Abstract
Small object detection in UAV imagery is crucial for applications such as search-and-rescue, traffic monitoring, and environmental surveillance, but it is hampered by tiny object size, low signal-to-noise ratios, and limited feature extraction. Existing multi-scale fusion methods help, but add computational burden and blur fine details, making small object detection in cluttered scenes difficult. To overcome these challenges, we propose the Multi-scale Global-detail Feature Integration Strategy (MGDFIS), a unified fusion framework that tightly couples global context with local detail to boost detection performance while maintaining efficiency. MGDFIS comprises three synergistic modules: the FusionLock-TSS Attention Module, which marries token-statistics self-attention with DynamicTanh normalization to highlight spectral and spatial cues at minimal cost; the Global-detail Integration Module, which fuses multi-scale context via directional convolution and parallel attention while preserving subtle shape and texture variations; and the Dynamic Pixel Attention Module, which generates pixel-wise weighting maps to rebalance uneven foreground and background distributions and sharpen responses to true object regions. Extensive experiments on the VisDrone benchmark demonstrate that MGDFIS consistently outperforms state-of-the-art methods across diverse backbone architectures and detection frameworks, achieving superior precision and recall with low inference time. By striking an optimal balance between accuracy and resource usage, MGDFIS provides a practical solution for small-object detection on resource-constrained UAV platforms.
Introduction
Detecting tiny objects from aerial platforms is a significant challenge for computer vision. When targets are smaller than 32x32 pixels, they often blend into background noise or are lost during the downsampling process of standard convolutional neural networks (CNNs). In Unmanned Aerial Vehicle (UAV) imagery, this problem is exacerbated by complex background textures and varying lighting conditions. While feature fusion—combining low-level spatial details with high-level semantic information—is a common remedy, traditional approaches often introduce excessive computational overhead or inadvertently blur the very details needed to identify small targets.
The MGDFIS Framework
To address these limitations, we introduce the Multi-scale Global-detail Feature Integration Strategy (MGDFIS). This framework is designed to be a "plug-and-play" enhancement for existing object detectors, focusing on preserving fine details while capturing broad context.
The strategy is built upon three core modules:
- FusionLock-TSS Attention (FTSSA): This module captures global context efficiently. It utilizes Token Statistics Self-Attention (TSSA) to achieve linear complexity, making it suitable for edge devices. It is paired with DynamicTanh normalization, a learnable activation function that adaptively filters noise by adjusting its range based on input variance.
- Global-detail Integration Module (GDIM): This module serves as the primary fusion engine. It includes a Global Mixing Module (GMM) for long-range pixel interactions and a Detail Mixing Module (DMM) that uses multi-directional convolutions to capture the structural signatures of small objects.
- Dynamic Pixel Attention Module (DPAM): To combat the imbalance between massive backgrounds and tiny foreground objects, DPAM generates pixel-wise weighting maps. This sharpens the network's focus on true object regions and suppresses background clutter.
Experimental Results
The MGDFIS framework was evaluated on the VisDrone 2019 benchmark, which contains over 10,000 images featuring high densities of small objects in diverse environments.
Quantitative Performance
MGDFIS demonstrated consistent improvements across various detection architectures:
- Faster R-CNN: Achieved a significant 12.5% increase in $AP_{50}$ (from 40.7% to 53.2%).
- YOLO Series: Improvements were observed across the board, including a 1.5% $AP_{50}$ boost for the modern YOLOv11s model.
Qualitative Analysis
Heatmap visualizations reveal that MGDFIS produces much tighter and more intense activations on small targets compared to baseline models. While standard models are often distracted by background textures like pavement or foliage, MGDFIS maintains a sharp focus on the actual objects of interest.
Conclusion and Future Work
MGDFIS offers a robust solution for small object detection on resource-constrained UAV platforms by balancing global context with local detail. By employing linear-complexity attention and directional convolutions, it achieves state-of-the-art performance without prohibitive computational costs. Future research will focus on optimizing the framework for specific edge hardware and extending its capabilities to handle extreme weather conditions and temporal consistency in video streams.
Summary updated