Lightweight unmanned aerial vehicle image target detection method and system

Abstract

The present disclosure provides a lightweight unmanned aerial vehicle image target detection method and system, relating to the technical field of image detection, comprising: obtaining target information labeled images in real-time video frames of the unmanned aerial vehicle; inputting the labeled images into an unmanned aerial vehicle image target detection model to output target information in real-time video frames; wherein in the unmanned aerial vehicle image target detection model, a C2fCA reconstruction backbone network is used to extract target feature maps of different scales in channels and spaces, the target feature maps of different scales are input into an enhanced multi-scale feature fusion network, feature information is introduced into a PAN structure for top-down transmission, target feature maps of different scales are fused, spatial features and semantic features are obtained, the spatial features and the semantic features are cross-feature fused to obtain fusion features, and the fusion features are used to output target information in real-time video frames. The present disclosure improves target detection accuracy and greatly reduces model size and memory consumption.

Description

Technical Field

[0001] This disclosure relates to the field of image detection technology, specifically to a lightweight UAV image target detection method and system. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] As a novel remote sensing platform, unmanned aerial vehicles (UAVs) have been widely applied in fields such as smart agriculture, environmental monitoring, and traffic management. Object detection in aerial imagery is a key component of these applications. UAV imagery offers a wider field of view and less spatial constraint, enabling the capture of more useful information. Combined with ground-based vehicle target detection systems, UAVs form an integrated air-ground intelligent transportation system, offering significant advantages in observing and locating vehicles on target road sections.

[0004] Drone images are typical small target images, with most targets occupying relatively small and densely packed pixels, making it more difficult to extract effective features. The detection accuracy of existing detection algorithms still needs to be improved, and the algorithms are not lightweight enough, making them unfriendly to some resource-constrained edge devices, which makes it difficult for the algorithms to achieve satisfactory performance on drone platforms. Summary of the Invention

[0005] To address the aforementioned issues, this disclosure proposes a lightweight UAV image target detection method and system. By constructing a lightweight UAV image target detection model, a network is reconstructed using a backbone adaptive structure including C2fCA, and a multi-scale deep feature fusion network, EMS-PAFPN (Enhanced Multi-scale Path-Aggregation Feature Pyramid Network), to fully extract target feature information and fuse multi-scale target feature information, thereby achieving higher detection accuracy for UAV image targets.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions:

[0007] Acquire real-time video frames from the drone and perform preprocessing to obtain target information labeled images;

[0008] The labeled image is input into the UAV image target detection model, which outputs target information in real-time video frames;

[0009] In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in the channel and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fused to obtain fused features. The target information in the real-time video frame is output using the fused features.

[0010] According to some embodiments, the present disclosure adopts the following technical solutions:

[0011] A lightweight UAV image target detection system includes:

[0012] The data acquisition module is used to acquire real-time video frames from the UAV and perform preprocessing to obtain target information labeled images;

[0013] The target detection module is used to input labeled images into the UAV image target detection model and output target information in real-time video frames;

[0014] In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in the channel and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fused to obtain fused features. The target information in the real-time video frame is output using the fused features.

[0015] According to some embodiments, the present disclosure adopts the following technical solutions:

[0016] A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned lightweight UAV image target detection method.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the aforementioned lightweight UAV image target detection method.

[0019] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0020] This disclosure presents a lightweight image target detection method. By constructing a lightweight UAV image target detection model, it utilizes a backbone adaptive structure reconstruction network including C2fCA and a multi-scale deep feature fusion network based on EMS-PAFPN (Enhanced Multi-scale Path-Aggregation Feature Pyramid Network). This achieves higher detection accuracy, lower computational cost and parameter count, smaller model size and memory consumption for UAV image target detection. It solves the problems of high model complexity, poor algorithm performance, and high memory consumption in existing UAV image target detection algorithms, realizing lightweight, high-precision UAV image target detection with wider applicability.

