Unmanned aerial vehicle (UAV) imagery has become an indispensable data source in modern remote sensing and intelligent surveillance. The rapid growth of low-altitude economy and the continuous advancement of UAV technology have enabled the acquisition of high-resolution aerial images from urban streets, agricultural fields, disaster zones, and natural environments. However, the unique characteristics of UAV-based imaging introduce severe challenges for object detection, particularly for small objects. These challenges include extreme scale variation, high density of tiny objects, complex background clutter, severe occlusion, and class imbalance. Traditional object detectors designed for ground-level images often fail to maintain satisfactory performance when applied directly to aerial scenes. In this paper, I present two complementary detection frameworks built upon YOLOv11 to address the challenges of small object detection in unmanned aerial vehicle imagery: a high-accuracy algorithm named PE-YOLO and a lightweight algorithm named JDP-YOLO. The proposed methods incorporate novel attention modules, feature fusion strategies, dynamic detection heads, and loss functions to improve both accuracy and efficiency. Extensive experiments on the VisDrone2019 and DOTAv1.0 datasets demonstrate the effectiveness of the proposed approaches.
The first approach, PE-YOLO, focuses on enhancing the feature extraction and fusion capabilities of the baseline network. A Multi-Dimensional Pixel Enhancement (MDPE) attention module is designed to allocate computational resources adaptively via dynamic channel segmentation and a lightweight smoothing network. A Dynamic Adaptive Feature Integration (DAFI) module is constructed to reduce feature loss during downsampling and achieve multi-scale feature fusion. A Feature Alignment Dynamic Head (FADH) embeds a block-based dynamic alignment convolution to improve the alignment and fusion of feature maps. Furthermore, an ARSIoU loss function is introduced to optimize bounding box regression, especially for small objects with irregular shape aspect ratios.
The second approach, JDP-YOLO, targets the deployment constraints of UAV onboard equipment by jointly optimizing network architecture and pruning. A Joint Shuffle Module (JSM) replaces the partial C3k2 modules in the backbone to break the feature isolation caused by grouped convolutions while reducing parameters and computation. A Dynamic Gated Convolution (DGConv) enhances the neck feature fusion with a dual-branch structure and SE-gated dynamic weighting, which can be re-parameterized into a single convolution during inference. A Bilevel Interactive Pruning Module (BIPM) performs end-to-end structured pruning with a temperature-annealed Sigmoid and a FLOPS adversarial mechanism, significantly compressing the model without sacrificing detection accuracy.
The remainder of this paper is organized as follows. Section 2 reviews the related work on object detection and small object detection in UAV imagery. Section 3 presents the PE-YOLO algorithm in detail. Section 4 describes the JDP-YOLO lightweight algorithm. Section 5 reports the experimental setup, results, and analyses. Section 6 concludes the paper.
1. Related Work
Object detection has evolved from traditional hand-crafted feature methods to deep learning based paradigms. Early methods such as HOG and SIFT combined with SVM classifiers achieved impressive results in specific scenarios, but their performance degraded significantly in complex aerial scenes. The emergence of deep convolutional neural networks (CNNs) brought a revolutionary change. Region-based methods like R-CNN, Fast R-CNN, and Faster R-CNN introduced two-stage detection architectures that achieve high accuracy but suffer from low inference speed. Single-stage detectors such as SSD and the YOLO series offer real-time processing while maintaining competitive accuracy. Recent transformer-based detectors including DETR and its variants have also shown strong performance, but they often require large amounts of training data and computational resources.
The specific problem of small object detection in UAV imagery has attracted considerable attention. Small objects are usually defined as those with pixel dimensions below 32×32 or requiring a high IoU threshold for proper evaluation. In aerial images, the proportion of small objects is extremely high, often exceeding 60% in datasets such as VisDrone2019. The limited spatial resolution and lack of contextual information make feature extraction for these objects particularly challenging. Some works focus on data augmentation strategies, such as slicing-aided hyper inference (SAHI), to improve recall. Others propose specialized feature pyramid networks or attention mechanisms to enhance multi-scale representation. Lightweight network designs have also been explored to meet the real-time requirements of UAV embedded systems. However, achieving a good balance between accuracy and efficiency remains an open problem.
