1. Introduction and Research Motivation
In recent years, the rapid advancement of unmanned aerial vehicle (UAV) technology has catalyzed its widespread adoption across numerous civilian and military domains, including intelligent transportation systems, ecological surveillance, disaster response, and precision agriculture. The ability to perform real-time, accurate object detection from UAV-captured imagery is of paramount importance for these applications. However, aerial images captured by UAVs present unique and formidable challenges that distinguish them from conventional ground-level imagery. Objects within these images are typically small in pixel footprint, often occupying fewer than 32×32 pixels, and are subject to significant scale variations due to fluctuating flight altitudes and camera angles. Furthermore, the presence of complex background clutter, varying illumination conditions, and atmospheric interference exacerbates the difficulty of reliable detection. Traditional detection algorithms, which rely heavily on manually crafted features such as Histogram of Oriented Gradients (HOG) or Scale-Invariant Feature Transform (SIFT), exhibit limited feature representation capacity and poor adaptability to such dynamic and complex scenes, resulting in suboptimal detection performance.
Parallel to the challenge of detection accuracy, the practical deployment of deep learning models on UAV platforms is constrained by severe limitations in onboard computing resources, memory capacity, and power consumption. High-performance detection models, while achieving impressive accuracy on powerful workstations, are often computationally prohibitive for real-time inference on embedded systems typically mounted on UAVs. Consequently, this creates a critical need for detection algorithms that not only excel in accuracy for small object detection but also demonstrate efficiency in model size and computational complexity, enabling seamless deployment on resource-constrained edge devices. To address these intertwined challenges, this thesis focuses on two principal research directions: enhancing the detection accuracy of small objects in UAV aerial imagery, and achieving model lightweighting to facilitate efficient deployment.
2. Related Work
The evolution of object detection has been primarily driven by the transition from traditional handcrafted feature-based methods to deep learning-based approaches. Traditional techniques, exemplified by the Viola-Jones detector and the Deformable Parts Model (DPM), established foundational principles for object localization and classification. However, their reliance on sliding windows and shallow features led to high computational costs and limited robustness in complex scenarios. The advent of deep learning revolutionized the field, giving rise to two dominant paradigms: two-stage detectors and one-stage detectors. Two-stage frameworks, such as the R-CNN family including Fast R-CNN and Faster R-CNN, adopt a sequential strategy of region proposal generation followed by classification and bounding box regression. These methods generally achieve superior accuracy but at the expense of speed. Conversely, one-stage detectors like the YOLO (You Only Look Once) series and SSD (Single Shot MultiBox Detector) perform classification and localization in a single forward pass, offering significant advantages in inference speed, making them more suitable for real-time applications. The YOLO series, in particular, has undergone continuous refinement, with each iteration introducing architectural innovations to improve the trade-off between speed and accuracy. YOLOv11, the latest iteration, integrates advanced modules such as C3k2 and C2PSA in its backbone to enhance feature extraction, and employs a more efficient detection head, establishing a strong baseline for further improvements.
In the specific context of UAV aerial imagery, research efforts have concentrated on overcoming the challenges of small object detection and model deployment. Strategies commonly explored include the enhancement of feature extraction through novel convolutional modules, the construction of multi-scale feature fusion networks to better aggregate contextual information, and the incorporation of attention mechanisms to focus on salient regions. For instance, some studies have introduced deformable convolutions to better adapt to geometric variations, while others have designed specialized feature pyramid networks to preserve fine-grained details from shallow layers. With respect to model lightweighting, techniques such as network pruning, knowledge distillation, and efficient architecture design are frequently employed. Structured pruning methods aim to remove redundant channels or layers, while knowledge distillation transfers knowledge from a cumbersome teacher model to a compact student model, mitigating the performance degradation often associated with model compression.
3. Proposed EGD-YOLO for Enhanced Small Object Detection
Our initial research objective was to address the deficiencies in detection accuracy for small objects within UAV aerial images. To this end, we proposed a novel detection model, termed EGD-YOLO, which introduces systematic improvements to the baseline YOLOv11 architecture across three key components: the backbone network, the neck network, and the detection head. A comprehensive overview of the EGD-YOLO architecture is presented conceptually, where the three enhancement modules work in a synergistic manner to improve feature extraction, fusion, and prediction.
