SDD-YOLO: A Small-Target Detection Framework for Ground-to-Air Anti-UAV Surveillance with Edge-Efficient Deployment

1. Problem Statement

The paper addresses the critical challenge of detecting small Unmanned Aerial Vehicles (UAVs) from a Ground-to-Air (G2A) perspective. This scenario presents three compounding difficulties:

• Sub-pixel Targets: At distances exceeding 100 meters, UAVs often occupy fewer than 16×16 pixels (sometimes <8 pixels) in standard images, resulting in a loss of geometric and texture details.
• Cluttered Backgrounds: Aerial backgrounds contain heterogeneous distractors (clouds, birds, buildings, backlighting) that cause high false-positive rates in generic detectors.
• Edge Deployment Constraints: Practical surveillance requires real-time inference (≥30 FPS) on resource-constrained edge hardware (e.g., domestic NPUs). Standard deep learning operators, particularly Distribution Focal Loss (DFL) and Non-Maximum Suppression (NMS), often fail under INT8 quantization or introduce unacceptable latency on these platforms.

Existing YOLO-based models typically rely on an 8× downsampling (P3) feature map, which collapses an 8-pixel target into a single feature point, rendering it undetectable.

2. Methodology: SDD-YOLO Architecture

The authors propose SDD-YOLO, a framework built on the YOLO26 architecture but heavily modified for micro-target detection and edge efficiency.

A. P2 High-Resolution Detection Head

• Innovation: Introduces a P2 detection head operating at 4× downsampling (compared to the standard P3 at 8×).
• Mechanism: For an 8-pixel target, P3 reduces it to a 1×1 feature response (losing all detail), whereas P2 preserves a 2×2 receptive window, retaining crucial edge textures.
• Implementation: The P2 feature map is generated by fusing shallow backbone outputs with upsampled P3 features via a C3 bottleneck.

B. Streamlined Inference (DFL-Free & NMS-Free)

• DFL Removal: The paper discards the Distribution Focal Loss (DFL) module (which uses Softmax and integral operations) due to its incompatibility with NPU quantization. It is replaced by a direct Wise-IoU v3 regression loss.
• NMS-Free Inference: Leveraging YOLO26's paradigm, the model uses a dual label assignment mechanism:
  • Training: One-to-Many (O2M) for rich gradient signals.
  • Inference: One-to-One (O2O) to ensure a single prediction per target without post-processing NMS.
  • Benefit: Reduces end-to-end latency by 20–30% on resource-constrained SoCs.

C. Dual Attention Mechanism

To suppress false positives from aerial clutter, a Dual Attention Module is embedded between the backbone and heads:

• Spatial Attention: Highlights motion regions and suppresses static background activation.
• Channel Attention: Re-weights channels to amplify UAV-discriminative frequencies while dampening background-heavy channels.

D. Advanced Training Strategies

• MuSGD Hybrid Optimizer: Adapts the Muon optimizer (originally for LLMs) for vision tasks. It applies Newton-Schulz iteration to orthogonalize gradient updates in high-dimensional weight matrices, preventing gradient oscillation on sparse micro-target data.
• ProgLoss & STAL: Uses Progressive Loss to dynamically re-weight loss components and Small-Target-Aware Label Assignment (STAL) to assign higher weights to micro-target anchors.
• Knowledge Distillation (KD): A multi-scale feature-aligned KD scheme transfers knowledge from a heavy teacher model (YOLO26x/l) to the compact student (SDD-YOLO-n), aligning feature representations at P2–P5 levels using KL divergence.

3. Key Contributions

1. DroneSOD-30K Dataset: A new large-scale G2A dataset containing ~30,000 annotated images covering diverse weather (fog, backlight, dawn/dusk) and "spot-level" targets (<20 pixels).
2. SDD-YOLO Architecture: The first integration of a P2 high-resolution head with DFL-free/NMS-free inference specifically for sub-pixel UAV detection.
3. Optimization Strategy: Successful adaptation of MuSGD and STAL to stabilize training on sparse, small-target signals.
4. Efficiency Validation: Comprehensive benchmarking proving the model's viability for edge deployment without sacrificing accuracy.

4. Experimental Results

The model was evaluated on the DroneSOD-30K dataset and compared against YOLOv5n and YOLO26n baselines.

• Accuracy:
  • SDD-YOLO-n (Final) achieved 86.0% mAP@0.5, surpassing the native YOLOv5n baseline (78.2%) by 7.8 percentage points.
  • It significantly improved recall on sub-16-pixel targets due to the P2 head.
• Efficiency (Inference Speed):
  • GPU (NVIDIA RTX 5090): 226 FPS.
  • CPU (Intel Xeon Platinum 8470Q): 35 FPS.
  • Despite the added P2 head, the model maintains 5.77 GFLOPs (lower than YOLOv5n's 7.18 GFLOPs) due to the removal of DFL and NMS.
• Ablation Studies:
  • The P2 Head alone provided a +6.8% mAP boost.
  • The DFL-free/NMS-free design improved CPU FPS from ~30.8 to 35.0 FPS while maintaining accuracy.

5. Significance and Impact

• Solving the "Sub-pixel" Problem: By shifting the detection head to 4× downsampling, SDD-YOLO recovers geometric details that standard 8× downsampling architectures inevitably lose, making it the first YOLO variant capable of reliably detecting UAVs at extreme distances.
• Edge-Ready Design: The removal of DFL and NMS makes the model uniquely suitable for domestic NPU platforms (e.g., Ascend, Rockchip) where these operators are poorly supported or cause accuracy collapse under INT8 quantization.
• Real-World Applicability: The combination of high accuracy (86% mAP) and real-time CPU performance (35 FPS) demonstrates that high-performance anti-UAV surveillance can be deployed on standard, low-cost hardware without requiring expensive GPU clusters.

Conclusion: SDD-YOLO represents a significant leap in small-target detection, bridging the gap between theoretical deep learning performance and the strict hardware constraints of real-world G2A surveillance systems.