[0021] This disclosure presents a lightweight image target detection method that introduces Focal EIoU loss at the output layer. FocalEIoU is an improved EIoU method that incorporates focus and embedding to further enhance target detection accuracy. The Focal EIoU loss employs a learnable embedding function and weights the EIoU calculation by introducing focus between overlapping and non-overlapping regions. This weighting allows Focal EIoU to better focus on hard-to-detect target boxes, thereby improving the quality of the detection results. Attached Figure Description

[0022] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0023] Figure 1 A flowchart illustrating a lightweight UAV image target detection method according to an embodiment of this disclosure;

[0024] Figure 2 This is a structural diagram of the improved RE-YOLOv8 in the lightweight UAV image target detection method in the embodiments of this disclosure;

[0025] Figures 3(a) and 3(b) are respectively the lightweight reconstruction network of the backbone feature extraction network and the C2fCA module structure diagram of the embodiment of this disclosure;

[0026] Figure 4 This is a network structure diagram of the Enhanced Multi-scale Path-Aggregation Feature Pyramid Network (EMS-PAFPN) according to an embodiment of this disclosure. Detailed Implementation

[0027] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0029] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0030] Example 1

[0031] One embodiment of this disclosure provides a lightweight UAV image target detection method. By constructing a lightweight UAV image target detection model, namely, reconstructing the network with a backbone adaptive structure including C2fCA and using EMS-PAFPN (Enhanced Multi-scale Path-Aggregation Feature Pyramid Network) as a multi-scale deep feature fusion network, higher detection accuracy of UAV image targets is achieved. The method includes:

[0032] Acquire real-time video frames from the drone and obtain target information labeled images.

[0033] The labeled image is input into the UAV image target detection model, which outputs target information in real-time video frames;

[0034] In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in the channel and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fused to obtain fused features. The target information in the real-time video frame is output using the fused features.

[0035] As one embodiment, the specific implementation process of the lightweight UAV image target detection method disclosed herein includes:

[0036] Step 1: Build and train a drone image target detection model, including:

[0037] 1. Acquire a large amount of drone-captured video, extract images containing pedestrians and vehicles frame by frame, and perform preprocessing.

[0038] 2. Label the pedestrian and vehicle targets in the extracted images and divide them into training, validation and test sets according to a certain ratio;

[0039] 3. Train the network model using the constructed dataset, and adjust the hyperparameters to obtain a fully converged model; 4. Deploy it into the drone's onboard computer to capture video frames in real time for target detection.

[0040] By using drones to capture video, images containing pedestrians and vehicles are extracted frame by frame. Image processing techniques are then used to enhance image clarity and improve image reading speed. The image sizes are normalized to a uniform size; in this embodiment, the drone image size is normalized to 640×640 for easier subsequent processing.

[0041] The extracted images of pedestrians and vehicles are labeled and divided into training, validation and test sets according to a certain ratio.

[0042] Specifically, a labeling tool is used to label targets such as pedestrians or vehicles in the UAV images. In this embodiment, ten target categories are selected for labeling: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motorcycle. For all labels of pedestrians or vehicles, five parameters are generated to determine their position and category in the image: category parameter, target center coordinates (x, y), and target width and height (w, h). The label files of the UAV image data are stored in txt format. Several UAV images and their corresponding labels are divided into an 8:1:1 ratio as training, validation, and test sample sets, respectively. These sets are used to train a pre-defined lightweight UAV image detection model, and the hyperparameters are adjusted to obtain a fully converged model.

[0043] Furthermore, the construction process of the lightweight UAV image detection model specifically includes:

[0044] The traditional YOLO v8 network structure includes a backbone feature extraction network and a multi-scale feature fusion network. This disclosure improves the traditional YOLO v8 network structure to obtain a lightweight UAV image detection model (RE-YOLO v8), such as... Figure 2 As shown:

[0045] 1) Construct a backbone network for reconstructing images that conforms to the characteristics of UAV images;

[0046] In the traditional YOLO v8 network structure, the backbone network possesses strong feature extraction capabilities. It mainly consists of four stages, employing downsampling connections with a kernel size of 3 and a stride of 2 to gradually reduce image resolution and increase the receptive field. Each stage contains n C2f modules, where n depends on the network depth parameter. C2f integrates CSP and residual ideas, enabling the preservation of more gradient information.

