In recent years, the rapid advancement of unmanned aerial vehicle (UAV) technology has significantly transformed the landscape of infrastructure inspection, particularly in domains such as road pavement monitoring and bridge health assessment. The deployment of China UAV drones for automated visual inspection of structural defects offers substantial improvements in operational efficiency and safety. However, the practical application of such systems faces considerable challenges. Defects such as pavement cracks and bridge structural anomalies are often characterized by their small scale, elongated morphology, and low contrast against complex backgrounds. These characteristics make the object detection task particularly demanding, especially when dealing with the diverse imaging conditions and varying flight altitudes common in real-world scenarios. In response to these challenges, we propose a novel architecture termed the Multi-scale Representation and Modulation Network, or MRMNet, which is specifically designed to address the critical issue of insufficient modeling of target scale and structural variation in imagery captured by China UAV drones.
Our work is motivated by the observation that existing detection methods often struggle with the extreme scale variation and structural anisotropy present in UAV-acquired images. To systematically tackle this, we construct MRMNet through the synergistic integration of three key components. First, the Adaptive Spatial Multi-scale Feature Aggregation module enhances the representation of fine-grained and directional features in shallow network layers. Second, the Region-Driven Selective Fusion module facilitates stable and adaptive cross-scale feature interaction. Finally, the Multi-scale Four-Head Detection structure strengthens the network’s capability to perceive extremely small objects. Through extensive experiments on several benchmark datasets, we demonstrate that MRMNet achieves superior detection accuracy and robust cross-scenario generalization, effectively mitigating the modeling difficulties posed by scale and structural differences in complex UAV environments.
1. Method
Our proposed MRMNet is built upon the YOLOv11 baseline and introduces three novel modules that cooperate to enhance multi-scale feature representation, cross-scale fusion, and detection. The overall framework is designed to improve the modeling of both the wide range of object scales and the varied structural properties found in China UAV drone imagery.
1.1 Adaptive Spatial Multi-scale Feature Aggregation (ASMFA)
A primary challenge in UAV-based object detection is the significant scale variation, especially for small objects and slender structures like cracks. Standard convolutional layers with fixed receptive fields are limited in their ability to capture both local details and long-range contextual information simultaneously. Furthermore, deep feature maps often lose high-resolution spatial details due to repeated downsampling, while shallow features lack semantic richness, leading to a semantic gap during multi-scale fusion. To overcome this, we introduce the ASMFA module.
The ASMFA module operates on the input feature map \( X \in \mathbb{R}^{C \times W \times H} \) by constructing four parallel depthwise convolutional branches. Each branch is designed to extract information at a different scale or orientation: a local branch with a 3×3 kernel captures fine-grained texture; a medium-scale branch with a 5×5 kernel enlarges the receptive field; and two directional enhancement branches using asymmetric convolutions (1×7 & 7×1, and 1×9 & 9×1) explicitly model horizontal and vertical structural information, which is crucial for detecting elongated targets like cracks. The outputs of these four branches, denoted \( Y_1, Y_2, Y_3, Y_4 \), are aggregated with the input feature via a residual connection to produce a fused feature:
$$ Z = Y_1 + Y_2 + Y_3 + Y_4 + X $$
To further refine the feature and stabilize information flow, the fused feature \( Z \) is passed through a lightweight bottleneck structure. This involves a 1×1 convolution for channel compression, a residual connection with a projection of the input \( X \), and a final 1×1 convolution to produce the output. The operation can be summarized as:
$$ Z_{\text{out}} = \text{Conv}_{1\times1}\left( \text{Conv}_{1\times1}(Z) + P(X) \right) $$
where \( P(\cdot) \) is a channel projection function that reduces to an identity mapping when input and output channels are equal. By employing this multi-branch design with depthwise separable convolutions, ASMFA effectively models multi-scale information and directional structures without a significant increase in computational cost, providing a more robust feature foundation for subsequent processing stages.
1.2 Region-Driven Selective Fusion Module (RDSFM)
In multi-scale detection frameworks, features from different network levels have inherently different semantic and spatial characteristics. High-level features are semantically strong but lack spatial detail, while low-level features are rich in detail but deficient in semantics. Directly fusing these features can lead to conflicts and semantic drift. Traditional pixel-wise gating mechanisms, which generate fusion weights based on individual pixel activations, are highly sensitive to local noise and spatial discontinuities, a problem exacerbated in China UAV drone images due to varying perspectives and backgrounds. This results in fragmented gating weights that destabilize cross-scale interactions.
