@Article{cmc.2026.086413,
AUTHOR = {Xinbin Hu, Mohd Shahrizal Sunar, Norhaida Mohd Suaib},
TITLE = {TAEPM-Net: An Accurate and Lightweight Detector for Multi-Class Small Object Detection in UAV Imagery},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/28364},
ISSN = {1546-2226},
ABSTRACT = {Unmanned aerial vehicle (UAV) imagery has made pedestrian and multi-class small-object detection a critical perception task, yet it remains difficult because targets often occupy fewer than 32 × 32 pixels. High accuracy on such targets is usually obtained with parameter-heavy models, whereas the compact detectors suited to a UAV’s limited memory budget tend to sacrifice accuracy, leaving a persistent accuracy-compactness gap. To narrow this gap, we propose TAEPM-Net, a lightweight detector built on YOLOv11n that raises small-object accuracy while keeping the parameter count low. TAEPM-Net introduces three components: a Partial Multi-Scale Feature Aggregation module (CSPMA) that replaces the C3k2 backbone block and supplies parallel multi-scale receptive fields with residual fusion; a Task-Aligned Dynamic Detection Head (TALDH) that couples classification and localization on a shared spatial representation through deformable convolutions (DCNv2) and a layer-level attention mechanism, reducing feature conflict on densely packed small targets while using fewer head parameters; and an Efficient Up-Convolution Block (EUCB) in the neck that learns to reconstruct spatial detail during skip-connection upsampling. On the VisDrone-2019 benchmark, TAEPM-Net attains an mAP@0.5 of 37.3% and an mAP@0.5:0.95 of 22.3%, exceeding the YOLOv11n baseline by 4.4 and 3.1 percentage points while reducing the parameter count by about 10% to 2.33 × 10<sup>6</sup>. Against twelve representative detectors, it achieves the highest accuracy among nano-scale models while remaining more compact than the YOLOv11n baseline, offering a favorable accuracy-parameter trade-off for UAV pedestrian and multi-class small-object detection.},
DOI = {10.32604/cmc.2026.086413}
}