[0047] In the embodiments of this disclosure, the C2f and backbone network structures are improved. First, an attention mechanism CA is introduced to form a module with the C2f module. In C2f, after Bottleneck extracts multi-gradient information and performs preliminary fusion, the attention mechanism is applied to enhance the representation of important features and suppress interference.

[0048] Given that the original network structure underutilizes shallow, high-resolution feature maps, this disclosure reconstructs the backbone network structure. This primarily involves adding a basic convolutional block before layer 0 of the original structure, adding a C2fCA layer after layer 1, changing the channel count of the first four layers, and removing the deep, low-resolution backbone structure. In UAV images, most targets can only have sufficient features extracted from high-resolution feature maps. The reconstructed backbone structure is better suited to the characteristics of UAV images, thus enhancing the algorithm's performance.

[0049] The UAV image labeled with pedestrians or vehicles is input into the reconstructed backbone feature extraction network. In this embodiment, the reconstructed backbone network has fewer parameters but can better meet the needs of UAV image feature extraction.

[0050] As shown in Figure 3(a), C2fCA integrates the attention mechanism CA on the basis of C2f, enhancing feature representation in terms of channels and space. After the initial concatenation of internal Bottleneck multi-gradient features, all feature maps are input into CA, and then adjusted to an appropriate number of channels through a basic convolution module.

[0051] As shown in Figure 3(b), the reconstructed backbone network consists of basic convolutional modules, C2fCA, and SPPF. To enhance the feature extraction capability of shallow layers, a basic convolutional layer (3,1,1) and a C2fCA layer are added before layer 0 and after layer 1 of the original network, respectively, and other C2f layers are replaced with C2fCA. After comprehensive consideration of the characteristics of UAV images, redundant deep low-resolution (20×20) backbone structures are removed. To balance the requirements of model lightweighting, the number of channels in each layer of the backbone network is redesigned. The reconstructed network maintains the same number of module layers, significantly reducing the number of backbone structure parameters. By changing the size of the feature maps through the above backbone feature extraction network, target feature maps of different scales are obtained.

[0052] 2) Construct an enhanced multi-scale deep feature fusion network (EMS-PAFPN) and introduce a multi-scale detection head with Focal EIoU;

[0053] In the traditional YOLO v8 network architecture, to maximize the efficiency of feature information utilization, the YOLO v8 Neck employs an FPN+PAN network structure. This connects feature maps from different stages of the backbone layers with the feature maps in the Neck, enabling the network to detect objects at different scales of feature maps more efficiently. The Neck part primarily aggregates features extracted from different stages by the backbone, achieving cross-layer feature fusion. In the early stages of the backbone structure, there are high-resolution feature maps, but their semantic information is weak. However, in later stages, the feature maps are semantically rich, but their spatial information is low. YOLO v8 uses a multi-scale architecture to construct the Neck structure, fusing cross-layer feature information to obtain feature maps with richer semantic and spatial information.

[0054] In embodiments of this disclosure, in order to further improve the effect of feature extraction and fusion and reduce the computational load, such as... Figure 4 As shown, an enhanced multi-scale path-aggregation feature network (EMS-PAFPN) was designed based on the FPN and PAN network structures.

[0055] First, based on YOLO v8-p2, the detection head for large targets was removed to adapt to the backbone network design. Second, to reduce feature loss and degradation during network propagation, this invention introduces feature information from the backbone structure into the bottom-up PAN structure to compensate for feature loss and degradation. Specifically, feature maps from the p1 and p3 feature layers in the backbone are fused at the p2 feature layer to obtain more comprehensive spatial and semantic information. The specific operation involves downsampling the p1 feature layer output and upsampling the p3 feature layer output, then concatenating them at the p2 feature layer of the PAN, followed by C2f feature extraction before proceeding to the next structure. At the p3 feature layer, fusion with the p4 feature layer in the backbone is performed only because the propagation after the upper-level fusion is not too long, and the spatial information of small targets is not significantly lost. The p4 feature layer of the PAN fuses the p3 feature layer in the backbone, supplementing the lost shallow and rich spatial information and deep and rich semantic information.

