RFHA-YOLO: Dynamic Receptive Field and Adaptive Hybrid Attention for Small-Object Detection in Remote Sensing Images

Small object detection in remote sensing images remains challenging due to limited feature resolution and complex backgrounds. Conventional detectors, due to fixed receptive fields and uniform attention, struggle to capture small-target features and suffer from background clutter. To address these limitations, we propose RFHA-YOLO, a novel method designed to enhance small-object feature extraction. RFHA-YOLO integrates two key innovations. Firstly, a dynamic kernel decomposition module with an embedded Receptive Field Attention Block (RFA-Block) is constructed within the backbone as an alternative to the conventional spatial pyramid pooling structure. The core of RFA-Block is the RFAConv layer, dynamically adjusts receptive fields across large-kernel branches to enable fine-grained detail preservation for small targets while suppressing irrelevant background regions. Secondly, a enhanced Hybrid Attention Transformer (AHAT) is introduced in the feature fusion stage, enabling the adaptive integration of channel attention, spatial attention, and window-based self-attention. By learning importance scores at each pyramid level, this approach dynamically balances accuracy and computational efficiency in feature fusion. Extensive experiments on the DIOR, USOD, and AI-TODV2 aerial benchmarks have been demonstrated to show that RFHA-YOLO outperforms several state-of-the-art detectors under similar model complexity. It achieves mAP50 scores of 84.8%, 72.3% and 78.1%, reflecting overall detection accuracy, and impressive mAP50scores of 28.8%, 10.0% and 33.8%, which specifically measure performance on small objects. These results confirm the effectiveness of RFHA-YOLO for smcall object detection in remote sensing images, as its integrated RFA-Block and AHAT synergistically enhance small object feature capture and representation refinement.