💻 computer science

GhostV2-YOLO: A Hardware-Aware Lightweight UAV Detection Network for Edge NPUs

This paper proposes GhostV2-YOLO, a hardware-aware lightweight UAV detection network that integrates DFC-Light attention, a full-link C2f_GhostV2 module, and strategic Coordinate Attention to achieve high accuracy and real-time inference on resource-constrained edge NPUs while demonstrating strong cross-dataset generalization.

Original authors: Jifa Zhang, Heyang Zhao, Linjiang Jia, Chen Ju, Ting Wei, Zonglin Li, Weiguang Zhang

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jifa Zhang, Heyang Zhao, Linjiang Jia, Chen Ju, Ting Wei, Zonglin Li, Weiguang Zhang

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

In the crowded skies above our cities and countryside, a new kind of invisible traffic is taking shape. Unmanned aerial vehicles, or drones, have moved beyond hobbyist toys to become essential tools for delivery, inspection, and surveillance. Yet, this proliferation brings a critical challenge: how do we reliably spot these small, fast-moving machines against complex backdrops like city skylines or forest canopies? The answer lies in computer vision, a field where software learns to "see" by analyzing patterns in images. For years, the most effective tools for this job have been deep learning networks, complex systems that mimic the human brain's ability to recognize objects. However, these systems are often too heavy and power-hungry to run on the small, battery-operated devices that need to watch the skies in real time. The goal for engineers has long been to build a detector that is both sharp enough to catch a tiny drone and light enough to run on a chip the size of a postage stamp.

A team of researchers at Xi'an Technological University has tackled this exact problem with a new system they call GhostV2-YOLO. Their work focuses on a specific type of hardware found in modern edge devices: neural processing units, or NPUs. These are specialized chips designed to speed up artificial intelligence tasks, but they have strict rules about what kinds of mathematical operations they can perform efficiently. Many existing drone-detection networks use complex mathematical steps that these chips cannot handle well, forcing the device to slow down and use its main processor instead. This creates a bottleneck where the theoretical speed of the chip is never reached in the real world. The researchers set out to design a detection network that respects these hardware limits from the very beginning, ensuring that every step of the process runs smoothly on the chip without needing to fall back to slower methods.

The core of their solution is a clever rethinking of how the network processes visual information. Instead of using large, heavy mathematical filters to scan an image for details, they broke these filters down into smaller, simpler steps that the hardware loves. Imagine trying to move a heavy piece of furniture through a narrow doorway; instead of forcing the whole object through at once, you break it down into manageable parts that fit easily. The researchers applied this logic to the network's "eyes," replacing a single large scanning operation with two smaller, cascaded ones. This change reduced the amount of data the system needed to handle by nearly 28 percent while keeping its ability to see fine details intact. Furthermore, they redesigned the network's internal structure so that this lightweight approach was used everywhere, from the initial image capture all the way to the final decision-making stage, rather than just in the beginning.

To make the system even better at spotting small objects, the team added a specific type of attention mechanism. In the same way a human might focus on a specific spot in a crowded room to find a friend, this addition helps the computer understand exactly where an object is located in space. By placing this focus tool at key layers of the network, the system became significantly better at distinguishing a small drone from a bird or a cloud. The result is a network that is not only smaller and faster but also more accurate than previous standard models. When tested on a custom collection of over 20,000 images featuring drones in various environments—from urban areas to forests—the new system correctly identified targets with high precision. It achieved a detection accuracy of 96.36 percent while using only 2.32 million parameters, a measure of the model's complexity, which is 23 percent fewer than the standard benchmark model.

The true test of this work, however, was not just in simulation but in physical deployment. The researchers took their design and installed it on a HiSilicon Hi3403 edge processor, a chip commonly used in real-world surveillance equipment. The system ran at a speed of 25.5 frames per second, meaning it could analyze more than 25 images every single second without lag. This performance met the strict requirement for real-time tracking, allowing the device to follow a moving drone instantly. Crucially, the system utilized 100 percent of the chip's specialized processing power, with no part of the calculation falling back to the slower main processor. This confirmed that their hardware-aware design successfully eliminated the bottlenecks that plague other systems. The researchers also tested the system on public datasets containing images of drones and birds, and it maintained its high performance without needing any extra training, proving it could generalize well to new situations.

This work demonstrates that high-performance artificial intelligence does not require massive, power-hungry computers. By carefully aligning the design of the software with the specific capabilities of the hardware it runs on, it is possible to create systems that are both lightweight and incredibly effective. The GhostV2-YOLO network offers a practical path forward for deploying reliable drone detection in the field, whether for securing airports, monitoring wildlife, or ensuring public safety. It shows that when engineers understand the constraints of their tools, they can build solutions that are not just theoretically sound, but practically ready for the real world.

Technical Summary: GhostV2-YOLO

Problem Statement

The rapid proliferation of Unmanned Aerial Vehicles (UAVs) has created an urgent need for reliable, real-time detection systems for airspace security and anti-drone operations. While deep learning-based object detection (specifically the YOLO series) offers high accuracy, deploying these models on resource-constrained edge computing platforms presents significant challenges.

A critical bottleneck identified in this work is the mismatch between standard lightweight network designs and the hardware constraints of edge Neural Processing Units (NPUs). Many edge NPUs, such as the HiSilicon Hi3403, are hardware-hardened to accelerate only specific operators (e.g., $1\times1$ and $3\times3$ convolutions). They often lack native support for non-standard operations common in advanced lightweight networks, such as $5\times5$ depthwise convolutions, channel shuffling, or specific activation functions (e.g., h-swish). When these unsupported operators are encountered, compilers often fall back to CPU execution, causing a drastic drop in inference speed that negates the theoretical efficiency gains of lightweight models.

