💻 computer science

YOLO-MAG: An edge-preserving and bilateral dual-gated YOLO11 network for small object detection in UAV aerial imagery

The paper proposes YOLO-MAG, an enhanced YOLO11-based network featuring edge-preserving modules, hybrid attention mechanisms, and a bilateral dual-gated fusion pyramid, which significantly improves small object detection accuracy and real-time performance in UAV aerial imagery by mitigating challenges like low resolution and complex backgrounds.

Original authors: Ruiqing Zhang, Hongtian Zhang, Tiezhu Li

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ruiqing Zhang, Hongtian Zhang, Tiezhu Li

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Drones have transformed the way we see the world from above, turning the sky into a new frontier for observation. Whether monitoring traffic, tracking wildlife, or searching for people in distress, these flying cameras rely on computers to identify objects in the images they capture. However, seeing small things from high up is a difficult task for artificial intelligence. When a car or a person appears as a tiny cluster of pixels in a vast landscape, the computer struggles to distinguish them from the background. The edges of these small objects often blur as the image is processed, and the sheer variety of sizes and the clutter of the environment can confuse standard detection systems. For drones to be truly effective, their "eyes" need to be sharp enough to spot these faint details without getting lost in the noise.

Researchers have developed a new approach to solve this problem, creating a system specifically designed to find small objects in drone footage. The team, led by scientists from Yellow River Conservancy Technical University and Henan University, built upon an existing, highly efficient detection framework known as YOLO11. While the original system is fast and capable, it tends to lose the delicate outlines of tiny targets as it analyzes an image, much like how a sketch might lose its fine lines if drawn too quickly. The researchers' goal was to preserve those critical edges while also filtering out the visual clutter that often hides small targets. They achieved this by weaving three specific improvements into the network's structure, resulting in a new model they call YOLO-MAG.

The first major change focuses on protecting the edges of objects. In standard image processing, as a computer zooms out to understand the bigger picture, the fine details of small items often fade away. The new system introduces a specialized module that acts like a safeguard for these boundaries. Instead of letting the edge information get diluted, this module actively extracts and strengthens the outlines of objects at every stage of the analysis. It ensures that even when an object is tiny and far away, its shape remains distinct and recognizable to the computer, preventing it from being swallowed up by the surrounding background.

To handle the complexity of the scene, the researchers also improved how the system pays attention to different parts of the image. Small objects often appear in groups or are surrounded by confusing patterns, making it hard to tell where one object ends and another begins. The new design combines two different ways of looking at the image: one that focuses on the immediate local details and another that understands the broader context. By blending these two perspectives, the system can better understand the relationship between objects and their surroundings. This allows it to pinpoint a small vehicle in a crowded street or a person in a field with greater precision, without requiring significantly more computing power.

The final piece of the puzzle involves how the system combines information from different layers of its analysis. Imagine a network that receives a blurry, wide view of a scene and a sharp, close-up view simultaneously. The new architecture uses a dual-gating mechanism to decide which parts of this information are useful. It acts like a sophisticated filter that opens the gate for clear, important details while closing the gate on background noise and irrelevant clutter. This ensures that the final decision on what an object is and where it is located is based on the strongest, most relevant evidence, rather than being confused by the visual chaos of the environment.

When tested on real-world datasets containing thousands of drone images, the new system demonstrated a significant leap in performance. On a standard benchmark used to evaluate drone vision, the improved model detected small objects with a success rate that was nearly seven percent higher than the original, unmodified system. It also maintained a high level of accuracy even when the criteria for a correct detection were made stricter. Despite these gains in precision, the system remained fast enough to process images in real time, capable of analyzing over one hundred frames per second on modern hardware. This speed is crucial for drones, which often need to make split-second decisions while in flight.

The researchers also tested the system on a different set of images to ensure it could handle various environments, from busy city streets to open roads. The results were consistent, showing that the improvements helped the system generalize well to new situations. The study suggests that by focusing on preserving edge details, blending attention mechanisms, and filtering out noise more effectively, it is possible to make drone vision much more reliable. This advancement offers a promising path forward for applications where spotting small, distant objects is critical, from emergency rescue operations to automated traffic monitoring, ensuring that the view from above is as clear and useful as possible.

Technical Summary: YOLO-MAG for Small Object Detection in UAV Imagery

Problem Statement

Small object detection in Unmanned Aerial Vehicle (UAV) aerial imagery faces significant challenges due to low spatial resolution, faint object boundaries, drastic scale variations, and complex background interference. While the state-of-the-art YOLO11 architecture offers high efficiency, it exhibits performance bottlenecks in this specific domain. Standard downsampling in the backbone network causes severe degradation of boundary details, attenuating critical spatial information for tiny targets. Furthermore, standard feature fusion paths in the neck network often suffer from inadequate cross-layer interaction, causing subtle small-object features to be overwhelmed by deep semantic information and background clutter. Existing solutions often apply edge extraction only in the backbone or rely on single-gated fusion mechanisms, leaving edge representations vulnerable to dilution or failing to sufficiently filter background noise across all feature pathways.

