Model training method and air low-slow small unmanned aerial vehicle detection method

Abstract

The invention discloses a model training method and an air low-slow small unmanned aerial vehicle detection method, and belongs to the technical field of target detection. The method comprises the following steps: firstly, acquiring a sample unmanned aerial vehicle image, a sample unmanned aerial vehicle annotation image and a YOLOv8 improved model comprising a backbone network OfficientRep which has a RepVGG style structure and can be reparameterized, a hybrid fusion module CGAFusion based on content guidance attention and a WassersteinLoss loss function; and then, based on the sample unmanned aerial vehicle image and the sample unmanned aerial vehicle annotation image, training a YOLOv8 improved model to obtain a target unmanned aerial vehicle detection model. The method can accurately and efficiently detect the low-slow small unmanned aerial vehicle in the air.

Description

Technical Field

[0001] The present application relates to the field of target detection technology, and in particular to a model training method and a method for detecting small, low-altitude, and slow-moving drones in the air. Background Art

[0002] As the application scenarios of drones continue to expand and drone technology continues to develop and improve, while bringing convenience and fun to people's lives, drones also pose a certain threat to social security and national defense security. If they are not operated properly, they may threaten public security and the safety of public life and property, and even be used as a tool for illegal crimes. In recent years, there have been many civilian drones "flying randomly", "flying illegally", injuring people and disturbing the public, and even worse, drones are used to spy on privacy. Therefore, it is crucial to effectively identify and detect illegal drones flying in the air with low and slow characteristics to ensure public safety.

[0003] At present, the one-stage target detection network is superior to the two-stage target detection network in terms of detection speed, so it is widely used in fields such as industrial detection with high real-time requirements. In the one-stage detection network, the YOLO series of models have always attracted much attention. However, the existing YOLOv8 model is difficult to be effectively detected and discovered in practical applications when applied to complex environmental backgrounds, such as high-rise buildings in cities and trees in forests, because drones usually have the characteristics of low flying altitude, slow flying speed and small size.

[0004] Therefore, it is urgent to develop an improved model in YOLOv8 to make it more suitable for the detection of low, slow and small drones in the air and improve the detection speed and accuracy. Summary of the invention

[0005] In view of this, the present application provides a model training method and a method for detecting small drones flying low and slow in the air, which can accurately and efficiently detect small drones flying low and slow in the air.

[0006] Specifically, the following technical solutions are included:

[0007] In a first aspect, an embodiment of the present application provides a model training method, the method comprising:

[0008] Obtain sample drone images, sample drone annotated images, and a YOLOv8 improved model, wherein the YOLOv8 improved model includes a reparameterizable backbone network EfficientRep with a RepVGG style structure, a content-guided attention-based hybrid fusion module CGAFusion, and a WassersteinLoss loss function;

[0009] Based on the sample drone image and the sample drone annotated image, the YOLOv8 improved model is trained to obtain a target drone detection model.

[0010] In some embodiments, the training of the YOLOv8 improved model based on the sample drone image and the sample drone annotated image to obtain the target drone detection model includes:

[0011] Based on the sample drone image and the sample drone annotated image, training a YOLOv8 improved model to obtain an intermediate YOLOv8 improved model;

[0012] In response to the performance of the intermediate YOLOv8 improved model satisfying the training completion condition, the intermediate YOLOv8 improved model satisfying the training completion condition is used as the target UAV detection model; in response to the performance of the intermediate YOLOv8 improved model not satisfying the training completion condition, the training weights and training parameters of the intermediate YOLOv8 improved model are adjusted until the performance of the intermediate YOLOv8 improved model satisfies the training completion condition, and the training parameters include one or more parameters of the initial learning rate, the number of training iterations, and the hyperparameters.

[0013] In some embodiments, the YOLOv8 improved model is a model improved on the basis of the YOLOv8 model, and the improvement process includes:

[0014] Replacing the combination of the Conv module and the C2f module in the backbone network of the YOLOv8 model with the combination of the RepVGGBlock module and the RepBlock module to obtain the reparameterizable backbone network EfficientRep with the RepVGG style structure, and using the backbone network EfficientRep as the backbone network of the YOLOv8 improved model;

[0015] In the neck network of the YOLOv8 model, a layer of hybrid fusion module CGAFusion based on content-guided attention is added before outputting the features to the three detection heads;

[0016] Select WassersteinLoss loss function as the loss function of the YOLOv8 improved model.

