To mitigate small-object feature loss and background interference in UAV imagery, we propose a lightweight YOLOv11n-based detector. As shown in Fig. 1, our method integrates LPConv for orthogonal feature alignment and geometric detail capture via a dynamic channel strategy. SG-PANet constructs a high-resolution P2 interaction space, while the GCA operator suppresses cross-scale fusion noise, ensuring semantic fidelity for faint objects. It is important to note that MFCI-YOLO refers to our overall end-to-end detection framework, whereas the SG-PANet specifically designates the novel neck architecture (feature fusion network) embedded within it.

In the refinement and detection stage, CSPStage is deployed before the head to denoise and reconstruct features via full-history state awareness, bridging semantic gaps and enhancing robustness in occlusions. Finally, the Wise-IoU v3 loss function prioritizes high-quality small object instances, significantly boosting performance while maintaining real-time efficiency.

UAV-detected objects show distinct anisotropic elongated features. Traditional 3 × 3 convolution with square receptive fields introduces excessive background noise when processing such features, so we propose a lightweight pinwheel context convolution (LPConv) based on orthogonal center alignment and dual-stream complementarity.

As shown in Fig. 2, the input feature tensor is defined as: X∈RC×H×W. Where X is the input tensor, C is channel count, and H,W are its height and width.

To improve detection and avoid channel dimension explosion, we split input features into a processing stream (Xp) for feature enhancement and an identity stream (Xi) for original feature preservation: Xp=X[1:C2,:,:],Xi=X[C2+1:C,:,:]. Subscripts denote channel slicing, with Xp,Xi∈R2C×H×W.

For Xp, we build a Local Windmill Extraction Unit with three anisotropic convolution branches: 3 × 3 depthwise separable convolution (core, high-frequency texture), 1 × k (horizontal context), and k×1 (vertical context). Specifically, the 1×k and k×1 orthogonal branches are selected to construct a cross-shaped receptive field that tightly aligns with the anisotropic, slender morphologies of typical aerial objects (e.g., vehicles, pedestrians) from a top-down view, effectively avoiding the diagonal background noise introduced by standard k×k square filters. Residual additive aggregation unifies these features:(1)Yp=ℱcore(Xp)⊕ℱhor(Xp)⊕ℱver(Xp)where Yp is the processing stream output, ℱcore/ℱhor/ℱver are the three branches, and ⊕ is element-wise addition.

When LPConv is specifically employed as a downsampling module to replace the standard Conv layer, a coordinated spatial reduction strategy is applied to both streams. The initial 1 × 1 splitting convolution maintains a stride of 1. Subsequently, the parallel depthwise convolutions (the 3×3, 1×k, and k×1 branches) in the processing stream operate with a stride of 2 (s=2) to halve the spatial dimensions while extracting multi-scale contextual features. Synchronously, a 2×2 Max Pooling operation is applied to the identity stream (Xi). This pooling step is essential to strictly align the spatial dimensions of Yp and Yi (both reduced to H/2×W/2) prior to the final channel concatenation. For standard feature extraction without downsampling, the stride remains s=1 and the Max Pooling operation is bypassed.

Finally, we remap and unify these complementary subspaces by concatenating Yp and Yi along the channel dimension. A 1×1 pointwise convolution then aggregates features and adjusts channel depth. The final LPConv output, Yout, is expressed as: Yout=ℱfuse(Concat(Yp,Yi)). Where Concat(⋅) is channel-wise splicing and ℱfuse is 1 × 1 fusion convolution.

In simpler terms, equations mathematically describe a split-and-conquer strategy. By dividing the feature maps, LPConv ensures that one half actively captures the orthogonal shapes of small objects, while the other half preserves the original uncorrupted details, achieving a highly efficient balance between contextual learning and parameter reduction.

Unlike conventional strip convolution methods that simply alternate 1×k and k×1 kernels to expand the receptive field at the cost of losing central high-frequency details, LPConv is driven by a dual-stream orthogonal alignment motivation. By explicitly decoupling the feature space into a processing stream for anisotropic contextual capture and an identity stream for fine-grained texture preservation, LPConv fundamentally avoids the detail degradation typical in extremely small objects during deep feature extraction.

For lightweight evaluation, assume input/output channels C, core kernel kcore=3, and strip kernel kstrip=k. The standard 3×3 convolution parameter count is Pstd=9C2. Because LPConv performs spatial convolution on only half the channels, its total parameter count is PLPC=C2+C2(9+2k). The parameter compression ratio is thus defined as ηparam=PLPC/Pstd. Setting C=128 and k=7, we obtain ηparam≈12.1%. This theoretically demonstrates that LPConv expands the effective receptive field to 7×7 while retaining robust feature extraction capabilities using only 12.1% of the parameters of a standard convolution.

