Unmanned aerial vehicle small target detection method based on lightweight multi-scale sensing network
By optimizing feature extraction and fusion through a lightweight multi-scale perception network architecture, the accuracy and computational complexity issues of small target detection in UAVs under complex backgrounds are solved, and efficient detection is achieved on resource-constrained platforms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-10
AI Technical Summary
Existing UAV small target detection methods have insufficient detection accuracy and high computational complexity in complex backgrounds, making it difficult to operate efficiently on resource-constrained UAV platforms.
A lightweight multi-scale perceptual network architecture is adopted, which combines the C3K2-GDF module, AMFAN network and LES detection head to optimize feature extraction and fusion, reduce computational complexity and improve detection accuracy.
It significantly improves the detection accuracy of small targets in complex backgrounds, reduces computational complexity, adapts to the computational resource limitations of UAV platforms, and achieves efficient real-time detection.
Smart Images
Figure CN121640310A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of computer vision and artificial intelligence, particularly the field of UAV image processing and small target detection, and specifically relates to a method for UAV small target detection based on a lightweight multi-scale perception network. Background Technology
[0002] Unmanned aerial vehicle (UAV) small target detection technology has been widely applied in recent years in various fields such as agricultural monitoring, environmental protection, and security patrol. However, from the perspective of UAVs, small targets have weak features and are often in complex backgrounds, which limits the effectiveness of traditional target detection methods in such tasks. To improve detection accuracy and efficiency, various improvement schemes have been proposed in existing research.
[0003] The YOLO (You Only Look Once) model family is widely used in general object detection tasks due to its simple structure and fast inference speed. However, in the scenario of detecting small targets on UAVs, the YOLO model still faces the problem of insufficient accuracy. Traditional YOLO models extract deep features by downsampling layer by layer. This process is prone to losing detailed information about small targets, especially in cases with complex backgrounds or low image resolution, where detection performance is significantly limited. To alleviate the above problems, existing research has introduced multi-scale feature fusion strategies, among which Feature Pyramid Network (FPN) is widely used to fuse features from different levels to enhance the model's ability to perceive small targets. However, FPN may introduce redundant information during feature fusion, while also increasing the number of model parameters and computational complexity, limiting its application in resource-constrained scenarios.
[0004] Furthermore, to enhance the feature representation capabilities of small targets, some methods introduce attention mechanisms, such as Convolutional Block Attention (CBAM) and Focus Modulation (FMM), which optimize the feature extraction process through adaptive weighting and contextual information enhancement. However, these modules often significantly increase network complexity and computational overhead, making efficient deployment difficult on UAV edge devices with limited computing resources. Summary of the Invention
[0005] To address the above problems, this invention provides a method for detecting small targets on unmanned aerial vehicles (UAVs) based on a lightweight multi-scale perception network, comprising the following steps:
[0006] S1. Input the image to be processed into the backbone network to obtain a feature image set, where,
[0007] The backbone network includes a first CBS block, a second CBS block, a first C3K2-GDF module, a third CBS block, a second C3K2-GDF module, a fourth CBS block, a third C3K2-GDF module, an SPPF module, and a C2PSA module, which are cascaded in sequence.
[0008] The feature image group includes the output images of the first C3K2-GDF module, the second C3K2-GDF module, the third C3K2-GDF module, and the C2PSA module;
[0009] S2. Input the feature image group into the neck network to obtain an aggregated image group, where,
[0010] The neck network comprises a first HFR module, a first C3k2 module, a first upsampling module, a second HFR module, a second C3k2 module, a second upsampling module, a third HFR module, a third C3k2 module, a fifth CBS block, a first PFE module, a fourth C3k2 module, a sixth CBS block, a second PFE module, and a fifth C3k2 module, all cascaded in sequence. The first C3k2 module is also connected to the second PFE module, and the second C3k2 module is also connected to the first PFE module.
