A divide-and-conquer multi-attention-guided small target efficient real-time detection method

Abstract

The application relates to a divide-and-conquer multi-attention-guided small target efficient real-time detection method.The divide-and-conquer multi-attention-guided small target efficient real-time detection method comprises the following steps: S1, inputting an unmanned aerial vehicle image to be detected into a backbone network to obtain a feature image group; S2, inputting the feature image group into a neck network, and adopting a DCPP strategy to perform multi-scale feature fusion; the DCPP strategy comprises a CASA module enhanced top-down feature propagation path and an SADA module enhanced bottom-up feature propagation path; and S3, inputting the fused feature image into a detection network to output a target detection result.The detection method decouples the processing of semantic and spatial information into special paths by adopting a divide-and-conquer guiding strategy in the neck network, solves the inherent conflict, and enhances the distinguishing and positioning of small targets without increasing the calculation cost.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a divide-and-conquer multi-attention-guided method for efficient real-time detection of small targets. Background Technology

[0002] With the rapid development and widespread application of UAV technology, real-time small target detection algorithms based on UAV platform real-world scenarios have become one of the research hotspots in the field of computer vision. Unlike target images in general scenarios: (1) Images from the perspective of UAVs are affected by flight altitude, and the target size in the image is extremely small. The discriminative information such as texture, structure and color available for learning is extremely limited, and it is very easy to be confused with background noise, resulting in a lack of feature information; (2) The flexible adjustment of UAV flight altitude and attitude leads to a large change in the scale of the same type of target in the image. At the same time, the overhead view will cause target deformation, making feature extraction more difficult; (3) UAV shooting covers a wide range, and the image often contains a large number of complex background features (such as buildings, vegetation, shadows, etc.). These background elements may be visually similar to small targets, forming serious interference. In addition, dynamic shooting will also introduce problems such as motion blur and lighting changes; (4) In a high-resolution large field of view image, targets (especially small targets) are often densely distributed in certain areas (such as crossroads), while most areas of the image may be completely empty. This uneven spatial distribution not only leads to a waste of computing resources, but also causes serious occlusion problems in dense areas, further increasing the difficulty of detection.

[0003] Current detection methods primarily focus on network optimization for small target detection: employing more efficient bidirectional feature pyramid networks to achieve effective multi-scale feature fusion, introducing various attention mechanisms to enhance the discriminative power of small target features, and reducing computational cost by adding shallower prediction layers for even smaller target detection and removing deep feature layers. However, these improvements neglect the functional differences in feature information; direct fusion can lead to information conflicts, which in turn weakens the discriminative power of small target features. Therefore, effectively fusing multi-scale features within a lightweight model structure to alleviate conflicts between different information and improve the detection accuracy of small targets from a UAV perspective remains a critical technical challenge that urgently needs to be addressed. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a divide-and-conquer multi-attention-guided method for efficient real-time detection of small targets. By embedding a lightweight feature extraction module in the feature extraction stage, a parallel spatial-channel attention mechanism is used to effectively associate cross-dimensional representation features and model the changes between channels, thereby extracting discriminative features of small targets with higher accuracy and minimal computational cost and enhancing the discriminative ability between target categories. At the same time, a divide-and-conquer guidance strategy is adopted in the neck network to decouple the processing of semantic and spatial information into dedicated paths to resolve their inherent conflicts, enhance features related to high-level semantics, and suppress responses from redundant or noisy channels. In the process of preserving and enhancing detailed information, fine-grained features that are crucial for accurate localization are highlighted, thereby enhancing the differentiation and localization of small targets without increasing computational cost.

[0005] This invention proposes a divide-and-conquer multi-attention-guided method for efficient real-time detection of small targets, comprising the following steps: S1, Input the image of the UAV under test into the backbone network to obtain the feature image group; S2, the feature image group is input into the neck network, and multi-scale feature fusion is performed using the Divide-and-Conquer Propagation Path (DCPP) strategy; S3 inputs the fused feature map into the detection network and outputs the target detection result.

[0006] Specifically, in step S2, the DCPP strategy includes a top-down feature propagation path enhanced by the ChannelAttention-guided Semantic Aggregation (CASA) module and a bottom-up feature propagation path enhanced by the Spatial Attention-guided Detail Aggregation (SADA) module.

[0007] The detection method provided by this invention decouples the processing of semantic and spatial information into a dedicated path through the DCPP strategy to resolve their inherent conflicts, thereby achieving enhanced recognition of small targets without increasing significant computational costs.

[0008] Furthermore, the backbone network includes multiple feature extraction stages, some of which embed a Light-weight Feature Extraction Module (LFEM) to improve feature representation capabilities while controlling computational cost. The detection method provided by this invention embeds an LFEM module in the feature extraction stage of the backbone network. This module employs a parallel spatial-channel attention mechanism to effectively correlate cross-dimensional representations, thereby efficiently extracting discriminative features of small targets with higher accuracy and minimal computational overhead.