The CARAFE operator proposed by Wang et al. [24] employs a dynamic kernel generation mechanism, yet its channel sharing approach exhibits limitations in UAV aerial photography scenarios. As background information dominates the scene, the shared kernel often functions as a smoothing filter, leading to the misidentification of small object edges as noise and their subsequent suppression. This results in the phenomenon of feature drowning.

To address this issue, we propose a group-content-aware (GCA) reconstruction method based on low-level decoupling. As illustrated in Fig. 3, we partition the channel space C into G independent semantic groups (Cg=C/G for each group) and process them through two parallel branches. In our implementation, the number of groups G is empirically set to 4. The choice of G significantly impacts both detection performance and computational efficiency. A smaller G (e.g., G = 1) regresses to a channel-shared mechanism, failing to decouple the background from the object and leading to feature drowning. Conversely, an excessively large G forces the network to generate too many independent kernels, which not only drastically increases the memory access cost (MAC) and computational overhead but also disrupts the synergistic feature representation within semantic channel groups. Setting G = 4 effectively strikes an optimal balance, ensuring sufficiently fine-grained semantic decoupling for small objects while adhering to strict lightweight deployment constraints.

In the group kernel prediction branch, input features undergo 1 × 1 convolution compression before being processed through group convolutions (content encoder), which concurrently output G sets of independent parameters. Following Pixel Shuffle and Softmax normalisation, this generates G distinct re-organised kernel tensors {W1,W2,…,WG}.

Simultaneously, within the grouping and reorganisation branch, the input X is decomposed into G sub-streams {X1,…,XG}. Each sub-stream undergoes grouped dot product reorganisation solely with its corresponding kernel Wg:(2)Xl′,c,g′=∑n=−rr∑m=−rrWl′(n,m)g⋅X(l′+n,j+m),c,ggwhere Xl′,c,g′ denotes the pixel at position l′, channel c, and group g. Wg is the reorganization kernel with radius r, while (n,m) represents the local coordinate offset.

To restore channel capacity and spatial alignment, high-frequency details are reaggregated by concatenating all G reconstructed subflows {X1′,…,XG′} along the channel dimension. The final output Yout is expressed as: Yout=Concat(X1′,X2′,…,XG′). Where Concat(⋅) denotes the concatenation operation along the channel dimension.

The recombined sub-streams are reshaped from B×G×(C/G)×oH×oW back to the standard B×C×H×W format. Unlike CARAFE’s shared smoothing kernels, GCA assigns dedicated high-frequency kernels to independent semantic groups. This prevents small-object textures from being suppressed by dominant backgrounds, effectively alleviating feature drowning and enhancing discriminative details for detection.

Intuitively, a globally shared kernel acts as a single “broad brush” that inevitably blurs fine object details into the dominant background. In contrast, our grouped GCA mechanism assigns dedicated “fine-tipped brushes” exclusively for the high-frequency textures of small objects, explicitly decoupling them from noise and completely avoiding the feature drowning effect.

To address the high-frequency feature decay of small objects in deep convolutional networks, we draw inspiration from the designs of GiraffeDet [25] and CSPNet [26] to propose the Cross-Stage Dense Feature Refinement module (CSPStage). This module enhances the model’s representational capability for low-pixel-count objects through full-spectrum feature reuse and structural reparameterization.

As shown in Fig. 4, CSPStage employs a dual-stream feature evolution system. The input tensor X is projected onto the baseline subspace Y1 and the evolutionary subspace Y2(0).

In the baseline branch, 1 × 1 convolutions construct truncated gradient highways, enabling raw high-resolution spatial features to bypass deep nonlinear transformations and reach the fusion layer directly, which ensures precise pixel-level localization of small objects.

Within the evolutionary branch, the dense aggregation mechanism integrates the outputs Y2(i) from n cascaded units. This unit employs a RepConv operator, which leverages a multi-branch topology during training to capture multi-scale features. During inference, it fuses equivalently into a single kernel, enhancing extraction robustness without increasing.

By fusing shallow spatial and deep semantic features, Yout mitigates small-object scarcity. Positioning CSPStage before the detection head aligns task features and bridges semantic gaps. Under dense occlusions, the shallow branch bypasses deep non-linearities to preserve high-resolution textures of unoccluded fragments, while the deep branch provides robust semantic context. Aggregating these streams enables MFCI-YOLO to infer partially visible objects from surviving local textures and global priors, significantly reducing missed detections.

