Small target detection method in unmanned aerial vehicle application scene

Abstract

The invention belongs to the technical field of target detection, and particularly relates to a small target detection method in an unmanned aerial vehicle application scene. On the basis of a YOLOv8 network, firstly, the characteristics of a PConv structure are utilized, a C2f-PConv module and a PConvModule module are introduced, a network feature extraction module is lightened, learning of the network on redundant features is reduced, and the calculation overhead of the model is reduced; secondly, an EMMB module is designed, a lightweight attention mechanism is integrated as a feature fusion module of the model, the fusion ability of shallow and deep features is enhanced, and the small target detection performance of the model is improved; finally, a PA-AFPN feature fusion module is designed, features of each layer can be gradually and effectively combined, and multi-scale features are effectively utilized; according to the method, the problems of small image targets and dense distribution of unmanned aerial vehicle visual angle target detection are effectively solved, and the model is lightened under the condition of ensuring that the recognition accuracy is superior to that of similar models.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a method for detecting small targets in an image under an unmanned aerial vehicle application scenario. Background Art

[0002] With the rapid development of drone technology, its advantages of wide field of view and high flight speed have gradually driven its application in target detection. Compared with ground surveillance images, drone perspective images provide a unique bird's-eye view. Highly maneuverable drones can effectively avoid obstacles such as ground buildings, and combined with intelligent detection algorithms, they can achieve efficient and flexible data collection. This wide, low-cost monitoring and analysis capability makes drone-based target detection commercially promising and widely used in various fields.

[0003] However, compared with traditional target detection, target detection in drone application scenarios faces many challenges: drones usually operate at higher altitudes, and from their perspective, targets appear relatively small, and fewer features can be extracted; in drone application scenarios, the background is relatively complex, including various backgrounds such as city streets, wilderness, forests, and green belts, which makes target recognition more complicated; the computing and storage resources of drone platforms are limited, making it difficult to implement high-demand target detection models.

[0004] Therefore, in order to solve the many limitations of target detection from the perspective of drones, the present invention proposes a drone image target detection algorithm YOLOv8-SOD which improves YOLOv8 to improve the detection performance of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a small target detection method in drone application scenarios with excellent detection performance, so as to solve the problem of poor detection performance caused by the high complexity of the background and the limitations of target recognition in drone use scenarios.

[0006] The small target detection method in the application scenario of unmanned aerial vehicles provided by the present invention comprises: designing a C2f-PConv module, using a partial convolutional layer (PConv) to replace the Bottleneck structure in the C2f module, such lightweight processing can effectively reduce the number of network parameters and the computational load; designing a lightweight EMMB module combining an inverted residual and an EMA attention mechanism, which is used to effectively capture the feature information of small objects under different receptive fields, improve the positioning and classification of small objects in complex environments, and effectively enhance the detection performance of the network; designing a new feature fusion structure PA-AFPN, by combining a progressive feature fusion network with an original PAFPN feature fusion network, efficient multi-scale information fusion is achieved, so as to effectively integrate the fine-grained information of a shallow feature map and the semantic information of a deep feature map; the specific steps are as follows:

[0007] S1: Obtain an image dataset; specifically, the public image dataset VisDrone2019 ( https: / / github.com / VisDrone / VisDrone-Dataset , Machine Learning and Data Mining Laboratory of Tianjin University), dividing the dataset VisDrone2019 into a training set, a validation set, and a test set;

[0008] Specifically, all collected images are resized to 640 × 640 pixels to meet the model input requirements and ensure the consistency of the model when processing images, which is convenient for feature extraction and calculation.

[0009] S2: Build the small target detection network model YOLOv8-SOD from the perspective of drones;

[0010] The YOLOv8-SOD network is an improved YOLOv8 network, which includes three parts: backbone network Backbone, neck structure Neck and prediction layer Head; the construction process of the YOLOv8-SOD network is as follows:

[0011] S2-1: Build C2f-PConv module;

[0012] In the YOLOv8-SOD model, the C2f-PConv module retains the extended feature extraction function of C2f and replaces the original network structure with a partial bottleneck structure PConvBottleneck; the partial bottleneck structure consists of two PConvModule modules; the first module is used to compress features, and the second module is responsible for high-dimensional feature mapping and fusion with the input; the PConvModule module contains a layer that performs partial convolution calculations, and the uncalculated part is concatenated with the calculated part. At the same time, batch normalization (BatchNorm2d) is used to accelerate network convergence, and the SiLU activation function is used to improve the optimization effect of the deep model.

