Cascaded Multi-Scale Aggregation With Hierarchical Task Alignment for Aerial Small Object Detection

Abstract

Small object detection in aerial images confronts three intertwined challenges that span the detection pipeline: the mismatch between the fixed receptive fields of backbone networks and the extreme scale variation of targets, the joint degradation of shallow-layer representations caused by gradient attenuation and spectral aliasing along feature pyramid pathways, and the spatial misalignment between classification and localization branches trained with independent objectives. These three bottlenecks reside at successive stages of the pipeline and propagate in a cascaded manner, so that optimizing any single stage alone yields limited returns. This paper introduces PATH-YOLO, a framework that deploys coordinated modules across all three stages. A Progressive Multi-Scale Aggregation Block (PMSABlock) replaces fixed-kernel feature extraction with cascaded heterogeneous convolutions to broaden the effective receptive field of the backbone. An Adaptive Multi-Branch Feature Pyramid Network (AMB-FPN) improves signal quality in the pyramid pathway from three complementary perspectives: preserving gradient flow through linear lateral connections, suppressing spectral aliasing via depthwise separable upsampling, and providing multi-scale dynamic feature enhancement through parallel heterogeneous convolution. A Hierarchical Task-Aligned Detection Head (HTA-Head) explicitly models the spatial coupling between classification and localization, and incorporates a Scale-Aware Progressive IoU loss (PSIoU) to strengthen regression on small targets. On the VisDrone-2019 test-dev, RSOD, and NWPU VHR-10 benchmarks, PATH-YOLO reaches mAP@50 scores of 33.60%, 94.30%, and 88.20%, the highest mAP@50 among nano-scale models at a comparable computational cost of 7.9 GFLOPs and 3.32 M parameters, while remaining competitive on mAP@50:95 and $\text {AP}_{\text {S}}$ . Ablation studies show that the three modules contribute consistent and complementary gains across all datasets.