[0017] In some embodiments, the training completion condition is that the accuracy, recall rate and mAP of the intermediate YOLOv8 improved model do not increase for 50 consecutive rounds.

[0018] In some embodiments, the reparameterizable backbone network EfficientRep having a RepVGG style structure includes a RepVGGBlock module and a RepBlock module;

[0019] The RepBlock module includes a plurality of cascaded RepVGGBlocks, and the RepVGGBlock module includes two 3×3 convolutional layers, a 1×1 convolutional layer and three BN units. In some embodiments, the hybrid fusion module CGAFusion based on content-guided attention includes a CGA unit and a 1×1 convolutional layer, and the hybrid fusion module CGAFusion based on content-guided attention is used to first weighted sum the output features of the 4th, 6th and 9th layers in the backbone network and the output features of the 15th, 18th and 21st layers in the neck network, respectively, and input them into the CGA unit to calculate the weights, and then combine them using the weighted summation method, and also add input features to the backbone network features and the neck through jump connections to alleviate the gradient vanishing problem and simplify the learning process. Finally, the fused features are projected by a 1×1 convolutional layer to obtain the final features. In some embodiments, the CGA unit includes two global average pooling layers, one global maximum pooling layer, two 1×1 convolutional layers, two 7×7 convolutional layers, one channel shuffle layer, and one sigmoid activation function layer.

[0020] In the CGA unit, the global context information of the feature map is generated through the global average pooling layer and the global maximum pooling layer. The global average pooling layer captures the average information in the feature map, and the global maximum pooling layer captures the most significant information of the feature map to obtain different feature representations.

[0021] In some embodiments, the WassersteinLoss loss function is to add the NWD loss to the regression loss function and adjust the ratio of IoU to NWD loss to 5:5.

[0022] In some embodiments, the initial training parameter settings of the YOLOv8 improved model include: setting the gradient descent algorithm to SGD, setting the initialization learning rate to 0.01, setting the momentum to 0.937, setting the weight decay to 0.0005, setting the size of the input image to be normalized to 640×640, setting the number of training rounds to 500, and setting the batch size to 16.

[0023] In a second aspect, an embodiment of the present application provides a method for detecting a low, slow, small drone in the air, the method comprising:

[0024] Obtaining a target drone detection model in the model training method as described in the first aspect;

[0025] The drone image to be tested is input into the target drone detection model to obtain a drone detection result, which includes a target category, a target bounding box position, a confidence level, and a category probability.

[0026] The beneficial effects of the technical solution provided by the embodiments of the present application include at least:

[0027] The embodiment of the present application provides a model training method and a method for detecting a small, low, slow, and aerial drone. The method includes first obtaining a sample drone image, a sample drone annotated image, and a YOLOv8 improved model including a reparameterizable backbone network EfficientRep with a RepVGG style structure, a hybrid fusion module CGAFusion based on content-guided attention, and a WassersteinLoss loss function. The reparameterizable backbone network EfficientRep has a RepVGG style structure, which is more friendly to hardware computing power, memory bandwidth, compilation optimization characteristics, network representation capabilities, etc. The consistency and efficiency of the calculation can be maintained through structural reparameterization. The hybrid fusion module CGAFusion based on content-guided attention can alleviate the gradient vanishing problem, simplify the learning process, and enhance the multi-layer feature fusion capability of the model. The WassersteinLoss loss function helps to improve the model's ability to detect tiny drones; then, based on the sample drone image and the sample drone annotated image, the YOLOv8 improved model is trained to obtain a target drone detection model. The method can accurately and efficiently detect small, low, slow, and aerial drones. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 A method flow chart of the model training method provided in an embodiment of the present application;

[0030] Figure 2 A schematic diagram of the structure of the YOLOv8 improved model in the model training method provided in the embodiment of the present application;

[0031] Figure 3 A schematic diagram of the structure of the RepVGGBlock module of the YOLOv8 improved model in the model training method provided in an embodiment of the present application;

[0032] Figure 4 A schematic diagram of the structure of the RepBlock module of the YOLOv8 improved model in the model training method provided in an embodiment of the present application;

[0033] Figure 5A schematic diagram of the structure of a content-guided attention hybrid fusion module CGAFusion of the YOLOv8 improved model in the model training method provided in an embodiment of the present application;