[0013] The C2f-PConv module achieves lightweight processing and reduces computational load by replacing the Bottleneck structure in the C2f module with a partial convolutional layer (PConv). PConv only processes a subset of the input channels c. p The convolution operation is performed while the remaining channels remain unchanged, thus reducing computation and memory access requirements.

[0014] S2-2: construct EMMB module;

[0015] The efficient multi-scale attention mechanism module (EMA) is integrated into the meta-mobile module (MMB) framework; the feature information under different receptive fields is obtained through multi-scale convolution in the EMA mechanism. Among them:

[0016] The meta-mobile module (MMB) consists of depth-wise separable convolution and point-wise convolution; among them:

[0017] Depthwise separable convolution is used to reduce computational complexity by decomposing standard convolution into depthwise convolution and pointwise convolution, significantly reducing the number of parameters and FLOPs.

[0018] Point-by-point convolution, through the formula Y = σ (W p *X) Adjust the number of channels of the input feature map to facilitate subsequent feature fusion operations; W p is a point-by-point convolution filter, σ is an activation function, and X is an input feature map.

[0019] The efficient multi-scale attention mechanism (EMA) is mainly composed of multi-scale convolution and cross-space attention mechanism. Multi-scale convolution uses convolution kernels of different sizes to apply to the input feature map to obtain a multi-scale feature map to capture features of different scales.

[0020] In the cross-spatial attention mechanism, the channel attention map is calculated through global average pooling and fully connected layers; the spatial attention map is calculated using convolutional layers and sigmoid activation functions; the channel and spatial attention maps are element-wise multiplied to obtain the final attention map, and then the attention map is applied to the input feature map to obtain the enhanced feature map.

[0021] Feature fusion is performed by fusing the output feature map of the efficient multi-scale attention mechanism module (EMA) with the feature map of the original meta-mobile module (MMB) through point-by-point convolution and jump connection to obtain the final output.

[0022] S2-3: Construct PA–AFPN feature fusion structure;

[0023] The concept of progressive fusion of AFPN is introduced, which first fuses the low-level features of adjacent levels and then progressively fuses the high-level features, so as to effectively avoid the problems caused by the large gap in semantic information of non-adjacent levels in PAFPN, while retaining more detailed information, enhancing the network's ability to fuse multi-scale features, and improving the model's robustness in handling target size changes in UAV scenarios, thereby significantly improving the accuracy and performance of target detection.

[0024] In the multi-scale feature fusion process, PA-AFPN uses adaptive spatial fusion (ASFF) technology to assign variable spatial weights to different feature levels, enhance the importance of key levels, and minimize the impact of conflicting information of different objects. For feature maps of different levels, ASFF makes them have the same size through upsampling and downsampling operations.

[0025] S3: Training and validation of small object detection models;

[0026] The training set and validation set in the VisDrone2019 dataset are sent to the small target detection model YOLOv8-SOD described in S2 for training and validation. The specific steps are as follows:

[0027] S3-1: Experimental environment settings: Windows Server 2019 Standard is used as the experimental operating system, the GPU is NVIDIA A40 GPU, the memory is 64GB, and torch 1.12.1 combined with cu113 is used as the deep learning framework;

[0028] S3-2: Setting of model training hyperparameters: The experiment sets the batch size to 16, the number of training rounds to 200, the initial learning rate to 0.01, the input image size to 640×640, the momentum coefficient to 0.937, the weight decay coefficient to 0.0005, the optimizer to SGD, the closed set mosaic enhancement probability to 10, the number of warm-up rounds to 10, and the warm-up training is performed in the first 10 rounds of training;

[0029] S3-3: Send the training set to the YOLOv8-SOD network for training, obtain the trained model, and use the trained model for testing.

[0030] S4: Use the test set in the VisDrone2019 dataset for testing, output the final detection results and compare and analyze them with the detection results of other methods.

[0031] The improved YOLOv8 network proposed in the present invention firstly utilizes the characteristics of the PConv structure to introduce the C2f-PConv and PConvModule modules to lightweight the network feature extraction module and reduce the network's learning of redundant features, thereby reducing the computational overhead of the model; secondly, an EMMB module is proposed to integrate a lightweight attention mechanism as a feature fusion module of the model to enhance the fusion capability of shallow and deep features and improve the small target detection performance of the model; finally, a PA-AFPN feature fusion structure is proposed to gradually and effectively combine the features of each layer and effectively utilize multi-scale features.