2. Baseline: YOLOv11
YOLOv11 is one of the most recent real-time object detectors from the Ultralytics family. It features a refined CSPGhost backbone, an improved C3k2 module, and a C2PSA attention module. The network structure is composed of three main parts: backbone, neck, and head. The backbone uses C3k2 modules to extract hierarchical features with reduced parameters. The neck integrates SPPF and C2PSA to fuse multi-scale features, while the head adopts an anchor-free detection strategy. YOLOv11 achieves promising results on COCO benchmark, but its performance on UAV aerial small objects is limited due to insufficient fine-grained feature extraction and ineffective feature fusion for extreme scale variations.
In this work, I select YOLOv11n as the baseline for both the high-accuracy and lightweight variants. The baseline achieves a mAP@0.5 of 33.9% on the VisDrone2019 validation set, with 2.59 million parameters and 6.4 GFLOPs. Although this is already a strong result, there remains room for improvement in terms of detecting very small and densely packed objects.
3. PE-YOLO: High-Accuracy Detection for UAV Imagery
PE-YOLO is designed to enhance the detection ability of YOLOv11 for small objects in UAV images. The overall architecture is shown conceptually in the previous sections. It consists of four major improvements: the MDPE attention module, the DAFI feature fusion module, the FADH detection head, and the ARSIoU loss function. Each module contributes to improving the feature extraction, fusion, and regression accuracy for small objects.
3.1 Multi-Dimensional Pixel Enhancement Attention Module (MDPE)
Standard channel attention mechanisms typically compute global channel statistics and assign scalar weights to each channel, ignoring pixel-level variations within a channel. Spatial attention focuses on regional information, but cannot precisely locate the tiny features of small objects. Pixel attention (PA) directly models the importance of each pixel, which is beneficial for capturing fine details, but it often lacks global context and is sensitive to background noise.
The Multi-Dimensional Pixel Enhancement (MDPE) attention module is proposed to overcome these limitations. First, it uses a channel attention mechanism to compute a channel weight vector for the input feature tensor. A dynamic threshold is then computed based on the mean and standard deviation of the channel weights, and the input channels are split into a high-importance branch and a low-importance branch. The threshold is defined as:
$$ \tau = \mu + \alpha \cdot \sigma $$
where $\mu$ is the mean channel weight, $\sigma$ is the standard deviation, and $\alpha$ is a hyperparameter controlling the dynamic range. The high-importance branch is processed by a self-attention module to enhance edge and texture details, while the low-importance branch is processed by a spatial attention module to suppress redundant background information. Finally, a pixel-channel synergistic enhancement mechanism is applied by concatenating the original feature and the fused feature, feeding them into a grouped convolution with a 7×7 kernel, and generating an attention weight map via a Sigmoid function. The output is computed as:
$$ F’ = F_w \odot F_i $$
where $F_w$ is the attention weight map and $F_i$ is the original input feature. The MDPE module enables the network to focus on the most informative pixel-level details while maintaining robustness to complex backgrounds.
3.2 Dynamic Adaptive Feature Integration Module (DAFI)
In the neck network of YOLOv11, feature fusion is typically performed by simple concatenation or element-wise addition. These operations neglect the semantic complementary information between features of different scales, causing small object features to be suppressed by strong features from larger objects. To address this problem, the DAFI module is introduced. It consists of a Feature Enhancement Module (FEM) and a dual dynamic weighting mechanism.
The FEM combines global context and local spatial information by using two parallel paths: a max-pooling path for downsampling and a bicubic interpolation path for upsampling. The outputs are concatenated and passed through a depthwise separable convolution to generate a feature map that captures both fine local details and global contextual information. Each path is enhanced by an Enhancement Unit (EU), which uses two depthwise convolutions with ReLU and Sigmoid activations to produce attention weights.
The dual dynamic weighting mechanism operates on two input features, a low-level feature $F_1$ and a high-level feature $F_2$. For the low-level feature, a Sigmoid-generated weight matrix $A$ focuses on edge and detail regions, while for the high-level feature, a spatial attention mechanism generates a complementary weight matrix $B$ that emphasizes the target body. The weighted fusion is expressed as:
$$ F’_1 = A \otimes F_1 + (1 – A) \otimes F_2 $$
$$ F’_2 = B \otimes F_1 + (1 – B) \otimes F_2 $$
where $\otimes$ denotes element-wise multiplication. The fused features are then passed through a 1×1 convolution, batch normalization, and SiLU activation to produce the final output. This adaptive fusion allows the network to dynamically adjust the blending strength according to the importance of each feature map, effectively reducing the missed detection of small objects.
