关键词: 跨域小样本目标检测, 域偏移, 原型学习, 特征校准, 背景干扰

Abstract: To address two issues in existing Cross-Domain Few-Shot Object Detection (CD-FSOD) methods, namely neglecting modality reliability differences under domain shift in multimodal prototype fusion and the susceptibility of Region of Interest (RoI) features to feature deviation and background confusion, a detection framework driven by domain-modulated hybrid prototypes and feature calibration, termed Domain-Modulated Hybrid-Prototype Calibration ViTO (DHC-ViTO), was proposed. First, a Domain-Modulated Hybrid Prototype Generation module (DM-HPG) was designed to construct visual prototypes encoding target-domain information and enhanced textual prototypes incorporating domain-related cues. A dynamic gating network was then used to adaptively adjust fusion weights according to the representation states of the two modalities in the current target domain, yielding domain-modulated hybrid prototypes with cross-domain robustness and fine-grained discriminability. Second, a Semantic Anchor-based Feature Calibration module (SAFC) was designed to use the hybrid prototypes as semantic anchors for guidance feature generation and RoI feature calibration, driving shifted RoI features toward the semantic centers of the corresponding categories. An Orthogonal Noise Decoupling (OND) strategy was further introduced to suppress background interference and improve RoI discriminability. Experimental results on six cross-domain benchmark datasets show that under the 1-shot, 5-shot, and 10-shot settings, the average mAP values of DHC-ViTO on the six target-domain datasets are 14.6%, 27.3%, and 30.8%, respectively. Compared with Cross-Domain Vision Transformer Object detector（CD-ViTO）, DHC-ViTO improves the average mAP by 0.7, 1.2, and 1.2 percentage points. These results verify the effectiveness of the proposed method.

Key words: Cross-Domain Few-Shot Object Detection （CD-FSOD）, domain shift, prototype learning, feature calibration, background interference