[0032] The present invention effectively solves the problem of small image targets and dense distribution faced by target detection from the perspective of unmanned aerial vehicles, and effectively lightweights the model while ensuring that the recognition accuracy is superior to that of similar models. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of the present invention.

[0034] Figure 2 This is the YOLOv8-SOD network model structure diagram.

[0035] Figure 3 This is the structure diagram of the C2f-PConv module.

[0036] Figure 4 This is the PConvBottleneck module structure diagram.

[0037] Figure 5 This is the PConvModule module structure diagram.

[0038] Figure 6 This is the EMMB module structure diagram.

[0039] Figure 7 Figure 1 is a schematic diagram of the PA-AFPN feature fusion structure. The middle block is mainly used to represent the convergence and interaction nodes of different scale features (C2, C3, C4) in the fusion process, which are represented by specific processing operations such as convolution and pooling in the actual model. DETAILED DESCRIPTION

[0040] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the present invention will be more comprehensive and complete and fully convey the concept of the example embodiments to those skilled in the art.

[0041] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solution of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0042] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0043] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0044] like Figure 1As shown, the small target detection method for drone perspective based on improved YOLOv8 provided by the present invention specifically includes the following steps:

[0045] S1: Obtain the public dataset of drone-collected images, VisDrone2019, and divide the dataset into a training set, a validation set, and a test set;

[0046] Specifically, all collected images are resized to 640 × 640 pixels to meet the model input requirements and ensure the consistency of the model when processing images, which is convenient for feature extraction and calculation.

[0047] S2: Build the UAV perspective small target detection network model YOLOv8-SOD;

[0048] The YOLOv8-SOD network is an improved YOLOv8 network, which includes three parts: backbone network Backbone, neck structure Neck and prediction layer Head. The network model can be found in Figure 2 The construction process of the YOLOv8-SOD network is as follows:

[0049] S2-1: C2f-PConv module constructed by the present invention. In the UAV scenario, since the UAV is a mobile platform, its computing power, storage space, battery life and other resources are limited. The lightweight characteristics of the C2f-PConv module make it very suitable for such resource-constrained environments.

[0050] In the YOLOv8-SOD model, the C2f-PConv module retains the extended feature extraction function of C2f. The module structure is shown in Figure 3 , which is mainly composed of PConvBottleneck module and PConvModule. For the PConvBottleneck module structure, see Figure 4 This part of the structure consists of two PConvModule modules, where the first module is used to compress features and the second module is responsible for high-dimensional feature mapping and fusion with input. For PConvModule module structure, see Figure 5 It contains the PConv module that performs partial convolution calculations, uses the BatchNorm2d module for batch normalization to accelerate network convergence, and uses the SiLU activation function to improve the optimization effect of the deep model.

[0051] The C2f-PConv module optimizes the original Bottleneck structure in the C2f module by using partial convolutional layers (PConv) to achieve lightweight processing and reduce computational load. PConv only processes a subset of the input channels c pThe convolution operation is performed, and the remaining channels remain unchanged, thereby reducing the amount of computation and memory access requirements. The floating point operations (FLOPs) calculation process of some convolutions is as follows:

[0052]

[0053] Among them, h is the input feature map height, w is the width, k is the kernel size, c p is the number of channels used for partial convolution. p When the number of channels is 1 / 4, the FLOPs of partial convolution is 1 / 16 of regular convolution.

[0054] The memory access amount of conventional convolution is relatively large, and the memory access calculation formula of PConv is:

[0055]

[0056] The memory access of PConv is greatly reduced compared with conventional convolution. p When the number of channels is 1 / 4, the memory access of PConv is about 1 / 4 of that of regular convolution.

[0057] S2-2: EMMB module constructed by the present invention, the module structure is shown in Figure 6 , its overall structure is to integrate the efficient multi-scale attention mechanism (EMA) into the meta-mobile module (MMB) framework. The feature information under different receptive fields is obtained through the multi-scale convolution in the EMA mechanism. This allows the model to perceive target features from multiple scales. In drone aerial images, small targets are small in size and have few feature details. Multi-scale convolution ensures that the key features of small targets at different scales are not missed, such as the local texture features of small vehicles (captured by small kernel convolution) and the overall contour features (captured by large kernel convolution), so as to describe the target more comprehensively and accurately, and effectively improve the detection accuracy of small targets.

