Cross-Domain Few-Shot Infrared Ship Segmentation with Class-Specific Adapters and SAM Refinement

Infrared ship semantic segmentation is pivotal for all-weather maritime surveillance and national early-warning applications. However, the significant domain gap between infrared (IR) and visible (VIS) images, combined with the extreme scarcity of annotated samples in the target IR domain, poses a dual challenge that severely restricts the cross-domain application of segmentation networks. Existing few-shot segmentation methods mostly assume similar data distributions, while domain adaptation methods rely on abundant unlabeled data; consequently, neither approach effectively addresses the combined problem of cross-domain and few-shot learning. To address this issue, a two-stage framework named CAS-Net (Class-specific Adapters and SAM refinement Network) is proposed for cross-domain few-shot infrared ship semantic segmentation. The framework first learns domain-invariant features and then leverages a handful of annotated IR samples for class-specific adaptation and mask refinement. Specifically, in the first stage, a dual-branch network integrating wavelet transform and convolution is designed to extract robust cross-domain features, employing a random convolution perturbation (RCP) strategy to stabilize adversarial training. In the second stage, independent, lightweight class-specific adapters are introduced for each target class and efficiently fine-tuned via self-supervised contrastive learning. Furthermore, the Segment Anything Model (SAM) is incorporated during inference through a prototype-driven prompt generation mechanism to refine initial segmentation results and enhance boundary accuracy. Experimental results on the VI-Ship and Agriculture-Vision datasets demonstrate that the proposed method outperforms state-of-the-art approaches across multiple evaluation metrics. Notably, under the extreme one-shot setting on the VI-Ship dataset, the proposed method achieves a mean Intersection over Union (mIoU) of 63.19%, exceeding the second-best method by 4.43%, thereby validating its effectiveness.