AI
New multi-scale fusion technique boosts small object detection in drone imagery
Researchers propose MGDFIS, a feature integration strategy that preserves fine details while using global context to detect tiny objects in aerial photos.
A research team has introduced a novel approach to spotting small objects in Unmanned Aerial Vehicle (UAV) photographs—a notoriously difficult task in computer vision. The technique, called MGDFIS (Multi-scale Global-detail Feature Integration Strategy), tackles the core tension between retaining subtle local clues and leveraging broader context to distinguish tiny targets from messy, complex backgrounds.
The core challenge#
Traditional multi-scale feature fusion methods attempt to improve detection by combining information across different image resolutions, but they typically come at a cost: added computational overhead and loss of fine-grained detail as features pass through multiple cross-scale blending operations. The real problem lies in balancing three competing demands: preserving weak signals from small objects, suppressing background clutter, and efficiently exchanging contextual information across image scales.
How MGDFIS works#
The proposed strategy operates at the neck level of object-detection networks—the intermediate processing stage between feature extraction and final predictions. It combines three coordinated components:
- FusionLock-TSS Attention stabilizes responses across spectral and spatial dimensions, preventing degradation during feature blending
- Global-detail Integration merges long-range context mixing with fine-grained local feature capture
- Dynamic Pixel Attention reweights compact foreground regions to highlight small objects
This coordinated design lets the system maintain both big-picture context and small-scale nuance without sacrificing computational efficiency.
Measured improvements#
When integrated into the YOLO26m detection framework on the VisDrone benchmark dataset, MGDFIS yielded substantial gains: AP50:95 scores jumped from 25.7 to 30.2, while AP50 improved from 37.2 to 44.2. The system operates at 96.1 GFLOPs, a reasonable computational load for real-world deployment.
Testing on additional datasets showed consistent results: 38.9 AP50 and 21.9 AP50:95 on UAVDT, and 97.4 AP50 on CARPK. The researchers have made their code publicly available, enabling other teams to build on their work.
