An unmanned aerial vehicle small target recognition method, system, device and storage medium

Abstract

The application discloses a kind of unmanned plane small target identification method and system, comprising: S1.unmanned plane collects image data, and image data is preprocessed;S2.YOLOv11 model network framework is constructed, SCSA module and SAFM module are introduced in main network and loss function is optimized, combined with Mosai c data enhancement technology, to enhance the recognition and extraction ability of small target features of model;S3.YOLOv11 model training learning optimization is carried out;S4.the model trained is deployed to embedded device, real-time processing unmanned plane collected image data is used to output target identification result;S5.lastly, the result is output and fed back to unmanned plane.The network structure of the application optimizes YOLOv11 algorithm, significantly improves the recognition accuracy of small target of unmanned plane, significantly improves the recognition accuracy and real-time performance of small target of unmanned plane, and adapts to the calculation resource limitation of embedded device of unmanned plane.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) target recognition technology, specifically to a method and system for small UAV target recognition based on an improved YOLOv11, as well as computer equipment and storage media. Background Technology

[0002] Unmanned aerial vehicle (UAV) target recognition technology is a core component of modern UAV technology applications. It relies on multiple sensors and advanced algorithms to detect, classify, and track specific targets. With the rapid development of UAV technology, the application of UAVs in the field of target recognition is becoming increasingly widespread.

[0003] However, existing UAV target recognition technologies have the following problems in small target recognition: (1) Low accuracy of small target recognition: Since small targets occupy fewer pixels in the image, traditional target detection algorithms are difficult to recognize accurately. (2) High computational complexity: Deep learning models usually require a lot of computing resources and are difficult to run in real time on UAV embedded devices. (3) Poor adaptability: Existing algorithms are poorly adaptable to complex environments (such as changes in lighting, occlusion, etc.).

[0004] Therefore, this application proposes a method and system for small target recognition using unmanned aerial vehicles (UAVs) to solve the aforementioned technical problems. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for small target recognition in unmanned aerial vehicles (UAVs). By optimizing the network structure and introducing new technical modules, the accuracy and efficiency of small target recognition are significantly improved, thereby solving the technical problems mentioned in the background art.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0007] A method and system for small target identification by unmanned aerial vehicles (UAVs) includes the following steps:

[0008] S1. Use a drone to collect image data through an onboard camera, and perform preprocessing operations on the image data, including denoising and normalization.

[0009] S2. Construct the YOLOv11 model network framework, introduce the SCSA and SAFM modules into the backbone network of the YOLOv11 model network framework and optimize the loss function, and combine Mosaic data augmentation technology to enhance the model's ability to identify and extract features of small targets.

[0010] S3. Use the COCO dataset to pre-train the YOLOv11 model, then use the Adam optimizer for learning and optimization, and finally use data augmentation techniques to improve the model's generalization ability.

[0011] S4. Real-time target recognition: Deploy the trained model onto an embedded device to process image data collected by the drone in real time and output target recognition results;

[0012] S5. Finally, the results are output and fed back to the drone.

[0013] Preferably, in step S2, the SCSA module consists of a shareable multi-semantic space attention SMSA module and a progressive channel self-attention PCSA module, wherein:

[0014] The SMSA module is used to extract spatial information at different semantic levels from the feature map using multi-scale convolution, and accelerates model convergence through group normalization, and fuses the extracted spatial information to generate spatial attention information.

[0015] The PCSA module is used to inject the spatial attention information generated by the SMSA module into the channel self-attention using a progressive compression strategy, and to generate channel attention weights to enhance the feature extraction capability of small targets.

[0016] Preferably, in step S2, the SAFM module consists of a multi-scale feature generation unit (MFGU) and a feature aggregation and modulation module, wherein:

[0017] MFGU contains at least two branches, one of which uses 3×3 depthwise convolution to process local features, and the other branches generate features at different scales through pooling operations, and then restore them to the original resolution through upsampling;

[0018] The feature aggregation and modulation module is used to concatenate features of different scales along the channel dimension, aggregate them through 1×1 convolution, normalize the generated attention map using the GELU function, and then modulate the input features by element-wise multiplication.

[0019] Preferably, the data processing flow of the SAFM module in step S2 includes:

[0020] L1 performs channel segmentation on the normalized input features to generate four component parts. One part is processed by depthwise convolution, while the other three parts are fed into the multi-scale feature generation unit (MFGU) and undergo downsampling, convolution, and upsampling operations to generate multi-scale features.

