DDG-Net: a dynamic detail-guided network for aerial small object detection

Detecting small objects in aerial VisDrone images remains a challenging task due to their small size, low resolution, and frequent occlusions. To address these issues, we propose DDG-Net, a dynamic detail-guided network designed for small object detection in aerial imagery. First, we introduce a Lightweight High-Frequency Direction Injection (LWHFDI) module that integrates a Haar-inspired frequency decomposition strategy to introduce directional high-frequency cues during downsampling. Next, we present a Detail-Guided Semantic Feature Pyramid Network (DGS-FPN), which leverages shallow spatial details to guide deep semantic learning, combined with a multi-kernel context aggregation module to improve contextual representations for small object detection. Additionally, we replace traditional interpolation with DySample, a content-aware upsampling method that promotes feature alignment and helps maintain fine-grained details during multi-scale fusion. Experimental results show that, compared to the baseline, DDG-Net improves by 5.9 in AP50 and 3.6 in AP on the VisDrone-test dataset. Furthermore, DDG-Net achieves improvements of 2.7% and 3.5% on the AI-TOD tiny object dataset, respectively.