3.3 Feature Alignment Dynamic Head (FADH)
The original YOLOv11 decoupled head processes classification and regression independently, which limits the interaction between semantic and localization information. Moreover, the standard convolution in the regression branch is not suitable for objects with deformed or rotated shapes. To overcome these issues, a new detection head called FADH is proposed. It first uses a task interaction module that extracts a global semantic vector via global average pooling and generates channel attention weights. The classification branch then applies spatial attention to enhance discriminative spatial regions. The regression branch integrates a Block-based Dynamic Alignment Convolution (BDAC) module.
BDAC divides the input feature map into non-overlapping 2×2 blocks and predicts a unified primary offset for each block, thereby significantly reducing the number of parameters compared to pixel-wise dynamic convolution. To further refine the alignment, a residual offset is predicted via a pointwise convolution and combined with the primary offset:
$$ O_{final} = O_{main} + \beta \cdot O_{res} $$
where $\beta$ is a learnable scalar parameter. The final offsets are then used to sample the feature via bilinear interpolation, and the sampled features are weighted fused. This approach makes the head more adaptive to scale and shape variations, improving the robustness of detection for small objects.
3.4 ARSIoU Loss Function
The CIoU loss used in YOLOv11 incorporates overlap area, center-point distance, and aspect ratio consistency. However, when the IoU is low, the aspect ratio penalty is weak, and small objects contribute less to the total loss. To enhance the regression performance for small objects, the ARSIoU loss is designed based on Inner IoU and an aspect ratio difference term. Let $AR_{pred}$ and $AR_{gt}$ denote the aspect ratios of the predicted and ground-truth boxes, respectively. The aspect ratio difference is defined as:
$$ \Delta AR = \frac{|AR_{pred} – AR_{gt}|}{AR_{gt} + \epsilon} $$
where $\epsilon$ is a small constant for numerical stability. The ARSIoU metric is then expressed as:
$$ ARS_{IoU} = IoU_{inner} – \lambda \cdot \Delta AR $$
where $\lambda$ controls the penalty strength. To emphasize small objects, a scale-dependent weight $\alpha$ is introduced:
$$ \alpha = \begin{cases} 0.7 & \text{if } S < T \\ 0.3 & \text{otherwise} \end{cases} $$
where $S = w \cdot h$ is the area of the bounding box and $T$ is an area threshold. Finally, the loss function is formulated as:
$$ L_{ARSIoU} = \alpha \cdot (1 – ARS_{IoU}) $$
The ARSIoU loss improves the sensitivity of the model to small objects and enhances the robustness of bounding box regression under viewpoint variations.
3.5 Implementation Details for PE-YOLO
The PE-YOLO model is trained on VisDrone2019 for 300 epochs with an input size of 640×640. The batch size is 12, and the SGD optimizer is used with an initial learning rate of 0.01 and a final learning rate of 0.001. The momentum is 0.937 and the weight decay is 0.0005. Ablation experiments are conducted by incrementally adding the proposed modules to the baseline YOLOv11n. The results are presented in the following tables.
4. JDP-YOLO: Lightweight Detection with Joint Dynamic Pruning
Although PE-YOLO significantly boosts detection accuracy, its model size and computational cost remain relatively high for resource-constrained UAV platforms. To address this issue, a lightweight detection algorithm named JDP-YOLO is proposed. It integrates three major components: the Joint Shuffle Module (JSM), the Dynamic Gated Convolution (DGConv), and the Bilevel Interactive Pruning Module (BIPM). These components work together to reduce the number of parameters and FLOPs while preserving detection accuracy.
4.1 Joint Shuffle Module (JSM)
The Joint Shuffle Module is inspired by ShuffleNet V2 and aims to reduce computational cost while maintaining cross-channel and cross-position information flow. In the backbone network, some C3k2 modules are replaced by JSM. For the case where no downsampling is required, the input tensor is divided into two branches along the channel dimension. One branch is passed through an identity mapping, while the other branch is processed by a refinement path that sequentially performs pointwise convolution, depthwise convolution, and pointwise convolution. The two branches are then concatenated. When downsampling is needed, the module uses two parallel paths: one path uses a stride-2 depthwise convolution, and the other uses the refinement path with stride-2 depthwise convolution to simultaneously reduce spatial resolution. After concatenation, a spatial-channel joint shuffle operation is applied to exchange information between different groups.