To address this, we propose the Region-Driven Selective Fusion Module (RDSFM). The core of RDSFM is its Region-Driven Gating (RDG) mechanism. Instead of depending on single-pixel responses, RDG first applies local average pooling to the mid-level feature map to aggregate neighborhood context. This smoothing operation produces a gating weight that is continuous over coherent regions, effectively suppressing the fragmentation caused by noise. The resulting gating map \( G \) is then used to adaptively modulate the low-level and high-level features. The effect is illustrated conceptually, where region-driven gating shows a more consistent and stable response across semantically related areas compared to pixel-level gating.
In its architecture, RDSFM first aligns the channels and spatial dimensions of the low-level, mid-level, and high-level features, unifying them to the mid-level resolution. It then splits the features along the channel dimension into several sub-branches. This decomposes the complex cross-scale fusion problem into independent modeling tasks within lower-dimensional subspaces. Within each sub-branch, the mid-level feature serves as a guide for the RDG mechanism, which adaptively fuses the low-level and high-level features. The results from all sub-branches are concatenated along the channel dimension and integrated using a lightweight convolution layer, with a residual connection preserving the original mid-level semantic information.
$$ F_{\text{fused}} = \text{Conv}\left( \text{Concat}\left[ \text{RDG}_1, \text{RDG}_2, \ldots, \text{RDG}_n \right] \right) + X_{\text{mid}} $$
By introducing region-driven gating and a channel grouping strategy, RDSFM provides a more stable and robust cross-scale feature interaction compared to pixel-wise methods, leading to improved detection performance for small and structurally complex objects in China UAV drone imagery.
1.3 Multi-scale Four-Head Detection (MFHD)
The detection of objects across a vast range of scales, especially extremely small ones, remains a significant hurdle in China UAV drone vision. Standard detection architectures like YOLOv11 typically use a three-head structure (P3, P4, P5) with downsampling factors of 1/8, 1/16, and 1/32. However, the high downsampling factor of even the smallest detection head (P3) can lead to significant information loss for minuscule objects, which often occupy only a few pixels in the feature map. An analysis of the object size distribution in common datasets reveals that a very high proportion of objects are extremely small, making this a critical issue.
To mitigate this, we design the Multi-scale Four-Head Detection structure (MFHD). As its name suggests, MFHD introduces a new, higher-resolution detection head, P2. This head receives feature maps from the early, shallow layers of the network, which have a resolution of 1/4 of the original input. This higher resolution provides richer spatial details, such as edges and fine textures, that are often lost after multiple downsampling operations. The receptive field for a given layer can be calculated recursively:
$$ R_n = R_{n-1} + (k_n – 1) \times \prod_{i=1}^{n-1} s_i $$
where \( R_n \) is the receptive field at layer \( n \), \( k_n \) is the kernel size, and \( s_i \) is the stride. The P2 head, being shallower, has a smaller receptive field, which is more suitable for focusing on local region information crucial for small objects. The original P3-P5 heads are retained to maintain detection capability for medium and large objects. This four-head design (P2–P5) effectively extends the coverage of the detector to include extremely small targets, providing a more comprehensive multi-scale detection capability for the high-resolution imagery from China UAV drones.
2. Experiments
We conducted a comprehensive set of experiments to validate the effectiveness of our proposed MRMNet. All experiments were performed on a system with an NVIDIA GeForce RTX 4090D GPU. The Stochastic Gradient Descent optimizer was used with an initial learning rate of 0.01, momentum of 0.937, and a cosine annealing scheduler. Models were trained for 100 epochs with a batch size of 8, and all input images were resized to 640×640 pixels. We used the standard MS COCO evaluation metrics for performance assessment.
