A counter unmanned aerial vehicle target identification method based on YOLOv11n

Abstract

This invention discloses a YOLOv11n-based method for countering drone targets, relating to the field of drone countermeasures technology. The method first acquires a dataset of counter-drone scene images and performs data augmentation. Then, based on the YOLOv11n network model, it integrates a SOCA spatial-channel attention module and a UAV-FEM drone-specific feature enhancement module at the output ends of the backbone and neck networks, respectively, to construct a YOLOv11n counter-drone recognition network model. The YOLOv11n counter-drone recognition network model is then trained using the augmented dataset to obtain a trained YOLOv11n counter-drone recognition network model. This model is then used to detect the counter-drone scene images, obtaining the position, confidence level, and category information of the drone targets in the images, and outputting visual annotation results adapted to countermeasure operations. This invention can significantly reduce the number of model parameters and computational load while maintaining accuracy, adapting to the computing power limitations of counter-drone airborne platforms and enabling real-time inference.

Description

Technical Field

[0001] This application relates to the field of drone countermeasures technology, and in particular to a method for identifying drone targets based on YOLOv11n. Background Technology

[0002] With the rapid popularization of drone technology, the illegal use of small and micro drones for activities such as surreptitious filming, cross-border reconnaissance, and dropping dangerous objects is increasing. Drone countermeasure technology has become a core requirement for ensuring low-altitude security. Target recognition, as a key preliminary step in the "discovery-identification-tracking-interception" process of drone countermeasures, directly determines the success or failure of the countermeasure mission in terms of its recognition accuracy, real-time performance, and anti-interference capabilities.

[0003] When traditional target recognition models (including conventional lightweight models) are deployed directly, they suffer from high inference latency and a surge in energy consumption, failing to meet the millisecond-level response requirements of "detection-identification-interception" in countermeasures. Deep learning-based target detection models have become mainstream due to their excellent real-time performance and recognition accuracy. While YOLOv11n, as a new generation of lightweight models, has the potential to adapt to low computing power, its feature extraction capabilities are insufficient when directly applied to UAV countermeasures scenarios, especially for low-altitude small targets, maneuvering evasive targets, and low signal-to-noise ratio images under electromagnetic interference. Target tracking is prone to interruption in dynamic adversarial scenarios, and its high parameter count and computational complexity make it prone to inference latency when adapted to the limited computing power of UAVs, failing to meet the millisecond-level response requirements of countermeasures.

[0004] In summary, when existing target recognition technologies are applied to the field of drone countermeasures, the contradiction between the high precision and robustness requirements of detection algorithms in drone countermeasure scenarios and the limited computing, storage, and power consumption constraints of front-end devices remains a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] Therefore, it is necessary to provide a counter-drone target identification method and system based on YOLOv11n to address the aforementioned technical problems.

[0006] The following technical solution is adopted in this specification: This specification provides a YOLOv11n-based method for countering drone target identification, including: Acquire images of the counter-drone scene; Construct a pre-trained YOLOv11n anti-drone identification network model; the YOLOv11n anti-drone identification network model includes a backbone network integrating SOCA spatial-channel attention module at the output end, a neck network integrating UAV-FEM drone-specific feature enhancement module at the output end, a functional module area, and a visualization annotation module; Based on the counter-drone scene image, multi-scale features of the counter-drone scene image are extracted through the backbone network to obtain a multi-layer feature map; the SOCA spatial-channel attention module performs spatial and channel weighted enhancement on the multi-scale feature map to obtain a multi-feature map focused on the target region; the neck network performs cross-scale fusion operations on the multi-feature map focused on the target region to obtain a fused feature map; the UAV-FEM drone-specific feature enhancement module performs drone-specific feature filtering on the fused features to obtain targeted enhanced drone features; the functional module area performs confidence level filtering on the targeted enhanced drone features and coordinate processing on the drone category detection boxes to obtain effective detection boxes; the visualization annotation module generates the position, confidence, and category information of the drone target on the image to be detected to obtain a visualization annotation result adapted to the counter-operation.