The spatial-channel joint shuffle can be mathematically described as follows. Given an input tensor $X \in \mathbb{R}^{B \times C \times H \times W}$, it is first split into $G_c$ channel groups and $G_s^2$ spatial grids, resulting in a high-dimensional tensor. By performing a sequence of dimension permutes, the channel-group dimension and spatial-block dimensions are interleaved, and then the spatial grid indices are swapped. Finally, the tensor is reshaped back to $B \times C \times H \times W$. This operation effectively mixes information from different groups without additional parameters, improving global context awareness.
4.2 Dynamic Gated Convolution (DGConv)
The original GSConv module is a lightweight convolution that reduces parameters and computational cost but suffers from a lack of spatial context in its fast branch. The proposed Dynamic Gated Convolution (DGConv) improves the fast branch by adding a depthwise convolution and an SE gate. The architecture consists of two branches: a fast branch and a cheap branch. In the fast branch, a 1×1 pointwise convolution compresses the channels by half and optionally performs downsampling. The resulting feature is then passed through a 3×3 depthwise convolution to capture local spatial context. A global average pooling is applied to aggregate spatial information, followed by two pointwise convolutions and a Sigmoid to produce channel weights $\alpha$. The original compressed feature and the spatial-enhanced feature are then combined element-wise:
$$ y_{fast} = (1 – \alpha) \odot y_1 + \alpha \odot y_2 $$
where $y_1$ is the initial compressed feature and $y_2$ is the spatial-enhanced feature. The cheap branch retains the original GSConv design, using a pointwise convolution and a depthwise convolution to produce a second half-channel feature. The outputs of the two branches are concatenated and then passed through a channel shuffle operation with 8 groups to mix information between the branches. During inference, the weighted fusion in the fast branch can be re-parameterized into a single depthwise convolution, eliminating extra computational overhead.
4.3 Bilevel Interactive Pruning Module (BIPM)
To further compress the model, a Bilevel Interactive Pruning Module (BIPM) is designed. It is embedded into the convolutional layers in a controlled manner to perform end-to-end structured pruning. The module starts with a spatial average pooling operation to compress the feature map into a channel descriptor. A bottleneck fully connected network with a reduction ratio $r$ is then used to enable information interaction among channels. A temperature-annealed Sigmoid is applied as a gating function:
$$ s = \sigma\left(\frac{x}{T}\right) $$
where $T$ is the temperature coefficient that linearly decays from 5 to 1 during training. This annealing schedule gradually sharpens the gate from a continuous state to a binary state, allowing the network to first learn with all channels and then focus on important channels. During training, a classification loss and a FLOPS penalty are jointly optimized. The FLOPS penalty is defined as:
$$ L_{FLOPs} = \lambda \cdot \mathbb{E}[s] $$
where $\lambda$ controls the compression strength. The gradient of this penalty with respect to $s$ is a positive constant, which pushes the gate values downward, while the classification gradient pushes upward. This adversarial mechanism ensures that only channels with a contribution larger than their computational cost are retained. After training, the model is evaluated on a calibration set to accumulate global channel significance. For each layer, the absolute weights of the gating layer are averaged over all calibration samples, and a top-k selection is performed according to the pruning ratio. The weights and biases of the corresponding output channels are then sliced to produce a structurally pruned model.
4.4 Integration of JDP-YOLO
The JDP-YOLO model is constructed by applying JSM, DGConv, and BIPM to the baseline YOLOv11n. Specifically, JSM replaces several C3k2 modules in the backbone to reduce parameters and enhance feature communication. DGConv replaces the standard convolutions in the neck to improve feature fusion with dynamic gating. BIPM is applied to the C3k2 layer in the middle of the backbone, where the parameter redundancy is high and the pruning risk is relatively low. Different pruning ratios are tested to evaluate the trade-off between accuracy and efficiency.
5. Experiments and Results
5.1 Datasets and Evaluation Metrics
I use the VisDrone2019 and DOTAv1.0 datasets to evaluate the proposed algorithms. VisDrone2019 is a large-scale UAV imagery dataset containing 10,209 static images with 2.6 million bounding-box annotations in 10 categories. The images are captured from altitudes ranging from 20 to 120 meters, covering various urban and rural scenes. The dataset is challenging because it contains a high proportion of small objects, severe density, and many truncated or obscured instances. DOTAv1.0 is an aerial image detection benchmark with 2,806 high-resolution images and 188,282 oriented bounding boxes in 15 categories. It includes diverse scenes such as harbors, airports, and residential areas, with large scale variations and arbitrary orientations.
For evaluation, I report precision (P), recall (R), mAP@0.5, and mAP@0.5:0.95. The model complexity is measured by the number of parameters (in millions) and the number of GFLOPs. The inference speed is reported as frames per second (FPS).