[0011] The first HFR module receives the output images of the third C3K2-GDF module and the C2PSA module; the second HFR module receives the output images of the second C3K2-GDF module and the first upsampling module; and the third HFR module also receives the output images of the first C3K2-GDF module and the second upsampling module.
[0012] The aggregated image group includes the output images of the third C3k2 module, the fourth C3k2 module, and the fifth C3k2 module;
[0013] S3. Input the aggregated image group into the head network to obtain the detection results, where,
[0014] The head network includes three LES detection heads with identical structures. The output images of the third, fourth, and fifth C3k2 modules are respectively sent to different LES detection heads for processing.
[0015] The beneficial effects of this invention are:
[0016] Improving Small Target Detection Accuracy: By introducing the C3K2-GDF module and the AMFAN network, this invention effectively enhances the feature extraction and fusion capabilities for small targets. The C3K2-GDF module can fully preserve the fine-grained features of small targets. The AMFAN network is composed of a top-down feature fusion HFR module and a bottom-up feature fusion PFE module. Through an adaptive multi-scale feature aggregation mechanism, it improves the detection accuracy of small targets at different scales, thereby significantly improving the recognition accuracy of small targets in complex backgrounds. Reducing Computational Complexity: A lightweight network architecture is adopted, combined with LES head design and optimized convolutional operations, significantly reducing the number of model parameters and computational overhead. This design allows the model to run efficiently on computationally limited UAV platforms while maintaining excellent detection performance, meeting real-time processing requirements.
[0017] Enhanced feature information preservation: By improving the multi-scale feature fusion strategy and reducing the number of downsampling times, this invention effectively suppresses the spatial information loss problem common in traditional feature extraction methods, thereby improving the detection effect of small targets, especially in low-resolution images and complex scenes.
[0018] High practicality for UAV platforms: This technical solution fully considers the limitations of UAVs in terms of computing power and power consumption, and adopts an efficient lightweight structural design. While ensuring detection accuracy, it enables rapid deployment and stable operation on resource-constrained devices, and has high engineering application value. Attached Figure Description
[0019] Figure 1 This is a diagram of the lightweight multi-scale sensing network structure of the present invention.
[0020] Figure 2 This is a structural diagram of the C3K2-GDF module of the present invention;
[0021] Figure 3 This is a structural diagram of the HFR module of the present invention;
[0022] Figure 4 This is a structural diagram of the PFE module of the present invention;
[0023] Figure 5 This is a diagram of the head network structure of the present invention. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] Existing technologies have proposed various methods to address the accuracy and efficiency issues in small target detection, but most methods still suffer from the following shortcomings:
[0026] 1) Loss of small target features: Existing methods are prone to losing fine-grained features of small targets when performing multiple downsampling and feature fusion, resulting in reduced detection accuracy.
[0027] 2) High computational complexity: While improving detection accuracy, existing models often bring high computational complexity, which is difficult to meet the needs of unmanned aerial vehicle platforms with limited resources.
[0028] Some embodiments of the present invention provide a method for small target detection on unmanned aerial vehicles (UAVs) based on a lightweight multi-scale sensing network. This method proposes, for example... Figure 1 The lightweight multi-scale perception network (LMGA Net) architecture is shown. By optimizing feature extraction and fusion mechanisms, LMGA Net significantly reduces computational complexity while maintaining detection accuracy, making it suitable for resource-constrained UAV platforms.
[0029] The specific process of the UAV small target detection method includes the following steps:
[0030] S1. Input the image to be processed into the backbone network to obtain a feature image set, where,
[0031] The backbone network includes a first CBS block, a second CBS block, a first C3K2-GDF module, a third CBS block, a second C3K2-GDF module, a fourth CBS block, a third C3K2-GDF module, an SPPF module, and a C2PSA module, which are cascaded in sequence.