3.1 Dual-Conv Efficient Layer Aggregation Network (ELAN-DC)
The backbone network is fundamental to extracting rich and discriminative features. In the standard YOLOv11, the C3k2 module is a core building block. We identified that while effective, its feature extraction capability could be further enhanced, particularly for small objects that contain limited pixel information. To address this, we proposed the Dual-Conv Efficient Layer Aggregation Network (ELAN-DC). This module strategically replaces the standard bottleneck convolutions within the C3k2 architecture with a Dual-Conv (DC) unit.
The Dual-Conv unit is designed based on two complementary convolutional operations: a 3×3 convolution and a 1×1 convolution. The 3×3 convolution is adept at capturing fine-grained spatial details and local context, which is crucial for delineating the boundaries and textures of small objects. Conversely, the 1×1 convolution focuses on reducing channel dimensionality and facilitating efficient cross-channel information exchange. This combination allows the module to extract rich spatial features while maintaining a relatively low computational cost. The structural layout of the Dual-Conv unit involves a parallel or sequential arrangement of these kernels, followed by a concatenation operation to fuse the multi-scale context. By integrating this unit into the ELAN module’s design, which itself promotes gradient flow and feature reuse through a partial-to-whole aggregation strategy, the ELAN-DC module effectively enhances the backbone’s feature extraction prowess. The architecture of the ELAN-DC module involves splitting the input feature map along the channel dimension into two branches. One branch is passed through a series of Dual-Conv blocks for deep feature transformation, while the other branch is passed through directly to preserve the original spatial information. Finally, these two branches are concatenated, fusing the shallow details with the deep semantic features.
3.2 Global-to-Local Bidirectional Feature Pyramid Network (GLBiFPN)
Multi-scale feature fusion is of utmost importance in object detection, especially when dealing with the significant scale variations observed in UAV imagery. The standard neck network in YOLOv11, based on a PAN-FPN structure, facilitates feature integration but may not be optimal. We therefore constructed a Global-to-Local Bidirectional Feature Pyramid Network (GLBiFPN). This architecture enhances the BiFPN by incorporating the Global-to-Local Spatial Attention (GLSA) module, leading to more effective multi-scale feature aggregation.
The GLSA module is a dual-branch attention mechanism designed to comprehensively model feature dependencies. It separately processes global and local contextual information to generate enhanced feature representations. In the global branch, a global spatial attention mechanism is employed to capture long-range dependencies and model the overall context. The attention computation can be formalized as follows:
$$Att_{G}(F^{1}_{i}) = Softmax(Transpose(C_{1\times1}(F^{1}_{i})))$$
$$G_{sa}(F^{1}_{i}) = MLP(Att_{G}(F^{1}_{i}) \otimes F^{1}_{i}) + F^{1}_{i}$$
where $C_{1\times1}$ denotes a 1×1 convolution, $\otimes$ represents matrix multiplication, and $MLP$ is a multi-layer perceptron. The local branch, on the other hand, focuses on extracting fine-grained local spatial features to capture detailed textures and edges. Its computation is given by:
$$Att_{L}(F^{2}_{i}) = \sigma(C_{1\times1}(F_{c}(F^{2}_{i}) + F^{2}_{i}))$$
$$L_{sa}(F^{2}_{i}) = Att_{L}(F^{2}_{i}) \cdot F^{2}_{i} + F^{2}_{i}$$
where $F_{c}$ is a cascaded structure of convolutional layers, and $\sigma$ is the Sigmoid function. By fusing the outputs of these two branches, the GLSA module effectively encodes both global semantic context and local spatial detail, significantly enhancing the discriminative power of the features. The GLBiFPN structure integrates these modules into the bidirectional cross-scale connections, enabling a richer and more adaptive feature fusion process.
