In recent years, UAV drones have become increasingly important in public security, disaster monitoring, urban governance, traffic surveillance, and emergency response. Because UAV drones can rapidly collect large amounts of aerial imagery from flexible viewpoints, they provide valuable situational awareness in scenarios where ground-based systems are difficult to deploy. However, object detection from UAV drones remains challenging because aerial images often contain extremely small objects, dense distributions, complex backgrounds, and severe occlusion. Moreover, UAV drones are usually constrained by limited onboard computation, memory, and power, which makes it difficult to directly deploy large and computationally expensive deep models. In this study, I focus on improving detection accuracy for small objects in UAV drone imagery while also achieving a lightweight model suitable for edge deployment. My work combines multi-scale feature enhancement with knowledge distillation, resulting in a framework that balances accuracy and efficiency for UAV drone applications.
The core motivation of my research is that UAV drones operate in highly dynamic environments. The altitude, viewing angle, lighting conditions, and weather can change rapidly, causing large variations in object scale and appearance. Small objects such as pedestrians, bicycles, and motorcycles may occupy only a few pixels, and their features can easily be lost during repeated downsampling in convolutional neural networks. In addition, background textures from buildings, roads, vegetation, and shadows often produce strong noise that interferes with feature learning. These characteristics make UAV drone object detection fundamentally different from generic object detection on ground-level images. I therefore design a detection framework that explicitly enhances small-object representations and suppresses background interference, while also transferring knowledge from a high-performance teacher model to a lightweight student model.
Related Work and Technical Background
Before the widespread adoption of deep learning, traditional machine learning methods for UAV drone object detection relied heavily on handcrafted features. Techniques such as histogram of oriented gradients, scale-invariant feature transform, speeded-up robust features, and local binary patterns were combined with classifiers such as support vector machines and AdaBoost. These methods achieved moderate performance in early UAV drone applications, but they struggled with complex backgrounds, small object scales, and dense object distributions. The main limitation is that handcrafted features cannot effectively represent high-level semantic information, especially when objects are tiny and visually ambiguous. As a result, traditional machine learning approaches gradually gave way to deep learning methods that can automatically learn hierarchical feature representations.
With the rapid development of deep learning, convolutional neural networks have become the dominant approach for UAV drone object detection. Two-stage detectors such as Faster R-CNN and Cascade R-CNN first generate region proposals and then refine them for classification and localization. These methods often achieve high accuracy, but their computational cost is usually too high for real-time UAV drone deployment. One-stage detectors such as SSD and the YOLO family directly predict bounding boxes and class probabilities from feature maps, offering a better trade-off between speed and accuracy. In the context of UAV drones, YOLO-based methods have been widely studied because they can operate in real time on embedded platforms. However, standard YOLO models still face difficulties in detecting small objects under complex aerial backgrounds. My work builds on the YOLOv10 architecture and introduces targeted improvements for UAV drone imagery.
Knowledge distillation is another important technique for compressing deep models. The idea is to transfer knowledge from a large teacher model to a smaller student model. In object detection, distillation can be applied at the output level, feature level, or relation level. Output-level distillation uses soft labels from the teacher to guide the student. Feature-level distillation aligns intermediate representations between teacher and student. Relation-level distillation transfers structural relationships among samples or channels. Despite their success, traditional distillation methods often treat all spatial locations and channels equally, which is problematic for UAV drone imagery where foreground objects are small and backgrounds are noisy. I therefore propose a selective distillation framework that focuses on informative foreground regions and decouples classification and regression supervision.
Proposed Teacher Model: YOLOv10s-WRS
My first contribution is a high-performance teacher model called YOLOv10s-WRS, which is built upon the YOLOv10s baseline. The goal is to enhance multi-scale feature representation for small objects in UAV drone images. I introduce three key modules: C2f_WTConv, RFCBAMConv, and SOEP. These modules work together to improve frequency-domain modeling, adaptive receptive field selection, and small-object feature fusion. The overall architecture follows the standard backbone-neck-head design of YOLOv10, but the backbone and neck are modified to better capture fine-grained details.