[0021] L2. Adaptive max pooling is applied to the input features to generate multi-scale features. These multi-scale features are then concatenated by convolution to obtain aggregated features. After that, the features are normalized by the GELU nonlinear function to estimate the attention map.

[0022] L3. Adaptively modulate the original input features based on the attention map estimation data.

[0023] Preferably, in step S2, the loss function is optimized by adjusting the weights of positive and negative samples, so that the model pays more attention to the difficult-to-classify positive samples, wherein the function is expressed as:

[0024]

[0025] Among them, P t The predicted value `pred` is the result after passing through the sigmoid function and is used to compare with the true value `true`. α and γ are hyperparameters that adjust the weights of negative samples. Hyperparameter α is used to prevent excessive suppression of negative samples, and hyperparameter γ is used to reduce the contribution of easily classified samples to the loss, thereby improving the accuracy of small object detection.

[0026] Preferably, the Mosaic data augmentation technique in step S2 specifically includes:

[0027] Four images are randomly selected from the dataset, and each image is randomly cropped and scaled to generate image patches of different sizes;

[0028] The cropped image blocks are stitched together according to a specified layout to form a new image;

[0029] Adjust the target bounding boxes in each image according to the position of the stitched images to adapt them to the new image positions.

[0030] Preferably, in the YOLOv11 model network framework constructed in step S2, depthwise separable convolutions are used instead of traditional convolutions in the detection head to reduce the number of model parameters and computational cost. The depthwise separable convolutions consist of depthwise convolutions and pointwise convolutions, wherein:

[0031] Depthwise convolution, for input C in Each channel of the multi-channel tensor is convolved independently using an h×w convolution kernel, and the number of output channels is the same as the number of input channels;

[0032] Point convolution uses a 1×1 convolution kernel to perform channel mixing on the output of depthwise convolution to integrate information from different channels.

[0033] Preferably, the specific operation process of step S5 includes:

[0034] S51. Transmit the identification results to the ground control station via the 4G / 5G module;

[0035] S52. The ground control station generates control commands based on the identification results and sends them to the UAV.

[0036] A small target recognition system for unmanned aerial vehicles (UAVs), used to execute any of the above-described small target recognition methods for UAVs, comprising:

[0037] The unmanned aerial vehicle (UAV) platform and ground control station are provided. The UAV platform is equipped with an airborne camera, an embedded processor, a wireless communication module, a data acquisition and preprocessing module, and a real-time inference and result output module. The data acquisition and preprocessing module is connected to the airborne camera. The embedded processor has a YOLOv11 model network framework embedded in it. The real-time inference and result output module is used to control data processing operations based on the YOLOv11 model network framework.

[0038] The ground control station is used to receive and process data transmitted by the UAV, and based on the data results, automatically adjust the UAV's flight path or perform specific tasks.

[0039] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0040] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0041] As can be seen from the above technical solution, the present invention provides a method and system for small target identification by unmanned aerial vehicles (UAVs).

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] 1. This invention optimizes the network structure of the YOLOv11 algorithm by introducing a spatial and channel collaborative attention module, a SAFM module, an improved Varifocal Loss function, and Mosaic data augmentation technology. This significantly improves the accuracy and real-time performance of UAVs in recognizing small targets, while also adapting to the computational resource limitations of UAV embedded devices.

[0044] 2. This invention sets up an SCSA module in the backbone network, first extracts cross-semantic space information through multi-scale convolution in the SCSA module, and then injects spatial attention into the channel dimension through a progressive compression strategy in the PCSA module, so as to enhance the model's response to the subtle features of small targets, thereby achieving the effect of accurately capturing the outline of small targets in complex backgrounds, and realizing the synergistic effect of multi-semantic space attention and progressive channel attention.

[0045] 3. This invention, by setting a depth-separable convolution structure in the detection head, decomposes traditional convolution into a cascaded operation of depthwise convolution and pointwise convolution, which can reduce the number of model parameters, improve inference speed, and achieve high-speed real-time processing capabilities on embedded devices of UAVs, making it easier to meet the deployment requirements of harsh mobile environments.