[0032] The feature image group includes the output images of the first C3K2-GDF module, the second C3K2-GDF module, the third C3K2-GDF module, and the C2PSA module;
[0033] S2. Input the feature image group into the neck network to obtain an aggregated image group, where,
[0034] The neck network comprises a first HFR module, a first C3k2 module, a first upsampling module, a second HFR module, a second C3k2 module, a second upsampling module, a third HFR module, a third C3k2 module, a fifth CBS block, a first PFE module, a fourth C3k2 module, a sixth CBS block, a second PFE module, and a fifth C3k2 module, all cascaded in sequence. The first C3k2 module is also connected to the second PFE module, and the second C3k2 module is also connected to the first PFE module.
[0035] The first HFR module receives the output images of the third C3K2-GDF module and the C2PSA module; the second HFR module receives the output images of the second C3K2-GDF module and the first upsampling module; and the third HFR module also receives the output images of the first C3K2-GDF module and the second upsampling module.
[0036] The aggregated image group includes the output images of the third C3k2 module, the fourth C3k2 module, and the fifth C3k2 module;
[0037] S3. Input the aggregated image group into the head network to obtain the detection results, where,
[0038] The head network includes three LES detection heads with identical structures. The output images of the third, fourth, and fifth C3k2 modules are respectively sent to different LES detection heads for processing.
[0039] In some embodiments, CBS includes a convolutional layer, a BN (BatchNorm) layer, and a SiLU activation function layer, wherein the convolutional kernel size of the convolutional layer is 3×3 and the stride is 2.
[0040] In some embodiments, the SPPF module and C2PSA module adopt the structure of the YOLOv11 network.
[0041] In some embodiments, such as Figure 2 As shown, the C3K2-GDF module includes multiple cascaded GDFR modules. The outputs of all GDFR modules are concatenated to obtain the output of the G3K2-GDF module. The processing procedure of the GDFR module includes:
[0042] S11. Input the input image into the BN layer to obtain the first batch of features. Input the first batch of features into the DWC (Dynamic Weighting Convolution) module for processing to obtain the output features of the DWC module. Add the output features of the DWC module to the input image to obtain the first feature image. The processing steps of the DWC module include:
[0043] The first batch of features is divided into first branch features and second branch features by the Split layer;
[0044] The first branch features and the second branch features are processed by a DPDC (Dynamic PyramidDepthwise Convolution) module to obtain the first DPDC output features and the second DPDC output features.
[0045] The first DPDC output features and the second DPDC output features are concatenated along the channel dimension, and the concatenation result is fused through a 1×1 convolutional layer to obtain the DWC module output features.
[0046] S12. The first feature image is processed through a BN layer to obtain a second batch of features. The second batch of features is then input into a GCU (Granular Channel Unit) module for processing. The processing result is added to the first feature image to obtain the output features of the GCU module. The processing procedure of the GCU module includes:
[0047] The second batch of features is input into a 1×1 convolutional layer and activated by Sigmoid to obtain a channel weight map;
[0048] The second batch of features is multiplied element by element with the channel weight map to obtain the weighted feature map;
[0049] The weighted feature maps are input into four MSCA (Multi-Scale Context Aggregation attention module) modules, and the four MSCA modules use convolutional kernels of different sizes.
[0050] The outputs of the four MSCA modules are concatenated along the channel dimension and then fused using a 1×1 convolutional layer to obtain the output features of the GCU module. In some embodiments, the convolutional kernel sizes used for the four MSCA modules are 9×9, 7×7, 5×5, and 3×3, respectively.
[0051] In some embodiments, the DPDC module is used for adaptive fusion of multi-scale features, and it includes a weight branch and a convolutional branch. The processing procedure of the DPDC module includes:
[0052] In the weight branch, the input features pass through the pooling layer and the convolutional layer in sequence to obtain pooled convolutional features. The pooled convolutional features are normalized in the channel dimension by softmax to generate a set of scale weight coefficients. The scale weight coefficients include the first weight, the second weight and the third weight, and the three weights correspond to different receptive fields.