[0034] Figure 6 A schematic diagram of the structure of a CGA unit included in a content-guided attention hybrid fusion module CGAFusion of an improved YOLOv8 model in a model training method provided in an embodiment of the present application;

[0035] Figure 7 Schematic diagram of the effect of the method for detecting a small, low-flying, and slow-flying drone in the air provided in an embodiment of the present application. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0037] In order to make the technical solutions and advantages of the present application more clear, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0038] In a first aspect, the present application provides a model training method, such as Figure 1 As shown, the method includes:

[0039] Step 101, obtain sample drone images, sample drone annotated images and a YOLOv8 improved model, where the YOLOv8 improved model includes a reparameterizable backbone network EfficientRep with a RepVGG style structure, a content-guided attention-based hybrid fusion module CGAFusion, and a WassersteinLoss loss function.

[0040] The re-parameterizable backbone network EfficientRep has a RepVGG style structure, which is more friendly to hardware computing power, memory bandwidth, compilation optimization characteristics, network representation capabilities, etc. The consistency and efficiency of the calculation can be maintained through structural re-parameterization. The hybrid fusion module CGAFusion based on content-guided attention can alleviate the gradient vanishing problem, simplify the learning process, and enhance the model's multi-layer feature fusion capability. The WassersteinLoss loss function helps to improve the model's detection capability for tiny drones.

[0041] In some embodiments, the sample drone images can come from the aerial drone dataset TIB-Net collected by Sun Han's team at Nanjing University of Aeronautics and Astronautics. TIB-Net is uniformly obtained from aerial drone images taken by fixed cameras in real environments. The main scenes are teaching buildings inside the school, and the main interference targets of the dataset are complex background information.

[0042] In some embodiments, the data set is enhanced using the Mosaic data enhancement method.

[0043] In some embodiments, the data set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0044] Step 102: Based on the sample drone images and the sample drone annotated images, the YOLOv8 improved model is trained to obtain a target drone detection model.

[0045] In some embodiments, Figure 2 As shown in the figure, the YOLOv8 improved model is a model improved on the basis of the YOLOv8 model. The improvement process includes:

[0046] (1) The combination of Conv module and C2f module in the backbone network of YOLOv8 model is replaced by the combination of RepVGGBlock module and RepBlock module to obtain the reparameterizable backbone network EfficientRep with RepVGG style structure. The backbone network EfficientRep is used as the backbone network of YOLOv8 improved model.

[0047] This setting makes the improved YOLOv8 model more friendly to hardware computing power, memory bandwidth, compilation optimization features, network representation capabilities, etc., and the consistency and efficiency of calculations can be maintained through structural reparameterization.

[0048] (2) In the neck network of the YOLOv8 model, a layer of content-guided attention-based hybrid fusion module CGAFusion is added after the 15th layer, the 18th layer, and the 21st layer before outputting the features to the three detection heads.

[0049] The hybrid fusion module CGAFusion based on content-guided attention can alleviate the gradient vanishing problem, simplify the learning process, and enhance the model's multi-layer feature fusion capability.

[0050] (3) The WassersteinLoss loss function is selected as the loss function of the YOLOv8 improved model. The WassersteinLoss loss function helps to improve the model's ability to detect micro-UAVs.

[0051] In some embodiments, the training completion condition is that the accuracy, recall and mAP of the intermediate YOLOv8 improved model do not increase for 50 consecutive rounds.

[0052] In some embodiments, Figure 2 As shown in FIG, the reparameterizable backbone network EfficientRep with a RepVGG style structure includes a RepVGGBlock module and a RepBlock module; the RepBlock module includes multiple cascaded RepVGGBlocks, such as Figure 3 As shown, the RepVGGBlock module consists of two 3×3 convolutional layers, one 1×1 convolutional layer, and three BN units.

[0053] It should be noted that if Figure 3 As shown in the left part, during the training phase, a 3×3 convolutional layer, a 1×1 convolutional layer, and three BN layers perform multi-branch parallel calculations, thereby improving the feature extraction capability of the model during the training process. Figure 3 As shown in the right part, after training is completed, these branches can be merged into a single 3×3 convolution kernel through reparameterization, thereby simplifying the computational complexity of the inference phase.