[0058] The MMB framework consists of depthwise separable convolution and pointwise convolution. Depthwise convolution performs independent convolution on each channel, greatly reducing the number of parameters; pointwise convolution only adjusts the number of channels, and the amount of calculation is relatively small. When processing large-scale drone aerial image data, it effectively reduces the demand for computing resources. For example, on resource-constrained drone computing units, it can process image data faster, improve the model running speed, and meet the needs of real-time target detection.

[0059] Depthwise separable convolution is used to reduce computational complexity by decomposing standard convolution into depthwise convolution and pointwise convolution, which significantly reduces the number of parameters and FLOPs. For example, for the input feature map (H, W, C are height, width and number of channels respectively), the calculation process of depth separable convolution is to first perform depth convolution Z = σ (W d *Y)(Wd is a deep convolution filter), and then perform point-by-point convolution to adjust the number of channels.

[0060] Point-by-point convolution is done by the formula Y = σ(W p *X)(W p is a point-by-point convolution filter, σ is an activation function, and X is an input feature map) to adjust the number of channels of the input feature map to facilitate subsequent feature fusion operations.

[0061] The efficient multi-scale attention mechanism mainly consists of multi-scale convolution and cross-space attention mechanism. Multi-scale convolution uses convolution kernels of different sizes to apply to the input feature map to obtain a multi-scale feature map to capture features of different scales.

[0062] In the cross-spatial attention mechanism, the channel attention map A is calculated by global average pooling (GAP) and a fully connected layer. c =σ(W c (GAP(M)))(W c is the weight matrix of the fully connected layer, M is the multi-scale feature map); the spatial attention map A is calculated using the convolutional layer and the sigmoid activation function s =σ(W s *M)(W s is a convolution filter); the channel and spatial attention maps are element-wise multiplied to obtain the final attention map A = A c ⊙A s , and then the attention map is applied to the input feature map to obtain the enhanced feature map F = A⊙M.

[0063] Feature fusion: The output feature map of the efficient multi-scale attention mechanism module is fused with the feature map of the original MMB through point-by-point convolution and jump connection to obtain the final output O = W g *F+X(W g is a point-by-point convolution filter, and X is the input feature map).

[0064] S2-3: PA-AFPN feature fusion module constructed by the present invention. The module structure is shown in Figure 7 Its input is the different scale features (C2, C3, C4) output by the backbone network (Backbone). After feature fusion and ASFF technology processing within the module, the fused feature map (P2, P3, P4) is output to the prediction layer (Head); the middle block is mainly used to represent the convergence and interaction nodes of different scale features (C2, C3, C4) in the fusion process, which is represented by specific processing operations such as convolution and pooling in the actual model. In order to highlight the entire fusion architecture, the middle block is not specifically labeled.

[0065] The problem of PAFPN structure in YOLOv8 in processing target detection in UAV scenes is solved, that is, the semantic information gap between non-adjacent levels is not fully considered, which may lead to the loss of semantic information and detail information in the multi-scale feature extraction process. The present invention introduces the concept of progressive fusion of AFPN, first fusing low-level features of adjacent levels, and then progressively fusing high-level features, effectively avoiding the problem caused by the large semantic information gap between non-adjacent levels in PAFPN, while retaining more detail information, enhancing the ability of the network to fuse multi-scale features, and improving the robustness of the model in handling target size changes in UAV scenes, thereby significantly improving the accuracy and performance of target detection.

[0066] The PA-AFPN structure integrates three different scale features C2, C3 and C4 extracted by the backbone network. Considering the large semantic information gap between non-adjacent levels of C2 and C4, a specific fusion strategy is adopted: first, the semantic information of C3 is passed to C2 through a top-down path, and then the features of C2 and C3 are gradually merged using a progressive fusion strategy to ensure the retention of detail information; then the feature fusion effect is enhanced through a bottom-up path, and then C4 is fused with C2 and C3 in the same way to achieve complete integration of features at all levels, and finally three feature maps P2, P3 and P4 of different scales are generated for target prediction. In this process, the semantic similarity between C2 and C3 is given priority to avoid the semantic gap problem caused by the direct fusion of C2 and C4. Since C3 and C4 are adjacent and the semantic gap is small, the semantic gap between C2 and C4 is further reduced, and the accuracy and effectiveness of feature fusion are improved.