To improve localization robustness against low-quality annotations, we employ Wise-IoU v3. It first constructs a geometric distance penalty RWIoU=exp⁡(ρ2(Wg2+Hg2)∗), where ρ is the Euclidean distance between predicted and ground-truth centers, and Wg,Hg are the minimum bounding box dimensions illustrated in Fig. 5. The asterisk (*) denotes detachment from the computation graph, ensuring it acts purely as an attention weight without hindering small-object convergence, yielding the base loss LWIoU_v1=RWIoU⋅LIoU.

To optimize gradient allocation across inconsistent annotation qualities, we introduce an outlier degree β=LWIoU_v1/LIoU¯ (with LIoU¯ as the batch’s moving average loss) and a dynamic focusing gain r=βδαβ−δ, producing the final loss LWIoU_v3=r⋅LWIoU_v1. As depicted by the non-monotonic curve in Fig. 6, this mechanism acts as an intelligent dynamic filter. Calculus analysis indicates r peaks at β=1/ln⁡α, effectively down-weighting both “easy” high-quality samples (β≪1) to prevent overfitting, and “hard” extreme outliers (β≫1) to avoid harmful gradient updates from noisy labels. Consequently, the network optimally prioritizes ordinary-quality, challenging small objects.

Although YOLOv11n [27] native path aggregation network enhances feature flow, its P3–P5 architecture causes scale mismatch and fusion loss in UAV aerial images: 32x downsampling loses small object information, and simple interpolation or convolution fails to distinguish objects from backgrounds, drowning small object features and amplifying background noise. To solve this, we propose Scale-Gradient and Context-Aware Feature Fusion Network (SG-PANet). As shown in Fig. 7, SG-PANet adopts a scale-down strategy: discarding P5 and extending fusion paths to shallow layers, constructing a P2-P3-P4 high-resolution fusion hierarchy. This preserves geometric textures and resolves small object information diffusion.

We introduce GCA as the fusion operator to address feature drowning. Unlike traditional interpolation (blending background and object features), GCA uses group decoupling to partition channels, dynamically enhancing object-related channels, thus injecting deep semantics into shallow features while preserving object details. Meanwhile, we replace standard convolutions with LPConv in the bottom-up localization enhancement path to reduce computational redundancy and texture noise in high-resolution features. LPConv’s orthogonal alignment uses strip convolution branches to capture small-object morphology and additive fusion to preserve high-frequency details, reducing computation and resolving localization deviations. In summary, SG-PANet achieves performance improvement via physical-scale reconstruction and bidirectional path optimization.

VisDrone2019 [28] (Fig. 8) by Tianjin University’s AISKYEYE team focuses on UAV-based computer vision tasks, comprising 10,209 static images (6471 training, 548 validation, 1610 testing) across 14 Chinese cities’ urban/rural scenes, with 10 core categories.

To validate MFCI-YOLO’s robustness for aerial small object detection, we adopted DOTA v1.5 [29] (Fig. 9)—an upgraded DOTA v1.0 retaining 2806 images, with finer annotations and abundant <10-pixel small instances.

Figs. 8 and 9 visually confirm that most UAV instances occupy sub-10-pixel ratios with severe class imbalance. This extreme distribution empirically justifies SG-PANet for preserving high-resolution details, and Wise-IoU v3 for dynamically balancing the gradients of small, low-quality samples.

The ablation and comparison experiments utilize the VisDrone2019 dataset (Fig. 8) to validate the model’s detection performance, while the model’s generalization capability is evaluated using the DOTA v1.5 dataset (Fig. 9).

The experiments were conducted using an NVIDIA RTX 5090 GPU (32 GB VRAM) with CUDA 12.8 and PyTorch 2.8.0. The model was trained for up to 200 epochs (with early stopping upon convergence) using the SGD optimizer (momentum: 0.937, weight decay: 0.0005) and a Cosine Annealing scheduler with an initial learning rate of 0.01. Furthermore, to ensure statistical stability and reproducibility, all reported evaluation metrics represent the average results from three independent training runs under identical hyperparameter settings.

We comprehensively evaluate the model’s detection performance using standard accuracy metrics: Precision (P), Recall (R), mAP50, and mAP50-95. Furthermore, to strictly assess the model’s deployability on resource-constrained UAV edge devices, we measure computational efficiency using total parameters (M), computational complexity (GFLOPs), and real-time inference speed (FPS).