[0054] like Figure 3 As shown in the figure, the RepVGGBlock module uses a structural reparameterization technique, which allows the use of complex multi-branch structures to capture more detailed features during the training phase and simplifies them to a single convolutional layer during the inference phase. This design not only maintains the high expressiveness of the model during training, but also greatly improves the detection speed and efficiency during inference. By using a more complex structure to enhance feature extraction capabilities during the training phase, the backbone network EfficientRep can better capture the detailed information of small targets, while the simplified structure during the inference phase ensures the needs of real-time detection.

[0055] like Figure 4As shown in the figure, the RepBlock module includes multiple cascaded RepVGGBlock modules. By cascading multiple RepConv modules, a deep convolutional network is formed to enhance the network's expressiveness while maintaining efficient computing power in the inference stage. The RepBlock module has a multi-level RepVGGBlock structure that can extract and fuse features of different scales more carefully, which is particularly critical for the detection of small targets. Small targets usually contain subtle and easily lost information. Through deeper and richer feature extraction, the backbone network EfficientRep can more accurately identify and locate small targets. Especially for small targets that appear in a large background, the multi-scale feature fusion design of the backbone network EfficientRep can effectively improve the detection accuracy of small targets. Compared with the YOLOv8 model, when processing small targets, the YOLOv8 model is often prone to missed detection or false detection due to the limitation of feature map resolution. The backbone network EfficientRep enhances the model's feature extraction ability during the training stage through a multi-branch convolutional structure, especially for those subtle and easily ignored features, which makes it perform well in small target detection. At the same time, through its efficient reasoning stage structure, it reduces the amount of calculation while maintaining high accuracy, thus supporting real-time small target detection, which is especially suitable for scenarios that require immediate response, such as monitoring and unmanned driving. In addition, the backbone network EfficientRep structure design also enables it to have stronger generalization ability and robustness, enhance the focus on effective information, reduce the interference of background noise, and still maintain high detection accuracy in complex backgrounds and diverse scenarios.

[0056] In some embodiments, Figure 5 As shown in the figure, the hybrid fusion module CGAFusion based on content-guided attention includes a CGA unit and a 1×1 convolution layer. The hybrid fusion module CGAFusion based on content-guided attention first performs weighted summation on the output features of the 4th, 6th and 9th layers in the backbone network and the output features of the 15th, 18th and 21st layers in the neck network and inputs them into the CGA unit to calculate the weights, and then combines them using the weighted summation method. It also adds input features to the backbone network features and the neck through jump connections to alleviate the gradient vanishing problem and simplify the learning process. Finally, the fused features are projected by a 1×1 convolution layer to obtain the final features.

[0057] The features of the 4th and 15th layers, the 6th and 18th layers, and the 9th and 21st layers in the structure of the YOLOv8 model backbone network of the improved backbone network EfficientRep are fused using CGA units. Because the small target information in the backbone network is relatively weak, and after the continuous convolution of the subsequent neck network, most of the small target information has been lost, the CGA unit is used to fuse the features of the backbone network and the neck network.

[0058] The hybrid fusion module CGAFusion calculates the spatial weights of feature modulation. The low-level features of the encoder part and the corresponding high-level features are input into CGA to calculate the weights, and then combined by weighted summation; then, the input features are increased through jump connections to alleviate the gradient vanishing problem and simplify the learning process; finally, the fused features are projected by a 1×1 convolutional layer to obtain the final features, and then the obtained features are passed to the detection head for detection.

[0059] In some embodiments, Figure 6 As shown in Figure 1, the CGA unit includes two global average pooling layers, one global maximum pooling layer, two 1×1 convolutional layers, two 7×7 convolutional layers, one channel shuffle layer, and one sigmoid activation function layer.

[0060] In the CGA unit, the global context information of the feature map is generated through the global average pooling layer and the global maximum pooling layer. The global average pooling layer captures the average information in the feature map, and the global maximum pooling layer captures the most significant information in the feature map to obtain different feature representations. The CGA unit uses global average pooling ( Figure 6 Global AveragePooling) and Global Maximum Pooling ( Figure 6 The Global Max Pooling in the feature map is used to generate the global context information of the feature map. Different feature representations can be obtained through these two operations. The global average pooling layer can capture the average information in the feature map, helping the network understand the overall feature distribution; while the global maximum pooling layer can capture the most significant information in the feature map, helping the network highlight the key parts.

