Apex-Net: An Hourglass-Apex Enhanced Multi-Scale Network for Remote Sensing Small Object Detection

Abstract

Detecting objects in aerial imagery is difficult because tiny targets occupy only a few pixels and cluttered backgrounds weaken discriminative cues. This paper presents Apex-Net, an efficient YOLOv11-based detector. Its Hourglass-Apex hierarchy combines a high-resolution P2 branch, scale-adaptive circular sampling, and channel reallocation to retain shallow spatial detail. Multi-Group Golden Cudgel Convolution strengthens grouped shallow features through re-parameterized multi-branch processing, while Partial Gated Fusion selectively balances cross-scale features with low overhead. On VisDrone-2019, Apex-Net achieves 40.45 percent mAP at 0.50 IoU with 18.19 million parameters, improving the YOLOv11-m baseline by 4.87 percentage points. On HRRSD, it reaches 90.99 percent mAP at 0.50 IoU, a 2.13-point gain, demonstrating effectiveness across tiny-object and complex aerial scenes.