[0007] Furthermore, the training process of the YOLOv11n anti-drone identification network model specifically includes: Acquire a dataset of images of counter-drone scenarios and perform data augmentation on the dataset; Using the augmented dataset, the YOLOv11n network model is iteratively trained with preset training parameters and pre-trained weight files; In the iterative training process, CIoU loss and distributed focus loss are used as localization losses, binary cross-entropy loss is used as classification losses, and the AdamW optimizer is used to update the network weights.

[0008] Furthermore, the data augmentation process for the counter-drone scene image dataset specifically includes: Obtain a dataset of images depicting counter-drone scenarios; Within a preset range, the image brightness and contrast of the counter-drone scene images in the dataset are randomly adjusted to obtain counter-drone scene images with enhanced environmental interference. Data augmentation is achieved by performing geometric transformation enhancement on anti-drone scene images with enhanced environmental interference and specific enhancement on small targets, and adjusting the coordinates of the target bounding boxes to obtain new training samples with effective annotations. The geometric transformation enhancement includes: horizontal flipping at a first preset ratio, random rotation within a range of ±15° at a second preset ratio, and proportional scaling within a range of 0.8 to 1.2 times at a third preset ratio. The small target is specifically enhanced by: super-resolution magnification of the target image region containing a size smaller than a preset threshold, and random local cropping of the region where the small target is located at a fourth preset ratio.

[0009] Furthermore, the construction of the fused feature map includes: Obtain multi-layer feature maps of the SOCA spatial-channel attention module output network; The multi-layer feature maps are input into the neck network of the improved PAN structure; the neck network performs spatial resolution scale alignment on the multi-layer feature maps respectively to obtain feature maps with the same resolution; Multiple feature maps with the same resolution are stitched together across layers to obtain a fused feature map; The spatial resolution scale alignment operation is performed through upsampling and downsampling.

[0010] Furthermore, the UAV-FEM drone-specific feature enhancement module matches and filters the fused features processed by the neck network using a preset drone-specific feature template to eliminate feature interference from confusing targets and to directionally enhance the feature expression of drone targets. The human-machine exclusive feature template includes: rotor texture, fuselage geometry and pixel gradient features.

[0011] Furthermore, the functional module area includes a multi-dimensional precision filtering module and an edge target correction module; The multi-dimensional precision filtering module is used to perform confidence-level screening of the targeted enhanced drone features and generate drone category detection boxes. The edge target correction module is used to perform coordinate restoration and boundary constraint processing on the UAV category detection box to obtain an effective detection box.

[0012] Furthermore, the configuration of the confidence level screening includes: Set a first confidence threshold and filter out detection boxes with a confidence level lower than the first confidence threshold; Perform category confidence verification on the remaining detection boxes, and retain only the detection boxes whose category is drone and whose category confidence is not lower than the second confidence threshold; Wherein, the first confidence threshold is lower than the second confidence threshold.

[0013] Furthermore, the coordinate restoration and boundary constraint processing of the drone category detection box specifically includes: Coordinate restoration processing, based on the fill parameters and scaling ratio in the image preprocessing stage, restores the coordinates of the detection box to the original image size; The formula for restoring the coordinates is: ; Where, x pad y pad pad is the original coordinate of the detection box. left ,pad topThe parameter is the fill parameter, and scale is the scaling ratio; Coordinates to restore the original image size; Boundary constraint processing: restoring the coordinates , Each bounded within the original image size range yields an effective detection box.

[0014] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This specification presents a YOLOv11n-based counter-drone target recognition method. By integrating the SOCA spatial-channel attention module and the UAV-FEM drone-specific feature enhancement module into the backbone and neck networks respectively, a lightweight model with a three-layer optimized structure of lightweight backbone + general attention enhancement + scene-specific feature filtering is formed for counter-drone recognition. Specifically, the integration of the SOCA spatial-channel attention module enhances the feature response of the target region and suppresses the noise impact of complex backgrounds. The UAV-FEM drone-specific feature enhancement module, based on preset feature templates such as rotor texture and fuselage geometry, selectively filters drone target features, further improving the model's detection capability for small, maneuverable evasive targets. Compared to traditional target detection models and the standard YOLOv11n model, the improved YOLOv11n model significantly improves the accuracy and robustness of drone recognition in complex environments through the synergistic effect of the multi-dimensional precision filtering module and the edge target correction module.

