HDST-YOLO for High-Density Small-Target Detection in Remote Sensing Imagery

Abstract

Aerial remote sensing object detection plays a critical role in fields such as traffic management, ecological monitoring, and public safety. While the YOLO series methods are widely adopted for their efficiency, existing approaches still face limitations when processing complex scenes containing dense, multiscale targets. Moreover, under constrained attention mechanisms, they often fail to adequately adapt feature extraction, leading to suboptimal detection performance for small objects under degraded imaging conditions. To address these issues, we propose HDST-YOLO. Our method introduces a multiattention higher receptive field module (MAHRM), which integrates cascaded multihead window attention and multilevel attention mechanisms to mitigate image degradation. For multiscale dense objects, a dense attention dilated receptive (DADR) block is employed, enhancing adaptability to multiscale features through adaptively weighted dilated convolutions. Furthermore, a space-to-depth convolution (SPDConv) is integrated to preserve fine-grained details of small objects, which are pixel sizes less than $32\times 32$ pixels. Evaluations on the NWPU VHR-10 and DIOR datasets show that HDST-YOLO achieves mean average precision (mAP) scores of 97.7% and 69.6%, respectively, reaching state-of-the-art performance. Ablation studies validate the contribution of each module.