5.2 Experimental Setup
All experiments are conducted on an NVIDIA GeForce RTX 4060 Laptop GPU with 16 GB memory, under the Windows operating system. The deep learning framework is PyTorch 2.5.1 with Python 3.12.4 and CUDA 12.4. The baseline model is YOLOv11n. For a fair comparison, the same training hyperparameters are used for both the baseline and the improved models. The input image size is set to 640×640, the batch size is 12, and the training is performed for 300 epochs. The SGD optimizer is used with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005.
5.3 Ablation Study of PE-YOLO
To verify the contribution of each module, I conduct ablation experiments on the VisDrone2019 dataset. The results are shown in Table 1.
| MDPE | DAFI | FADH | ARSIoU | P/% | R/% | mAP@0.5/% | Params/M | GFLOPs |
|---|---|---|---|---|---|---|---|---|
| × | × | × | × | 46.2 | 34.4 | 33.9 | 2.62 | 6.4 |
| √ | × | × | × | 47.1 | 35.0 | 34.5 | 2.68 | 6.5 |
| √ | √ | × | × | 48.6 | 35.7 | 35.9 | 2.82 | 7.1 |
| √ | √ | √ | × | 49.3 | 36.2 | 36.3 | 2.86 | 7.2 |
| √ | √ | √ | √ | 49.7 | 36.3 | 37.1 | 2.86 | 7.2 |
From Table 1, the baseline YOLOv11 achieves a mAP@0.5 of 33.9%. Adding the MDPE module increases the mAP to 34.5%, while improving both precision and recall. The DAFI module further improves the mAP to 35.9%, indicating that the adaptive feature fusion effectively reduces the loss of fine-grained information. The FADH detection head contributes a further 0.4% increase in mAP, reaching 36.3%, with only a small increase in parameters. Finally, the ARSIoU loss function improves the mAP to 37.1%, achieving a total gain of 3.2 percentage points over the baseline. The precision and recall are also improved by 3.5 and 1.9 percentage points, respectively, demonstrating the effectiveness of the proposed modules for small object detection in UAV imagery.
5.4 Comparison with State-of-the-Art Methods for PE-YOLO
A comprehensive comparison between PE-YOLO and other mainstream detectors is presented in Table 2. The comparison includes two-stage detectors, single-stage detectors, and recent UAV-specific improvements.
| Model | mAP@0.5/% | P/% | R/% | Params/M | FPS |
|---|---|---|---|---|---|
| Faster R-CNN | 33.8 | 44.5 | 37.5 | 43.75 | 141.3 |
| SSD | 23.3 | 37.0 | 26.9 | 24.15 | 171.2 |
| YOLOv3n | 38.5 | 49.8 | 39.7 | 61.4 | 158.1 |
| YOLOv5n | 32.0 | 38.0 | 33.5 | 7.05 | 158.5 |
| HIC-YOLOv5n | 33.6 | 39.4 | 34.1 | 7.33 | 152.4 |
| YOLOv8n | 33.0 | 42.0 | 32.5 | 3.05 | 177.4 |
| UAV-YOLOv8n | 34.2 | 43.1 | 33.0 | 2.65 | 185.7 |
| ARF-YOLOv8n | 35.7 | 45.6 | 34.2 | 3.12 | 175.6 |
| PC-YOLOn | 34.3 | 46.9 | 35.1 | 2.48 | 189.4 |
| DMF-YOLOv11n | 36.7 | 48.2 | 36.2 | 3.18 | 172.5 |
| YOLOv12n | 33.2 | 45.3 | 33.8 | 2.45 | 195.4 |
| YOLOv11n | 33.9 | 46.2 | 34.4 | 2.57 | 191.2 |
| PE-YOLO | 37.1 | 49.7 | 36.3 | 2.86 | 180.8 |
The results show that PE-YOLO outperforms all the compared methods in terms of mAP@0.5. Compared to YOLOv3n, which yields a high mAP of 38.5%, PE-YOLO is slightly lower but is significantly lighter (2.86M vs. 61.4M parameters). This suggests that PE-YOLO achieves a better balance between accuracy and model complexity. Compared to the baseline YOLOv11n, PE-YOLO improves the mAP by 3.2 percentage points with only a modest increase in parameters and a small reduction in FPS. The improvements are particularly notable in categories with many small objects, such as pedestrian and vehicle, owing to the pixel-level enhancement and dynamic feature fusion.