3.3 Dynamic Head (DyHead)
The detection head is responsible for the final classification and localization predictions. The standard head in YOLOv11, while efficient with the use of depthwise convolutions in the classification branch, may not be sufficiently adaptive to the large variations in object scale and appearance. To this end, we adopted the Dynamic Head (DyHead) framework, which unifies object detection heads with attention mechanisms.
DyHead applies three types of attention sequentially to the feature tensor $F \in \mathbb{R}^{L \times S \times C}$, corresponding to the number of scales ($L$), spatial locations ($S$), and channels ($C$). The core computation is:
$$W(F) = \pi_{C}(\pi_{S}(\pi_{L}(F) \cdot F) \cdot F) \cdot F$$
Here, $\pi_{L}$, $\pi_{S}$, and $\pi_{C}$ represent the scale-aware, spatial-aware, and task-aware attention functions, respectively. The scale-aware attention dynamically aggregates features from different scales, ensuring robust handling of scale variation. The spatial-aware attention focuses on the most informative spatial locations, emphasizing key regions. The task-aware attention modulates the features based on the specific requirements of the classification and regression tasks. By jointly leveraging these attention mechanisms, DyHead significantly enhances the model’s capacity to detect small objects in complex environments, moving beyond the limitations of fixed-parameter convolution heads.
3.4 Experimental Validation of EGD-YOLO
We conducted extensive experiments to validate the effectiveness of the proposed EGD-YOLO model, primarily utilizing the VisDrone2019 dataset, a large-scale benchmark for UAV object detection. This dataset presents numerous challenges, including small object scales, dense distributions, and complex backgrounds. We also used the DOTAv1.0 dataset for validation. The experimental environment was configured with an NVIDIA RTX 4060-TI 16GB GPU, and the model was trained for 300 epochs on VisDrone2019 and 200 epochs on DOTAv1.0.
To systematically evaluate the contribution of each proposed module, we conducted a series of ablation studies. The results on the VisDrone2019 dataset are summarized in the table below, where ‘A’ represents ELAN-DC, ‘B’ represents GLBiFPN, and ‘C’ represents DyHead.
| Model | P (%) | R (%) | mAP@50 (%) | mAP@50-95 (%) | Param (M) | GFLOPs |
|---|---|---|---|---|---|---|
| Baseline (YOLOv11n) | 43.8 | 33.3 | 33.3 | 19.4 | 2.58 | 6.3 |
| Baseline + A | 43.5 | 34.1 | 33.8 | 19.6 | 2.43 | 6.9 |
| Baseline + B | 44.2 | 34.0 | 34.0 | 20.2 | 2.07 | 6.7 |
| Baseline + C | 46.7 | 35.1 | 35.7 | 21.1 | 3.10 | 7.5 |
| Baseline + A + B | 46.1 | 35.6 | 36.2 | 21.3 | 2.46 | 9.7 |
| Baseline + A + C | 47.8 | 35.8 | 36.7 | 21.7 | 2.95 | 8.0 |
| Baseline + B + C | 47.1 | 36.0 | 36.0 | 21.4 | 2.70 | 8.0 |
| Baseline + A + B + C (EGD-YOLO) | 47.7 | 37.7 | 38.4 | 22.9 | 3.01 | 10.5 |
The ablation study indicates that each module contributes positively to the overall performance. The integration of all three modules in EGD-YOLO resulted in a substantial improvement in mAP@50 and mAP@50-95 by 5.1% and 3.5%, respectively, compared to the baseline YOLOv11n. This significant gain underscores the effectiveness of the proposed enhancements in feature extraction, fusion, and prediction. The PR curves and confusion matrix comparisons further confirmed the improved precision and recall characteristics of the EGD-YOLO model, particularly in distinguishing small and easily confusable categories like pedestrians and people. We further compared EGD-YOLO with several mainstream object detectors, and the results are presented in the following table.