[0046] 4. This invention, by setting up Mosaic data augmentation technology during the model training stage and adopting a four-image stitching and dynamic annotation adjustment strategy, can simulate the multi-scale target distribution under the top-down view of a UAV. Combined with the multi-scale feature generation unit and adaptive max pooling technology of the SAFM module, it can improve the model's recognition robustness in strong lighting and partial occlusion scenarios, and achieve stable detection results in complex environments.

[0047] It should be understood that the descriptions in this section are not intended to identify key or essential features of embodiments of the invention, nor are they intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Of course, implementing any product of the invention does not necessarily require achieving all of the advantages described above simultaneously. Attached Figure Description

[0048] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0049] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0050] Figure 2 This is a schematic diagram illustrating the composition and function of the SAFM module in the YOLOv11 model network framework of this invention;

[0051] Figure 3 This is a schematic diagram of the system structure of the system of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] For details in the embodiments, please refer to Figures 1 to 3 .

[0054] like Figure 1 As shown. The UAV small target recognition method and system proposed in this embodiment of the invention includes the following steps:

[0055] S1. Data Acquisition and Preprocessing: Image data was acquired using a drone via its onboard camera at a resolution of 1920×1080. The image data was then preprocessed, including denoising, normalization, and data augmentation.

[0056] S2. Construct the YOLOv11 model network framework. Introduce the SCSA and SAFM modules into the backbone network of the YOLOv11 model network framework and optimize the loss function. Combine with Mosaic data augmentation technology to enhance the model's ability to identify and extract features of small targets.

[0057] The improved YOLOv11 algorithm design includes:

[0058] (1) Based on the YOLOv11 algorithm, a spatial and channel collaborative attention module (SCSA) is introduced to enhance the model's ability to extract features of small targets. The SCSA module consists of two parts: a shareable multi-semantic spatial attention (SMSA) module and a progressive channel self-attention (PCSA) module, wherein:

[0059] The SMSA module uses multi-scale convolutions (such as 3×3, 5×5, 7×7, etc.) to extract spatial information at different semantic levels from the feature map. It accelerates model convergence through group normalization and fuses the extracted spatial information to generate spatial attention information.

[0060] The PCSA module uses a progressive compression strategy to inject the spatial attention information generated by the SMSA module into the channel self-attention;

[0061] By exploring the similarity between channels through a single-head self-attention mechanism, robust feature interaction is achieved, and finally channel attention weights are generated, which can enhance the feature extraction capability of small targets.

[0062] Specifically, by setting up an SCSA module in the backbone network, multi-scale convolution is first used in the SCSA module to extract cross-semantic space information, and then a progressive compression strategy is used in the PCSA module to inject spatial attention into the channel dimension, which enhances the model's response to subtle features of small targets, thereby achieving the effect of accurately capturing the outline of small targets in complex backgrounds.

[0063] The following table illustrates the improvement in mean accuracy (mAP) for typical small target recognition before and after introducing the SCSA module in the VisDrone dataset in a specific embodiment:

[0064] Target Category mAP without SCSA module mAP with SCSA module Increase (percentage) pedestrian 65.2％ 73.8％ +8.6％ bike 58.7％ 67.4％ +8.7％ car 70.5％ 78.9％ +8.4％ tricycle 52.3％ 60.7％ +8.4％ motorcycle 61.4％ 69.8％ +8.4％ Overall mAP 63.7％ 72.2％ +8.5％

[0065] The table above compares the average accuracy of the model in recognizing small targets before and after introducing the SCSA module.

[0066] (2) Replace the existing feature fusion module (such as SPPF or C2f) in the neck of YOLOv11 with the SAFM module to optimize the neck structure of the model and improve the model's ability to recognize small targets.

[0067] At this point, the SAFM module can effectively utilize non-local feature interactions through multi-scale feature representation and dynamic modulation mechanism, thereby improving the detection capability of small targets.

[0068] Furthermore, the SAFM module consists of a multi-scale feature generation unit (MFGU) and a feature aggregation and modulation module, and its structure is as follows: Figure 2 As shown:

[0069] MFGU contains multiple branches. One branch uses 3×3 depthwise convolution to process local features, while the other branches generate features at different scales through pooling operations and then restore them to the original resolution through upsampling.

[0070] The feature aggregation and modulation module concatenates features of different scales along the channel dimension, aggregates them through 1×1 convolution, normalizes the generated attention map using the GELU function, and then modulates the input features through element-wise multiplication.