[0015] Furthermore, the functional module area effectively reduces false detections and missed detections by setting a dual verification mechanism of basic confidence threshold and UAV category confidence, as well as coordinate restoration and boundary constraint mechanisms. Especially for targets in the edge areas of the image, the positioning accuracy can be improved by correcting the coordinates, thereby improving the overall detection accuracy and enabling real-time inference. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 This document provides a flowchart illustrating a YOLOv11n-based method for countering drone target identification. Figure 2 This document provides a schematic diagram of the entire object detection process using the initial YOLOv11n model. Figure 3This document provides a schematic diagram illustrating the entire process of detection using the YOLOv11n+SOCA+UAV-FEM model provided in this manual, following the steps of "input → trunk → neck → functional module → visualization → output". Figure 4 A schematic diagram of the PR curve for a YOLOv11n-based counter-drone target recognition method provided in this specification; Figure 5 The images show a comparison of the detection results before and after. The left side shows the detection results of the original YOLOv11n model, and the right side shows the detection results of the YOLOv11n+SOCA+UAV-FEM integrated model. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0019] Current drone target recognition technologies are mainly divided into traditional computer vision methods and deep learning methods. Traditional methods rely on manually designed feature extraction operators. In complex low-altitude environments (such as building obstruction, vegetation interference, sudden changes in lighting, weather clutter, etc.), they have poor robustness in identifying small intruding drones and are easily confused with non-threatening targets such as birds and kites. The false negative and false positive rates are high, making it difficult to meet the needs of countermeasures.

[0020] Current drone target recognition technologies are mainly divided into traditional computer vision methods and deep learning methods. Traditional methods rely on manually designed feature extraction operators. In complex low-altitude environments (such as building obstruction, vegetation interference, sudden changes in lighting, weather clutter, etc.), they have poor robustness in identifying small intruding drones and are easily confused with non-threatening targets such as birds and kites. The false negative and false positive rates are high, making it difficult to meet the needs of countermeasures. Deep learning methods, especially target detection models based on the YOLO series, have become mainstream due to their excellent real-time performance and recognition accuracy. However, they still have significant shortcomings in drone countermeasure scenarios: First, existing models (including some lightweight models) have a high number of parameters and computational complexity, which can easily lead to inference delays when adapted to the limited computing power of drones, failing to meet the millisecond-level response requirements of countermeasure tasks and drastically consuming onboard power. Second, they lack feature extraction capabilities for low-altitude small targets, maneuvering evasive targets, and low signal-to-noise ratio images under electromagnetic interference, which can easily cause target tracking interruptions in dynamic adversarial scenarios. Third, most models can only classify targets and cannot simultaneously determine the threat level of multiple intruding drones, resulting in unreasonable allocation of countermeasure resources. Fourth, the recognition system and countermeasure payloads (electromagnetic jammers, net capture devices, etc.) are mostly independent modules, resulting in high data interaction latency, making it difficult for recognition results to drive countermeasure actions in real time, and exhibiting poor collaborative adaptability.

[0021] To address core issues such as limited onboard computing power for countering drones, insufficient robustness in complex environments, distortion of edge targets, and poor visualization adaptability, this invention provides a drone target recognition method and system based on YOLOv11n. The core solution consists of "lightweight backbone + standardized preprocessing + multi-dimensional filtering + adaptive visualization." Through the collaborative efforts of these five core modules, accurate, real-time, and interference-resistant drone recognition is achieved, overcoming the challenges of accurate recognition, real-time response, and coordinated countermeasures in complex adversarial environments, thereby improving the reliability and efficiency of drone countermeasure missions.

[0022] The technical problems to be solved by this invention mainly include: 1. The contradiction between limited onboard computing power and the real-time recognition requirements of counter-drones; Countermeasure drone onboard platforms inherently suffer from limited computing power and short endurance. When traditional target recognition models (including conventional lightweight models) are deployed directly, inference latency is high and energy consumption surges, failing to meet the millisecond-level response requirements of the "detection-identification-interception" process in countermeasure missions. This invention aims to solve the problems of deep lightweighting of the YOLOv11n model and precise allocation of computing resources, reducing the number of model parameters and computational complexity while ensuring the accuracy of intrusion drone recognition, thus adapting to the real-time inference requirements of embedded hardware for countermeasure drones.