[0009] Furthermore, the backbone network includes a first CBS module, a second CBS module, a first LFEM module, a third CBS module, a second LFEM module, a fourth CBS module, a third LFEM module, a fifth CBS module, a fourth LFEM module, a spatial pyramid pooling module (SPPF) module, and a cross-stage local pyramid compressed attention (C2PSA) module, which are cascaded in sequence. The feature image group includes the output images of the second LFEM module, the third LFEM module, and the C2PSA module.

[0010] Furthermore, the LFEM module splits the input features into two branch paths through convolution operations. One branch path is fed into n consecutive MDFR modules for processing; the other branch path is retained and fused with the integrated feature representation output by the MDFR module.

[0011] Furthermore, the MDFR module processing procedure includes a first stage; The first stage uses a 3×3 partial convolutional layer to process the feature map corresponding to one-quarter of the total number of channels. The processed channels are then connected to the remaining original channels. The connected features are then passed through two pointwise convolutional layers to generate the main branch features. Then, the main branch features are added element-wise to the input feature map to generate the output features of the first stage.

[0012] Furthermore, the MDFR module processing also includes a second stage; The second stage divides the output features of the first stage into two independent first branches and second branches along the channel dimension. The first branch and the second branch respectively receive input features F of size H×W×C / 2, where H, W and C are the height, width and number of channels, respectively. Each branch uses a multi-scale kernel to divide the feature map into k groups along the channel dimension. In the first branch, sub-features Spatial attention module is used for processing, and the sub-features are obtained through the following calculations. Spatial attention characteristics : ; Where σ is the sigmoid activation function; In terms of spatial dimension Perform global max pooling to extract spatial statistics; and Let H and W represent the convolution weight matrix and bias of size H×W×1, respectively. In the second branch, sub-features Channel attention module is used for processing, and the sub-features are obtained through the following calculations. Channel attention features : ; in, It is in the channel dimension Perform global average pooling to aggregate channel-level statistics; and These represent the convolution weight matrix and bias, respectively, which are of size 1×1×C / 2k; Then, the first branch and the second branch generate k sets of spatial attention and channel attention feature maps, which are then reconstructed into a new integrated feature representation through shuffling and rearrangement.

[0013] Furthermore, in the top-down feature propagation path, the deep features are enhanced by the CASA module and then extracted by the LFEM module. They are then gradually passed to the shallow layers through the upsampling module and fused with the CASA module enhancement features of the corresponding shallow layers to output the perceptual channel features.

[0014] Furthermore, in the bottom-up feature propagation path, shallow features are enhanced by the SADA module and fused with the perceptual channel features output by the corresponding shallow CASA module. Then, the aggregated features are extracted and output by the LFEM module. Finally, the features are gradually passed to deeper layers through the downsampling module and fused across layers with the enhanced features of the corresponding deep SADA module and the perceptual channel features.

[0015] As an example, the top-down feature propagation path includes a first CASA module, a fifth LFEM module, a first upsampling module, a second CASA module, a sixth LFEM module, a second upsampling module, and a third CASA module cascaded in sequence; the first CASA module receives the output features of the C2PSA module; the second CASA module receives the output features of the third LFEM module and the first upsampling module; the third CASA module receives the output features of the second LFEM module and the second upsampling module. The bottom-up feature propagation path includes a first SADA module, a seventh LFEM module, a first downsampling module, a second SADA module, an eighth LFEM module, a second downsampling module, a third SADA module, and a ninth LFEM module, which are cascaded in sequence. The first SADA module receives the output features from the second LFEM module and the third CASA module. The second SADA module receives the output features from the first downsampling module, the third LFEM module, and the second CASA module. The third SADA module receives the output features from the second downsampling module, the C2PSA module, and the first CASA module.

[0016] Furthermore, the processing procedure of the CASA module includes: First, a dual-pooling channel attention module is used to process the input features by applying average pooling and max pooling operations in the spatial dimension. The pooled feature maps are first compressed through a 1×1 shared convolutional layer, and then the channel dimension is restored through another 1×1 shared convolutional layer. The two feature maps are then combined by summing element-wise, and the sigmoid function is applied to generate the channel attention weights. : ; in, This indicates max pooling. Indicates average pooling; It consists of two 1 × 1 convolutional layers and a ReLU activation function; σ is the sigmoid activation function; Then the input features and deep feature mapping The channel sensing features are obtained through the following calculations. : .

[0017] Furthermore, the processing procedure of the SADA module includes: First, a dual-pooling spatial attention module is used to aggregate spatial information from the input features. Then, the pooling graph is fused by element-wise summation, and a spatial attention weight map is generated by applying the sigmoid function. : ; in, This indicates max pooling. Indicates average pooling; σ represents a 3 × 3 convolution and the ReLU activation function; σ represents the sigmoid activation function. Then the input features Shallow features and channel sensing features The final aggregated feature is obtained through enhanced fusion using the following calculations. : .

[0018] Furthermore, the detection network performs classification prediction and bounding box regression on the fused feature map output by the neck network, and outputs the target's category, confidence level and location information to complete the target detection.