| Model | Params (M) | GFLOPs | P (%) | R (%) | mAP@50 (%) | mAP@50-95 (%) |
|---|---|---|---|---|---|---|
| Faster R-CNN | 41.0 | 207.0 | 45.5 | 32.2 | 33.1 | 16.0 |
| YOLOv5s | 9.1 | 24.1 | 50.0 | 38.5 | 33.9 | 23.2 |
| YOLOv8n | 3.0 | 8.1 | 45.4 | 32.6 | 33.1 | 19.4 |
| YOLOv10s | 8.04 | 24.5 | 50.5 | 38.3 | 39.1 | 23.5 |
| YOLOv11n | 2.58 | 6.3 | 43.8 | 33.3 | 33.3 | 19.4 |
| YOLOv11s | 9.41 | 21.3 | 50.5 | 38.7 | 39.3 | 23.6 |
| EGD-YOLO | 3.01 | 10.5 | 47.7 | 37.7 | 38.4 | 22.9 |
The comparison demonstrates that EGD-YOLO achieves a superior balance between detection accuracy and model complexity. While its mAP is slightly lower than YOLOv11s and YOLOv10s, it operates with a significantly lower parameter count and computational cost (roughly one-third and one-half of YOLOv11s, respectively). This makes EGD-YOLO a highly efficient and accurate model for UAV-based detection. To assess its generalization capability, we evaluated the model on DOTAv1.5 and COCO datasets. The results show consistent improvements over the baseline YOLOv11n, with mAP@50 increasing from 36.2% to 37.7% on DOTAv1.5 and from 49.0% to 50.3% on COCO, confirming the model’s robustness and cross-domain applicability.
4. Proposed EGD-YOLO-Lite for Lightweight Deployment
While EGD-YOLO successfully improved detection accuracy, its increased computational overhead poses challenges for deployment on resource-constrained UAV platforms. To address this, we proposed a lightweight variant, termed EGD-YOLO-Lite, which integrates efficient network design, structured pruning, and knowledge distillation to achieve an optimal balance between performance and efficiency.
4.1 Lightweight Architecture Design
The core of EGD-YOLO-Lite’s efficiency comes from a two-pronged approach to architecture modification. First, we replaced the heavy GLBiFPN in the neck with a more computationally economical Bidirectional Feature Pyramid Network (BiFPN). BiFPN simplifies the feature fusion paths by removing nodes with only one input edge and adding an extra edge from the original input to the output node, enabling more efficient cross-scale feature mixing with fewer parameters. Second, we integrated the efficient ELAN-DC backbone and DyHead from EGD-YOLO. The overall structure of EGD-YOLO-Lite is designed for efficient feature processing while minimizing redundancy.
4.2 LAMP-based Structured Pruning
To further compress the model, we employed the Layer-Adaptive Magnitude-based Pruning (LAMP) algorithm for structured pruning. LAMP leverages a global perspective to assign pruning ratios across layers based on a score that reflects the importance of each weight within its own layer. The LAMP score for the $u$-th element of a flattened weight vector $W$ is defined as:
$$Score(u; W) = \frac{(W[u])^2}{\sum_{v \geq u} (W[v])^2}$$
where the weights are sorted by magnitude in ascending order. This score assigns higher importance to weights that contribute more significantly to the output norm in the context of the remaining larger weights. The pruning process involves computing these scores, sorting them, and then removing the weights with the lowest global scores until the target sparsity is achieved. The LAMP pruning flow consists of four main steps: training the model, computing the LAMP scores, pruning the network, and fine-tuning the pruned model to recover performance. We conducted experiments to select the appropriate pruning strategy and ratio. Local pruning, which prunes each layer independently based on its own weight distribution, was found to be more effective than global pruning, as it preserved more of the model’s performance. Through a comparative analysis under different speedup ratios (defined as the ratio of pre-pruning to post-pruning GFLOPs), we selected a speedup of 1.5× as the optimal operating point, achieving a significant reduction in parameters and computation with only a minor impact on accuracy.