[0023] 2. Issues related to target tracking continuity and resistance to evasion interference in dynamic adversarial scenarios; During countermeasures, intruding drones maneuver to evade attacks (dive, hover, and circle around), while the countermeasure drones themselves are also in high-speed flight, causing rapid shifts in the target's perspective and dynamic changes in its relative position. Some intruding drones also release false targets and smoke for interference. Existing models rely solely on single-frame image recognition, lacking the fusion of motion trajectory and attitude data, which easily leads to target tracking interruptions and abrupt changes in recognition results. This invention aims to solve the problem of coordinated adaptation between YOLOv11n and the flight attitude and motion parameters of countermeasure drones, achieving continuous and stable recognition and anti-interference tracking of intruding drones in dynamic adversarial scenarios.

[0024] 3. The issue of integrated and coordinated adaptation between the identification system and the countermeasure payload; Existing UAV identification systems and countermeasure payloads (electromagnetic jammers, net capture devices, laser interceptors) are mostly independent modules with inconsistent interfaces, high data interaction latency, and the inability to drive countermeasure actions in real time. Furthermore, model updates rely on specialized personnel, hindering rapid iterative optimization in battlefield environments. This invention aims to address the modular design and linkage adaptation issues of YOLOv11n-based identification systems with countermeasure payloads, achieving low-latency interaction between identification results and countermeasure commands, while simplifying the model update process and improving the system's battlefield operational convenience.

[0025] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating a YOLOv11n-based counter-drone target identification method described in this specification, which specifically includes the following steps: S101: Acquire a dataset of images of counter-drone scenarios and perform data augmentation.

[0027] First, a dataset of counter-drone scene images is acquired. For the counter-drone scene images in the acquired image dataset, basic data augmentation operations are performed to improve the model's adaptability to complex environments. Then, specific augmentation operations are performed for small target drones to prevent the model from missing detection of low-altitude micro drones.

[0028] Data augmentation methods specifically include: Basic environmental interference enhancement: Randomly adjust the brightness and contrast of the image, with an adjustment range of ±20%; Geometric transformation enhancement: The image is horizontally flipped with a probability of 0.5, randomly rotated within a range of ±15° with a probability of 0.3, and scaled proportionally by 0.8 to 1.2 times with a probability of 0.2, thereby improving the model's robustness to changes in target pose. Small target enhancement: For extremely small drone samples with pixel size ≤30×30, perform a 2x super-resolution magnification operation to supplement the detailed features of small targets; at the same time, perform random local cropping of small target regions with a probability of 0.4 to enhance the model's ability to perceive the features of small targets; Synchronous adaptation of bounding boxes: For images that have undergone geometric transformations and cropping operations, the coordinates of their corresponding target bounding boxes are adjusted synchronously to ensure that the annotation information is consistent with the image content, thus obtaining new training samples with effective annotations.

[0029] In this embodiment, the parameter range for basic environmental interference enhancement is: brightness adjustment range [-0.2, 0.2], contrast adjustment range [-0.2, 0.2]; the probability threshold for geometric transformation enhancement is: horizontal flip 0.5, random rotation 0.3, and proportional scaling 0.2; the super-resolution amplification factor for small target enhancement is fixed at 2, and the local clipping probability is 0.4.

[0030] S102: Construct a YOLOv11n counter-drone identification network model.

[0031] For example Figure 2 The traditional YOLOv11n network shown is used as the base network. The SOCA spatial-channel attention module and the UAV-FEM drone-specific feature enhancement module are integrated at the output ends of the backbone network and neck network, respectively, resulting in a YOLOv11n anti-drone recognition network model that includes the backbone network, SOCA spatial-channel attention module, neck network, UAV-FEM drone-specific feature enhancement module, multi-dimensional precise filtering module, edge target correction module, and visualization annotation module. This forms a three-layer optimized structure of "base network + general attention enhancement + scene-specific feature selection" and a complete technical process of "original image input → standardized preprocessing → feature extraction (backbone network) → feature enhancement (SOCA module) → feature fusion (neck network) → specific feature selection (UAV-FEM module) → initial target detection → multi-dimensional filtering → edge correction → visualization output → result saving".

