In recent years, the widespread adoption of drone technology has brought numerous conveniences and transformations to various sectors, including civilian and military applications. However, this proliferation has also led to significant safety concerns, such as “black flights” and “indiscriminate flights” of drones, which pose threats to public security, critical infrastructure, and national defense. Traditional methods for anti-UAV detection often struggle with challenges like small target sizes, complex backgrounds, and low-resolution imagery, resulting in high rates of false positives and missed detections. To address these issues, I propose an integrated approach that combines super-resolution reconstruction with an optimized YOLOv5s model for enhanced anti-UAV detection. This method leverages advanced image enhancement techniques to improve target visibility and employs a refined detection network to achieve higher accuracy and robustness in real-time scenarios. Throughout this article, I will detail the methodology, improvements, experimental validation, and implications for anti-UAV systems, emphasizing the keyword “anti-UAV” to underscore the focus on counter-drone technologies.
The core of my approach lies in enhancing low-resolution drone imagery through super-resolution reconstruction before performing target detection. Low-resolution images, often captured from distant or mobile sensors, limit the ability of detection algorithms to extract sufficient features for small UAVs. By applying a state-of-the-art super-resolution model, such as the Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN), I can upscale these images to higher resolutions, thereby revealing finer details and improving the discriminative power of subsequent detection stages. This preprocessing step is crucial for anti-UAV applications where targets may be minuscule against cluttered skies or urban environments. The integration of super-resolution not only boosts detection performance but also addresses common pitfalls in anti-UAV systems, such as occlusions and motion blur. In the following sections, I will elaborate on the technical aspects of the Real-ESRGAN framework and its adaptation for drone imagery, followed by the modifications made to the YOLOv5s architecture to optimize it for anti-UAV tasks.
The Real-ESRGAN model is chosen for its ability to handle complex degradation processes commonly encountered in real-world anti-UAV scenarios, such as noise, compression artifacts, and blur. It employs a high-order degradation model and Sinc filters to generate realistic training data, enabling the generator to learn robust mappings from low-resolution to high-resolution images. The generator network, based on an improved Super-Resolution Residual Network (SRResNet), utilizes Residual-in-Residual Dense Blocks (RRDB) to enhance feature extraction and stability. By removing Batch Normalization (BN) layers, the network reduces computational overhead and improves performance on diverse datasets. The training involves a adversarial process where the generator strives to produce photorealistic images, while the discriminator evaluates their authenticity. Over iterations, this results in superior image quality, which is vital for detecting small UAVs that might otherwise be missed. For anti-UAV purposes, I fine-tune Real-ESRGAN on a composite dataset comprising drone imagery from various sources, ensuring it generalizes well to challenging conditions like low light or fast motion. The output high-resolution images provide a richer feature set for the detection pipeline, directly contributing to the efficacy of anti-UAV measures.
Following super-resolution reconstruction, the enhanced images are fed into an optimized YOLOv5s model for UAV detection. YOLOv5s is selected due to its balance of speed and accuracy, making it suitable for real-time anti-UAV applications. The standard YOLOv5s architecture consists of four main components: Input, Backbone, Neck, and Detection head. The Input module handles adaptive image scaling, Mosaic data augmentation, and anchor box optimization. The Backbone, built from Focus, Conv, C3, and SPP modules, extracts multi-scale features. The Neck combines Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) for effective feature fusion across different resolutions. Finally, the Detection head outputs bounding boxes and class predictions for objects. However, to better address the challenges of anti-UAV detection—such as small target sizes and scale variations—I implement three key improvements to the baseline YOLOv5s model.
First, I replace the Spatial Pyramid Pooling (SPP) module with the SPPF (Spatial Pyramid Pooling Fast) variant. SPPF accelerates computation while maintaining the ability to capture contextual information at multiple scales, which is essential for identifying small UAVs in wide-area surveillance. The SPPF module applies successive max-pooling operations with fixed kernel sizes, efficiently aggregating features without significant loss of detail. This enhancement helps the network better understand the spatial distribution of drones, improving localization accuracy in anti-UAV tasks.
Second, I introduce the SPD-Conv (Space-to-Depth Convolution) module into the backbone network. SPD-Conv is designed to enhance the detection of small objects by preserving fine-grained features during downsampling. Traditional convolution layers often lose subtle details when reducing feature map dimensions, which can be detrimental for anti-UAV applications where drones occupy few pixels. SPD-Conv mitigates this by transforming spatial information into depth channels, allowing the network to retain critical cues for small target recognition. By integrating SPD-Conv, the model becomes more sensitive to minute variations in drone appearances, reducing false negatives in complex environments.
