YOLO-FEDER FusionNet: Drone & Federated Detection

• The paper demonstrates that a dual-backbone architecture, merging YOLO’s generic features with FEDER’s camouflage detection through attention-based fusion, significantly improves detection metrics.
• YOLO-FEDER FusionNet integrates complementary information by combining feature decomposition with Federated Averaging to overcome challenges posed by fragmented and heterogeneous training data.
• The system leverages attention mechanisms and optimized training protocols to boost mAP while reducing false negatives and false discovery rates in both drone surveillance and industrial scenarios.

Searching arXiv for the cited and closely related papers to ground the article. arxiv_search.query({"search_query":"id:(Lenhard et al., 2024) OR id:(Hegiste et al., 2023) OR id:(Hegiste et al., 2023) OR all:\"Feature Decomposition and Edge Reconstruction\" OR all:Federated Averaging YOLOv5 object detection", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"}) Retrying with a narrower arXiv query. arxiv_search.query({"search_query":"id:(Lenhard et al., 2024)", "max_results": 5, "sort_by": "relevance", "sort_order": "descending"}) YOLO-FEDER FusionNet denotes a family of YOLO-centered object-detection systems in which complementary information sources are fused to improve robustness under distribution shift. In the most specific and formal usage, it is the name of a dual-backbone architecture for drone detection that combines a generic YOLO detector with the camouflage object detector FEDER, using attention-based fusion to improve performance in complex, highly textured scenes (Lenhard et al., 2024). In a related but distinct usage, the same expression is used descriptively for a federated, YOLO-based object-detection framework in industrial quality inspection, where multiple clients train a shared YOLO model via Federated Averaging without exchanging raw images (Hegiste et al., 2023). The two usages differ in mechanism—feature fusion versus parameter aggregation—but both address a common problem: standard YOLO detectors degrade when training data are fragmented, visually heterogeneous, or poorly matched to deployment conditions.

1. Terminology and scope

The literature uses the expression in two overlapping senses.

Usage	Core mechanism	Representative context
Drone-detection YOLO-FEDER FusionNet	Dual backbones, feature-level fusion, attention in the neck	Camouflaged drones near trees, branches, or urban clutter
Federated YOLO-based “FusionNet”	FedAvg aggregation of local YOLO models	Shared production and industrial quality inspection

In the drone-detection paper, YOLO-FEDER FusionNet is a named architecture in which a YOLOv5l detector is fused with FEDER, a camouflage object detection model that outputs a segmentation map of camouflaged regions (Lenhard et al., 2024). In the manufacturing papers, the same label is used as a convenient description of a federated YOLO system built with YOLOv5 and FedAvg, sometimes interpreted as a “fusion” of multiple local detectors into a single global model via weight averaging (Hegiste et al., 2023). A recurrent source of confusion is therefore terminological rather than technical: in the drone-detection line, “FEDER” refers to the camouflage branch; in the industrial line, the emphasis is on federated learning rather than a FEDER backbone.

This overlap is significant because it links two different research agendas. One agenda studies architectural fusion inside a single network; the other studies cross-client fusion during training. The former is motivated by camouflage and textured backgrounds, while the latter is motivated by privacy, non-IID data, and generalization across factories.

2. Federated YOLO object detection in shared production

In shared-production quality inspection, the underlying problem is that multiple factories must inspect related products but do not want to share raw images. The industrial formulation uses YOLOv5 as the detector and Federated Averaging as the optimization protocol, with a neutral server that distributes global weights, collects local updates, and aggregates them into a new global model (Hegiste et al., 2023). The motivating settings include USB-stick inspection across three clients and cabin-plus-windshield inspection across two clients, both deliberately non-IID because product appearance, defect type, background, and illumination differ across clients.

The shared-production scenario is defined as a multi-factory ecosystem in which multiple companies contribute to a common product family while preserving data privacy. In the cabin setting, each client initially sees only a subset of windshield types, yet the eventual deployment requires both clients to handle unseen combinations of cabin color and windshield type. In the USB setting, each client’s Not_OKAY class corresponds to a different error mode—small sticker or scratch marks, physically damaged or bent ports, or rusted ports—so the federated model must combine heterogeneous defect modes without access to the underlying images (Hegiste et al., 2023).