[0056] Finally, after the PAN structure is completed and the three scale feature maps are output before being fed into the Head, cross-feature fusion is performed again to fully integrate the rich semantic features from the deep layer and the rich spatial features from the shallow layer. This allows for further fusion of the rich spatial information from the shallow layer and the rich semantic information from the deep layer. After such thorough feature extraction and fusion, the loss and degradation of features during the transmission process in the neural network are suppressed, and the model extracts more fully integrated feature information.

[0057] 3) The constructed backbone network is fused with the multi-scale deep feature fusion network, and the network is trained using the training sample set to obtain a fully converged network model.

[0058] After constructing the reconstructed backbone feature extraction network and EMS-PAFPN, the backbone network is connected to the EMS-PAFPN. Specifically, the EMS-PAFPN uses the feature maps output from the backbone network P2-4 as input, with output feature map sizes of 40×40, 80×80, and 160×160. After passing through the EMS-PAFPN, three detection feature maps are output: a 40×40 feature map for detecting large targets, an 80×80 feature map for detecting medium-sized targets, and a 160×160 feature map for detecting small targets.

[0059] The above-mentioned pre-set UAV image target detection model is trained based on the training sample set. Under the premise of ensuring that there is no overfitting or underfitting, the model is trained fully until the loss function fully converges, the optimal weight file is retained, and the model training is completed.

[0060] Step 2: Capture real-time video frames from the UAV and obtain target information labeled images. Input the target information labeled images into the UAV image target detection model and output the target information in the real-time video frames.

[0061] In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in channels and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fused to obtain fused features. The fused features are used to output the target information in the real-time video frame.

[0062] The UAV image target detection model described in this embodiment consists of three parts: an input layer, an intermediate layer, and an output layer. Images from the training sample set are input into the input layer of the detection model. The intermediate layer includes a backbone network and an EMS-PAFPN. The backbone network has been reconstructed and a CA attention mechanism has been incorporated. The EMS-PAFPN includes enhanced multi-scale feature fusion methods from top to bottom and from bottom to top. Feature extraction and fusion are performed through the intermediate layer. Finally, Focal EIoU loss is introduced in the output layer to output three more accurate feature maps at different scales, which are used to determine the probability of whether the input training image contains pedestrian or vehicle targets.

[0063] Finally, video frames captured by the drone are acquired in real time, and the trained drone image target detection model is used for detection. The final output is the target information such as pedestrians and vehicles present in the video frames, along with the corresponding probability of the targets.

[0064] YOLO v8 uses CIoU to calculate the loss between predicted and ground truth bounding boxes, but its calculation scheme needs improvement. Focal EIoU is an improved method based on EIoU, introducing focal and embedding to further enhance object detection accuracy. The Focal EIoU loss employs a learnable embedding function and weights the EIoU calculation by introducing focal points between overlapping and non-overlapping regions. This weighting allows Focal EIoU to better focus on hard-to-detect bounding boxes, thus improving the quality of detection results. The formula for calculating the Focal EIoU loss is shown below:

[0065] (1)

[0066] (2)

[0067] in, bp and b gt These represent the coordinates of the center points of the predicted bounding box and the target bounding box, respectively. d (b p ，b gt ） This represents the Euclidean distance between two parameters. （w c ） 2 and （h c ） 2 This represents the width and height of the smallest outer rectangle of the predicted bounding box and the target bounding box. w p , w gt , h p and h gt This represents the width and height of the ground truth bounding box and the predicted bounding box.

[0068] IoU=|A∩B| / |A∪B|, where γ is a parameter that controls the degree of outlier suppression.

