MDFR-Net for enhanced feature refinement in aerial small object detection

Detecting tiny objects in aerial imagery remains challenging due to their minimal scale, diverse orientations, and complex background interference, limiting the ability of existing methods to capture discriminative features. To address these limitations, we propose MDFR-Net with the following key contributions. Firstly, the Multi-scale Hierarchical Attentional Fusion (MHAF) module leverages hierarchical convolution layers and dual-path attention to jointly extract salient features while suppressing background noise. Next, the Orientation Decoupled Feature Enhancer (ODFE) module incorporates direction-separated paths with lightweight attention to focus on orientation-specific target details. Additionally, the Hierarchical Convolution Pooling Fusion (HCPF) module employs parallel dilated convolutions and pooling operations to capture cross-scale contextual information. Finally, a high-resolution P2 detection layer preserves spatial details of tiny objects by incorporating high-resolution features into the detection head. Experiments on the VisDrone and AI-TOD datasets show that MDFR-Net improves detection performance under the evaluated settings. Multi-seed experiments on VisDrone provide evidence of the stability of these improvements, while the single-run results on AI-TOD offer preliminary evidence of cross-dataset applicability.