4.3 Channel-Wise Distillation for Performance Recovery
Structured pruning inevitably leads to a drop in detection accuracy. To mitigate this issue, we introduced a Channel-Wise Distillation (CWD) method. Knowledge distillation facilitates the transfer of knowledge from a complex, high-performance teacher model to a simpler student model. CWD is a feature-based distillation method that operates by aligning the channel-wise probability distributions of the teacher’s and student’s intermediate feature maps. In this method, the spatial features of each channel are first normalized using a softmax function to convert them into a probability distribution, which highlights the important spatial regions within each channel. Then, a Kullback-Leibler (KL) divergence loss is calculated to minimize the difference between these distributions. The distillation loss can be expressed as:
$$\varphi(y)_c = \frac{e^{y_{c,i}/\tau}}{\sum_{i=1}^{W \cdot H} e^{y_{c,i}/\tau}}$$
$$KL(\varphi(y^T), \varphi(y^S)) = \frac{1}{C} \sum_{c=1}^{C} \sum_{i=1}^{W \cdot H} \varphi(y^T_{c,i}) \log \frac{\varphi(y^T_{c,i})}{\varphi(y^S_{c,i})}$$
where $y^T$ and $y^S$ are the feature maps of the teacher and student models, respectively, $\tau$ is a temperature parameter, and $C$, $W$, $H$ are the dimensions of the feature map. This approach encourages the student to mimic the teacher’s focus on significant spatial regions, leading to enhanced feature representation and improved detection performance. For our experiments, the student model was the pruned 1.5× EGD-YOLO-Lite, and the teacher model was the unpruned EGD-YOLO-Lite (s version), which achieved the highest mAP@50 and served as an effective guide for the student.
4.4 Experimental Validation of EGD-YOLO-Lite
We conducted a comprehensive set of experiments to validate the effectiveness of the EGD-YOLO-Lite. The ablation study on the VisDrone2019 dataset is presented in the table below.
| Model | mAP@50 (%) | Weight (MB) | Param (M) | GFLOPs |
|---|---|---|---|---|
| EGD-YOLO (Baseline) | 38.4 | 6.1 | 3.01 | 10.5 |
| + BiFPN | 38.2 | 5.8 | 2.86 | 10.0 |
| + LAMP Pruning | 37.4 | 4.0 | 1.94 | 6.6 |
| + CWD Distillation | 37.6 | 4.0 | 1.94 | 6.6 |
The ablation study demonstrates the efficacy of the lightweighting pipeline. Replacing GLBiFPN with BiFPN slightly reduced mAP@50 by 0.2% but decreased parameters by 0.15M. The subsequent LAMP pruning significantly reduced the model’s weight, parameters, and computational cost to 4.0MB, 1.94M, and 6.6 GFLOPs, respectively. However, this compression resulted in a 0.8% drop in mAP@50. The introduction of CWD distillation successfully recovered 0.2% of this lost accuracy, achieving a final mAP@50 of 37.6% while maintaining the reduced model complexity.
We further compared the pruned and distilled model against the original EGD-YOLO and other state-of-the-art lightweight models.
| Model | Params (M) | GFLOPs | mAP@50 (%) | mAP@50-95 (%) |
|---|---|---|---|---|
| YOLOv5s | 9.1 | 24.1 | 33.9 | 23.2 |
| YOLOv7tiny | 6.01 | 13.1 | 32.7 | 16.0 |
| YOLOv8n | 3.0 | 8.1 | 33.1 | 19.4 |
| YOLOv11n | 2.58 | 6.3 | 33.3 | 19.4 |
| EGD-YOLO | 3.01 | 10.5 | 38.4 | 22.9 |
| EGD-YOLO-Lite | 1.94 | 6.6 | 37.6 | 22.1 |
Compared to the original EGD-YOLO, EGD-YOLO-Lite achieves a 36% reduction in parameters and a 37% reduction in GFLOPs, with only a 0.8% decrease in mAP@50. The model also significantly outperforms other lightweight models like YOLOv8n and YOLOv11n in terms of detection accuracy, while maintaining a comparable or even lower computational cost. To further validate our design choices, we conducted comparative experiments on pruning methods and distillation techniques, as shown in the tables below.