[0032] The traditional YOLOv11n network consists of three parts: the backbone network, the neck network, and the detection head. The backbone network extracts multi-scale features from the input image through a series of convolution and pooling operations, resulting in multi-layered feature maps from large to small. The neck network adopts a Feature Pyramid Network (FPN) structure to transfer feature information from bottom to top and from top to bottom on feature maps of different scales. The detection head part predicts the target location and category based on the feature maps processed by the neck network.

[0033] The backbone network of the YOLOv11n anti-drone identification network model constructed in this invention uses the C3k2 lightweight convolutional unit built into YOLOv11n to perform multi-level downsampling and feature extraction on the input image, obtaining multi-layer feature maps of 80×80, 40×40, and 20×20. The neck network is based on an improved PAN structure and is used to perform cross-scale fusion of multi-layer feature maps. The multi-dimensional precise filtering module is used to filter the preliminary detection results based on confidence threshold and category confidence verification, excluding low-confidence and redundant detection boxes. The edge target correction module is used to restore the coordinates and boundary constraints of targets in the edge areas of the image. The visualization annotation module is used to generate "red background with white text" annotation results adapted to the countermeasure operation.

[0034] The core modules of the YOLOv11n anti-drone identification network model include the SOCA spatial-channel attention module and the UAV-FEM drone-specific feature enhancement module; SOCA Spatial-Channel Attention Module: The core is to enhance the feature response of the target area and suppress background noise by weighted fusion of spatial attention branch (capturing the location information of the target area) and channel attention branch (filtering effective feature channels). The spatial attention branch is used to capture the location information of the UAV target area, and the channel attention branch is used to filter the feature channels that are effective for UAV identification.

[0035] UAV-FEM Drone-Specific Feature Enhancement Module: This module is a self-developed feature selection module adapted to counter-drone scenarios. The code is independently written based on the PyTorch framework. The core is to match and filter the input feature map through preset drone-specific feature templates (rotor texture, fuselage geometry, pixel gradient features) to eliminate feature interference from confusing targets such as birds and kites, and to enhance the feature expression of drone targets in a targeted manner.

[0036] The YOLOv11n anti-drone identification network model of this invention is based on a traditional YOLOv11n network (such as a backbone network and neck network based on C3k2 units, a multi-dimensional precise filtering module, an edge target correction module, and a visualization annotation module). Figure 2 Improvements were made based on the model shown. The improved YOLOv11n anti-drone identification network model specifically includes a backbone network and neck network based on C3k2 units, an SOCA spatial-channel attention module, a UAV-FEM drone-specific feature enhancement module, a multi-dimensional precise filtering module, an edge target correction module, and a visualization annotation module.

[0037] The YOLOv11n backbone network includes multiple standard convolutional layers, multiple C3k2 layers, one SPPF layer, and one C2PSA layer. These layers are connected serially and numbered sequentially from layer 0 to layer 10. Layers 0, 2, 4, and 6 are standard convolutional layers; layers 1, 3, 5, and 7 are C3k2 layers; layer 8 is an SPPF layer; and layer 9 is a C2PSA layer. The backbone network of this invention retains the connection structure of each layer. Through the channel compression characteristics of the C3k2 units, the number of channels in the multi-layer feature maps output by the backbone network is controlled between 128 and 512, while maintaining the feature map resolution (80×80, 40×40, 20×20). An SOCA module is integrated at the output of the C2PSA layer (layer 9) of the backbone network to perform spatial-channel weighted enhancement on the multi-scale feature maps output by the backbone network, ensuring that the features received by the subsequent neck network are more focused on the target region.