Wavelet Convolution for Fine-Grained Feature Extraction
In UAV drone images, small objects are often characterized by high-frequency edge and texture information. Standard convolution tends to emphasize low-frequency semantic content and may lose fine details during downsampling. To address this issue, I replace the standard C2f module in the backbone with C2f_WTConv, which integrates wavelet convolution into the bottleneck. The wavelet convolution decomposes input features into frequency sub-bands using discrete wavelet transform. For an input feature map \(X\), the 2D discrete wavelet transform produces four components:
$$X_{LL}, X_{LH}, X_{HL}, X_{HH} = \text{DWT}(X)$$
Here, \(X_{LL}\) represents the low-frequency approximation component, while \(X_{LH}\), \(X_{HL}\), and \(X_{HH}\) represent high-frequency components in different directions. These sub-bands capture global semantics and directional edge information, respectively. The sub-band features are concatenated along the channel dimension:
$$X_{freq} = \text{Concat}(X_{LL}, X_{LH}, X_{HL}, X_{HH})$$
A depthwise convolution is then applied to model frequency-specific patterns, followed by a learnable scaling function:
$$X’_{freq} = S^{(l)}(\text{DWConv}(X_{freq}))$$
After frequency-domain enhancement, an inverse wavelet transform reconstructs the spatial feature map:
$$X_{wavelet} = \text{IWT}(X’_{freq})$$
The final output of WTConv combines spatial and frequency-domain branches:
$$F_{WTConv}(X) = F_{spatial}(X) + F_{wavelet}(X)$$
By integrating WTConv into C2f, the model preserves high-resolution details while maintaining a lightweight structure. This is particularly beneficial for UAV drone imagery, where small objects such as pedestrians and bicycles rely on subtle edge and texture cues.
Receptive Field and Attention Based Convolution
Complex backgrounds in UAV drone images often contain distracting textures that resemble targets. To improve robustness, I design RFCBAMConv, a convolution module that combines variable receptive fields with channel and spatial attention. The module first generates multi-receptive-field candidate features using depthwise convolution. Given an input feature map \(X \in \mathbb{R}^{C \times H \times W}\), a depthwise convolution with kernel size \(k \times k\) produces:
$$F_{depth} = \phi(\text{BN}(\text{ReLU}(\text{DWConv}_{k}(X))))$$
where \(\phi(\cdot)\) denotes the activation function. The feature is then rearranged to expose receptive field information along the spatial dimension. A squeeze-and-excitation (SE) block is applied to model channel dependencies:
$$s = \sigma(W_2 \delta(W_1 \text{GAP}(F_{depth})))$$
$$F_{se} = s \odot F_{depth}$$
Next, spatial attention is computed by aggregating max-pooled and average-pooled features:
$$A_r = \sigma(\text{Conv}_{3\times3}([\text{MaxPool}(F_{se}); \text{AvgPool}(F_{se})]))$$
The final output is obtained by applying a strided convolution to the attention-weighted features:
$$F_{out} = \text{Conv}_{k,s}(F_{se} \otimes A_r)$$
This design allows the network to adaptively focus on informative regions and channels, which is essential for suppressing background noise in UAV drone scenes.
Small-Object Enhanced Feature Pyramid: SOEP
Feature pyramids are widely used to detect objects at different scales. However, in UAV drone images, small objects are often lost after multiple downsampling operations. Adding an extra P2 detection layer can improve small-object detection, but it also increases computation and memory. To avoid this, I design SOEP, a feature pyramid structure that enhances high-resolution features without adding a new detection layer. SOEP uses SPDConv to reorganize spatial information into channel dimensions. Given the P2 feature \(F_{P2} \in \mathbb{R}^{H \times W \times C}\), the space-to-depth operation is defined as:
$$F_{P2}^{SPD} = \text{SPD}(F_{P2}, r)$$
where \(r\) is the rearrangement factor. A convolution and nonlinear activation are then applied:
$$F_{P2}’ = \phi(\text{Conv}(F_{P2}^{SPD}))$$
For multi-scale integration, I adopt a CSP-based design with OmniKernel. The input feature is split into two parts:
$$F_1, F_2 = \text{Split}(F)$$
One part is passed through a cross-stage connection, while the other is processed by OmniKernel, which combines global, large-scale, and local branches:
$$F_{omni} = \sum K_i(F_2)$$
The final output concatenates the cross-stage feature and the OmniKernel output:
$$F_{out} = \text{Concat}(F_1, F_{omni})$$
This structure enhances small-object features while keeping the computational overhead manageable. In my experiments, SOEP significantly improves recall for tiny objects such as bicycles and pedestrians.
