A newly developed lightweight neural network architecture significantly improves the detection and localization of small targets within remote sensing imagery while preserving real-time processing speeds. Built upon the baseline YOLOv11n detector, the proposed model, designated as FFLE-Net, addresses long-standing challenges associated with identifying compact, densely distributed objects that exhibit ambiguous boundaries and substantial variations in scale across aerial and drone surveys.
To overcome limitations in conventional visual feature extraction, the researchers designed a Coordinate-Shuffle Attention Fusion module. This core mechanism integrates two distinct attention pathways, known as Coordinate Attention and Shuffle Attention, into a unified feature extraction pipeline.
Within this fusion module, the Coordinate Attention pathway preserves direction-aware positional cues by encoding horizontal and vertical spatial features separately. Concurrently, the Shuffle Attention branch promotes channel and spatial interaction by grouping feature channels and executing shuffling operations. The network subsequently merges the complementary outputs from both branches through a weighted fusion process, directly sharpening the representation of small and densely clustered structures.
Beyond feature extraction enhancements, the architecture incorporates an IoU-based Localization Enhancement strategy directly into the bounding-box regression objective. This training optimization strengthens supervision over target localization during the learning phase. Crucially, because this supervisory adjustment modifies only the training loss calculations, it delivers higher geometric precision without imposing any computational overhead or memory burden during inference operations.
Researchers evaluated the performance of FFLE-Net across three established benchmark datasets: DOTA-v1, DOTA-v1.5, and VisDrone2019-DET. Experimental results demonstrated notable improvements over the baseline YOLOv11n model. Specifically, 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.
The detector also demonstrated robust performance under more demanding precision criteria across the higher evaluation threshold range. For the metric measuring mean Average Precision across intersection-over-union thresholds from 0.5 to 0.95, FFLE-Net achieved substantial gains over the baseline detector, recording 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.
Remarkably, these accuracy gains required only a marginal expansion in network complexity and computational demands. The parameter count of the model grew from 2.59 million in the baseline detector to 2.60 million in FFLE-Net, while computational volume increased from 6.46 to 6.51 GFLOPs.
High operational throughput was maintained throughout testing across all benchmark environments. FFLE-Net achieved 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. The experimental findings confirm that pairing feature fusion with localization loss optimization successfully enhances small-object detection across diverse aerial and unmanned aerial vehicle image distributions without compromising operational velocity.
Subject of Research: Lightweight remote sensing small object detection
Article Title: FFLE-Net: Feature fusion and loss enhancement network for small object detection in remote sensing images
Article References: Bao, J., Huang, H., & Wang, J. (2027). FFLE-Net: Feature fusion and loss enhancement network for small object detection in remote sensing images. Cluster Computing, 30(1), Article 2. https://doi.org/10.1007/s10586-026-06791-6
Image Credits: AI Generated
DOI: 10.1007/s10586-026-06791-6
Keywords: remote sensing, small object detection, FFLE-Net, YOLOv11n, Coordinate Attention, Shuffle Attention, Coordinate-Shuffle Attention Fusion, bounding-box regression, IoU-based localization enhancement, DOTA-v1, DOTA-v1.5, VisDrone2019-DET, aerial imagery, unmanned aerial vehicles, mean average precision, feature representation, channel shuffling, inference speed
News Source: Denise Maddox. (November 16, 2026). FFLE-Net Improves Small Object Detection in Remote Sensing Across Aerial Datasets. Scienmag.