[0061] In order to make up for the deficiency of IoU loss in small target detection, in some embodiments, the WassersteinLoss loss function adds the normalized Gaussian Wasserstein distance (NWD) loss to the regression loss function, and adjusts the ratio of IoU and NWD loss to 5:5, thereby improving the model's detection capability for tiny drones.

[0062] It should be noted that the WassersteinLoss loss function calculates the similarity between the real box and the predicted box through the Gaussian distribution corresponding to the real box and the predicted box. The introduced NWD metric can be easily embedded in the allocation, non-maximum suppression and loss function of any anchor-based detector to replace the commonly used IoU metric. Since the NWD loss is insensitive to the scale of the target, it is more suitable for measuring the similarity between small targets.

[0063] In some embodiments, based on the sample drone images and the sample drone annotated images, the YOLOv8 improved model is trained to obtain the target drone detection model, including:

[0064] (1) Based on the sample drone images and sample drone annotated images, the YOLOv8 improved model is trained to obtain the intermediate YOLOv8 improved model.

[0065] (2) In response to the performance of the intermediate YOLOv8 improved model satisfying the training completion condition, the intermediate YOLOv8 improved model satisfying the training completion condition is used as the target UAV detection model; in response to the performance of the intermediate YOLOv8 improved model not satisfying the training completion condition, the training weights and training parameters of the intermediate YOLOv8 improved model are adjusted until the performance of the intermediate YOLOv8 improved model satisfies the training completion condition, and the training parameters include one or more parameters of the initial learning rate, the number of training iterations, and the hyperparameters.

[0066] In some embodiments, the initial training parameter settings of the YOLOv8 improved model include: setting the gradient descent algorithm to SGD, setting the initialization learning rate to 0.01, which can achieve better results in balancing the training speed and convergence stability. Setting the momentum to 0.937 reduces shocks during training. Setting the weight decay to 0.0005 helps avoid overfitting, making the model more robust on new data. Setting the size of the input image to be normalized to 640×640, setting the number of training rounds to 500, and setting the batch size to 16 can make the model update more frequent and help the model find the optimal solution faster. Through a reasonable combination of learning rate, momentum, weight decay, and batch size, overfitting can be avoided and training efficiency can be improved while ensuring model performance.

[0067] It should be noted that precision refers to the ratio of the correct results predicted by the model to all the predicted results. Recall refers to the ratio of the correct results predicted by the model to all the correct results. mAP is the average precision.

[0068] As shown in Table 1, the accuracy of the target drone detection model increased by 3.2%, the recall rate increased by 1.4%, and the mAP increased by 2.7%. All three important performance indicators have been improved, indicating that the target drone detection model obtained by the model training method provided in the embodiment of the present application has indeed improved the model's detection ability for micro drones.

[0069] Table 1 Performance comparison of target drone detection model and YOLOv8 model

[0070]

[0071] In summary, the model training method provided in the embodiment of the present application is more friendly to hardware computing power, memory bandwidth, compilation optimization characteristics, network characterization capabilities, etc., has strong computing consistency and high efficiency, can enhance the model's multi-layer feature fusion capabilities, and helps to improve the model's detection capabilities for tiny drones, thereby accurately and efficiently detecting low and slow small drones in the air.

[0072] In a second aspect, an embodiment of the present application provides a method for detecting a low, slow, small drone in the air, the method comprising:

[0073] Obtaining a target drone detection model in the model training method of the first aspect;

[0074] The drone image to be tested is input into the target drone detection model to obtain the drone detection result, which includes the target category, target bounding box position, confidence and category probability.

[0075] The drone detection results are as follows Figure 7 As shown, a rectangular frame is used to mark the drone detection image ( Figure 7 Confidence level not indicated).

[0076] In summary, the method for detecting low, slow and small drones in the air provided by the embodiment of the present application is more friendly to hardware computing power, memory bandwidth, compilation optimization characteristics, network characterization capabilities, etc., has strong computing consistency and high efficiency, can enhance the model's multi-layer feature fusion capabilities, and help improve the model's detection capabilities for tiny drones, thereby accurately and efficiently detecting low, slow and small drones in the air.

[0077] In the present application, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The term "plurality" refers to two or more than two, unless otherwise clearly defined.

[0078] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the present application disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only.