Experiments for the Teacher Model
I evaluate the proposed teacher model on the VisDrone2019 dataset, which contains 8,629 aerial images captured by UAV drones across 14 cities. The dataset includes ten object categories: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor. The images are divided into 6,471 training images, 548 validation images, and 1,610 test images. I also use the UAVDT dataset for generalization experiments. The implementation is based on PyTorch 2.0.0 with CUDA 11.8 on an NVIDIA RTX 4090 GPU. The input resolution is 640×640, batch size is 8, and training runs for 200 epochs with SGD optimizer, initial learning rate 0.01, momentum 0.937, and weight decay 0.0005. The evaluation metrics include precision \(P\), recall \(R\), mean average precision \(mAP@0.5\), \(mAP@0.5:0.95\), parameters, GFLOPs, and FPS.
The precision and recall are computed as:
$$P = \frac{TP}{TP + FP}, \quad R = \frac{TP}{TP + FN}$$
Average precision is the area under the precision-recall curve:
$$AP = \int_0^1 P(R) \, dR$$
Mean average precision is the average over all categories:
$$mAP = \frac{1}{N} \sum_{i=1}^{N} AP_i$$
Frames per second is defined as:
$$FPS = \frac{1}{t_{avg}}$$
Table 1 compares the baseline YOLOv10s and my improved YOLOv10s-WRS. My model achieves 54.6% precision, 41.6% recall, 43.6% mAP@0.5, and 26.0% mAP@0.5:0.95. Compared with the baseline, this represents improvements of 6.0%, 3.0%, 4.3%, and 2.7%, respectively. The results demonstrate that the proposed modules effectively enhance detection accuracy for UAV drone imagery.
| Model | P (%) | R (%) | mAP@0.5 (%) | mAP@0.5:0.95 (%) |
|---|---|---|---|---|
| YOLOv10s | 48.6 | 38.6 | 39.3 | 23.3 |
| YOLOv10s-WRS | 54.6 | 41.6 | 43.6 | 26.0 |
Table 2 provides per-category results on VisDrone2019. The improvements are consistent across all categories. For large objects such as trucks and buses, the model benefits from better context modeling. For small objects such as bicycles and pedestrians, the gains are particularly notable. For example, bicycle mAP@0.5 increases from 13.2% to 17.3%, and pedestrian mAP@0.5 increases from 43.0% to 47.7%. These results validate the effectiveness of wavelet convolution for fine-grained feature extraction and SOEP for small-object preservation.
| Category | YOLOv10s mAP@0.5 (%) | YOLOv10s-WRS mAP@0.5 (%) |
|---|---|---|
| pedestrian | 43.0 | 47.7 |
| people | 34.7 | 37.9 |
| bicycle | 13.2 | 17.3 |
| car | 80.3 | 82.6 |
| van | 45.7 | 50.1 |
| truck | 34.2 | 41.2 |
| tricycle | 27.1 | 33.2 |
| awning-tricycle | 16.6 | 18.2 |
| bus | 53.1 | 58.8 |
| motor | 45.6 | 49.6 |
Table 3 compares my model with other mainstream detectors. Two-stage detectors such as Faster R-CNN and Cascade R-CNN have high parameter counts and computational costs, making them unsuitable for UAV drones. Lightweight one-stage detectors such as YOLOv5s and YOLOv7-tiny are faster but less accurate. My YOLOv10s-WRS achieves the best mAP@0.5 and mAP@0.5:0.95 among the compared models, while maintaining a moderate parameter size of 9.5M. This makes it an excellent teacher model for knowledge distillation.
