TITLE:
A Lightweight Weighted Multi-Scale Fusion Network for Small-Person Detection in Post-Disaster UAV Imagery
AUTHORS:
Danah Alotaibi
KEYWORDS:
UAV Imagery, Post-Disaster Search and Rescue, Person Detection, Small-Object Detection, Weighted Feature Fusion, YOLO11n, Lightweight Neural Networks
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.11,
November
5,
2026
ABSTRACT: Person detection in post-disaster unmanned aerial vehicle imagery requires sensitivity to small targets and a compact deployment budget. This study presents LMF-YOLO, a YOLO11n variant that retains the pretrained backbone and neck and adds three residual positive-normalized fusion blocks at P3, P4, and P5. A single P3 projection increases the model from 2,582,347 to 2,590,609 parameters and from 6.4 to 6.5 GFLOPs. The PDD audit produced a group-isolated split of 619 training images with 1783 person instances, 112 validation images with 313 instances, and 146 frozen-test images with 481 instances. Across seeds 17, 43, and 97, A0 and A2 achieved validation mAP50-95 of 0.53095 ± 0.01065 and 0.53053 ± 0.03128, a paired mean change of −0.04 percentage points. Frozen-test mAP50-95 changed from 0.45675 ± 0.00389 to 0.44530 ± 0.03469, a paired mean change of −1.15 points. Small-object test AR increased descriptively by 4.44 ± 4.60 points. A2 seed 97 completed 97 epochs and reached zero mosaic-free epochs, whereas the planned schedule assigned the final 15 epochs to mosaic-free training. The evidence therefore characterizes LMF-YOLO as a low-cost mechanism for shifting scale-sensitive coverage, with fixed-schedule repetitions providing the next confirmatory test.