[0038] The specific working process of the neck network and UAV-FEM module shown in Figure 2 is as follows: Based on the feature order of the multi-layer feature maps output by the backbone network, P1 (80×80), P2 (40×40), and P3 (20×20) are selected as feature maps to be fused. The multi-layer feature maps are input into the neck network of the improved PAN structure. The neck network performs cross-scale fusion operations on the feature maps output by the three C3k2 layers of the backbone network. Specifically, for each feature map, the neck network first scales the feature map with its neighboring feature maps by upsampling the smaller feature maps and downsampling the larger feature maps, ensuring that the feature map resolutions are the same and maintaining a consistent number of channels across the three feature maps. Taking the fusion of feature map P1 as an example, the scale alignment process can be represented as follows:

[0039] X1₊1 = UpSample (P1₊1); Where X1₊1 represents the scale-aligned feature map, and UpSample represents the upsampling operation.

[0040] After spatial resolution adjustment, the feature maps aligned to the three scales are concatenated across layers to obtain a fused feature map. This process can be represented as: X_fuse = Concat (P1, X1₊1); A UAV-FEM module is integrated at the output of the last C3k2 layer (number 18) of the neck network to perform UAV-specific feature filtering on the fused feature map, further eliminating interference from non-target features and ensuring that the features received by the detection head are the UAV features after targeted enhancement.

[0041] Furthermore, the multi-dimensional precision filtering module processes the detection results output by the neck network. The specific process is as follows: Confidence-based tiered filtering: Set multi-level confidence thresholds, with a basic filtering threshold T_conf=0.1. First, filter out detection boxes with confidence levels below this threshold. In response to the core requirement of "only identifying drone targets" in countermeasure scenarios, an additional category confidence check is added. Only detection boxes with the category "drone" and a category confidence level ≥ 0.85 are retained. This dual filtering mechanism excludes false detection results of confusing targets such as birds and kites.

[0042] Where Area (Box1∩Box2) is the intersection area of ​​the two detection boxes, and Area (Box1∪Box2) is the union area of ​​the two detection boxes.

[0043] Furthermore, the edge target correction module performs coordinate processing on the filtered detection boxes. The specific process is as follows: Coordinate Restoration: Based on the fill parameters and scaling ratios from the image preprocessing stage, the coordinates of the 640×640 detection box are restored to the original image size. The formula is as follows: Where x_pad and y_pad are the coordinates output by the model, pad_left and pad_top are the padding parameters, and scale is the scaling ratio; Boundary constraints: The restored coordinates are constrained within the boundaries of the original image by max (0, int (x_raw)) and min (w-1, int (x_raw)), to prevent the coordinates from exceeding the valid range.

[0044] S103: Use the augmented dataset to train the YOLOv11n anti-drone identification network model to obtain the trained YOLOv11n anti-drone identification network model.

[0045] This embodiment uses the Ultralytics framework to train the YOLOv11n model. The specific training process and parameter configuration are as follows: Pre-trained weight loading: Load the official YOLOv11n pre-trained weight file (yolo11n.pt) as the model initialization weights, and use the general feature extraction capabilities of the pre-trained weights to reduce the difficulty of model convergence; Dataset configuration: Divide the augmented dataset into training, validation, and test sets in an 8:1:1 ratio. Specify the dataset path, number of categories (UAV only), category names, and other core information through the train_yolo11n_SOCA_UAV-FEM.yaml configuration file. Training parameter settings: Input image size is uniformly scaled to 640×640 pixels; batch size is set to automatic adaptation mode (batch=-1), which allows the framework to automatically allocate the optimal batch based on hardware computing power; memory caching (cache="ram") is enabled to accelerate dataset reading; the number of data loading threads is set to 1 (workers=1) to adapt to the thread resource limitations of onboard embedded devices; the number of training epochs is set to 100 epochs to control training time while ensuring model convergence; Loss function and optimizer: CIoU loss and distributed focus loss (DFL) are used as localization loss, and binary cross-entropy loss (BCE) is used as classification loss; the framework's default Adam W optimizer is used to update network weights, with an initial learning rate of 0.01 and a weight decay coefficient of 0.0005; Model saving: After training is completed, the trained network weights are automatically saved to the default path. In this embodiment, the final weight file storage path is (runs / detect / train / weights / best.pt).