| Model | mAP@0.5 (%) | mAP@0.5:0.95 (%) | Params (M) | GFLOPs |
|---|---|---|---|---|
| Faster R-CNN | 29.5 | 15.2 | 41.5 | 206 |
| Cascade R-CNN | 31.8 | 16.5 | 69.1 | 235 |
| YOLOv5s | 36.2 | 19.2 | 7.2 | 16.5 |
| YOLOv7-tiny | 35.8 | 18.9 | 6.2 | 13.8 |
| YOLOv8s | 40.5 | 24.1 | 11.2 | 28.6 |
| YOLOv10s | 39.3 | 23.3 | 7.2 | 21.6 |
| YOLOv10s-WRS | 43.6 | 26.0 | 9.5 | 35.8 |
Table 4 presents the ablation study for the teacher model. Starting from the YOLOv10s baseline, I progressively add C2f_WTConv, RFCBAMConv, and SOEP. Each module brings consistent improvements. The baseline achieves 39.3% mAP@0.5. Adding C2f_WTConv improves it to 41.8%. Adding RFCBAMConv further improves it to 42.9%. Finally, adding SOEP achieves 43.6%. The results confirm that the three modules are complementary and jointly enhance small-object detection in UAV drone imagery.
| Experiment | Baseline | C2f_WTConv | RFCBAMConv | SOEP | P (%) | R (%) | mAP@0.5 (%) |
|---|---|---|---|---|---|---|---|
| 1 | √ | 48.6 | 38.6 | 39.3 | |||
| 2 | √ | √ | 51.4 | 40.2 | 41.8 | ||
| 3 | √ | √ | √ | 53.2 | 41.1 | 42.9 | |
| 4 | √ | √ | √ | √ | 54.6 | 41.6 | 43.6 |
I also evaluate generalization on the UAVDT dataset. As shown in Table 5, my model achieves 42.6% mAP@0.5 and 25.5% mAP@0.5:0.95, outperforming the baseline by 4.0% and 2.3%, respectively. This indicates that the improvements are not dataset-specific but address common challenges in UAV drone object detection.
| Model | P (%) | R (%) | mAP@0.5 (%) | mAP@0.5:0.95 (%) |
|---|---|---|---|---|
| YOLOv10s | 49.4 | 37.7 | 38.6 | 23.2 |
| YOLOv10s-WRS | 51.4 | 41.3 | 42.6 | 25.5 |
Knowledge Distillation for Lightweight UAV Drone Detection
Although YOLOv10s-WRS achieves strong accuracy, its parameter size and computational cost may still be too high for resource-constrained UAV drones. To address this, I propose a knowledge distillation framework that transfers knowledge from the teacher YOLOv10s-WRS to a lightweight student YOLOv10n-WRS. The student model has only 2.8M parameters and 11.7 GFLOPs, and it can run at 126 FPS. The distillation framework consists of three components: cross-dimensional multi-scale semantic feature reconstruction, dynamic mask focused distillation, and adaptive decoupled distillation loss.
Cross-Dimensional Multi-Scale Semantic Feature Reconstruction
Teacher and student models often have different channel dimensions. Direct feature alignment can cause information loss or structural mismatch. To bridge this gap, I design a cross-dimensional feature reconstruction module. I select the neck feature pyramid as the distillation hub because it contains multi-scale semantic information. For each scale \(l\), the teacher feature is \(F_t^{(l)} \in \mathbb{R}^{H_l \times W_l \times C_t^{(l)}}\), and the student feature is \(F_s^{(l)} \in \mathbb{R}^{H_l \times W_l \times C_s^{(l)}}\). The spatial resolution is identical, but the channel dimensions differ. I introduce a \(1 \times 1\) convolution to project the student feature to the teacher dimension:
$$W^{(l)} \in \mathbb{R}^{C_t^{(l)} \times C_s^{(l)} \times 1 \times 1}$$
The projected student feature is:
$$\hat{F}_s^{(l)}(i,j) = \sum_{c=1}^{C_s^{(l)}} F_s^{(l)}(i,j,c) \cdot W^{(l)}(c) + b^{(l)}$$
Since the convolution kernel is \(1 \times 1\), the spatial resolution is preserved. The projected student feature \(\hat{F}_s^{(l)}\) has the same shape as \(F_t^{(l)}\), enabling point-wise supervision. This module allows the student to learn high-order semantic representations from the teacher without destroying spatial details, which is crucial for small objects in UAV drone images.
