Bao, Huang and Wang Develop FFLE-Net for Real-Time Small Aerial Target Detection

Researchers built a compact neural network framework to detect small, densely packed targets in aerial imagery at real-time speeds.

The system is named FFLE-Net. Authors Jiangyan Bao, Haisong Huang, and Junhao Wang created it based on the YOLOv11n baseline.

Research published Nov. 16, 2026, in Cluster Computing details the work. It combines a Coordinate-Shuffle Attention Fusion module with an IoU-based localization enhancement strategy.

The network integrates Coordinate Attention and Shuffle Attention into a unified pipeline. The Coordinate Attention pathway encodes horizontal and vertical spatial features separately to preserve direction-aware positional cues.

The Shuffle Attention branch groups feature channels and executes shuffling operations. A weighted fusion process then merges these complementary outputs.

An IoU-based Localization Enhancement strategy integrates directly into the bounding-box regression objective. This supervisory adjustment modifies training loss calculations to deliver higher geometric precision.

Bao, Huang and Wang Develop FFLE-Net for Real-Time Small Aerial Target Detection

The parameter count grew from 2.59 million in the baseline detector to 2.60 million. Computational volume increased from 6.46 to 6.51 GFLOPs.

Testing across benchmark environments demonstrated sustained inference speeds of 92.70 frames per second on DOTA-v1, 92.25 frames per second on DOTA-v1.5, and 92.88 frames per second on VisDrone2019-DET, as detailed in Cluster Computing.

Evaluations across established benchmark datasets showed improvements over the baseline YOLOv11n model. Mean Average Precision at an intersection-over-union threshold of 0.5 increased by 9.7 percentage points on DOTA-v1, 14.0 percentage points on DOTA-v1.5, and 24.9 percentage points on the VisDrone2019-DET dataset.

Mean Average Precision across intersection-over-union thresholds from 0.5 to 0.95 recorded improvements of 16.0 percentage points on DOTA-v1, 16.9 percentage points on DOTA-v1.5, and 22.5 percentage points on VisDrone2019-DET.

Dataset mAP50 Improvement mAP50–95 Improvement
DOTA-v1 9.7 percentage points 16.0 percentage points
DOTA-v1.5 14.0 percentage points 16.9 percentage points
VisDrone2019-DET 24.9 percentage points 22.5 percentage points

The baseline model achieved lower precision scores before these architectural updates.