[0046] S104: Utilize the trained YOLOv11n anti-drone identification network model to detect the anti-drone scene image to be detected, obtain the position, confidence level and category information of the drone target on the image to be detected, and output the visual annotation results adapted to the anti-drone operation.

[0047] Specifically, such as Figure 3 As shown, based on the trained YOLOv11n+SOCA+UAV-FEM model and weights, the following process is performed on the image of the countermeasure scene to be detected: First, image preprocessing is used to obtain a 640×640 standardized input; then, the backbone network extracts features, the SOCA module enhances target features, the neck network fuses multi-scale features, and the UAV-FEM module filters drone features; subsequently, multi-dimensional precise filtering (confidence ≥ 0.1 + drone category confidence ≥ 0.85) and edge target correction (coordinate restoration + boundary constraints + aspect ratio verification) are used to obtain effective detection boxes; finally, the visualization annotation module generates the annotation result of "green detection box (line width 3px) + red background and white text inside the box (RGB (0,0,255), transparency 0.9 + white UAV confidence text)", and the annotated image and YOLO format label file are saved to the specified directory simultaneously.

[0048] like Figure 1 The YOLOv11n-based counter-drone target identification method shown has the following advantages compared with existing technologies: 1. Lightweight design: The lightweight backbone network of YOLOv11n+SOCA+UAV-FEM is adopted, and its C3k2 unit is used for efficient feature extraction. While ensuring accuracy, the number of model parameters and computation is greatly reduced, which is suitable for the computing power limitations of the UAV-borne embedded platform and enables real-time inference.

[0049] 2. High-precision recognition: Through a multi-dimensional precision filtering module (confidence screening and category confidence verification) and an edge target correction module, false detections and missed detections are effectively reduced, especially for targets in the edge areas of the image. After correcting the coordinates, the positioning accuracy is improved, thereby improving the overall detection accuracy.

[0050] 3. Intuitive visual output: such as Figure 5 As shown, the red background and white text are used to mark the detection box in the lower left corner, making the recognition results clear and eye-catching on the anti-drone operation interface, which facilitates the operator to quickly locate and make decisions.

[0051] 4. End-to-end system integration: The entire process, from image input to detection result output, is modularly designed, facilitating integration with other anti-drone systems (such as flight control and net capture equipment) to achieve integrated identification-tracking-interception.

[0052] 5. Experimental verification showed that, on the self-made counter-drone dataset, such as... Figure 4 As shown, the model of this invention improves accuracy (mAP) by about 10% compared to the baseline model (such as YOLOv5n), reduces the number of parameters by about 20%, and achieves an inference speed of 45 FPS on the Jetson Xavier embedded platform, meeting the real-time requirements.

[0053] The above describes one or more embodiments of a UAV target recognition method based on YOLOv11n. Based on the same approach, this specification also provides a corresponding UAV target recognition system, including: Image normalization preprocessing module: realizes image loading, noise reduction, scale normalization and proportional padding, and eliminates distortion; YOLOv11n backbone network module: Extracts multi-scale features with C3k2 units as the core, adapting to the requirements of lightweight computing power for airborne applications; Multi-dimensional precision filtering module: Filters targets based on confidence level to eliminate low-confidence and redundant targets; Edge target coordinate correction module: restores the coordinates and sets boundary constraints for targets in the edge areas of the image to avoid distortion; Adaptive visualization annotation module: Outputs recognition results in the form of "red background and white text in the lower left corner of the detection box", adapting to the observation needs of countermeasure operations.

[0054] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A counter-drone target identification method based on YOLOv11n, characterized in that, include: Acquire images of the counter-drone scene; Construct a pre-trained YOLOv11n anti-drone identification network model; the YOLOv11n anti-drone identification network model includes a backbone network integrating SOCA spatial-channel attention module at the output end, a neck network integrating UAV-FEM drone-specific feature enhancement module at the output end, a functional module area, and a visualization annotation module; Based on the counter-drone scene image, multi-scale features of the counter-drone scene image are extracted through the backbone network to obtain a multi-layer feature map; the multi-scale feature map is enhanced by spatial and channel weighting through the SOCA spatial-channel attention module to obtain a multi-feature map focused on the target region. The neck network performs cross-scale fusion operations on the multi-feature maps focused on the target region to obtain a fused feature map. The UAV-FEM drone-specific feature enhancement module performs drone-specific feature filtering on the fused features to obtain targeted enhanced drone features. The functional module area performs confidence level filtering on the targeted enhanced drone features and performs coordinate processing on the drone category detection boxes to obtain effective detection boxes. The visualization annotation module generates the position, confidence level, and category information of the drone target on the image to be detected, and obtains the visualization annotation results for the adaptive countermeasure operation.

2. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 1, characterized in that, The training process of the YOLOv11n anti-drone identification network model specifically includes: Acquire a dataset of images of counter-drone scenarios and perform data augmentation on the dataset; Using the augmented dataset, the YOLOv11n network model is iteratively trained with preset training parameters and pre-trained weight files; In the iterative training process, CIoU loss and distributed focus loss are used as localization losses, binary cross-entropy loss is used as classification losses, and the AdamW optimizer is used to update the network weights.

3. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 2, characterized in that, The data augmentation process for the counter-drone scene image dataset specifically includes: Obtain a dataset of images depicting counter-drone scenarios; Within a preset range, the image brightness and contrast of the counter-drone scene images in the dataset are randomly adjusted to obtain counter-drone scene images with enhanced environmental interference. Data augmentation is achieved by performing geometric transformation enhancement on anti-drone scene images with enhanced environmental interference and specific enhancement on small targets, and adjusting the coordinates of the target bounding boxes to obtain new training samples with effective annotations. The geometric transformation enhancement includes: horizontal flipping at a first preset ratio, random rotation within a range of ±15° at a second preset ratio, and proportional scaling within a range of 0.8 to 1.2 times at a third preset ratio. The small target is specifically enhanced by: super-resolution magnification of the target image region containing a size smaller than a preset threshold, and random local cropping of the region where the small target is located at a fourth preset ratio.

4. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 1, characterized in that, The construction of the fused feature map includes: Obtain multi-layer feature maps of the SOCA spatial-channel attention module output network; The multi-layer feature maps are input into the neck network of the improved PAN structure; the neck network performs spatial resolution scale alignment on the multi-layer feature maps respectively to obtain feature maps with the same resolution; Multiple feature maps with the same resolution are stitched together across layers to obtain a fused feature map; The spatial resolution scale alignment operation is performed through upsampling and downsampling.

5. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 1, characterized in that, The UAV-FEM drone-specific feature enhancement module uses a preset drone-specific feature template to match and filter the fused features processed by the neck network, so as to eliminate feature interference from confusing targets and enhance the feature expression of drone targets in a targeted manner. The human-machine exclusive feature template includes: rotor texture, fuselage geometry and pixel gradient features.

6. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 1, characterized in that, The functional module area includes a multi-dimensional precision filtering module and an edge target correction module; The multi-dimensional precision filtering module is used to perform confidence-level screening of the targeted enhanced drone features and generate drone category detection boxes. The edge target correction module is used to perform coordinate restoration and boundary constraint processing on the UAV category detection box to obtain an effective detection box.

7. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 6, characterized in that, The configuration for the confidence level-based filtering includes: Set a first confidence threshold and filter out detection boxes with a confidence level lower than the first confidence threshold; Perform category confidence verification on the remaining detection boxes, and retain only the detection boxes whose category is drone and whose category confidence is not lower than the second confidence threshold; Wherein, the first confidence threshold is lower than the second confidence threshold.

8. The method for countering unmanned aerial vehicle (UAV) target identification based on YOLOv11n as described in claim 6, characterized in that, The coordinate restoration and boundary constraint processing of the UAV category detection box specifically includes: Coordinate restoration processing, based on the fill parameters and scaling ratio in the image preprocessing stage, restores the coordinates of the detection box to the original image size; The formula for restoring the coordinates is: ； Where, x pad y pad pad is the original coordinate of the detection box. left ,pad top The parameter is the fill parameter, and scale is the scaling ratio; Coordinates to restore the original image size; Boundary constraint processing: restoring the coordinates , Each bounded within the original image size range yields an effective detection box.

Images