LSOAF-YOLO: A Lightweight Small Object Detector With Attention Fusion for Aerial Remote Sensing Imagery

Abstract

Small object detection (SOD) remains a fundamental challenge in computer vision due to the limited spatial resolution of target regions and their susceptibility to background clutter. Three structural bottlenecks collectively constrain detection performance: systematic underrepresentation of fine-grained features in shallow network layers, the inability of spatial-only attention to disambiguate targets in spectral space, and inadequate semantic continuity modeling across heterogeneous feature scales. To address these challenges, we propose LSOAF-YOLO, a lightweight small object attention fusion detector built upon YOLOv11n, incorporating three targeted contributions. First, a lightweight SOD neck (LN) refocuses the detection hierarchy onto shallow feature layers (P1–P2) that preserve critical spatial detail while pruning redundant deep-scale branches. Second, a volumetric self-attention module (C2PSA_V) extends attention from the spatial domain to joint channel–spatial dual dimensions, enhancing target discriminability against background clutter via interchannel feature correlations. Third, a dual aggregation attention fusion convolutional module replaces conventional upsampling-concatenation with a joint self-attention and cross-attention mechanism that explicitly captures both intralayer feature consistency and interlayer semantic complementarity. Experiments on the VisDrone2019 benchmark demonstrate that the LN-based variant reduces parameter count by 23% and increases inference throughput from 250 to 714 FPS relative to YOLOv11n, while improving small object mAP50 by 3.1%. The full LSOAF-YOLO framework achieves a small object mAP50 of 25.4%, surpassing YOLOv11n by 4.7% and YOLOv26n by 3.4%, with generalization further validated on TinyPerson and UCAS-AOD datasets.