To systematically validate the effectiveness of each component within the MFCI-YOLO framework and their respective contributions to small object detection performance in UAV aerial photography, we designed a layer-by-layer ablation study on the VisDrone2019 dataset. Using YOLOv11n as the baseline model, we progressively integrated the Wise-IoU v3, SG-PANet, LPConv, GCA, and CSPStage modules. We quantitatively analyzed the trade-offs among accuracy (mAP), number of parameters, and FPS for each module. The quantitative results are shown in Table 1.

Table 1 demonstrates that the baseline YOLOv11n struggles with deep downsampling-induced feature diffusion, yielding only 32.1% mAP50. Integrating SG-PANet drastically boosts mAP50 to 47.3% by preserving high-resolution details. Replacing standard convolutions with LPConv effectively cuts parameters to 3.38M and GFLOPs to 38.5 with negligible accuracy drop, as its anisotropic receptive field perfectly matches elongated aerial objects. Furthermore, introducing GCA suppresses background noise, and CSPStage mitigates deep feature decay via historical state awareness. Ultimately, MFCI-YOLO achieves an optimal 48.2% mAP50 and 29.8% mAP50-95 at a real-time 97 FPS, significantly outperforming the baseline across all extremely small object categories (e.g., Pedestrian and Bicycle AP improved by 21.3% and 14.1%, respectively).

The training process visualized in Fig. 10 shows that MFCI-YOLO achieves a higher growth rate and superior final steady-state values for both mAP50 and mAP50-95 compared to YOLOv11n. Regarding loss functions, MFCI-YOLO demonstrates faster convergence and more stable final values in classification (Cls Loss) and distribution focal loss (DFL Loss). This confirms that the synergy of SG-PANet and Wise-IoU v3 effectively optimizes gradient propagation and enhances learning efficiency for small objects.

Fig. 11’s visualization offers intuitive proof. The baseline has many blue missed detections in dense crowds and distant vehicles, while MFCI-YOLO corrects these to green true detections. In summary, MFCI-YOLO optimally balances accuracy and robustness for UAV aerial scenarios under limited computation.

To validate robustness, we evaluated the model on the challenging DOTA v1.5 dataset. As shown in Table 5, our method achieves an outstanding 80.3% mAP50.

DOTA v1.5 presents extreme challenges: arbitrary orientations, massive scale variations, and dense sub-10-pixel clusters. MFCI-YOLO directly addresses these: LPConv’s orthogonal branches capture slender geometries, while GCA’s independent high-frequency kernels suppress severe background interference, effectively preventing dense objects from drowning.

Fig. 13 corroborates this analysis. Even under extreme lighting and occlusion, MFCI-YOLO achieves 0.894 precision for swimming pools and minimizes missed detections in dense vehicle areas. This confirms that CSPStage effectively bridges the semantic gap via historical state awareness. Overall, MFCI-YOLO demonstrates exceptional robustness and practical value for complex aerial deployment.

To highlight our network’s advantages, Table 5 compares MFCI-YOLO against mainstream baselines and recent state-of-the-art detectors on the DOTAv1.5 dataset. Earlier models like YOLOv5n and YOLOv8n achieve only 62.5% and 65.6% mAP50. Recent advanced networks like BRSTD, AG-YOLO, and AESOD also struggle with dense, arbitrary-oriented targets, scoring 65.4%, 72.8%, and 48.9%, respectively. While the baseline YOLOv11n reaches 75.3%, it remains limited by feature drowning in complex scenes. In contrast, MFCI-YOLO achieves an exceptional 80.3% mAP50. Notably, it outperforms AG-YOLO by 7.5% using nearly one-fourth of its parameters. Although this precision involves a parameter increase (from 2.6M to 4.5M), our model strictly satisfies the 5M lightweight constraint. This trade-off confirms its superior accuracy, robust generalization, and essential lightweight profile for complex aerial imagery.

Although MFCI-YOLO exhibits exceptional generalization in dense scenarios, it still occasionally encounters missed detections and false positives under extreme physical conditions. Specifically, during severe motion blur caused by high-speed UAV maneuvering, the structural contours of extremely small objects become severely distorted, causing the anisotropic receptive field of LPConv to fail in feature alignment. Additionally, under heavy nighttime glare or extreme low-light conditions, background noise and object textures become virtually indistinguishable, occasionally bypassing the GCA operator’s decoupling mechanism. Addressing these extreme multi-degradation scenarios via multi-modal fusion or semi-supervised learning remains an important avenue for our future research.