[0053] In the convolution branch, the input features are fed into three deep convolutional layers (DWConv) with different receptive fields to obtain the first deep convolutional feature, the second deep convolutional feature, and the third deep convolutional feature.
[0054] The first depthwise convolutional feature is multiplied element-wise with the first weight, the second depthwise convolutional feature is multiplied element-wise with the second weight, and the third depthwise convolutional feature is multiplied element-wise with the third weight; the products of the three are added together to obtain the weighted multiscale feature.
[0055] The weighted multi-scale features are input into the BS layer, which consists of a BN layer and a SiLU activation function, to obtain the output features of the DPDC module.
[0056] In some embodiments, the MSCA module includes:
[0057] The input features are passed through max pooling and average pooling layers respectively to obtain the first pooling feature and the second pooling feature; the first pooling feature and the second pooling feature are fused (e.g., concatenated or added) to obtain the pooling fused feature; the pooling fused feature is input into a convolutional layer and a ReLU activation function layer to obtain the first branch feature;
[0058] The input features are passed sequentially through a deep convolutional layer (DWConv) and a 1×1 convolutional layer to obtain the second branch features;
[0059] The first branch features are multiplied element-wise with the second branch features to obtain the output features of the MSCA module. Throughout the process, the C3K2-GDF module adaptively adjusts the receptive field, fully preserves local details, and highlights key channels, effectively improving the extraction and fusion capabilities of multi-scale features for small targets.
[0060] In some embodiments, the present invention proposes an Adaptive Multi-Scale Feature Aggregation Network (AMFAN) in the neck network, which achieves bidirectional adaptive fusion through a Hierarchical Feature Refinement (HFR) module and a Progressive Feature Enhancement (PFE) module.
[0061] In some embodiments, the HFR module includes two inputs, one from the C3K2-GDF module and the other from the upsampling module or the C2PSA module. For example... Figure 3 As shown, the processing procedure of the HFR module includes:
[0062] S21. Perform a 1×1 convolution on the input from the C3K2-GDF module to expand the channels and obtain the expanded features. Add the expanded features to the input from the upsampling module or the C2PSA module to obtain the first fused features.
[0063] S22. The input from the upsampling module or C2PSA module is processed through Channel Attention to obtain key region features. The key region features are multiplied with the first fusion feature to obtain the second fusion feature.
[0064] S23. The input from the C3K2-GDF module is processed through Spatial Attention to obtain long-range dependency features. The long-range dependency features are multiplied with the second fusion feature to obtain the output of the HFR module.
[0065] In some embodiments, the Channel Attention process includes:
[0066] S221. The input is processed through a 5×5 depthwise separable convolution (DSConv) and a ReLU activation function to obtain separable convolution features;
[0067] S222. The separated convolutional features are processed by max pooling and average pooling respectively, and the outputs of max pooling and average pooling are added together to obtain the summed features;
[0068] S223. The summed features are processed through a 1×1 Dynamic Expandable Separable Convolution (DES-conv) and a 1×1 Attention-Enhanced Channel-Spatial Convolution (ACS-conv) to obtain spatial channel weights. The spatial channel weights are then multiplied by the input to obtain the key region features.
[0069] In some embodiments, Spatial Attention includes two branch paths. The outputs of the two branch paths are fused and then processed by a sigmoid function to obtain long-range dependency features. Each branch path includes a sequentially cascaded Batch Normalization (BN) layer, a convolutional block, a 1×1 convolutional layer, an attention block, and a 1×1 convolutional layer. The convolutional block includes parallel 1×7, 1×11, and 1×21 convolutional layers.
[0070] The attention block in the first branch path calculates the first Q matrix, the first K matrix, and the first V matrix based on the output of the first 1×1 convolutional layer;
[0071] The attention block in the second branch path calculates the second Q matrix, the second K matrix, and the second V matrix based on the output of the first 1×1 convolutional layer.
