With the development of simulation technologies and deep learning, utilizing simulated images to train Synthetic Aperture Radar (SAR) target recognition model has become a research hotspot. However, SAR target recognition is hindered by the domain gap between simulated and real images, particularly under few-shot settings which mean only a limited number of unlabeled target domain samples are available for training. Regarding this issue, this paper proposed a new few-shot Unsupervised Domain Adaptation (UDA) method based on input alignment and class confusion loss (named IACC). Our method consists of two modules. The input alignment module (IAM) utilizes a special image-to-image translation model which can realize realistic image translation in few-shot settings, while the Class Confusion Module (CCM) employs a novel loss function to align the feature space and enhance classifier performance. Experimental results on the SAMPLE dataset demonstrate the superior performance of our method in both UDA and few-shot UDA scenarios compared with other existing methods. IACC offers a promising solution for SAR target recognition in challenging domain adaptation scenarios with limited data.

I. Introduction

In recent years, with the development of deep learning, the Synthetic Aperture Radar (SAR) image interpretation has made tremendous progress. SAR Target recognition, which aims to accurately recognize the fine-grained categories of the target is a fundamental problem. To achieve effective target recognition on some SAR images, the most common approach is to acquire and label a considerable amount of homologous training data with identical sensor parameters, similar background information, and analogous pose information. An efficient classification neural network is then trained by adjusting hyperparameters on the training data. However, due to the fact that SAR imaging is highly sensitive to the operating conditions of the platform and the state of the target, the entire process needs to be repeated every time new test data is collected from other platforms or under different working conditions. Furthermore, as SAR is predominantly used for military purposes, the uniqueness of the targets makes it extremely difficult or even impossible to acquire a large amount of homologous data in real condition. Additionally, SAR images differ significantly from optical images, making them difficult to distinguish for the human eye. Even for experts with specialized knowledge, fine-grained annotation tasks are time-consuming, labor-intensive, and difficult to ensure accuracy [1]. With the advancement of simulation technologies such as Computer-Aided Design (CAD), generating a large number of high-quality SAR images through simulation algorithms has become feasible. Simulation allows for costless modification of platform operating conditions and target states, and detailed label information can be easily obtained. Consequently, training deep models on simulated data is emerging as a viable alternative. Despite efforts to optimize simulation algorithms, which can indeed improve image quality, the inaccuracies inherent in modeling and the various types of noise present in the real world still lead to significant differences in the distribution between the simulated and real domains. As a result, models trained directly on simulated data often struggle to achieve good performance on real SAR images, underscoring the challenge of cross-domain recognition.