[0071] Therefore, the SAFM module processing steps are as follows:

[0072] First, the normalized input features are segmented into four components. One component is processed using depthwise convolution, while the other three are fed into a multi-scale feature generation unit. These components undergo downsampling, convolution, and upsampling to generate multi-scale features. Next, adaptive max pooling is applied to the input features to generate more multi-scale features. These multi-scale features are then concatenated using convolution to obtain aggregated features. These aggregated features are then normalized using the GELU nonlinear function to estimate the attention map. Finally, the original input features are adaptively modulated based on the estimated attention.

[0073] Furthermore, the following table illustrates the improvement in mean accuracy (mAP) for typical small target recognition before and after introducing the SAFM module in the VisDrone dataset, in a specific embodiment:

[0074]

[0075]

[0076] The table above compares the average accuracy of the model in recognizing small targets before and after introducing the SCSA module.

[0077] (3) Optimize the loss function to improve the accuracy of small target detection.

[0078] Specifically, an improved Varifocal Loss function is used, which adjusts the weights of positive and negative samples to make the model pay more attention to the difficult-to-classify positive samples. The function is expressed as:

[0079]

[0080] Among them, P t This is the result of passing the predicted value `pred` through the sigmoid function, used for comparison with the ground truth value `true`. `α` and `γ` are hyperparameters that adjust the weights of negative samples. Hyperparameter `α` prevents excessive suppression of negative samples, while hyperparameter `γ` reduces the contribution of easily classified samples to the loss, thus improving the accuracy of small object detection. This formula adjusts the weights of positive and negative samples, making the model focus more on difficult-to-classify positive samples. Therefore, the use of the α-γ dual-parameter control mechanism enhances the model's ability to focus on difficult small objects, ultimately achieving a breakthrough in performance metrics on the small object subset of the COCO dataset.

[0081] Furthermore, in specific experiments on the VisDrone dataset, the introduction of the improved Varifocal Loss function significantly improved the model's detection accuracy for small targets. The following is a comparison of the specific data:

[0082]

[0083]

[0084] The table above compares the average accuracy values ​​for small target recognition after introducing the improved Varifocal Loss function.

[0085] (4) Since Mosaic data augmentation can enhance the diversity of training data, especially the frequency of small targets, this diversity helps the model learn more target features and background information, thereby improving the detection ability of small targets. Therefore, Mosaic data augmentation is used here to improve the generalization ability of the model. Specifically, Mosaic data augmentation includes:

[0086] First, four images are randomly selected from the dataset. Each image is randomly cropped and scaled to generate image blocks of different sizes. The cropped image blocks are then stitched together into a new image according to a certain layout (such as top left, top right, bottom left, and bottom right). Based on the position of the stitched image, the target bounding boxes in each image are adjusted to adapt to the new image position.

[0087] At this point, by setting up Mosaic data augmentation technology during the training phase and employing a four-image stitching and dynamic annotation adjustment strategy, it is possible to simulate the multi-scale target distribution under the drone's top-down view. Combining the multi-scale feature generation unit (MFGU) and adaptive max pooling technology of the SAFM module, the robustness of the model in strong lighting and partially occluded scenes is improved, achieving stable detection results in complex environments.

[0088] Furthermore, experiments on the VisDrone dataset showed that introducing Mosaic data augmentation significantly improved the model's detection accuracy for small targets. The following is a detailed data comparison:

[0089]

[0090] The table above compares the average accuracy of small target recognition after introducing Mosaic data augmentation technology.

[0091] (5) Lightweight design (depth separable convolution) is adopted to reduce the computational complexity of the model and improve the efficiency of recognition.

[0092] In the YOLOv11 detection head, depthwise separable convolutions are used instead of traditional convolutions to reduce the number of model parameters and computational cost, thereby improving recognition efficiency. Depthwise separable convolutions consist of two parts: depthwise convolutions and pointwise convolutions.

[0093] Depthwise convolution performs a convolution operation on each channel of the input tensor independently, without involving interactions between channels. Assume the input tensor has C... in Each channel uses an h×w convolution kernel for convolution, and the number of output channels is the same as the number of input channels;

[0094] Point convolution uses a 1×1 convolution kernel to perform channel mixing on the output of depthwise convolution, integrating information from different channels.