Methodology

To address these constraints, the authors propose GhostV2-YOLO, a hardware-aware lightweight detection network built upon the YOLOv8 framework. The architecture is designed to achieve an optimal balance between accuracy and efficiency while ensuring 100% compatibility with edge NPU operator sets.

The methodology relies on three core architectural innovations:

DFC-Light Attention Mechanism:
- Problem: The original Dynamic Filter Convolution (DFC) attention uses a $5\times5$ depthwise convolution to capture long-range spatial dependencies. This operator is unsupported or inefficient on target NPUs.
- Solution: The authors decompose the $5\times5$ depthwise convolution into two cascaded $3\times3$ depthwise convolutions. This maintains an equivalent receptive field while reducing parameters by 28% (from $25C$ to $18C$ ) and ensuring full NPU compatibility.
C2f GhostV2 Module (Full-Link Lightweighting):
- Problem: Previous works often replaced only the backbone with lightweight modules (e.g., GhostNet) while retaining standard, heavier modules in the neck (feature fusion) and detection head.
- Solution: The authors replace all standard Bottleneck blocks within the YOLOv8 C2f module with GhostBottleneckV2 structures. This extends the lightweight GhostNetV2 paradigm from the backbone through the neck to the detection head, achieving true end-to-end parameter efficiency. The GhostBottleneckV2 integrates SE (Squeeze-and-Excitation) channel attention, while the DFC-Light spatial attention is applied to the concatenated output of the GhostModuleV2 instances within the bottleneck when the stride is 1.
Strategic CoordAtt Integration:
- Problem: Lightweighting often leads to a loss of feature representation capability, particularly for small and medium-sized targets common in UAV detection.
- Solution: Coordinate Attention (CoordAtt) modules are strategically inserted at the P3 ( $80\times80$ ) and P4 ( $40\times40$ ) feature layers of the backbone. This provides independent position-aware encoding to enhance the detection of small targets, complementing the channel and spatial attention mechanisms already present in the GhostBottleneck.

Training and Deployment:
The model was trained on a custom dataset of 20,256 UAV images using the AdamW optimizer and a composite loss function (CIoU + BCE). For deployment, the PyTorch model was exported to ONNX, compiled via the Ascend Tensor Compiler (ATC) into a dedicated .om format for the HiSilicon Hi3403 with INT8 quantization, and executed on a Linux-based edge device.

Key Contributions

Hardware-Aware Design: The paper introduces a detection network specifically optimized for the operator constraints of domestic edge NPUs, avoiding performance cliffs caused by CPU fallbacks.
Full-Link Lightweighting: Unlike prior art that limits lightweighting to the backbone, GhostV2-YOLO extends the GhostNetV2 structure throughout the entire network (backbone and neck), reducing parameters by 23% compared to the YOLOv8n baseline.
Synergistic Attention: The combination of DFC-Light, SE, and CoordAtt mechanisms compensates for the representational loss of the lightweight backbone, resulting in accuracy improvements despite the reduced parameter count.
Comprehensive Validation: The study includes extensive ablation studies, comparisons against mainstream architectures (MobileNetV3, ShuffleNetV2), and zero-shot transfer tests on public benchmarks (Drone-vs-Bird, DUT Anti-UAV).

Results

Experiments were conducted on a custom UAV dataset and two public benchmarks.

Accuracy-Efficiency Trade-off: On the custom dataset, GhostV2-YOLO achieved 96.36% mAP@50 and 67.05% mAP@50:95 with only 2.32M parameters and 6.3 GFLOPs.
- Compared to YOLOv8n, it reduced parameters by 23% while improving mAP@50:95 by 0.53%.
- It outperformed other lightweight models like MobileNetV3-YOLO and ShuffleNetV2-YOLO in both accuracy and efficiency.
Generalization: Zero-shot transfer tests showed significant improvements over the baseline on public datasets:
- Drone-vs-Bird: +4.35 percentage points in mAP@50 (42.29% to 46.64%) and +4.48 percentage points in mAP@50:95 (21.68% to 26.16%).
- DUT Anti-UAV: +0.41 percentage points in mAP@50:95 (50.07% to 50.34%).
Edge Deployment: Deployed on the HiSilicon Hi3403 edge NPU, the model achieved a real-time inference speed of 25.5 FPS with an end-to-end latency of 39.2 ms.
- The NPU utilization was 95.9%, and all operators were natively supported, confirming 100% compatibility without CPU fallback.
- The model size was 9.0 MB.

Significance and Claims

The paper claims that GhostV2-YOLO successfully addresses the "accuracy-efficiency-hardware" trilemma in edge UAV detection. Its primary significance lies in demonstrating that hardware-aware design—specifically tailoring operator choices to the target NPU's capabilities—is more critical than simply adopting the latest lightweight architecture.

The authors assert that their approach enables:

Real-time performance on low-cost, resource-constrained edge devices where standard lightweight models might fail due to unsupported operators.
Superior generalization for small target detection through the strategic use of coordinate attention, which is crucial for distinguishing UAVs from birds or background clutter.
Engineering feasibility, validated by a complete deployment pipeline from model training to on-board inference on a specific domestic NPU platform.

The work concludes that while single attention mechanisms offer limited gains, the synergistic integration of multiple attention types within a hardware-constrained architecture allows the model to surpass heavier baselines with significantly fewer parameters.