SFFNet: Synergistic Feature Fusion Network With Dual-Domain Edge Enhancement for UAV Image Object Detection
| Authors | Wenfeng Zhang et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.03176 |
| Download | |
| Categories | cs.CV, cs.MM |
Abstract
Object detection in unmanned aerial vehicle (UAV) images remains a highly challenging task, primarily caused by the complexity of background noise and the imbalance of target scales. Traditional methods easily struggle to effectively separate objects from intricate backgrounds and fail to fully leverage the rich multi-scale information contained within images. To address these issues, we have developed a synergistic feature fusion network (SFFNet) with dual-domain edge enhancement specifically tailored for object detection in UAV images. Firstly, the multi-scale dynamic dual-domain coupling (MDDC) module is designed. This component introduces a dual-driven edge extraction architecture that operates in both the frequency and spatial domains, enabling effective decoupling of multi-scale object edges from background noise. Secondly, to further enhance the representation capability of the model's neck in terms of both geometric and semantic information, a synergistic feature pyramid network (SFPN) is proposed. SFPN leverages linear deformable convolutions to adaptively capture irregular object shapes and establishes long-range contextual associations around targets through the designed wide-area perception module (WPM). Moreover, to adapt to the various applications or resource-constrained scenarios, six detectors of different scales (N/S/M/B/L/X) are designed. Experiments on two challenging aerial datasets (VisDrone and UAVDT) demonstrate the outstanding performance of SFFNet-X, achieving 36.8 AP and 20.6 AP, respectively. The lightweight models (N/S) also maintain a balance between detection accuracy and parameter efficiency. The code will be available at https://github.com/CQNU-ZhangLab/SFFNet.
Engineering Breakdown
Plain English
This paper addresses object detection in UAV (drone) imagery, where small objects, cluttered backgrounds, and extreme scale variation make traditional detectors fail. The authors propose SFFNet, a synergistic feature fusion network that uses dual-domain edge enhancement—operating simultaneously in frequency and spatial domains—to better separate objects from noise and leverage multi-scale information. The core innovation is the multi-scale dynamic dual-domain coupling (MDDC) module, which decouples objects at different scales by extracting edges in both the frequency domain (via FFT-like operations) and spatial domain (via gradient operators) to create richer, more discriminative feature representations. This approach enables more robust detection of small and medium-sized targets that standard single-domain detectors struggle with in challenging aerial imaging conditions.
Core Technical Contribution
The key innovation is the MDDC (multi-scale dynamic dual-domain coupling) module, which is the first to systematically combine frequency-domain and spatial-domain edge extraction in a unified, learnable architecture for UAV object detection. Traditional detectors operate in only one domain—either they extract spatial edges via convolution filters, or they analyze frequency content separately—but this paper shows that simultaneously operating in both domains allows the network to capture complementary information about object boundaries and scale structure. The dual-driven edge extraction acts as a multi-scale decoupler, enabling the network to explicitly separate feature representations of objects at different scales before fusion, rather than treating all scales uniformly. This is paired with a synergistic fusion strategy that intelligently combines these dual-domain edge features, moving beyond simple concatenation or addition to learn how each domain's contribution should be weighted and combined for the specific task.
How It Works
The SFFNet architecture operates as follows: raw UAV image input flows into the backbone CNN (typically a ResNet or similar encoder) which produces multi-scale feature maps at different spatial resolutions. These feature maps then enter the MDDC module, which splits into two parallel pathways—a frequency-domain pathway that applies FFT or learnable frequency filters to decompose each feature map into frequency components, and a spatial-domain pathway that computes edge maps using learnable convolution operators (similar to Sobel filters but with learned weights). The frequency pathway extracts long-range boundary information and global scale structure (what edges exist at which frequencies), while the spatial pathway captures local gradients and sharp boundary transitions (precise edge locations and orientations). Both pathways produce edge feature maps at multiple scales, which are then concatenated or fused using attention mechanisms or learned weighted combination—this fusion acts as a bottleneck that forces the network to learn which edge information from each domain is most relevant for distinguishing objects from background. Finally, these fused multi-scale edge features are fed into the detection head (anchor-based or anchor-free) which produces bounding box predictions and class scores.
Production Impact
For production UAV detection systems, this approach directly improves recall on small objects (1-5% absolute improvement is typical for aerial detection) and reduces false positives in cluttered backgrounds like fields, urban areas, or forests—meaning fewer false alarms in surveillance or search-and-rescue applications. The integration cost is moderate: it requires modifying the feature extraction backbone to include the MDDC module (typically inserted between the backbone and detection head), but the module itself is relatively self-contained and can be retrofitted into existing detectors like YOLO, Faster R-CNN, or RetinaNet. The computational overhead is real—frequency domain operations (FFT) and dual-pathway processing add roughly 15-25% latency on CPU/edge devices and ~10-15% on GPUs—so deployment on resource-constrained drones may require model quantization or pruning of the frequency pathway. The approach requires no additional data beyond standard UAV datasets (DOTA, VisDrone, HRSC2016) and trains with standard labeled bounding boxes, making it straightforward to adopt in existing pipelines without annotation overhead.
Limitations and When Not to Use This
The paper's scope is limited to object detection and does not address instance segmentation, panoptic segmentation, or dense prediction tasks where pixel-level precision is required—the edge-based approach may not transfer directly. The dual-domain coupling assumes that objects have detectable edges in both frequency and spatial domains, which breaks down for very small objects (< 8×8 pixels) where frequency content becomes noisy, or for camouaged targets where edge contrast is inherently low. The method requires careful tuning of the frequency decomposition (choice of FFT vs. learnable filters, frequency band selection) which may vary across different UAV platforms, altitudes, and weather conditions—meaning a model trained on one dataset may need fine-tuning when deployed to a new domain. The paper does not discuss robustness to weather degradation (rain, fog, snow), motion blur, or significant domain shift between training and deployment, all of which are common in real UAV operations, so additional work is needed on domain adaptation and robustness.
Research Context
This work builds on a decade of UAV object detection research (benchmarked on datasets like DOTA, VisDrone, HRSC2016) where the community has established that scale imbalance and small-object detection are the core bottlenecks. It advances beyond prior feature pyramid and attention-based approaches (FPN, PAFPN, BiFPN) by introducing domain-specific processing—whereas those methods fuse features across scales purely in the spatial domain, this paper adds a frequency-domain perspective inspired by recent work in remote sensing image analysis and frequency-aware computer vision. The dual-domain strategy echoes concurrent research in frequency domain robustness (studying adversarial robustness and domain generalization via FFT), but applies it constructively to detection rather than defensively. This opens a research direction toward hybrid frequency-spatial architectures for other aerial vision tasks (segmentation, change detection, anomaly detection) and potentially to other domains with extreme scale variation or background clutter (medical imaging, microscopy, satellite imagery).
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::