5.5 Generalization Experiments for PE-YOLO
To assess the generalization ability, I also evaluate PE-YOLO on the DOTAv1.0 dataset. The results are reported in Table 3.
| Model | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% |
|---|---|---|---|---|
| YOLOv3 | 51.2 | 25.5 | 32.4 | 18.9 |
| YOLOv5 | 53.6 | 30.2 | 33.2 | 19.3 |
| YOLOv8 | 55.3 | 31.4 | 34.8 | 20.5 |
| YOLOv11 | 58.4 | 33.5 | 36.3 | 21.5 |
| YOLOv12 | 57.8 | 31.6 | 36.1 | 21.1 |
| PE-YOLO | 59.5 | 34.7 | 39.2 | 23.7 |
PE-YOLO achieves the best performance across all metrics. Compared with YOLOv11, PE-YOLO improves mAP@0.5 by 2.9 percentage points and mAP@0.5:0.95 by 2.2 percentage points. This indicates that the proposed modules are not overfitting to the VisDrone2019 dataset, but rather offer general improvements for complex aerial imagery. In particular, categories with small instances such as vehicles and ships benefit significantly, as shown in Table 4.
| Class | YOLOv11 P/% | PE-YOLO P/% | YOLOv11 R/% | PE-YOLO R/% |
|---|---|---|---|---|
| plane | 59.1 | 63.2 | 60.6 | 60.8 |
| baseball-diamond | 53.5 | 54.6 | 37.0 | 38.2 |
| bridge | 46.8 | 47.5 | 39.8 | 33.3 |
| ground-track-field | 20.5 | 23.1 | 11.7 | 12.9 |
| small-vehicle | 49.3 | 59.1 | 39.6 | 43.8 |
| large-vehicle | 38.2 | 40.5 | 40.1 | 42.6 |
| ship | 43.3 | 47.3 | 50.3 | 54.0 |
| tennis-court | 47.1 | 47.9 | 47.2 | 51.1 |
| basketball-court | 24.6 | 26.1 | 23.3 | 23.3 |
| storage-tank | 46.8 | 54.1 | 32.9 | 37.6 |
| soccer-ball-field | 26.4 | 25.8 | 48.7 | 41.3 |
| roundabout | 35.6 | 36.9 | 24.1 | 24.3 |
| harbor | 33.5 | 34.1 | 22.3 | 26.8 |
| swimming-pool | 41.2 | 39.4 | 26.1 | 29.4 |
| helicopter | 54.5 | 56.2 | 36.9 | 38.2 |
Notably, for the small-vehicle class, PE-YOLO improves the precision from 49.3% to 59.1% and the recall from 39.6% to 43.8%, which confirms that the MDPE attention and DAFI fusion effectively preserve subtle features that are critical for detecting tiny objects. The ship class also shows a recall improvement from 50.3% to 54.0%, demonstrating the model’s ability to suppress background clutter in marine scenes.
5.6 Pruning Experiments for JDP-YOLO
For the lightweight approach, I first evaluate the BIPM pruning module on the intermediate C3k2 layer. Different pruning ratios are tested, and the results are shown in Table 5.
| Pruning Ratio | mAP@0.5/% | mAP@0.5:0.95/% | F1/% | Pruned Module Params | Total Params/M | GFLOPs |
|---|---|---|---|---|---|---|
| 0% | 34.1 | 20.0 | 37.0 | 272.71K | 2.202 | 5.9 |
| 25% | 34.1 | 20.0 | 37.1 | 254.10K | 2.183 | 5.9 |
| 50% | 34.1 | 20.0 | 37.3 | 235.49K | 2.165 | 5.8 |
| 75% | 34.0 | 19.9 | 36.9 | 216.88K | 2.146 | 5.8 |
At a pruning ratio of 50%, the mAP@0.5 remains exactly at 34.1%, while the module parameters are reduced by 13.65%. Even at an aggressive pruning ratio of 75%, the accuracy drop is only 0.1 percentage points. This demonstrates that the BIPM module can effectively identify and remove redundant channels without degrading performance, thanks to the temperature annealing and bidirectional adversarial mechanisms.