[0095] At this point, by setting a depth-separable convolution structure in the detection head, the traditional convolution is decomposed into a cascaded operation of depth convolution and point convolution. This design can reduce the number of model parameters, improve inference speed, and ultimately achieve high-speed real-time processing capabilities on UAV embedded devices, meeting the stringent computing power constraints of mobile deployment.

[0096] S3. Model Training and Optimization: First, the model is trained using the labeled COCO dataset, and then fine-tuned on a custom dataset. Then, the Adam optimizer is used for transfer learning to improve the model's generalization ability. At this time, the learning rate is set to 0.001 and the training cycle is 100 rounds. Finally, data augmentation techniques are used to improve the model's generalization ability.

[0097] S4. Real-time target recognition: Deploy the trained model onto the NVIDIA Jetson Xavier embedded device to process image data collected by the drone in real time, with an inference speed of 30 FPS, and finally output the target recognition results, including target category, location and confidence level.

[0098] S5. Result Output and Feedback: The recognition results are transmitted to the ground control station via the 4G / 5G module. The ground control station then generates control commands based on the recognition results and sends them to the UAV. The UAV can then automatically adjust its flight path or perform specific tasks.

[0099] On the other hand, such as Figure 3 As shown, the present invention also discloses a small target recognition system for unmanned aerial vehicles (UAVs), comprising hardware and software components, for executing any of the aforementioned small target recognition methods for UAVs, including:

[0100] The hardware components include:

[0101] (1) Unmanned aerial vehicle platform: including airborne camera, embedded processor (such as NVIDIA Jetson series), wireless communication module, etc.

[0102] (2) Ground control station: used to receive and process data transmitted by UAVs, and automatically adjust the flight path of UAVs or perform specific tasks based on the data results.

[0103] The software component includes:

[0104] (1) Data acquisition and preprocessing module.

[0105] (2) Improve the YOLOv11 algorithm model.

[0106] (3) Real-time reasoning and result output module.

[0107] The data acquisition and preprocessing module is connected to the airborne camera, the embedded processor has a YOLOv11 model network framework embedded in it, and the real-time inference and result output module is used to control data processing operations based on the YOLOv11 model network framework.

[0108] In summary, the "end-edge collaborative" processing architecture constructed in this application, based on the lightweight model of YOLOv11 in the embedded processor, combined with the ground station deployment decision module, 4G / 5G dual-mode transmission and adaptive command generation technology, can reduce the overall system response latency and ultimately achieve closed-loop optimization from image acquisition to flight control, significantly improving mission execution efficiency.

[0109] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0110] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0111] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the UAV small target recognition methods described above.

[0112] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above method.

[0113] This application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus.

[0114] Memory, used to store computer programs;

[0115] The processor, when executing the program stored in the memory, implements the above-mentioned UAV small target recognition method.

[0116] The communication bus mentioned in the above-mentioned electronic devices can be a standard bus for interconnecting peripheral components or an extended industrial standard structure bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0117] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0118] The memory may include random access memory or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0119] The processors mentioned above can be general-purpose processors, including central processing units, network processors, etc.; they can also be digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0120] It should also be noted that electronic devices include terminal devices, which can also be called terminals, user equipment, mobile stations, mobile terminals, etc. Terminal devices can be mobile phones, smart TVs, wearable devices, tablets, computers with wireless transceiver capabilities, virtual reality terminal devices, augmented reality terminal devices, wireless terminals in industrial control, wireless terminals in autonomous driving, wireless terminals in remote surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, and so on. The embodiments of this application do not limit the specific technologies or device forms used in the terminal devices.

[0121] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0123] Furthermore, it should be noted that if any directional indication (such as up, down, left, right, front, back, etc.) is involved in the embodiments of the present invention, the directional indication is only used to explain the relative positional relationship and movement of each component in a specific posture. If the specific posture changes, the directional indication will also change accordingly.

