Few-Shot Object Detection in Remote Sensing Images via Dynamic Adversarial Contrastive-Driven Semantic-Visual Fusion

The acquisition of remote sensing images (RSIs) requires expensive equipment, such as satellites or aircraft, along with advanced sensing technologies. Additionally, the complexity of terrains and the inaccessibility of certain regions further exacerbate the difficulty of obtaining RSI data, particularly for novel classes. Therefore, this article focuses on two core challenges in few-shot object detection (FSOD) for RSIs: 1) the difficulty in feature learning caused by insufficient samples of novel classes and 2) large intraclass variability and small interclass differences in novel classes, referring to the feature similarity between different categories and the feature variability within the same category. To address the above two major challenges, we propose an FSOD framework in RSIs that integrates the semantic-visual fusion module (SVFM) with the dynamic adversarial contrastive loss (DACL) function. First, to tackle the issue of insufficient feature representation for novel classes, we introduce a text inference branch that utilizes the CLIP text encoder to generate rich semantic representations. Furthermore, to promote deep integration and synergistic effects between textual information and image features, we design a multilevel cross attention (MLCA) mechanism and a dual-head co-attention (DHCA) guidance feature fusion module to enhance the semantic understanding and visual representation capabilities for novel classes, effectively compensating for the lack of samples. Second, to address interclass similarities and intraclass variabilities, we maximize interclass distances and minimize intraclass feature differences by generating adversarial features and dynamically optimizing feature distributions, thereby enhancing the model’s discriminative capability. The experimental results demonstrate that our proposed framework significantly outperforms existing methods across various shot settings on the NWPU VHR-10 and DIOR benchmark datasets for RSIs, showcasing its wide application potential in the field of FSOD in RSIs.