Dynamic Mask Focused Distillation
UAV drone images contain complex backgrounds that can dominate the distillation signal. Traditional mean squared error treats all spatial locations equally, forcing the student to fit irrelevant background information. To solve this, I propose a dynamic mask focused distillation mechanism. First, I compute a spatial response map from the teacher feature by averaging absolute activations across channels:
$$R(h,w) = \frac{1}{C} \sum_{c=1}^{C} |F_t(h,w,c)|$$
Then I normalize the response map to generate a soft mask:
$$M(h,w) = \frac{R(h,w) – \min R}{\max R – \min R + \epsilon}$$
The mask assigns high weights to foreground regions and low weights to background. The pixel-wise MSE between teacher and projected student features is:
$$L_{MSE}(h,w) = \frac{1}{C} \sum_{c=1}^{C} (F_t(h,w,c) – \hat{F}_s(h,w,c))^2$$
The masked distillation loss is:
$$L_{mask} = \frac{1}{H W} \sum_{h=1}^{H} \sum_{w=1}^{W} M(h,w) \cdot L_{MSE}(h,w)$$
This mechanism suppresses background gradients and encourages the student to focus on high-value foreground targets. It is especially effective for dense small objects in UAV drone scenes.
Adaptive Decoupled Distillation Loss
Detection heads usually have separate classification and regression branches. I design a decoupled distillation loss to provide task-specific supervision. For classification, I use temperature-scaled softmax to smooth the teacher and student logits:
$$p_t^{(i)} = \frac{\exp(Z_t^{(i)}/T)}{\sum_j \exp(Z_t^{(j)}/T)}, \quad p_s^{(i)} = \frac{\exp(Z_s^{(i)}/T)}{\sum_j \exp(Z_s^{(j)}/T)}$$
The classification distillation loss is based on Kullback-Leibler divergence:
$$L_{cls} = T^2 \cdot \frac{1}{N} \sum_{i=1}^{N} \sum_{c=1}^{C} p_t^{(i,c)} \log \frac{p_t^{(i,c)}}{p_s^{(i,c)}}$$
For bounding box regression, I use distribution focal loss representations. The teacher and student produce probability distributions \(\beta_t\) and \(\beta_s\) over discretized box coordinates. The regression distillation loss is:
$$L_{reg} = \frac{1}{N} \sum_{i=1}^{N} KL(\beta_t^{(i)} || \beta_s^{(i)})$$
The total distillation loss combines detection loss and distillation terms:
$$L_{total} = L_{det} + \lambda_{feat} L_{mask} + \lambda_{cls} L_{cls} + \lambda_{reg} L_{reg}$$
This decoupled design allows the student to learn both category relationships and localization uncertainty from the teacher, improving robustness in complex UAV drone environments.
Experiments for Knowledge Distillation
I conduct distillation experiments on VisDrone2019. The teacher is YOLOv10s-WRS, and the student is YOLOv10n-WRS. The teacher weights are frozen during training. The temperature is set to 3, and the distillation loss weight is 0.5. Table 6 shows the main results. The student baseline achieves 38.8% mAP@0.5 with 2.8M parameters, 11.7 GFLOPs, and 126 FPS. The teacher achieves 43.6% mAP@0.5 but has 9.5M parameters, 35.8 GFLOPs, and 91 FPS. After distillation, my method improves the student to 41.5% mAP@0.5 while keeping the same parameter count, computational cost, and inference speed. This is a 2.7% absolute improvement over the student baseline and narrows the gap to the teacher significantly.