[0069] Furthermore, the labeled image is input into the backbone feature extraction network, which outputs three feature maps P2, P3, and P4 at different scales. These four feature maps are then input into EMS-PAFPN to obtain feature maps of different scales: 40×40, 80×80, and 160×160. The most suitable prediction boxes are selected through non-maximum suppression (NMS) and Focal EIoU. EMS-PAFPN is a feature fusion concept, rather than a fixed-module network structure. The specific input or fusion method of the four images in this example can be seen in Figure 3.

[0070] Method Validation

[0071] To fully demonstrate the effectiveness of the detection method described in this embodiment, a portion of the YOLO v8-p2 algorithm was trained under the same experimental environment. The corresponding mAP@0.5 (the IoU threshold is greater than 0.5; the higher the value, the larger the threshold and the lower the accuracy), mAP@0.5:0.95 (the average mAP across multiple IoU thresholds from 0.5 to 0.95 with a step size of 0.05), computational cost, and parameter count (lower parameter count and computational cost indicate simpler computation and lower computational cost) were obtained to prove the effectiveness of the proposed improvement scheme. Table 1 shows the specific data.

[0072] Table 1 Data

[0073]

[0074] As shown in Table 1, compared with the YOLO v8n-p2 model, the proposed RE-YOLO v8n model improves accuracy (P) by 4.9%, recall (R) by 7%, mAP@0.5 by 7.4%, and mAP@0.5-0.95 by 5.3%, while reducing the number of parameters by 50%. The proposed model significantly improves algorithm performance and substantially reduces the number of model parameters while sacrificing some computational resources. Compared with the YOLOv8s-p2 model, performance metrics only show slight improvements, but the number of parameters decreases by 86.6% and computational cost decreases by 49.1%. The proposed model significantly reduces the number of model parameters and computational cost while maintaining slight improvements in performance metrics, making it more suitable for devices with limited computing resources.

[0075] To demonstrate the superior overall performance of the UAV image target detection algorithm (RE-YOLO v8) described in this embodiment, experiments verified its advantages in detection effect and computational cost compared with the YOLOv5 series, YOLO v7, YOLO v8 series and other improved algorithms by researchers. The comparison results of the UAV image target detection algorithm (RE-YOLOv8) in this embodiment with other algorithms are shown in Table 2.

[0076] Table 2 Comparison results of various algorithms

[0077]

[0078] YOLO v5(U) represents the latest version in the Ultralytics library, which incorporates some advanced ideas from YOLO v8. As shown in Table 2, considering the computational complexity and performance metrics of the algorithms, the proposed RE-YOLO v8 model has the best overall performance. Compared with other algorithms, the RE-YOLOv8n model shows significant improvements in various performance metrics, while having lower computational cost and parameter count. The deeper and wider RE-YOLO v8s models have significantly higher performance metrics than other models, and their computational cost is also lower than models with similar performance. This experiment demonstrates that the UAV image target detection algorithm proposed in this embodiment has better technical effects, not only being very lightweight in terms of computational cost but also achieving higher detection accuracy.

[0079] In summary, this embodiment provides a lightweight UAV image target detection method that solves the problems of high model complexity, poor algorithm performance, and high memory consumption in existing UAV image target detection algorithms. It achieves lightweight, high-precision UAV image target detection and has wider applicability.

[0080] Example 2

[0081] One embodiment of this disclosure provides a lightweight UAV image target detection system, comprising:

[0082] The data acquisition module is used to acquire real-time video frames from the UAV and to obtain target information labeled images.

[0083] The target detection module is used to input labeled images into the UAV image target detection model and output target information in real-time video frames;

[0084] In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in the channel and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fused to obtain fused features. The target information in the real-time video frame is output using the fused features.

[0085] Example 3

[0086] One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the aforementioned lightweight UAV image target detection method.

[0087] Example 4

[0088] One embodiment of this disclosure provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the lightweight UAV image target detection method.

