Few-shot object detection aims to accurately identify and localize novel categories using only a small number of labeled samples. In remote sensing images, however, this task faces significant challenges due to substantial variations in target scale and complex backgrounds. To address these issues, this paper proposes a dual-attention guided few-shot object detection framework, DAFSDet. Specifically, a dual-attention strategy is implemented across the feature modeling and proposal generation stages. For feature fusion, the Content-Aware Strip Pyramid (CASP) is designed to enhance multi-scale feature representation by modeling spatial and contextual information. In the detection stage, a Deformable Attention RPN (DA-RPN) is proposed to improve the localization quality of candidate regions. With these designs, the proposed method effectively mitigates the challenges posed by multi-scale variations and complex backgrounds. Experimental results on the DIOR and NWPU VHR-10 datasets demonstrate consistent improvements over baseline methods, with notable gains of 7.54 mAP on DIOR Split 2 under the 10-shot setting and 2.09 mAP on NWPU VHR-10 under the 3-shot setting. These results indicate that the proposed method offers an effective solution for few-shot object detection in complex remote sensing scenarios.

Few-shot object detection focuses on accurately identifying and locating novel categories with just a small set of annotated examples. However, accurate object detection in remote sensing images still faces challenges such as scale variations, small objects, and complex background interferences. To alleviate these issues, we propose DAFSDet, a dual attention guided few-shot object detection framework, to effectively mine discriminative key features. Specifically, we propose a Content-Aware Strip Pyramid (CASP) module that uses content-aware upsampling for spatial attention and bidirectional strip convolution for long-range context, forming a joint spatial–semantic attention mechanism. CASP produces robust multi-scale features, highlighting key regions while preserving semantic and contextual information for few-shot detection. Subsequently, we design a Deformable Attention Region Proposal Network (DA-RPN) to produce high-quality candidate regions from the enhanced features. Through the collaborative optimization of CASP and DA-RPN, our method significantly improves the accuracy and robustness of few-shot object detection in remote sensing images, particularly in complex scenarios with large number of small objects and messy backgrounds. Experimental results on two large-scale datasets, DIOR and NWPU VHR-10, demonstrate that the proposed model achieves strong performance and clear advantages.

Few-shot object detection in optical remote sensing imagery remains challenging due to large scalevariations, complex background clutters, and severe sample scarcity. Existing few-shot learning methods often suffer from distorted prototype representations, over-smoothed similarity distributions, andinsufficient discriminative feature space under extremely limited training samples. To address theseissues, this paper proposes a Self-Calibrated Prototype Distribution Network (SCPD-Net) for fewshot object detection in remote sensing images. Specifically, a Self-Calibrated Prototype DistributionLearning (SCPDL) strategy is developed to represent class prototypes as Gaussian distributionsrather than deterministic points, while a query-aware scene calibration mechanism is introduced todynamically alleviate domain shifts between support and query samples. Furthermore, a TemperatureScaled Similarity Enhancement (TSSE) module is introduced to mitigate the over-smoothing effect ofcosine similarity computed on L2-normalized features, thereby enhancing foreground-backgrounddiscrimination in complex scenes. In addition, a Meta-Feature-based Contrastive Optimization(MFCO) branch employing foreground-guided supervised contrastive learning is leveraged to enhanceintra-class compactness and inter-class separability. Notably, the MFCO branch is removed duringinference, enabling improved feature representation without introducing any additional computationaloverhead. Extensive experiments on the DIOR and NWPU VHR-10 benchmarks demonstrate thatSCPD-Net consistently surpasses state-of-the-art models across most few-shot settings.

Detecting tiny objects in remote sensing images is challenging, as it suffers from severe semantic ambiguity across scales, high sensitivity to spatial misalignment, and intrinsically weak feature responses. These issues are further exacerbated in lightweight detectors, where aggressive model simplification often leads to significant performance degradation.To address these challenges, we propose a lightweight Deformable Adaptive Feature Reconstruction Network (DAFR-Net), specifically designed to enhance tiny object perception while maintaining a favorable trade-off between detection accuracy and computational efficiency. Rather than indiscriminately compressing model capacity, DAFR-Net explicitly breakthroughs three critical bottlenecks in lightweight tiny objects detectors.Firstly, to improve multi-scale semantic discriminability of tiny objects during feature fusion, a Gram-Guided Upsample (GGU) module is introduced to model channel-wise semantic correlations, enabling informative channels to be adaptively emphasized while preserving fine-grained details.Secondly, to enhance spatial deformation modeling without incurring excessive overhead, a deformable convolution-based detection head (DCDH) is presented to refine spatial deformation modeling at high-resolution feature levels for tiny objects, which are most sensitive to spatial sampling deviations, while standard grouped convolutions are used to reduce computational overhead. Thirdly, to strengthen weak feature representations degraded by aggressive lightweight operations, a lightweight Cross-Spatial-Channel (CSC) module is developed to jointly model spatial and channel responses, effectively compensating for tiny object information loss introduced by grouped convolutions without extra parameter increase. Extensive experiments on the large-scale AI-TOD-v2 and LEVIR-SHIP datasets demonstrate the effectiveness of the proposed model and superiority compared with state-of-the-arts. In particular, compared with the baseline method, it reduces 35.8\% and 16.26\% in terms of computation complexity, meanwhile unexpected lifting the detection accuracy by 1.5\% on the AI-TOD-V2 dataset. Code will be released at github https://github.com/chibaobao000/DARF-Net upon the publication.