[0079] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A model training method, characterized in that: The method comprises: Obtain sample drone images, sample drone annotated images, and a YOLOv8 improved model, wherein the YOLOv8 improved model includes a reparameterizable backbone network EfficientRep with a RepVGG style structure, a content-guided attention-based hybrid fusion module CGAFusion, and a WassersteinLoss loss function; Based on the sample drone image and the sample drone annotated image, the YOLOv8 improved model is trained to obtain a target drone detection model.

2. The model training method according to claim 1, characterized in that: The step of training the YOLOv8 improved model based on the sample drone image and the sample drone annotated image to obtain a target drone detection model includes: Based on the sample drone image and the sample drone annotated image, training a YOLOv8 improved model to obtain an intermediate YOLOv8 improved model; In response to the performance of the intermediate YOLOv8 improved model satisfying the training completion condition, the intermediate YOLOv8 improved model satisfying the training completion condition is used as the target UAV detection model; in response to the performance of the intermediate YOLOv8 improved model not satisfying the training completion condition, the training weights and training parameters of the intermediate YOLOv8 improved model are adjusted until the performance of the intermediate YOLOv8 improved model satisfies the training completion condition, and the training parameters include one or more parameters of the initial learning rate, the number of training iterations, and the hyperparameters.

3. The model training method according to claim 1, characterized in that: The YOLOv8 improved model is a model improved on the basis of the YOLOv8 model, and the improvement process includes: Replacing the combination of the Conv module and the C2f module in the backbone network of the YOLOv8 model with the combination of the RepVGGBlock module and the RepBlock module to obtain the reparameterizable backbone network EfficientRep with the RepVGG style structure, and using the backbone network EfficientRep as the backbone network of the YOLOv8 improved model; In the neck network of the YOLOv8 model, a layer of content-guided attention-based hybrid fusion module CGAFusion is added before outputting the features to the three detection heads; Select WassersteinLoss loss function as the loss function of the YOLOv8 improved model.

4. The model training method according to claim 2, characterized in that: The training completion condition is that the accuracy, recall rate and mAP of the intermediate YOLOv8 improved model do not increase for 50 consecutive rounds.

5. The model training method according to any one of claims 1 to 4, characterized in that: The reparameterizable backbone network EfficientRep with RepVGG style structure includes RepVGGBlock module and RepBlock module; The RepBlock module includes multiple cascaded RepVGGBlocks, and the RepVGGBlock module includes two 3×3 convolutional layers, one 1×1 convolutional layer and three BN units.

6. The model training method according to any one of claims 1 to 4, characterized in that: The hybrid fusion module CGAFusion based on content-guided attention includes a CGA unit and a 1×1 convolutional layer. The hybrid fusion module CGAFusion based on content-guided attention is used to first perform weighted summation on the output features of the 4th, 6th and 9th layers in the backbone network and the output features of the 15th, 18th and 21st layers in the neck network, respectively, and input them into the CGA unit to calculate the weights, and then combine them using the weighted summation method. Input features are also added to the backbone network features and the neck through jump connections to alleviate the gradient vanishing problem and simplify the learning process. Finally, the fused features are projected by a 1×1 convolutional layer to obtain the final features.

7. The model training method according to claim 6, characterized in that: The CGA unit includes two global average pooling layers, one global maximum pooling layer, two 1×1 convolutional layers, two 7×7 convolutional layers, one channel shuffle layer, and one sigmoid activation function layer; In the CGA unit, the global context information of the feature map is generated through the global average pooling layer and the global maximum pooling layer. The global average pooling layer captures the average information in the feature map, and the global maximum pooling layer captures the most significant information of the feature map to obtain different feature representations.

8. The model training method according to any one of claims 1 to 4, characterized in that: The WassersteinLoss loss function adds NWD loss to the regression loss function and adjusts the ratio of IoU to NWD loss to 5:

5.

9. The model training method according to claim 2, characterized in that: The initial training parameter settings of the YOLOv8 improved model include: setting the gradient descent algorithm to SGD, setting the initialization learning rate to 0.01, setting the momentum to 0.937, setting the weight decay to 0.0005, setting the size of the input image to be normalized to 640×640, setting the number of training rounds to 500, and setting the batch size to 16.

10. A method for detecting a small drone flying low and slowly in the air, characterized in that: The method comprises: Obtaining a target drone detection model in the model training method as described in any one of claims 1 to 9; The drone image to be tested is input into the target drone detection model to obtain a drone detection result, which includes a target category, a target bounding box position, a confidence level, and a category probability.

Images