| Model | mAP@0.5 (%) | Params (M) | GFLOPs | FPS |
|---|---|---|---|---|
| Student YOLOv10n-WRS | 38.8 | 2.8 | 11.7 | 126 |
| Teacher YOLOv10s-WRS | 43.6 | 9.5 | 35.8 | 91 |
| Distilled student (my method) | 41.5 | 2.8 | 11.7 | 126 |
Table 7 compares my distillation method with several state-of-the-art distillation algorithms. The classic KD method achieves 39.8% mAP@0.5. FitNet and AT achieve 40.1% and 40.6%, respectively. CWD, FGD, and MGD achieve 40.8%, 41.0%, and 41.2%, respectively. My method achieves the best result of 41.5%. This demonstrates that my selective distillation framework is more effective for UAV drone object detection, where background noise and small object scales require focused knowledge transfer.
| Teacher | Student | Distillation Method | mAP@0.5 (%) |
|---|---|---|---|
| YOLOv10s-WRS | YOLOv10n-WRS | KD | 39.8 |
| YOLOv10s-WRS | YOLOv10n-WRS | FitNet | 40.1 |
| YOLOv10s-WRS | YOLOv10n-WRS | FGD | 41.0 |
| YOLOv10s-WRS | YOLOv10n-WRS | CWD | 40.8 |
| YOLOv10s-WRS | YOLOv10n-WRS | MGD | 41.2 |
| YOLOv10s-WRS | YOLOv10n-WRS | AT | 40.6 |
| YOLOv10s-WRS | YOLOv10n-WRS | My method | 41.5 |
Table 8 presents the ablation study for the distillation framework. Starting from the student baseline at 38.8% mAP@0.5, adding feature reconstruction improves it to 40.1%. Adding dynamic mask focused distillation further improves it to 40.5%. Finally, adding the decoupled distillation loss achieves 41.5%. Each component contributes positively, and their combination yields the best performance. This confirms that feature-level alignment, spatial focus, and task-specific output supervision are all essential for effective knowledge transfer in UAV drone detection.
| Experiment | Feature Reconstruction | Mask Focused | Decoupled Loss | mAP@0.5 (%) |
|---|---|---|---|---|
| 1 | 38.8 | |||
| 2 | √ | 40.1 | ||
| 3 | √ | √ | 40.5 | |
| 4 | √ | √ | √ | 41.5 |
Visual analysis further confirms the effectiveness of my method. In dense parking lots and intersections, the baseline student often misses small vehicles and pedestrians. The teacher detects more objects but is too heavy for real-time deployment. After distillation, the student detects more small objects, produces tighter bounding boxes, and maintains high confidence in complex backgrounds. These improvements are consistent with the quantitative results and demonstrate that my distillation framework effectively transfers knowledge from the teacher to the lightweight student without increasing inference cost.
Conclusion and Future Work
In this study, I proposed a comprehensive framework for UAV drone object detection that combines multi-scale feature enhancement and knowledge distillation. First, I designed a high-performance teacher model, YOLOv10s-WRS, which integrates wavelet convolution, receptive field attention, and a small-object enhanced feature pyramid. This model achieves 43.6% mAP@0.5 on VisDrone2019, significantly outperforming the baseline and other mainstream detectors. Second, I developed a knowledge distillation method that transfers knowledge from the teacher to a lightweight student model. The method includes cross-dimensional feature reconstruction, dynamic mask focused distillation, and adaptive decoupled distillation loss. The distilled student achieves 41.5% mAP@0.5 while maintaining 2.8M parameters, 11.7 GFLOPs, and 126 FPS. This demonstrates a strong balance between accuracy and efficiency for UAV drone deployment.
However, several challenges remain. First, UAV drone imagery often suffers from domain shifts caused by weather, illumination, and geographic differences. Future work could integrate unsupervised, semi-supervised, and domain adaptation methods to reduce reliance on labeled data. Second, extreme lightweight deployment on edge devices remains difficult. Combining my distillation framework with pruning and quantization could further reduce model size and power consumption. Third, dense and overlapping objects in high-altitude views still pose challenges. Incorporating multimodal data fusion, such as visible-light and infrared imagery, could improve robustness in complex environments. I believe these directions will further advance the practical application of UAV drone object detection.