5.7 Ablation Study of JDP-YOLO
The contributions of JSM, DGConv, and BIPM are analyzed incrementally. The results are presented in Table 6.
| JSM | DGConv | BIPM | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | Params/M | GFLOPs |
|---|---|---|---|---|---|---|---|---|
| × | × | × | 46.2 | 34.4 | 33.9 | 19.8 | 2.59 | 6.4 |
| √ | × | × | 45.8 | 34.2 | 33.6 | 19.6 | 2.46 | 6.1 |
| × | √ | × | 46.0 | 34.7 | 34.0 | 19.9 | 2.40 | 6.2 |
| √ | √ | × | 45.7 | 34.6 | 33.8 | 19.7 | 2.28 | 5.9 |
| √ | √ | √ | 46.1 | 35.4 | 34.1 | 20.0 | 2.20 | 5.9 |
Replacing some C3k2 modules with JSM alone reduces the parameters from 2.59M to 2.46M and the GFLOPs from 6.4 to 6.1, while reducing mAP@0.5 by 0.3 percentage points. Adding DGConv improves the mAP to 34.0% with fewer parameters than JSM alone, showing that the dynamic gating mechanism recovers the accuracy loss through enhanced feature fusion. Combining JSM and DGConv yields a parameter count of 2.28M and a mAP of 33.8%, which is only 0.1% lower than the baseline. Finally, adding BIPM not only reduces the parameters further to 2.20M but also improves the mAP to 34.1% and the recall to 35.4%. The final JDP-YOLO model achieves a 15.1% reduction in parameters and a 7.8% reduction in GFLOPs compared with the baseline, while improving the mAP by 0.2 percentage points. This clearly demonstrates the effectiveness of the lightweight design.
5.8 Comparison of JDP-YOLO with Other Methods
The comparison of JDP-YOLO with state-of-the-art detectors is shown in Table 7.
| Model | mAP@0.5/% | P/% | R/% | Params/M | FPS |
|---|---|---|---|---|---|
| Faster R-CNN | 33.8 | 44.5 | 37.5 | 43.75 | 141.3 |
| SSD | 23.3 | 37.0 | 26.9 | 24.15 | 171.2 |
| YOLOv3n | 38.5 | 49.8 | 39.7 | 61.4 | 158.1 |
| YOLOv5n | 32.0 | 38.0 | 33.5 | 7.05 | 158.5 |
| HIC-YOLOv5n | 33.6 | 39.4 | 34.1 | 7.33 | 152.4 |
| YOLOv8n | 33.0 | 42.0 | 32.5 | 3.05 | 177.4 |
| UAV-YOLOv8n | 34.2 | 43.1 | 33.0 | 2.65 | 185.7 |
| ARF-YOLOv8n | 35.7 | 45.6 | 34.2 | 3.12 | 175.6 |
| PC-YOLOn | 34.3 | 46.9 | 35.1 | 2.48 | 189.4 |
| DMF-YOLOv11n | 36.7 | 48.2 | 36.2 | 3.18 | 172.5 |
| YOLOv12n | 33.2 | 45.3 | 33.8 | 2.45 | 195.4 |
| YOLOv11n | 33.9 | 46.2 | 34.4 | 2.59 | 191.2 |
| JDP-YOLO | 34.1 | 46.1 | 35.4 | 2.20 | 203.5 |
JDP-YOLO achieves a mAP of 34.1%, slightly outperforming the baseline YOLOv11n, while using only 84.9% of its parameters. The inference speed reaches 203.5 FPS, which is the fastest among all compared methods. Compared to other lightweight UAV-specific methods such as UAV-YOLOv8n and PC-YOLOn, JDP-YOLO offers a favorable trade-off: it has a higher recall than UAV-YOLOv8n (35.4% vs. 33.0%) and a comparable mAP to PC-YOLOn while being lighter. This makes JDP-YOLO highly suitable for real-time applications on embedded UAV platforms.
5.9 Generalization of JDP-YOLO on DOTAv1.0
To further verify the generalization of the lightweight model, I evaluate JDP-YOLO on DOTAv1.0. The results are shown in Table 8.
| Model | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% |
|---|---|---|---|---|
| YOLOv3 | 51.2 | 25.5 | 32.4 | 18.9 |
| YOLOv5 | 53.6 | 30.2 | 33.2 | 19.3 |
| YOLOv8 | 55.3 | 31.4 | 34.8 | 20.5 |
| YOLOv11 | 58.4 | 33.5 | 36.3 | 21.5 |
| YOLOv12 | 57.8 | 31.6 | 36.1 | 21.1 |
| JDP-YOLO | 58.9 | 33.4 | 36.7 | 21.3 |
JDP-YOLO improves the mAP@0.5 from 36.3% to 36.7% compared with YOLOv11, while maintaining a similar mAP@0.5:0.95. The per-class results in Table 9 indicate that the lightweight model remains robust across most categories.