[0124] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, in the embodiments of this invention, "multiple" refers to two or more. Moreover, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

Claims

1. A method for identifying small targets using a drone, characterized in that, Includes the following steps: S1. The drone collects image data through its onboard camera and performs preprocessing operations, including noise reduction and normalization, on the image data; S2. Construct the YOLOv11 model network framework, introduce the SCSA and SAFM modules into the backbone network of the YOLOv11 model network framework and optimize the loss function, and combine Mosaic data augmentation technology to enhance the model's ability to identify and extract features of small targets. S3. Use the COCO dataset to pre-train the YOLOv11 model, and then use the Adam optimizer to learn and optimize it, thereby improving the model's generalization ability. S4. Deploy the trained model onto an embedded device to process the image data collected by the drone in real time and output the target recognition results; S5. Finally, the results are output and fed back to the drone; In step S2, the loss function is optimized by adjusting the weights of positive and negative samples, making the model pay more attention to the difficult-to-classify positive samples. The function is expressed as: in, It is the result of passing the predicted value `pred` through the sigmoid function, used for comparison with the true value `true`. and To adjust the hyperparameters for negative sample weights, the hyperparameters... Hyperparameters used to prevent over-suppression of negative samples This is used to reduce the contribution of easily classified samples to the loss, thereby improving the accuracy of small target detection; In step S2, the SCSA module consists of a shareable multi-semantic space attention SMSA module and a progressive channel self-attention PCSA module, wherein: The SMSA module is used to extract spatial information at different semantic levels from the feature map using multi-scale convolution, and accelerates model convergence through group normalization, and fuses the extracted spatial information to generate spatial attention information. The PCSA module is used to inject the spatial attention information generated by the SMSA module into the channel self-attention using a progressive compression strategy, and to generate channel attention weights to enhance the feature extraction capability of small targets.

2. The UAV small target identification method as described in claim 1, characterized in that, In step S2, the SAFM module consists of a multi-scale feature generation unit (MFGU) and a feature aggregation and modulation module, wherein: MFGU contains at least two branches, one of which uses 3×3 depthwise convolution to process local features, and the other branches generate features at different scales through pooling operations, and then restore them to the original resolution through upsampling; The feature aggregation and modulation module is used to concatenate features of different scales along the channel dimension, aggregate them through 1×1 convolution, normalize the generated attention map using the GELU function, and then modulate the input features by element-wise multiplication.

3. The UAV small target identification method as described in claim 1, characterized in that, The data processing flow of the SAFM module in step S2 includes: L1 performs channel segmentation on the normalized input features to generate four component parts. One part is processed by depthwise convolution, while the other three parts are fed into the multi-scale feature generation unit (MFGU) and undergo downsampling, convolution, and upsampling operations to generate multi-scale features. L2. Adaptive max pooling is applied to the input features to generate multi-scale features. These multi-scale features are then concatenated by convolution to obtain aggregated features. After that, the features are normalized by the GELU nonlinear function to estimate the attention map. L3. Adaptively modulate the original input features based on the attention map estimation data.

4. The UAV small target identification method as described in claim 1, characterized in that, The Mosaic data augmentation technique in step S2 specifically includes: Four images are randomly selected from the dataset, and each image is randomly cropped and scaled to generate image patches of different sizes; The cropped image blocks are stitched together according to a specified layout to form a new image; Adjust the target bounding boxes in each image according to the position of the stitched images to adapt them to the new image positions.

5. The UAV small target identification method as described in claim 1, characterized in that, In the YOLOv11 model network framework constructed in step S2, depthwise separable convolutions are used instead of traditional convolutions in the detection head to reduce the number of model parameters and computational cost. These depthwise separable convolutions consist of depthwise convolutions and pointwise convolutions, wherein: Depthwise convolution, on the input Each channel of the channel tensor uses a Each convolutional kernel performs convolution operations independently, and the number of output channels is the same as the number of input channels; Point convolution uses a 1×1 convolution kernel to perform channel mixing on the output of depthwise convolution to integrate information from different channels.

6. A small target recognition system for unmanned aerial vehicles (UAVs), used to execute the small target recognition method for UAVs according to any one of claims 1-5, characterized in that, include: The unmanned aerial vehicle (UAV) platform and ground control station are provided. The UAV platform is equipped with an airborne camera, an embedded processor, a wireless communication module, a data acquisition and preprocessing module, and a real-time inference and result output module. The data acquisition and preprocessing module is connected to the airborne camera. The embedded processor has a YOLOv11 model network framework embedded in it. The real-time inference and result output module is used to control data processing operations based on the YOLOv11 model network framework. The ground control station is used to receive and process data transmitted by the UAV, and based on the data results, automatically adjust the UAV's flight path or perform specific tasks.

7. A computer-readable storage medium, characterized in that, The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 5.

8. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 5.

Images