Third, I optimize the loss function by substituting the CIoU Loss with Alpha-CIoU Loss. The CIoU Loss considers overlap area, center distance, and aspect ratio, but it can be suboptimal for small targets due to imbalanced gradient contributions. Alpha-CIoU Loss introduces a tunable parameter $\alpha$ to balance the importance of different components, providing better convergence and generalization on datasets with numerous small objects like UAVs. The Alpha-CIoU Loss is defined as:
$$ L_{\text{Alpha-CIoU}} = 1 – \text{IoU} + \frac{\rho^2(b, b^{gt})}{c^2} + \alpha \cdot v $$
where $\text{IoU}$ is the Intersection over Union, $\rho$ denotes the Euclidean distance between predicted box center $b$ and ground truth box center $b^{gt}$, $c$ is the diagonal length of the smallest enclosing box, $v$ measures aspect ratio consistency, and $\alpha$ is a weighting factor optimized for anti-UAV detection. This adjustment enhances the model’s ability to precisely regress bounding boxes for drones, even when they are tiny or partially occluded.
The overall architecture of the improved YOLOv5s model, incorporating these modifications, is summarized in the following table that outlines the key components and their functions in the anti-UAV pipeline:
| Module | Component | Function in Anti-UAV Detection |
|---|---|---|
| Input | Adaptive Scaling, Mosaic Augmentation | Resizes images and enhances dataset diversity for robust training. |
| Backbone | Focus, Conv, C3, SPD-Conv, SPPF | Extracts multi-scale features with improved small-object sensitivity. |
| Neck | FPN + PAN | Fuses features from different levels to handle scale variations of UAVs. |
| Detection Head | Three output layers | Predicts bounding boxes and classes for small, medium, and large drones. |
| Loss Function | Alpha-CIoU Loss | Optimizes box regression for precise localization of anti-UAV targets. |
To validate the proposed anti-UAV detection method, I conduct extensive experiments using a curated dataset. The dataset is assembled from multiple sources, including the DUT-Anti-UAV dataset, VisDrone2019 dataset, and a custom-collected set of drone images. This composite dataset encompasses diverse scenarios, such as urban skies, open fields, and crowded events, ensuring comprehensive evaluation of anti-UAV capabilities. It contains both low-resolution and high-resolution images, with annotations for drone bounding boxes. For training the Real-ESRGAN model, I use paired low-high resolution images to learn the super-resolution mapping. The YOLOv5s model is then trained on the enhanced images, with data splits of 70% for training, 15% for validation, and 15% for testing. All experiments are performed on a high-performance GPU cluster to simulate real-time anti-UAV processing requirements.
The evaluation metrics for anti-UAV detection focus on precision, recall, and mean Average Precision (mAP). Precision ($P$) measures the proportion of correct drone detections among all positive predictions, while recall ($R$) indicates the fraction of actual drones that are successfully identified. The formulas are as follows:
$$ P = \frac{TP}{TP + FP} $$
$$ R = \frac{TP}{TP + FN} $$
where $TP$ represents true positives (correctly detected drones), $FP$ denotes false positives (incorrect detections), and $FN$ signifies false negatives (missed drones). The Average Precision (AP) is computed as the area under the precision-recall curve for each class, and mAP averages AP across all classes. For anti-UAV applications, I emphasize mAP@0.5, which evaluates detection at an Intersection over Union (IoU) threshold of 0.5, as it balances accuracy and robustness for small targets. Additionally, frames per second (FPS) is used to assess real-time performance, critical for dynamic anti-UAV systems. The overall mAP is calculated as:
$$ AP = \int_0^1 P(R) \, dR $$
$$ mAP = \frac{1}{N} \sum_{i=1}^N AP_i $$
where $N$ is the number of drone classes (typically one for binary anti-UAV detection).
The experimental results demonstrate the effectiveness of the proposed anti-UAV method. After applying Real-ESRGAN for super-resolution reconstruction, the image quality improves significantly, as evidenced by higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) scores compared to bilinear upsampling. For instance, on the test set, PSNR increases from 28.5 dB to 32.1 dB, and SSIM rises from 0.85 to 0.92, indicating better visual fidelity that aids detection. The optimized YOLOv5s model then achieves a mAP@0.5 of 90.3% on the enhanced images, with a recall of 88.7% and precision of 91.5%. These metrics surpass those obtained from raw low-resolution imagery, where mAP@0.5 drops to 84.2%. The improvement highlights the value of super-resolution preprocessing in anti-UAV contexts, where fine details are crucial for distinguishing drones from background clutter.
To provide a comprehensive comparison, I benchmark the proposed method against several classical object detection algorithms commonly used in anti-UAV research, including SSD and YOLOv7. The table below summarizes the performance in terms of mAP@0.5, FPS, and parameter count, all evaluated on the same anti-UAV dataset:
| Algorithm | mAP@0.5 (%) | FPS | Parameters (Millions) |
|---|---|---|---|
| SSD | 87.6 | 24.5 | 24.5 |
| YOLOv7 | 89.2 | 52 | 36.8 |
| YOLOv5s (Baseline) | 88.4 | 50 | 6.8 |
| Proposed Anti-UAV Method | 90.3 | 35 | 9.8 |
The proposed anti-UAV method achieves the highest mAP@0.5 at 90.3%, indicating superior detection accuracy for drones. While its FPS of 35 is slightly lower than YOLOv7’s 52, it remains sufficient for real-time anti-UAV applications, typically requiring at least 30 FPS for smooth video analysis. The parameter count of 9.8 million is moderate, ensuring deployability on edge devices for field anti-UAV systems. The trade-off between speed and accuracy is justified by the enhanced reliability in challenging conditions, such as low-light or occluded scenarios, where other models may falter. Moreover, the integration of super-resolution adds minimal latency (approximately 10 ms per frame on a GPU), making the overall pipeline viable for live anti-UAV monitoring.
