DTSFNet: A Lightweight Network Based on Dynamic Sampling and Scale Sequence Fusion for Aerial Image Object Detection

Unmanned aerial vehicle (UAV) and remote sensing (RS) object detection play vital roles in modernization and ensuring public safety. However, deploying existing large-scale networks on resource-constrained devices remains a challenge. Furthermore, aerial images pose unique difficulties due to varying object scales, which are not adequately addressed by conventional feature fusion methods. To address the above issues, we suggest a lightweight deep network based on dynamic sampling and scale sequence fusion (DTSFNet). Our approach comprises a multiscale feature extractor (MSFE) module, which employs diverse convolutional kernels to reduce model complexity while capturing multiscale features effectively. Additionally, we devise a dynamic scale sequence fusion (DSSF) module, which enables comprehensive exploration and efficient integration of multiscale features across different levels. The proposed approach is evaluated across three publicly accessible datasets: VisDrone2019, UVADT and DIOR. The results demonstrate that our approach achieves lightweight models while maintaining high detection accuracy.