Abstract
Ship detection in synthetic aperture radar (SAR) images holds significant importance for marine management and maritime navigation. Current methods typically train directly on a large number of labeled SAR images or transfer knowledge from labeled optical images, but in practice, obtaining labeled-rich data is often challenging. The source-free domain adaptation setting addresses this challenge by adapting a source-trained model to the target domain without accessing any labeled data, which generally uses pseudolabels based on a fixed threshold to guide the model for training. Nevertheless, these methods are predominantly designed for homogeneous images with the same imaging mechanism, rather than for heterogeneous images acquired by different sensors such as optical and SAR images, where a fixed threshold for pseudolabels is usually ineffective. To this end, we propose a novel pseudolabels’ training strategy for this cross-sensor scenario. Specifically, we introduce an adaptive pseudolabel generation (APG) module to efficiently use the source-trained model from the optical domain to generate pseudolabels for the SAR domain without relying on a fixed threshold. In addition, we design a region-based contrastive feature enhancement (RCFE) module aimed at exploring more valuable information in the target SAR domain to optimize feature representations and reduce the influence of noisy pseudolabels on detector training. The experimental results demonstrate that our proposed method effectively enables SAR ship detection without accessing any labeled data.
Funding Information
  • Fundamental Research Funds for the Central Universities (QNTD202308)