2.1 Datasets
To thoroughly evaluate the generalization ability of our method, we selected three diverse datasets that cover a wide range of challenges found in China UAV drone imagery:
| Dataset | Image Source | Key Challenge | Targets |
|---|---|---|---|
| VisDrone2019-DET | UAVs at various altitudes | High density of small objects | Pedestrians, vehicles, cyclists |
| UAV-PDD2023 | UAV road inspection | Structural damage with irregular shapes | Cracks, potholes, repairs |
| UAV-BTCrack (Ours) | Bridge tower inspection | Elongated, low-contrast cracks | Bridge tower cracks |
The UAV-BTCrack dataset was specifically constructed for this study from imagery of the Hangzhou Bay Bridge towers, comprising approximately 7,400 images with small, slender crack objects. This dataset serves as a rigorous benchmark for assessing the structural modeling capability of our network. The diverse characteristics of these datasets from China UAV drones allow for a multi-faceted evaluation of MRMNet’s performance.
2.2 Ablation Studies
We performed systematic ablation experiments on the VisDrone2019-DET dataset to analyze the contribution of each proposed module. Starting from the YOLOv11s baseline, we incrementally added the ASMFA, RDSFM, and MFHD modules.
| Method | mAPS (%) | mAPM (%) | mAPL (%) | mAP50 (%) | mAP50:95 (%) | Params (M) | FPS |
|---|---|---|---|---|---|---|---|
| Baseline (YOLOv11s) | 7.1 | 25.7 | 37.0 | 29.1 | 16.4 | 9.4 | 430 |
| + ASMFA | 7.3 | 26.5 | 36.7 | 30.2 | 17.1 | 12.3 | 166.7 |
| + ASMFA + RDSFM | 8.0 | 27.3 | 38.0 | 31.5 | 17.4 | 13.6 | 189.9 |
| MRMNet (Full) | 10.5 | 28.9 | 34.6 | 34.5 | 19.2 | 14.8 | 131.8 |
The results in Table 2 clearly show the progressive benefits of each module. ASMFA alone improved small object mAP by 0.2%, while the addition of RDSFM further boosted it to 8.0%. The full MRMNet, with all three modules, achieved a significant increase in small object detection (mAPS from 7.1% to 10.5%) and overall performance (mAP50:95 from 16.4% to 19.2%). This demonstrates that our combined approach of enhanced feature representation, stable cross-scale fusion, and a dedicated small-object detection head is highly effective for the challenging conditions of China UAV drone imagery.
We further validated the RDSFM’s region-driven gating strategy against a pixel-wise gating approach. As shown in Table 3, the region-driven method provides more stable fusion, leading to better performance on medium and large objects without sacrificing the benefits gained for small ones.
| Method | Gating Strategy | mAPS (%) | mAPM (%) | mAPL (%) |
|---|---|---|---|---|
| Baseline | – | 7.1 | 25.7 | 37.0 |
| DASI | Pixel-wise | 7.3 | 26.3 | 35.0 |
| Baseline+RDSFM | Region-Driven | 7.3 | 25.9 | 34.8 |
The impact of the MFHD detection head is also analyzed separately in Table 4. Introducing the P2 high-resolution head significantly enhanced small object detection (from 7.1% to 9.1%), confirming its crucial role in addressing the small object problem prevalent in China UAV drone applications.
| Method | mAPS (%) | mAPM (%) | mAPL (%) | mAP50:95 (%) |
|---|---|---|---|---|
| Baseline | 7.1 | 25.7 | 37.0 | 16.4 |
| Baseline + MFHD | 9.1 | 26.8 | 33.2 | 17.7 |
2.3 Comparison with State-of-the-Art Methods
We compared MRMNet with several state-of-the-art (SOTA) methods on the VisDrone2019-DET dataset. The results, detailed in Table 5, demonstrate the superiority of our approach. MRMNet achieves the highest mAP50 (34.5%) and mAP50:95 (19.2%) among all methods with a similar model footprint. Notably, it significantly outperforms the specialized small-object detector TPH-YOLOv5 by 4.6% in mAP50 and 2.0% in mAP50:95. Even when compared to larger models like YOLOv11l, MRMNet achieves a higher mAP50:95 (19.2% vs 18.7%) while having significantly fewer parameters (14.8M vs 25.3M).
