Abstract
Infrared target segmentation models trained on one spectral band generalize poorly to other bands: short-wave infrared (SWIR), mid-wave infrared (MWIR), and long-wave infrared (LWIR) images are formed by different physical mechanisms, and labeled data are often limited to a few images of a single band. We study a strict source-only setting: a model is trained on a few labeled SWIR images, without target-band images, target-band labels, or auxiliary infrared data, and is tested on unseen bands. We propose KDConFormer, a SegFormer-based framework with two training-time mechanisms. Target mask-guided contrastive learning (TMGCL) uses the ground-truth mask to sample target-centered and background-centered patches and contrasts target patches against a momentum-encoded queue. Multiscale knowledge distillation (MKD) regularizes the feature hierarchy through intra-network self-distillation, cross-scale alignment, and adaptive fusion. The deployed model needs no mask and increases the parameter count by 9.4% and the FLOPs by 7.8% relative to SegFormer. With 25 labeled SWIR images, KDConFormer reaches % mIoU on LWIR images of the real-world SDU-IRSD ship dataset, 5.59 points above the strongest competitor, with a ship IoU of 16.5% against at most 5.5% for the competing methods; on the simulated MIRAT aircraft dataset, it improves on SegFormer by 4.74 points on MWIR.