[0072] In the attention block of the first branch path, the second Q matrix is multiplied by the first K matrix and then multiplied by the first V matrix to obtain the first attention feature;
[0073] In the attention block of the second branch path, the first Q matrix is multiplied by the second K matrix and then multiplied by the second V matrix to obtain the second attention feature.
[0074] In some embodiments, the PFE module includes two inputs, one from the C3k2 module and the other from the CBS block; such as Figure 4 As shown, the processing procedure of the PFE module includes:
[0075] S31. Add the two inputs to obtain the hierarchical features;
[0076] S32. The input from the CBS block is passed through the Shared Multi-Semantic Space Attention (SMSA) module to generate a directional attention map that enhances fine-grained spatial details;
[0077] S33. The input from the C3k2 module is passed through the channel attention mechanism module to obtain semantic channel attention;
[0078] S34. Multiply the hierarchical features, directional attention map to enhance fine-grained spatial details, and semantic channel attention to obtain the output of the PFE module.
[0079] In some embodiments, the processing procedure of the SMSA module includes:
[0080] S321. Pool the input features horizontally into a one-dimensional sequence. Divide the one-dimensional sequence according to the channel dimension and pass it through MS-DWConv (Multi-Receptive Field Shared Depthwise Convolutions) to obtain horizontal multi-scale features. Concatenate the horizontal multi-scale features and pass them through a GN (GroupNorm) layer and a sigmoid function to obtain horizontal spatial pattern features.
[0081] S322. Pool the input features into a one-dimensional sequence along the vertical direction. Divide the one-dimensional sequence according to the channel dimension and pass it through MS-DWConv to obtain vertical multi-scale features. Concatenate the vertical multi-scale features and pass them through the GN layer and sigmoid function to obtain vertical spatial pattern features.
[0082] S323. The horizontal spatial pattern features are fused with the vertical spatial pattern features to obtain a directional attention map that enhances fine-grained spatial details.
[0083] In some embodiments, MS-DWConv includes four parallel deep convolutional layers with kernel sizes of 3×3, 5×5, 7×7, and 9×9, respectively.
[0084] In some embodiments, the processing procedure of the channel attention mechanism module includes:
[0085] The input features are processed through global average pooling and a multilayer perceptron (MLP) to obtain the weights for the first channel.
[0086] The input features are processed by global max pooling and MLP to obtain the weights of the second channel.
[0087] The modulation coefficients are obtained by adding the weights of the first channel and the weights of the second channel and then passing the sum through the Sigmoid function.
[0088] Semantic channel attention is obtained by multiplying the input features with the modulation coefficients channel by channel and element by element.
[0089] The PFE module first adds the two inputs to obtain hierarchical features, and then processes the two inputs spatially and in channels respectively. It uses the spatial and channel features obtained from the processing to modulate the fused features, so that the output can take into account both high-resolution localization and rich semantics, and can efficiently capture multi-scale information of small targets.
[0090] In some embodiments, the head network includes three structurally identical LES detection heads, such as... Figure 5 As shown, each LES detection head includes a first CGS module, a feature enhancement (FE) module, a second CGS module, a third CGS module, and a detection module, all cascaded in sequence. The CGS module comprises a convolutional layer, a GN layer, and a SiLU activation function layer, all cascaded in sequence. The first CGS module has a 1×1 kernel, while the second and third CGS modules have 3×3 kernels. The FE module comprises two parallel depthwise separable convolutions with kernel sizes of 1×1 and 5×5, respectively. The second and third CGS modules share a common structure and jointly serve the three LES detection heads.
