SAR Image Recognition Using ViT Network and Contrastive Learning Framework With Unlabeled Samples
Jianping Deng, Yuying Zhu, Shuning Zhang, Si Chen
Abstract
We propose an innovative vision transformer (ViT)-based architecture for synthetic aperture radar (SAR) automatic target recognition (ATR), which trains models in a self-supervised learning fashion. Compared with convolution neural networks (CNNs)-based models, transformer-based architectures further focus on locational information among features, enabling models to understand images globally. However, the integral challenge with transformer is that they commonly demand more samples for training than the CNN-based models. Furthermore, securing substantial labeled SAR images is typically a daunting task, particularly for noncooperative targets. To address these issues, the proposed model combines the ViT architecture with a contrastive learning framework. The process begins by pretraining the model using substantial unlabeled samples, followed by the execution of fine-tuning with limited labeled data. Besides, A data augmentation mechanism is designed for contrastive learning to enhance diversity and amounts of samples, simultaneously learning robust representations. Experiments conducted on MSTAR datasets demonstrate that the proposed model can perform very well on SAR image classification tasks even without sufficient labeled training samples.