| Pruning Method | P (%) | R (%) | mAP@50 (%) | Weight (MB) | GFLOPs |
|---|---|---|---|---|---|
| Unpruned | 50.6 | 36.7 | 38.2 | 5.8 | 10.0 |
| Group-Taylor | 39.2 | 25.6 | 25.8 | 3.1 | 5.0 |
| LAMP | 44.4 | 32.7 | 33.1 | 3.2 | 5.0 |
The comparison clearly indicates the superiority of LAMP over a conventional method like Group-Taylor. Under a similar computational constraint (5.0 GFLOPs), LAMP achieves a considerably higher mAP@50 (33.1% vs 25.8%), demonstrating its effectiveness in identifying and preserving important weights, leading to a better compression-performance trade-off.
| Distillation Method | Weight (MB) | mAP@50 (%) | Param (M) | GFLOPs |
|---|---|---|---|---|
| Student Model | 4.0 | 37.4 | 1.94 | 6.6 |
| Logit_L1 | 4.0 | 36.9 | 1.94 | 6.6 |
| Logit_L2 | 4.0 | 36.8 | 1.94 | 6.6 |
| CWD | 4.0 | 37.6 | 1.94 | 6.6 |
| Logit_L1 + CWD | 4.0 | 37.1 | 1.94 | 6.6 |
The results of the distillation comparison show that our chosen CWD method is more effective than logit-based distillation methods. Logit-based methods (L1, L2) even caused a slight degradation in performance, while CWD successfully improved the student’s mAP@50 to 37.6%. This confirms that for dense prediction tasks like object detection, feature-based knowledge transfer provides a richer supervisory signal. The visualization of feature maps and detection results on test images further confirmed the superior performance of EGD-YOLO-Lite over the original YOLOv11, showcasing its ability to focus on relevant small targets and suppress background clutter.
5. Conclusion and Future Work
This thesis has systematically addressed the challenge of balancing detection accuracy and computational efficiency for object detection in unmanned aerial vehicle (UAV) aerial imagery. We successfully proposed and validated two complementary models: EGD-YOLO and EGD-YOLO-Lite.
The EGD-YOLO model was designed to push the boundaries of detection accuracy. Through the synergistic integration of the ELAN-DC module for enhanced feature extraction, the GLBiFPN for improved multi-scale feature fusion, and the DyHead for adaptive prediction, the model demonstrated significant improvements in mAP@50 and mAP@50-95 on the challenging VisDrone2019 dataset, outperforming the baseline YOLOv11n by 5.1% and 3.5%, respectively. This established a strong, accurate baseline for UAV-based detection.
Recognizing the real-world deployment constraints on UAV platforms, we extended our work to develop the EGD-YOLO-Lite model. This model adopts a holistic approach to lightweighting. An efficient BiFPN structure, the LAMP structured pruning algorithm, and a CWD feature distillation method were integrated into a cohesive optimization pipeline. The final model achieved a Pareto improvement, reducing the parameter count and computational cost by approximately 36% and 37% compared to EGD-YOLO, while incurring only a minor 0.8% drop in mAP@50. This resulted in a compact model with a weight of only 4.0MB and 1.94M parameters, making it exceptionally well-suited for deployment on resource-constrained edge devices commonly found on UAVs.
Despite the promising results, there are several avenues for future work. First, the object categories in existing datasets are unevenly distributed, which could be addressed in future research by expanding the dataset with more multi-scene and multi-scale samples, especially for very small objects and under-represented classes. Second, in scenarios with extremely dense small objects and strong background noise, the model still exhibits some false positives and missed detections. Future work could explore higher-resolution feature learning and more advanced context modeling to further improve robustness. Third, while we validated our models on public datasets, actual deployment on real UAV hardware has not yet been performed. A critical next step is to conduct on-device experiments on platforms like the Jetson series, exploring techniques such as TensorRT acceleration and INT8 quantization to further optimize inference latency and throughput, thereby fully realizing the practical potential of the proposed EGD-YOLO-Lite in real-world unmanned aerial vehicle applications.