[0091] Specifically, such as Figure 5 As shown, the three feature maps P2~P4 of the aggregated image group are processed along independent but weight-shared paths to balance scale specificity and parameter efficiency. Each feature map first passes through its corresponding scale's independent first CGS module. The 1×1 convolution in the first CGS module, combined with group normalization and SiLU activation, achieves basic feature transformation and non-linear enhancement, ensuring gradient stability while improving representational power. Subsequently, it passes through an independent enhancement module, which uses a 5×5 depthwise separable convolution to provide a larger receptive field to capture broad context, and a 3×3 depthwise separable convolution to focus on refining local details. Finally, the outputs are element-wise summed and fused. Next, the outputs of the three parallel FE modules are fed into a shared second CGS module, using the same weights across all scales to reduce redundant parameters. This module uses a 3×3 group convolution to achieve group processing, efficiently capturing spatial structure. It then inputs into a shared third CGS module, where a 3×3 standard convolution completes the comprehensive integration. Throughout the processing, features at each scale remain independent and unmixed, but consistent enhancement is achieved through a unified processing framework. Thus, the head network significantly reduces the number of detection head parameters and computational burden while enhancing the ability to express multi-scale details, achieving lightweight design and optimized computational efficiency.
[0092] This invention features a lightweight network structure, reducing redundancy in traditional convolution operations. Furthermore, by optimizing the upsampling and convolution modules, the computational complexity of the model is significantly reduced. This enables the model to run efficiently on resource-constrained platforms such as drones, meeting real-time detection requirements.
[0093] This invention proposes an efficient UAV small target detection technology by introducing a lightweight network design, an improved small target feature extraction method, and an optimized multi-scale feature fusion strategy. This technology can reduce computational burden and improve robustness in complex environments while ensuring high accuracy.
[0094] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for detecting small targets of a UAV based on a lightweight multi-scale perception network, characterized in that, The method comprises the following steps: S1. inputting an image to be processed into a backbone network to obtain a feature image group, wherein, the backbone network comprises a first CBS block, a second CBS block, a first C3K2-GDF module, a third CBS block, a second C3K2-GDF module, a fourth CBS block, a third C3K2-GDF module, an SPPF module and a C2PSA module which are sequentially cascaded; the feature image group comprises output images of the first C3K2-GDF module, the second C3K2-GDF module, the third C3K2-GDF module and the C2PSA module; S2. inputting the feature image group into a neck network to obtain an aggregated image group, wherein, the neck network comprises a first HFR module, a first C3k2 module, a first up-sampling module, a second HFR module, a second C3k2 module, a second up-sampling module, a third HFR module, a third C3k2 module, a fifth CBS block, a first PFE module, a fourth C3k2 module, a sixth CBS block, a second PFE module and a fifth C3k2 module which are sequentially cascaded; the first C3k2 module is further connected with the second PFE module, and the second C3k2 module is further connected with the first PFE module; the first HFR module receives output images of the third C3K2-GDF module and the C2PSA module, the second HFR module receives output images of the second C3K2-GDF module and the first up-sampling module, and the third HFR module further receives output images of the first C3K2-GDF module and the second up-sampling module; the aggregated image group comprises output images of the third C3k2 module, the fourth C3k2 module and the fifth C3k2 module; S3. inputting the aggregated image group into a head network to obtain a detection result, wherein, the head network comprises three LES detection heads which are the same in structure, and the output images of the third C3k2 module, the fourth C3k2 module and the fifth C3k2 module are respectively sent into different LES detection heads for processing.