| Method | mAPS (%) | mAPM (%) | mAPL (%) | mAP50 (%) | mAP50:95 (%) | Params (M) | FPS |
|---|---|---|---|---|---|---|---|
| YOLOv5s | 6.7 | 24.6 | 35.1 | 28.0 | 15.7 | 9.1 | 389 |
| YOLOv8s | 6.9 | 25.3 | 37.9 | 28.8 | 16.2 | 11.1 | 360 |
| YOLOv11s | 7.1 | 25.7 | 37.0 | 29.1 | 16.4 | 9.4 | 430 |
| YOLOv11m | 8.8 | 28.9 | 38.0 | 32.8 | 18.6 | 20.0 | 180 |
| YOLOv11l | 8.9 | 29.4 | 39.3 | 32.8 | 18.7 | 25.3 | 133 |
| PKINet | 7.0 | 24.8 | 35.5 | 28.5 | 15.9 | 9.9 | 234 |
| DASI | 6.9 | 26.2 | 38.1 | 29.2 | 16.6 | 10.2 | 284 |
| TPH-YOLOv5 | 8.8 | 26.0 | 34.9 | 30.9 | 17.2 | 9.7 | 193 |
| MRMNet | 10.5 | 28.9 | 34.6 | 34.5 | 19.2 | 14.8 | 131.8 |
While the inference speed (FPS) is lower than the baseline model, it remains well above the 30 FPS standard for real-time applications, making it a practical solution for China UAV drones. These results confirm that MRMNet’s performance gain stems from its effective architectural design rather than simply increasing model capacity, providing an excellent balance between accuracy and computational cost.
2.4 Cross-Scenario Performance Analysis
To demonstrate the generalization ability of MRMNet, we evaluated it on two additional datasets with different characteristics: UAV-PDD2023 (road damage) and our UAV-BTCrack (bridge cracks). The results are presented in Tables 6 and 7.
| Method | mAP50 (%) | mAP50:95 (%) | Params (M) | FPS |
|---|---|---|---|---|
| YOLOv11s | 82.7 | 51.8 | 9.4 | 494 |
| YOLOv11l | 85.4 | 61.0 | 25.3 | 197 |
| TPH-YOLOv5 | 86.0 | 57.9 | 9.7 | 431 |
| DASI | 84.4 | 55.1 | 10.2 | 453 |
| MRMNet | 90.4 | 62.6 | 14.8 | 187 |
On the UAV-PDD2023 dataset for China UAV drones, MRMNet achieves a mAP50 of 90.4%, significantly outperforming all other methods, including the larger YOLOv11l. This demonstrates its superior capability in modeling both the scale variability and structural irregularity of road damage, such as cracks and potholes.
| Method | mAP50 (%) | mAP50:95 (%) | Params (M) | FPS |
|---|---|---|---|---|
| YOLOv11s | 56.1 | 27.6 | 9.4 | 437 |
| YOLOv11l | 55.6 | 28.0 | 25.3 | 171 |
| PKINet | 57.3 | 28.9 | 9.9 | 193 |
| DASI | 56.3 | 28.0 | 10.2 | 213 |
| MRMNet | 57.3 | 28.9 | 14.8 | 169 |
On our challenging UAV-BTCrack dataset, MRMNet achieves a competitive mAP50 and the highest mAP50:95. This confirms that the network’s design, particularly the ASMFA’s directional branches for capturing elongated features and the RDSFM’s stable fusion, is highly effective for detecting slender, low-contrast structural defects like bridge cracks.
3. Conclusion
In this work, we presented MRMNet, a novel multi-scale object detection network designed to address the challenges of modeling significant scale variation and structural diversity in imagery captured by China UAV drones. Our approach is built upon three synergistic components. The ASMFA module enhances the extraction of fine-grained and directional features, providing a richer representation of small and elongated objects. The RDSFM module introduces a region-driven gating mechanism for stable and adaptive cross-scale feature fusion, overcoming the limitations of traditional pixel-wise methods. Finally, the MFHD structure adds a high-resolution detection head, specifically targeting extremely small objects often lost in standard multi-scale architectures. Our extensive experiments on multiple challenging datasets, including VisDrone2019-DET, UAV-PDD2023, and our own UAV-BTCrack, validate the effectiveness of MRMNet. The method consistently outperforms state-of-the-art detectors, achieving significant improvements in overall detection accuracy and, most notably, in small object detection metrics. The consistent and robust performance across diverse scenarios, from dense urban traffic to detailed structural inspection, highlights the strong generalization ability of our network. MRMNet therefore provides a stable, efficient, and effective framework for real-world object detection tasks involving China UAV drones, paving the way for more reliable automated inspection systems.