Infrared Small Target Detection via Semantic-Unbiased Domain Adaptation

Abstract

Domain-induced feature discrepancies between cross-scenario training and testing data significantly degrade detection performance for infrared small target detection. However, infrared small target sparsity and the semantic differences between cross-domain infrared data cause domain adaptation object detection algorithms to perform poorly. This article proposes a knowledge distillation-based, semantic-unbiased hybrid domain adaptation method to address these distinctive challenges. In particular, a steady-style transfer is used to expand the original cross-domain infrared data, bridging the semantic gap between the source and target domains. Therefore, a self-training teacher model, which uses unlabeled target domain data, is employed to provide high-quality pseudo-labels. The student model uses these pseudo- labels to perform hybrid training that effectively includes detection, multilevel domain confusion, and contrastive learning tasks, thereby acquiring domain-invariant knowledge of the infrequent instance features in the target domain. The experimental findings demonstrate that this method achieves latent instance mining from unlabeled target domains, outperforming one-stage domain adaptive detectors in benchmark performance while maintaining computational frugality.