2. The unmanned aerial vehicle small target detection method based on a lightweight multi-scale perception network according to claim 1, characterized in that, The C3K2-GDF module comprises a plurality of cascaded GDFR modules, and the outputs of all the GDFR modules are spliced to obtain the output of the G3K2-GDF module, and the processing process of the GDFR module comprises: S11. inputting an input image into a BN layer to obtain a first batch of features, inputting the first batch of features into a DWC module to obtain DWC module output features, and adding the DWC module output features and the input image to obtain a first feature image; wherein, the processing process of the DWC module comprises: dividing the first batch of features into a first branch feature and a second branch feature through a Split layer; inputting the first branch feature and the second branch feature into a DPDC module respectively for processing to obtain a first DPDC output feature and a second DPDC output feature; splicing the first DPDC output feature and the second DPDC output feature in the channel dimension, and performing channel fusion on the splicing result through a 1x1 convolution layer to obtain the DWC module output feature; S12. The first feature image is input into a BN layer to obtain a second batch of features, and the second batch of features is input into a GCU module for processing, and a processing result is added to the first feature image to obtain a GCU module output feature; wherein the processing process of the GCU module comprises: The second batch of features is input into a 1x1 convolutional layer and is activated by a Sigmoid function to obtain a channel weight map; The second batch of features and the channel weight map are multiplied element by element to obtain a weighted feature map; The weighted feature map is input into four MSCA modules respectively, and the four MSCA modules use convolution kernels of different sizes; The outputs of the four MSCA modules are spliced in the channel dimension and are fused in the channel dimension by a 1x1 convolutional layer to obtain the GCU module output feature.
3. The unmanned aerial vehicle small target detection method based on a lightweight multi-scale perception network according to claim 2, characterized in that, The DPDC module comprises a weight branch and a convolution branch, and the specific processing process thereof comprises: In the weight branch, the input feature is sequentially input into a pooling layer and a convolutional layer to obtain a pooled convolutional feature, and the pooled convolutional feature is normalized by a softmax in the channel dimension to generate a set of scale weight coefficients; the scale weight coefficients comprise a first weight, a second weight and a third weight; In the convolution branch, the input feature is input into three deep convolutional layers with different receptive fields to obtain a first deep convolutional feature, a second deep convolutional feature and a third deep convolutional feature; The first deep convolutional feature is multiplied element by element with the first weight, the second deep convolutional feature is multiplied element by element with the second weight, and the third deep convolutional feature is multiplied element by element with the third weight; the product results of the three are added to obtain a weighted multi-scale feature; The weighted multi-scale feature is input into a BS layer composed of a BN layer and a SiLU activation function to obtain a DPDC module output feature.
4. The unmanned aerial vehicle small target detection method based on a lightweight multi-scale perception network according to claim 2, characterized in that, The MSCA module comprises: The input feature is input into a maximum pooling layer and an average pooling layer respectively to obtain a first pooled feature and a second pooled feature; the first pooled feature and the second pooled feature are fused to obtain a pooled fusion feature; and the pooled fusion feature is input into a convolutional layer and a ReLU activation function layer to obtain a first branch feature; The input feature is sequentially input into a deep convolutional layer and a 1x1 convolutional layer to obtain a second branch feature; The first branch feature and the second branch feature are multiplied element by element to obtain an MSCA module output feature.
5. The unmanned aerial vehicle small target detection method based on a lightweight multi-scale perception network according to claim 1, characterized in that, The HFR module comprises two inputs, one from the C3K2-GDF module and the other from the up-sampling module or the C2PSA module; and the processing process of the HFR module comprises: S21. The input from the C3K2-GDF module is expanded in the channel dimension by a 1x1 convolution to obtain an expanded feature, and the expanded feature is added to the input from the up-sampling module or the C2PSA module to obtain a first fusion feature; S22. The input from the up-sampling module or the C2PSA module is input into a Channel Attention to obtain a key region feature, and the key region feature is multiplied with the first fusion feature to obtain a second fusion feature; S23. The input from the C3K2-GDF module is input into the Spatial Attention to obtain long-distance dependency features, and the long-distance dependency features are multiplied by the second fusion features to obtain the output of the HFR module.
6. The unmanned aerial vehicle small target detection method based on a lightweight multi-scale perception network according to claim 5, characterized in that, The processing process of the Channel Attention includes: S221. The input is input into a 5*5 DSConv and a ReLU activation function to obtain separated convolution features; S222. The separated convolution features are respectively input into a max-pooling and an average-pooling, and the outputs of the max-pooling and the average-pooling are added to obtain added features; S223. The added features are input into a 1*1 DES-conv and a 1*1 ACS-conv to obtain spatial channel weights, and the spatial channel weights are multiplied by the input to obtain key region features.