| Class | YOLOv11 P/% | JDP-YOLO P/% | YOLOv11 R/% | JDP-YOLO R/% |
|---|---|---|---|---|
| plane | 59.1 | 59.2 | 60.6 | 59.9 |
| baseball-diamond | 53.5 | 57.3 | 37.0 | 37.1 |
| bridge | 46.8 | 50.1 | 39.8 | 37.1 |
| ground-track-field | 20.5 | 19.6 | 11.7 | 11.3 |
| small-vehicle | 49.3 | 50.5 | 39.6 | 37.3 |
| large-vehicle | 38.2 | 38.2 | 40.1 | 38.5 |
| ship | 43.3 | 45.3 | 50.3 | 49.2 |
| tennis-court | 47.1 | 46.9 | 47.2 | 49.9 |
| basketball-court | 24.6 | 24.7 | 23.3 | 23.8 |
| storage-tank | 46.8 | 45.5 | 32.9 | 32.2 |
| soccer-ball-field | 26.4 | 27.2 | 48.7 | 49.9 |
| roundabout | 35.6 | 35.1 | 24.1 | 23.0 |
| harbor | 33.5 | 35.6 | 22.3 | 23.2 |
| swimming-pool | 41.2 | 41.5 | 26.1 | 25.8 |
| helicopter | 54.5 | 56.8 | 36.9 | 35.3 |
The small-vehicle precision improves from 49.3% to 50.5%, while the recall remains essentially constant. This suggests that the lightweight architecture with channel shuffling and dynamic gating retains most of the critical features needed for small object detection, whereas pruning removes only redundant information. Overall, JDP-YOLO demonstrates satisfactory generalization ability across different aerial scenarios.
5.10 Visualization and Qualitative Analysis
To gain more insight into the behavior of the proposed models, I compare the heatmaps and detection results of YOLOv11, PE-YOLO, and JDP-YOLO on representative UAV images. In complex scenes with dense vehicles, YOLOv11 tends to focus on large foreground objects and often misses distant small vehicles. PE-YOLO generates more concentrated heat responses on small targets, thanks to the pixel-level enhancement and adaptive feature fusion. JDP-YOLO, despite being lighter, also produces satisfactory heatmap activations, confirming that the pruning process does not harm the feature representational power.
In low-light conditions, YOLOv11 struggles to distinguish objects from the dark background. PE-YOLO is able to detect more objects because the block-based dynamic alignment convolution in the FADH head better adapts to the feature shifts caused by blur and noise. JDP-YOLO also performs reasonably well in low-light settings, benefiting from the dynamic gating in DGConv, which adaptively enhances the spatial context.
Quantitatively, the Precision-Recall curves and the Precision-Confidence curves further confirm the improvements. The P-R curve of PE-YOLO consistently lies above that of YOLOv11, especially in the high-recall region, indicating better trade-offs. The P-C curve of JDP-YOLO reaches a precision of 100% at a confidence of 0.943, while YOLOv11 requires a lower threshold of 0.924, indicating lower false-alarm rates.
6. Conclusion
In this paper, I have presented two deep learning based detection algorithms specifically tailored for unmanned aerial vehicle imagery: PE-YOLO and JDP-YOLO. PE-YOLO improves the detection accuracy of small objects through a combination of pixel-level attention, adaptive feature fusion, a dynamic alignment head, and an aspect-ratio-aware loss function. Extensive experiments on VisDrone2019 and DOTAv1.0 show that PE-YOLO achieves a mAP@0.5 of 37.1% on VisDrone2019, outperforming the baseline YOLOv11 by 3.2 percentage points while maintaining a moderate model complexity.
JDP-YOLO is designed for lightweight deployment. By introducing a joint shuffle module, a dynamic gated convolution, and a bilevel interactive pruning module, the model reduces the number of parameters by 15.1% and the computational cost by 7.8% compared with YOLOv11, while slightly improving the mAP. The final model runs at 203.5 FPS on an RTX 4060 laptop GPU, making it suitable for real-time UAV applications. Both algorithms have been validated on an additional remote sensing dataset, demonstrating good generalization.
Future research will focus on extending the proposed methods to oriented bounding box detection, exploring multi-modal fusion (e.g., visible and infrared images), and adapting the models to edge devices with more extreme resource constraints. These directions will further improve the reliability and applicability of object detection in unmanned aerial vehicle scenarios.