[0067] In the multi-scale feature fusion process, PA-AFPN uses adaptive spatial fusion (ASFF) technology to assign variable spatial weights to different feature levels, enhancing the importance of key levels while minimizing the impact of conflicting information of different objects. For example, for the feature vector from level n to level l at position (i, j) The final feature vector is obtained by adaptive spatial fusion

[0068]

[0069] in, and are the spatial weights of the three-level features, and

[0070] For feature maps of different levels, ASFF uses upsampling and downsampling operations to make them have the same size. When upsampling, 1×1 convolution is first used to keep the number of channels of other levels consistent with the number of channels of the current level, and then interpolation is used to increase the resolution; when downsampling, 3×3 convolution with a stride of 2 is used to modify the number of channels to reduce the resolution. For secondary downsampling, maximum pooling with a stride of 2 is first performed, and then 3×3 convolution with a stride of 2 is used to modify the number of channels.

[0071] S3: Send the training set and validation set in the VisDrone2019 dataset to the YOLOv8-SOD target detection model described in S2 for training and validation. The specific implementation steps are as follows:

[0072] S3-1: Experimental environment settings: Windows Server 2019 Standard is used as the experimental operating system, the GPU is NVIDIA A40 GPU, the memory is 64GB, and torch 1.12.1 combined with cu113 is used as the deep learning framework;

[0073] S3-2: Setting of model training hyperparameters: The experiment sets the batch size to 16, the number of training rounds to 200, the initial learning rate to 0.01, the input image size to 640×640, the momentum coefficient to 0.937, the weight decay coefficient to 0.0005, the optimizer to SGD, the closed set mosaic enhancement probability to 10, the number of warm-up rounds to 10, and the warm-up training is performed in the first 10 rounds of training;

[0074] S3-3: Send the training set to the network for training, obtain the trained model, and use the trained model for testing.

[0075] S4: Tested using the test set in the VisDrone2019 dataset, output the final detection results and compared them with the detection results of other methods. Compared with YOLOv8, the experimental results show that YOLOv8-SOD achieves a 3.2% increase in mAP, a 25% reduction in parameters, a 28.3% reduction in computation, and a 30% increase in FPS.

[0076] In order to further verify the effectiveness of the improved model YOLOv8-SOD in target detection, the existing YOLO series target detection models are selected for comparison on the VisDrone2019 dataset and the SODA-A dataset. The experimental results are shown in Table 1. Compared with other single-stage target detection algorithms, the method proposed in this paper has advantages over other target detection algorithms in detection accuracy and detection speed.

[0077] Table 1 Comparative experiments of the present invention and the YOLO series models in VisDrone2019 and SODA-A datasets

[0078]

[0079] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the embodiments disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary technical means in the art that are not disclosed in the present invention.

[0080] It should be understood that the present invention is not limited to the exact construction that has 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 invention is limited only by the appended claims.

Claims

1. A small target detection method in a drone application scenario, characterized in that: include: Design the C2f-PConv module, and use the partial convolution layer (PConv) to replace the Bottleneck structure in the C2f module; design a lightweight EMMB module that combines the inverted residual and EMA attention mechanism to effectively capture the feature information of small objects under different receptive fields, improve the positioning and classification of small objects in complex environments, and enhance the detection performance of the network; design a feature fusion module PA-AFPN, and combine the progressive feature fusion network with the original PAFPN feature fusion network to achieve efficient multi-scale information fusion, so as to effectively integrate the fine-grained information of the shallow feature map and the semantic information of the deep feature map; the specific steps are as follows: S1: Obtain an image dataset; specifically, use the public image dataset VisDrone2019 collected by drones, and divide the dataset VisDrone2019 into a training set, a validation set, and a test set; S2: Build the small target detection network model YOLOv8-SOD from the perspective of drones; The YOLOv8-SOD network is an improved YOLOv8 network, which includes three parts: backbone network Backbone, neck structure Neck and prediction layer Head; the construction process of the YOLOv8-SOD network is as follows: S2-1: Build C2f-PConv module; In the YOLOv8-SOD model, the C2f-PConv module retains the extended feature extraction function of C2f and replaces the original network structure with a partial bottleneck structure PConvBottleneck module; S2-2: construct EMMB module; Integrate the efficient multi-scale attention mechanism module (EMA) into the meta-mobility module (MMB) framework; obtain feature information under different receptive fields through multi-scale convolution in the EMA mechanism; S2-3: Construct PA–AFPN feature fusion structure; The concept of progressive fusion of AFPN is introduced, which first fuses the low-level features of adjacent levels and then gradually fuses the high-level features, so as to effectively avoid the problem caused by the large gap in semantic information of non-adjacent levels in PAFPN, while retaining more detailed information, enhancing the ability of the network to fuse multi-scale features, and improving the robustness of the model in handling target size changes in UAV scenarios, thereby significantly improving the accuracy and performance of target detection; S3: Training and validation of small object detection models; The training set and validation set in the VisDrone2019 dataset are sent to the small target detection model YOLOv8-SOD described in S2 for training and validation; S4: Use the test set in the VisDrone2019 dataset for testing, output the final detection results and compare and analyze them with the detection results of other methods.