Methodology

The authors propose YOLO-MAG, an edge-preserving and bilateral dual-gated network built upon the YOLO11 framework. The architecture introduces three primary modifications to address the limitations of standard YOLO11:

1. Edge-Preserving Pipeline: C3k2_MEIS Module

To prevent the loss of edge details as network depth increases, the authors replace the standard residual units in the C3k2 modules (found in both the backbone and neck) with a Multi-scale Edge Information Selection (MEIS) module.

EdgeEnhancer: This sub-component explicitly extracts edge information by subtracting a smoothed feature map (obtained via adaptive average pooling) from the original input. This difference operation highlights high-frequency edge components, which are then convolved and added back to the original features.
Dual-Domain Selection Mechanism (DSM): Following edge enhancement, the DSM filters features in parallel using a Spatial Selection Module (SSM) to highlight key regions and a Frequency Selection Module (FSM) to optimize frequency-domain patterns. These are fused via learnable parameters ( $\alpha$ and $\beta$ ) to dynamically balance original and enhanced features.
Deployment: Unlike prior works that limit edge modeling to the backbone, YOLO-MAG deploys C3k2_MEIS throughout both the backbone and neck, establishing an end-to-end edge-aware pipeline.

2. Hybrid Attention: C2PSA_ACmix Module

To address the uneven spatial distribution of objects and the limitations of pure self-attention, the C2PSA block is redesigned into C2PSA_ACmix by integrating the ACmix module.

Mechanism: ACmix unifies self-attention and convolution within a single framework. It shares intermediate projections (Query, Key, Value) for both paths.
Dual-Path Aggregation: The Self-Attention path captures long-distance dependencies and global context, while the Convolution path utilizes shift and summation operations to retain local inductive bias and edge/texture details.
Fusion: The outputs of both paths are weighted by learnable scalars ( $\alpha$ and $\beta$ ), allowing the model to dynamically adjust the contribution of global context versus local detail with minimal computational overhead.

3. Bilateral Dual-Gated Fusion: GACFPN

The standard Path Aggregation Feature Pyramid Network (PA-FPN) is replaced by the Gated Activation Convolutional Fusion Pyramid Network (GACFPN) to handle scale variations and background clutter.

Reconstructed Fusion Path: The network routes features from specific backbone layers (4, 6, and 10) to corresponding neck layers (19, 14, and 26) to better preserve shallow spatial details.
Bilateral Dual-Gating: Unlike previous asymmetric single-gated designs (where only an auxiliary branch is gated), GACFPN employs a bilateral dual-gating mechanism. Both the deep context pathway and the auxiliary representation pathway are equipped with dynamic gating masks.
Cross-Calibration: The gating masks from both branches interact via element-wise multiplication to cross-calibrate spatial saliency and channel selectivity. This rigorously suppresses background clutter in both pathways while preserving critical multi-scale target representations.

Key Contributions

Full-Pipeline Edge Enhancement: The integration of the C3k2_MEIS module across both backbone and neck networks ensures that weak edge cues are shielded from semantic dilution throughout the entire detection pipeline.
Efficient Hybrid Attention: The C2PSA_ACmix module combines self-attention and convolution to improve small object precision with negligible computational cost, addressing the lack of local inductive bias in standard attention mechanisms.
Bilateral Dual-Gating: The GACFPN introduces a novel bilateral dual-gating activation mechanism that dynamically filters background noise across parallel pathways, overcoming the limitations of single-gated fusion in complex aerial scenarios.

Experimental Results

The model was evaluated on the VisDrone2019 and UAVDT datasets using an NVIDIA GeForce RTX 4080 GPU.

Performance on VisDrone2019:
- YOLO-MAG-n achieved 40.4% mAP@0.5 and 24.5% mAP@[0.5:0.95].
- This represents an improvement of 6.9% and 5.0% respectively over the baseline YOLO11n.
- The model maintained real-time inference speeds of 136 FPS, despite a slight increase in parameters (from 2.6M to 3.7M) and GFLOPs (from 6.5G to 8.7G).
- Ablation studies confirmed that each module (MEIS, ACmix, GACFPN) contributed positively, with the combination yielding the highest gains.
Generalization on UAVDT:
- YOLO-MAG demonstrated superior generalization, outperforming baseline YOLO11 variants across all scales (n, s, m, l) on the UAVDT dataset. For instance, YOLO-MAG-n exceeded YOLO11n by 3.6% in mAP@0.5.
Comparison with SOTA:
- YOLO-MAG-n outperformed other mainstream models (YOLOv5, YOLOv8, SRTSOD-YOLO, TA-YOLO) on the VisDrone2019 benchmark, achieving the highest accuracy among models with comparable parameter counts.

Significance and Claims

The paper claims that YOLO-MAG provides a solid technical foundation for practical UAV-based vision systems by effectively solving the specific challenges of small object detection in aerial imagery. By strengthening the extraction of edge features and refining multi-scale fusion through bilateral gating, the model achieves a superior balance between detection accuracy and computational efficiency. The authors emphasize that the method is particularly well-suited for edge devices with constrained computing resources, as demonstrated by the high performance of the lightweight YOLO-MAG-n variant. The work concludes that the proposed architecture offers a promising framework for real-world applications such as personnel rescue, traffic supervision, and wildlife tracking, while future work will focus on further model compression and adaptation to highly dynamic environments.