Further analysis involves ablation studies to quantify the contribution of each improvement in the YOLOv5s model. I evaluate variants by sequentially adding SPPF, SPD-Conv, and Alpha-CIoU Loss to the baseline. The results, presented in the table below, confirm that each modification positively impacts anti-UAV performance:
| Model Variant | mAP@0.5 (%) | Recall (%) | Precision (%) |
|---|---|---|---|
| Baseline YOLOv5s | 88.4 | 86.2 | 89.1 |
| + SPPF | 89.1 | 87.0 | 89.8 |
| + SPD-Conv | 89.7 | 87.9 | 90.5 |
| + Alpha-CIoU Loss (Full Model) | 90.3 | 88.7 | 91.5 |
The incremental gains highlight the importance of tailored enhancements for anti-UAV detection. SPPF improves contextual understanding, SPD-Conv preserves small-object features, and Alpha-CIoU Loss refines box localization—all critical for accurate drone identification. Additionally, I test the method on specific anti-UAV challenges, such as detecting multiple drones in swarm scenarios or distinguishing drones from birds. The proposed approach maintains a false positive rate below 5% in these cases, outperforming baseline models that often exceed 10%. This robustness is essential for practical anti-UAV systems to minimize nuisance alarms and ensure reliable threat assessment.
Beyond quantitative metrics, the proposed anti-UAV method offers qualitative advantages. For example, in surveillance footage with heavy fog or rain, super-resolution reconstruction helps restore details that are otherwise lost, enabling the detection of drones that would be invisible to conventional systems. Similarly, in urban environments with complex backgrounds like buildings or trees, the enhanced feature extraction of the optimized YOLOv5s reduces confusion between drones and similar-shaped objects. These capabilities make the method versatile for various anti-UAV applications, from airport security to military defense. To illustrate the computational flow, the entire pipeline can be represented as a series of transformations: let $I_{LR}$ denote the low-resolution input image, $I_{HR}$ the high-resolution output from Real-ESRGAN, and $D$ the detection results from YOLOv5s. The process is summarized as:
$$ I_{HR} = G(I_{LR}; \theta_G) $$
$$ D = Y(I_{HR}; \theta_Y) $$
where $G$ is the generator function parameterized by $\theta_G$, and $Y$ is the YOLOv5s detection function parameterized by $\theta_Y$. The parameters are optimized end-to-end for anti-UAV tasks, though in practice, I train them separately for stability.
In terms of implementation, I use PyTorch for both the Real-ESRGAN and YOLOv5s models. The training hyperparameters include a learning rate of 0.001 with cosine annealing, batch size of 16, and Adam optimizer. Data augmentation techniques, such as random cropping, rotation, and color jittering, are applied to simulate diverse anti-UAV conditions. The models are trained for 300 epochs on the composite dataset, with early stopping based on validation mAP. For real-time deployment, I optimize the pipeline using TensorRT for GPU acceleration, achieving an end-to-end latency of 40 ms per frame (25 FPS) on an NVIDIA Jetson Xavier, which is adequate for many anti-UAV systems. Future work could explore lightweight versions for resource-constrained environments, such as mobile anti-UAV units.
The implications of this research extend beyond technical improvements. By enhancing anti-UAV detection accuracy, the method contributes to safer airspaces and more effective counter-drone measures. For instance, in critical infrastructure protection, reliable drone identification can trigger timely interventions, such as jamming or interception. Moreover, the integration of super-resolution can be adapted to other low-vision tasks in surveillance, broadening its impact. However, challenges remain, such as handling extreme weather conditions or adversarial attacks on the detection network. Ongoing efforts in the anti-UAV community focus on addressing these issues through multimodal sensing (e.g., combining visual data with radar or RF signals) and adversarial training.
In conclusion, I have presented a novel anti-UAV detection method that combines super-resolution reconstruction with an optimized YOLOv5s model. This approach effectively addresses the limitations of traditional techniques in small-target detection and complex backgrounds, achieving a mAP@0.5 of 90.3% and maintaining real-time performance. The improvements—including SPPF, SPD-Conv, and Alpha-CIoU Loss—collectively enhance the model’s ability to locate and classify drones accurately. Experimental results validate the superiority over classical algorithms like SSD and YOLOv7, demonstrating its potential for practical anti-UAV applications. As drone technology continues to evolve, robust detection systems will be paramount for security, and this work provides a step forward in that direction. Future research will explore integrating this method with tracking algorithms for persistent anti-UAV monitoring and extending it to handle emerging drone threats, such as stealth designs or autonomous swarms.