2. The small target detection method in the drone application scenario according to claim 1 is characterized in that: In S1, all collected images are resized to 640 × 640 pixels to meet the model input requirements, ensuring consistency when the model processes images and facilitating feature extraction and calculation.

3. The small target detection method in the drone application scenario according to claim 2 is characterized in that: In S2-1, the PConvBottleneck module consists of two PConvModule modules; the first module is used to compress features, and the second module is responsible for high-dimensional feature mapping and fusion with the input; The PConvModule module contains layers that perform partial convolution calculations. The uncalculated part is concatenated with the calculated part. Batch normalization (BatchNorm2d) is used to accelerate network convergence, and the SiLU activation function is used to improve the optimization effect of the deep model.

4. The small target detection method in the drone application scenario according to claim 3 is characterized in that: In S2-2, the meta-mobile module (MMB) consists of depth-wise separable convolution and point-wise convolution; where: Depthwise separable convolution is used to reduce computational complexity by decomposing standard convolution into depthwise convolution and pointwise convolution, significantly reducing the number of parameters and FLOPs. Point-by-point convolution, through the formula Y = σ (W p *X) Adjust the number of channels of the input feature map to facilitate subsequent feature fusion operations; W p is a point-by-point convolution filter, σ is an activation function, and X is an input feature map; Efficient Multi-scale Attention Mechanism (EMA), which consists of multi-scale convolution and cross-spatial attention mechanism; multi-scale convolution uses convolution kernels of different sizes to apply to the input feature map to obtain a multi-scale feature map to capture features of different scales; In the cross-spatial attention mechanism, the channel attention map is calculated through global average pooling and fully connected layers; the spatial attention map is calculated using convolutional layers and sigmoid activation functions; the channel and spatial attention maps are element-wise multiplied to obtain the final attention map, and then the attention map is applied to the input feature map to obtain the enhanced feature map; Feature fusion is performed by fusing the output feature map of the multi-scale attention mechanism module (EMA) with the feature map of the original meta-mobile module (MMB) through point-by-point convolution and jump connection to obtain the final output.

5. The small target detection method in the drone application scenario according to claim 4 is characterized in that: In S2-3, during the multi-scale feature fusion process, PA-AFPN uses adaptive spatial fusion (ASFF) technology to assign variable spatial weights to different feature levels, enhance the importance of key levels, and minimize the impact of conflicting information of different objects; for feature maps of different levels, ASFF uses upsampling and downsampling operations to make them have the same size.

6. The small target detection method in the drone application scenario according to claim 5 is characterized in that: In S3, the specific steps for training and validating the small object detection model are as follows: S3-1: Experimental environment settings: Windows Server 2019 Standard is used as the experimental operating system, the GPU is NVIDIA A40 GPU, the memory is 64GB, and torch 1.12.1 combined with cu113 is used as the deep learning framework; S3-2: Setting of model training hyperparameters: The experiment sets the batch size to 16, the number of training rounds to 200, the initial learning rate to 0.01, the input image size to 640×640, the momentum coefficient to 0.937, the weight decay coefficient to 0.0005, the optimizer to SGD, the closed set mosaic enhancement probability to 10, the number of warm-up rounds to 10, and the warm-up training is performed in the first 10 rounds of training; S3-3: Send the training set to the YOLOv8-SOD network for training, obtain the trained model, and use the trained model for testing.

Images