Prototype-Guided Multilayer Alignment Network for Few-Shot Object Detection in Remote Sensing

Few-shot object detection (FSOD) addresses the challenge of limited labeled data by enabling detectors to learn from minimal annotations. Recent work on image–text fusion has shown promise in overcoming data scarcity issues. However, these models suffer from catastrophic forgetting due to two distinct challenges during few-shot adaptation. First, feature space distortion disrupts established relationships between image–text modalities and blurs decision boundaries, leading to inaccuracies in object classification. Second, attention drift impairs the precise visual–textual alignments learned during base training, limiting the model’s ability to focus on relevant regions for accurate object localization. To overcome these limitations, we propose Protoalign, a novel prototype-guided multilayer alignment network that maintains robust cross-modal relationships across multiple network layers while mitigating catastrophic forgetting. Specifically, cross-modal prototype guidance (CPG) enables stable feature learning through class-specific prototype fusion to mitigate feature space distortion. Multimodal feature aggregation (MFA) strengthens feature relationships through channel-level interactions to overcome attention drift. Moreover, we propose an integrator that facilitates consistent information flow between network layers. Protoalign progressively refines multimodal features, enabling effective novel class adaptation while preserving crucial base knowledge. The refined representations are then processed by a detection transformer (DETR) decoder for accurate object detection. Extensive experiments on iSAID, DIOR, FAIR1M-Airplane, and NWPU VHR-10 datasets demonstrate that Protoalign achieves superior performance while significantly reducing catastrophic forgetting compared to existing methods.