7. The unmanned aerial vehicle small target detection method based on a lightweight multi-scale perception network according to claim 5, characterized in that, The Spatial Attention includes two branch paths, and the outputs of the two branch paths are fused and input into a Sigmoid function to obtain long-distance dependency features; each branch path includes a BN layer, a convolution block, a 1*1 convolution layer, an attention block and a 1*1 convolution layer which are sequentially cascaded, and the convolution block includes a 1*7 convolution layer, a 1*11 convolution layer and a 1*21 convolution layer which are connected in parallel; wherein, The attention block in the first branch path calculates a first Q matrix, a first K matrix and a first V matrix according to the output of the first 1*1 convolution layer; and the attention block in the second branch path calculates a second Q matrix, a second K matrix and a second V matrix according to the output of the first 1*1 convolution layer; In the attention block of the first branch path, the first Q matrix is multiplied by the first K matrix, and then the first V matrix is multiplied by the product of the first Q matrix and the first K matrix to obtain first attention features; In the attention block of the second branch path, the second Q matrix is multiplied by the second K matrix, and then the second V matrix is multiplied by the product of the second Q matrix and the second K matrix to obtain second attention features. 8.The UAV small target detection method based on lightweight multi-scale perception network according to claim 1, characterized in that, The PFE module includes two inputs, one of which is from the C3k2 module and the other of which is from the CBS block; and the processing process of the PFE module includes: S31. The two inputs are added to obtain hierarchical features; S32. The input from the CBS block is input into the SMSA module to generate a direction attention map to strengthen fine-grained spatial details; wherein, the processing process of the SMSA module includes: S321. The input features are pooled into a one-dimensional sequence along the horizontal direction, the one-dimensional sequence is divided along the channel dimension, and then the MS-DWConv is used to obtain horizontal multi-scale features, the horizontal multi-scale features are spliced, and then the GN layer and the sigmoid function are used to obtain horizontal spatial mode features; S322. The input features are pooled into a one-dimensional sequence along the vertical direction, the one-dimensional sequence is divided along the channel dimension, and then the MS-DWConv is used to obtain vertical multi-scale features, the vertical multi-scale features are spliced, and then the GN layer and the sigmoid function are used to obtain vertical spatial mode features; S323. The horizontal spatial mode features and the vertical spatial mode features are fused to obtain a direction attention map to strengthen fine-grained spatial details; S33. The input from the C3k2 module is input into the Channel Attention mechanism module to obtain semantic channel attention; wherein, the processing process of the Channel Attention mechanism module includes: The input feature is subjected to global average pooling and a multi-layer perceptron (MLP) to obtain first channel weights; The input feature is subjected to global maximum pooling and a MLP to obtain second channel weights; The first channel weights and the second channel weights are added and subjected to a Sigmoid function to obtain a modulation coefficient; The input feature and the modulation coefficient are multiplied element by element in each channel to obtain semantic channel attention; S34. The hierarchical feature, the direction attention map, the reinforced fine-grained spatial details and the semantic channel attention are multiplied to obtain the output of the PFE module. 9.The UAV small target detection method based on lightweight multi-scale perception network according to claim 1, characterized in that, Each LES detection head comprises a first CGS module, an FE module, a second CGS module, a third CGS module and a Detect module which are sequentially cascaded; the CGS module comprises a convolution layer, a GN layer and a SiLU activation function layer which are sequentially cascaded, the convolution kernel size of the first CGS module is 1x1, and the convolution kernel size of the second CGS module and the third CGD module is 3x3; the FE module comprises two depth separable convolutions with convolution kernel sizes of 1x1 and 5x5 which are connected in parallel; wherein, the second CGS module and the third CGS module are shared structures and serve the three LES detection heads together.