The optimization rule follows standard FedAvg: $w^{t+1} = \sum_{k=1}^K \frac{n_k}{n}\, w_k^t,$ where $w^t$ is the global model, $w_k^t$ is the locally trained model at client $k$, and $n_k$ is the local sample count. Each communication round sends $w^t$ to all clients, performs local YOLO training for a fixed number of epochs, and averages the returned weights on the server. In the cabin use case, the reported schedule is 15 local epochs per round and 10 communication rounds, with a stopping condition of average accuracy $>96\%$ across clients. In the USB use case, the federated model uses 5 communication rounds with 15 local epochs per round (Hegiste et al., 2023).

A second federated line interprets this procedure through an ensemble lens. “Federated Ensemble YOLOv5” uses YOLOv5m across three clients, with a centralized dataset shuffled and partitioned into mutually exclusive client subsets. The authors explicitly characterize federated training as analogous to a synergistic blend of Bagging and Boosting: Bagging because clients train on different subsets, and Boosting because the global model is redistributed to clients after each round as a stronger initialization (Hegiste et al., 2023). In this interpretation, the “fusion” is a parameter-space ensemble rather than a multi-branch network.

3. Dual-backbone YOLO-FEDER FusionNet for drone detection

The drone-detection architecture addresses a different failure mode: standard YOLO models work well against simple, high-contrast backgrounds but often fail when drones are small, far away, or embedded in highly textured backgrounds such as foliage and dense urban structure (Lenhard et al., 2024). The central idea is to fuse a generic object detector with a camouflage object detector so that the detection pipeline is guided by cues that are specifically informative when object-background separability is low.

The original architecture contains two parallel backbones: a YOLOv5l backbone, identified as CSPDarkNet53, and a FEDER backbone for camouflage object detection (Lenhard et al., 2024). The input is an RGB image

$\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$

with $W=H$ and resized to $640\times 640$ in practice. The YOLO branch extracts hierarchical detection features. The FEDER branch encodes the same image with a camouflage feature encoder, decomposes features into high-frequency and low-frequency components via DWD, and decodes them with SED to produce a segmentation map $w^t$0 and an edge map $w^t$1. Only $w^t$2 is used in the original FusionNet.

Fusion occurs in the neck. YOLO feature maps are concatenated with the FEDER segmentation map at specific neck layers, channel attention is applied to reweight the fused representation, and CBAM is inserted inside C3 blocks to refine features through channel and spatial attention (Lenhard et al., 2024). The fusion step is written as

$w^t$3

and CBAM refinement as

$w^t$4

The detection head remains a standard YOLOv5l head that predicts $w^t$5, objectness, and class scores at three scales.

Training in the original model freezes both backbones. The YOLOv5l backbone is pre-trained on COCO, the FEDER branch is initialized with COD10K weights, and only the neck and head are trained (Lenhard et al., 2024). The detection loss is the standard YOLOv5 loss, decomposed into localization, objectness, and classification terms. This design isolates the contribution of fusion and attention rather than jointly re-optimizing both heavy backbones.

4. FEDER, feature decomposition, and the optimized iteration

The 2025 optimized iteration substantially broadens the architectural scope while retaining the central idea of combining generic YOLO features with camouflage-aware representations (Lenhard et al., 17 Sep 2025). In this version, the YOLO backbone can be selected from YOLOv5, YOLOv8, YOLOv9, or YOLOv11 variants; the neck integrates YOLO and FEDER features via four Attention Fusion Modules; and the head is YOLOv8-like and anchor-free, using a Distribution Focal Loss block for refined box regression.

FEDER itself is described as a 1,186-layer, 44.1M-parameter camouflage object detection model with three main components: a Camouflaged Feature Encoder, a Deep Wavelet-like Decomposition module, and a Segmentation-oriented Edge-assisted Decoder (Lenhard et al., 17 Sep 2025). DWD separates high-frequency and low-frequency content, reflecting a division between texture or edge cues and more global color or illumination structure. SED then refines these cues and produces both hierarchical decoder features and the final camouflage segmentation map. The optimized paper evaluates whether FusionNet should use only the final segmentation map, intermediate SED features, intermediate DWD features, or combinations thereof.