[0089] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A lightweight UAV image target detection method, characterized in that, include: Acquire real-time video frames from the drone, and obtain target information labeled images after preprocessing; The labeled image is input into the UAV image target detection model, which outputs target information in real-time video frames; In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in the channel and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fusion to obtain fused features. The fused features are used to output the target information in the real-time video frame. The UAV image target detection model is an improvement on the YOLO v8 network structure. The backbone network is reconstructed, and the reconstructed backbone network consists of basic convolutional modules, C2fCA modules, and SPPF. This includes: introducing the attention mechanism CA and combining it with the C2f modules to form the C2fCA module; adding a basic convolutional block before layer 0 of the original structure and adding a C2fCA module after layer 1; changing the number of channels in the first 4 layers; and deleting the deep, low-resolution backbone structure to obtain the C2fCA reconstructed backbone network. Focal EIoU loss is introduced into the output layer for training. The formula for calculating Focal EIoU loss is as follows: (1) (2) in, b p and b gt These represent the coordinates of the center points of the predicted bounding box and the target bounding box, respectively. d ( b p ，b gt ) represents the Euclidean distance between two parameters; w c ) 2 and( h c ) 2 Represents the width and height of the smallest outer rectangle of the predicted bounding box and the target bounding box; w p , w gt , h p and h gt Represents the width and height of the ground truth bounding box and the predicted bounding box; IoU=|A∩B| / |A∪B|, where γ is a parameter that controls the degree of outlier suppression.

2. The lightweight UAV image target detection method as described in claim 1, characterized in that, The preprocessing method is as follows: the real-time video frames of the acquired UAV are extracted to obtain an image containing target information. The size is normalized, and then the target in the normalized image is labeled using a labeling tool to obtain a target information labeled image.

3. The lightweight UAV image target detection method as described in claim 1, characterized in that, The enhanced multi-scale feature fusion network is based on the FPN and PAN network structure, using the YOLO v8 network. It removes the detection head for large targets, reconstructs the feature information in the backbone structure, and introduces it into the bottom-up PAN structure to compensate for feature loss and degradation. This includes fusing the feature maps from the p1 and p3 feature layers in the backbone at the p2 feature layer. That is, the output of the p1 feature layer is downsampled, and the output of the p3 feature layer is upsampled and then concatenated at the p2 feature layer of the PAN. After C2f feature extraction, it enters the next structure.

4. The lightweight UAV image target detection method as described in claim 1, characterized in that, After the PAN structure is completed, the output feature maps at multiple scales are fused together again before being fed into the Head, and then further fused to obtain multiple fused feature maps at different scales.

5. The lightweight UAV image target detection method as described in claim 1, characterized in that, Focal EIoU is used for prediction loss. An improved EIoU-based method is introduced to further improve the accuracy of object detection by incorporating focus and embedding. A learnable embedding function is used, and EIoU is calculated by weighting the focus between overlapping and non-overlapping regions. This allows Focal EIoU to better focus on target boxes that are difficult to detect. Feature maps of multiple different scales are fused, and the most suitable prediction boxes are selected by non-maximum suppression and Focal EIoU.

6. The lightweight UAV image target detection method as described in claim 5, characterized in that, Focal EIoU loss is introduced into the output layer to output multiple feature maps of different scales, which are used to determine the probability that the input labeled image contains target information.

7. A lightweight unmanned aerial vehicle (UAV) image target detection system, characterized in that, Specifically, the lightweight UAV image target detection method described in any one of claims 1-6 includes: The data acquisition module is used to acquire real-time video frames from the UAV and obtain target information labeled images after preprocessing. The target detection module is used to input labeled images into the UAV image target detection model and output target information in real-time video frames; In the UAV image target detection model, the C2fCA reconstruction backbone network is used to extract target feature maps of different scales in the channel and space. The target feature maps of different scales are input into the enhanced multi-scale feature fusion network, so that the feature information is introduced into the bottom-up PAN structure. The target feature maps of different scales are fused to obtain spatial features and semantic features. Then, the spatial features and semantic features are cross-feature fused to obtain fused features. The target information in the real-time video frame is output using the fused features.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a lightweight UAV image target detection method as described in any one of claims 1-6.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a lightweight UAV image target detection method as described in any one of claims 1-6.

Images