LPASU-SwinDet: A Multiscale Transformer Framework for Maritime Infrared Small Target Detection

Abstract

To address the challenges of severe background clutter interference, low signal-to-clutter ratio, and weak long-range semantic dependence modeling in maritime infrared small target detection, this letter proposes lightweight pose-aware subpixel upsampling (LPASU)-SwinDet, a multiscale Transformer-based detection framework. Built on a modified YOLO26 backbone, the network introduces the LPASU module to replace conventional nearest-neighbor upsampling, enhancing detail reconstruction in feature fusion while suppressing background clutter. To further capture long-range contextual information and distinguish small targets from sea clutter, Swin Transformer-based detection heads are employed to improve modeling of long-range semantic dependencies, boosting detection robustness and localization accuracy in complex marine scenes. In addition, the normalized Wasserstein distance (NWD) loss and Wasserstein distance loss (WDL) are introduced to stabilize gradients for tiny targets and alleviate the positive–negative sample imbalance problem in bounding box regression. Extensive experiments on maritime infrared datasets demonstrate that the proposed method achieves superior detection accuracy and strong robustness against various challenging sea conditions.