The principal result of that ablation is that Config. 4—using intermediate DWD features $w^t$6 alone—provides the best overall trade-off in mAP, FNR, FDR, and fitness, with an average fitness of $w^t$7 (Lenhard et al., 17 Sep 2025). This shifts the interpretation of FusionNet away from “segmentation-map injection” toward multi-scale spectral-feature fusion. The AFM operation is formalized as

$w^t$8

so the network learns how strongly to trust channels from the generic detector and from the camouflage branch at each stage of the neck.

The optimization paper also revises backbone choice. It compares YOLOv5l, YOLOv8m, YOLOv8l, YOLOv9c, YOLOv9e, YOLOv11l, and YOLOv11x inside the FusionNet framework and reports that YOLOv8l provides the best accuracy-efficiency balance, with 184 layers, 19.8M parameters, and 86.0 GFLOPs, outperforming the original YOLOv5l baseline in fitness while using fewer parameters than YOLOv5l’s 26.6M (Lenhard et al., 17 Sep 2025). The final loss in this version is

$w^t$9

and the selected weights after ablation are $w_k^t$0, $w_k^t$1, and $w_k^t$2.

5. Data regimes, evaluation protocols, and reported performance

The industrial federated line and the drone-detection line use markedly different data regimes. In shared production, the cabin dataset contains approximately 600 images per client with a 70/15/15 train/validation/test split, three backgrounds, and differing lighting, shadows, and blur; the USB use case distributes three different product types and three different defect patterns across three clients (Hegiste et al., 2023). The federated ensemble paper instead partitions centralized manufacturing datasets into three mutually exclusive client subsets, keeping validation and test sets identical across clients and the global model (Hegiste et al., 2023).

The original drone FusionNet is trained on a synthetic dataset S1 generated with Unreal Engine 4.27 and AirSim, using $w_k^t$3 training images, $w_k^t$4 validation images, and $w_k^t$5 test images, and is evaluated on two real datasets, R1 and R2, captured with a Basler acA200-165c camera (Lenhard et al., 2024). R1 is building-dominated and comparatively easier; R2 is tree-dominated and explicitly described as the harder, highly textured, camouflaging scenario. The optimized iteration expands the training corpus to SynDroneVision, with $w_k^t$6 images and 13 drone models, plus DUT Anti-UAV, with $w_k^t$7 images and 35 drone models, while evaluation remains real-only on R1 and R2 (Lenhard et al., 17 Sep 2025).

The reported metrics likewise reflect different operational priorities. Industrial papers emphasize mAP, AP across IoU thresholds, AP for medium and large objects, recall, and qualitative bounding-box precision (Hegiste et al., 2023). The drone papers emphasize mAP at IoU $w_k^t$8 and $w_k^t$9, False Negative Rate, and False Discovery Rate, because missed detections and false alarms are operationally central in surveillance (Lenhard et al., 2024). The optimized paper combines these into a scalar fitness: $k$0 thereby assigning greatest weight to recall-oriented behavior and false-alarm control (Lenhard et al., 17 Sep 2025).

Performance claims in the industrial line focus on generalization to unseen combinations. In the cabin experiment, the global federated model reaches $k$1, $k$2, and $k$3 on blue-cabin/red-cabin combinations that were not present in any client’s local training data, whereas local models obtain $k$4 and $k$5 on the same combined test set (Hegiste et al., 2023). In the USB setting, the federated model detects a rust defect from client 3 on client 1’s Huawei USB stick, which is an unseen combination, and the authors report improved bounding boxes relative to local-only training (Hegiste et al., 2023). The federated ensemble paper does not report numeric mAP but repeatedly attributes more precise bounding boxes, fewer false positives, and stronger generalization to unseen trailer and cabin configurations to the global federated model (Hegiste et al., 2023).

Performance claims in the drone line are more granular. On R1 at $k$6, the original YOLO-FEDER FusionNet reports $k$7, $k$8, $k$9, and $n_k$0, compared with YOLOv5l at $n_k$1, $n_k$2, $n_k$3, and $n_k$4, respectively (Lenhard et al., 2024). On the more difficult R2 at $n_k$5, FusionNet reports $n_k$6, $n_k$7, $n_k$8, and $n_k$9, compared with YOLOv5l at $w^t$0, $w^t$1, $w^t$2, and $w^t$3 (Lenhard et al., 2024). The same paper also introduces a labeling-bias correction for real-data annotations by enlarging predicted boxes with fixed or adaptive scale factors; on R2 at $w^t$4, this raises FusionNet’s $w^t$5 from $w^t$6 to $w^t$7 with fixed scaling and to $w^t$8 with variable scaling (Lenhard et al., 2024).

The optimized iteration extends these gains. Relative to the original baseline, the final configuration—YOLOv8l plus DWD features and optimized training data—reaches $w^t$9, $>96\%$0, and $>96\%$1 on R2 at $>96\%$2, compared with the baseline’s $>96\%$3, $>96\%$4, and $>96\%$5 (Lenhard et al., 17 Sep 2025). On R2 at $>96\%$6, it reaches $>96\%$7, $>96\%$8, and $>96\%$9, compared with $\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$0, $\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$1, and $\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$2 for the baseline (Lenhard et al., 17 Sep 2025). The abstract summarizes the aggregate effect as an FNR reduction of up to 39.1 percentage points and a mAP increase of up to 62.8 percentage points at an IoU threshold of $\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$3 (Lenhard et al., 17 Sep 2025).

6. Limitations, misconceptions, and broader research context

Both research lines explicitly identify limitations. In the federated industrial setting, the main constraints are the small number of clients, the assumption that all clients participate in every round, the absence of personalization, the lack of explicit treatment of class imbalance and rare defects, and the absence of formal privacy guarantees such as differential privacy or secure aggregation (Hegiste et al., 2023). The federated ensemble study adds that its partitions are derived from a centralized dataset and therefore do not instantiate a true privacy-driven FL scenario, and it relies primarily on qualitative rather than quantitative evaluation (Hegiste et al., 2023).

In the drone-detection setting, the main limitations are computational cost, domain specificity, and residual difficulty with very small or extremely challenging targets. The original FusionNet does not report exact FPS, even though it is intended for surveillance drone detection and is run on a single NVIDIA Quadro RTX-8000 GPU (Lenhard et al., 2024). The optimized paper states that dual backbones increase inference cost, that objects smaller than approximately $\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$4 to $\mathbf{X} \in \mathbb{R}^{W \times H \times 3},$5 pixels remain difficult, and that some false positives arise systematically from reflections, clutter, or small structural details that mimic drone shapes (Lenhard et al., 17 Sep 2025). The use of minimum size thresholds is shown to reduce FDR, but it remains a heuristic rather than a complete solution.

The broader literature around these systems spans federated object detection, camouflage object detection, synthetic-to-real training, and real-time detection in resource-constrained networks. FedPylot, for example, studies federated optimization of original anchor-based YOLOv7 in the Internet of Vehicles, with full client participation, FedAvg and FedAvgM, realistic heterogeneity through concept drift and label skew, and hybrid RSA-AES encryption for server-client communication (Quéméneur et al., 2024). That work is not a YOLO-FEDER FusionNet implementation, but it provides a concrete systems template for federated real-time object detection at scale. A plausible implication is that future work may combine the architectural ideas of camouflage-aware FusionNet with the distributed optimization and communication machinery of federated YOLO systems.

The principal misconception to avoid is that all papers using similar terminology describe the same architecture. They do not. One line studies a federated training protocol for YOLO in industrial inspection; the other studies a dual-backbone fusion architecture for drone detection. What unifies them is not a single canonical implementation, but a shared methodological premise: generic object detection benefits when YOLO is fused with information that standard end-to-end training does not already capture, whether that information comes from other clients through FedAvg or from a camouflage-specialized branch through attention-based feature fusion.