[0019] The beneficial effects of this invention are as follows: (1) A multi-scale feature aggregation network for efficient real-time detection of small targets guided by multi-attention and divide-and-conquer is proposed. It integrates a novel feature extraction module and a fusion strategy to improve the accuracy and real-time stability of small target detection in UAV applications. (2) Design a lightweight feature extraction module and adopt a parallel spatial-channel attention mechanism to effectively associate cross-dimensional representations, thereby extracting discriminative features of small targets with higher accuracy and minimal computational overhead; (3) A divide-and-conquer propagation path strategy is proposed, which decouples the processing of semantic information and spatial information into a dedicated path to resolve their inherent conflict and enhance the differentiation and localization of small targets without increasing computational cost. (4) Both paths of the divide-and-conquer propagation path strategy adopt a progressively focused fusion strategy to aggregate the refined features, which maintains the semantic consistency of multi-scale features in the top-down path and ensures the integrity of spatial details in the bottom-up propagation process, thereby improving the efficiency and fidelity of spatial information flow.

[0020] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0021] Figure 1 This invention provides a divide-and-conquer multi-attention guided architecture for efficient real-time detection of small targets; Figure 2 The LFEM module structure diagram provided by this invention; Figure 3 This is a structural diagram of the MDFR module provided by the present invention; Figure 4 The overall architecture diagram of the DCPP strategy provided by this invention; Figure 5 This is a structural diagram of the CASA module provided by the present invention; Figure 6 The diagram shows the structure of the SADA module provided by this invention. Detailed Implementation

[0022] 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.

[0023] In the description of this invention, it should be noted that the terms "vertical direction," "up," "down," and "horizontal," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0024] Traditional object detection methods typically rely on hand-designed features and sliding window classification strategies. However, these methods are limited by the expressive power of hand-designed features and often exhibit poor generalization ability and insufficient robustness when faced with small targets, occlusions, and lighting variations prevalent in UAV images, making them unsuitable for practical applications. Furthermore, while two-stage detectors based on convolutional neural networks (CNNs) possess high accuracy potential in small target detection due to region proposal and secondary fine-tuning mechanisms, their high computational cost and slow inference speed make them unsuitable for deployment on resource-constrained and real-time-critical UAV embedded platforms. In contrast, single-stage detectors based on CNNs have become the mainstream solution for real-time UAV detection due to their fast end-to-end inference capabilities. However, because small targets themselves lack sufficient texture details and semantic information, coupled with complex background interference in UAV images, existing single-stage algorithms, while balancing speed and accuracy, still face bottlenecks such as high false negative rates and inaccurate localization of small targets.

[0025] To enhance the feature representation capabilities of small targets, various existing attention-based methods, while introducing attention mechanisms to improve feature extraction, have not fundamentally solved the problem of weak feature representation capabilities for small targets, and further optimization is still needed to enhance feature discriminativeness. Existing technologies utilize feature pyramid networks with multi-feature fusion to attempt to combine deep semantic information with shallow spatial details through cross-layer feature fusion. However, existing feature fusion strategies often neglect the functional differences between the top-down semantic information flow and the bottom-up spatial information flow. Direct fusion may lead to information conflicts, insufficient semantic consistency in the top-down path, and inadequate preservation and highlighting of spatial details in the bottom-up path, thereby weakening the feature discriminativeness of small targets.

[0026] Based on this, please refer to Figure 1-6 This invention proposes a divide-and-conquer multi-attention-guided method for efficient real-time detection of small targets, comprising the following steps: S1, Input the image of the UAV under test into the backbone network to obtain the feature image group; S2, input the feature image group into the neck network, and use the divide-and-conquer propagation path DCPP strategy to perform multi-scale feature fusion; S3 inputs the fused feature map into the detection network and outputs the target detection result.

[0027] Specifically, in step S1, the backbone network includes multiple feature extraction stages, some of which embed a lightweight feature extraction (LFEM) module to improve feature representation capabilities while controlling computational load.

[0028] Specifically, in step S2, the DCPP strategy includes a top-down feature propagation path enhanced by the CASA module and a bottom-up feature propagation path enhanced by the SADA module.

[0029] This invention proposes a divide-and-conquer, multi-attention-guided, multi-scale feature aggregation network for efficient real-time detection of small targets. It integrates a novel feature extraction module and a fusion strategy to improve the accuracy and real-time stability of small target detection in UAV applications. The detection method embeds an LFEM module in the feature extraction stage of the backbone network. This module employs a parallel spatial-channel attention mechanism to effectively correlate cross-dimensional representations, thereby efficiently extracting discriminative features of small targets with higher accuracy and minimal computational overhead. Simultaneously, the proposed DCPP strategy decouples the processing of semantic and spatial information into dedicated paths to resolve their inherent conflicts, thus achieving enhanced recognition of small targets without significantly increasing computational costs.

[0030] Please see Figure 1 Furthermore, in some embodiments, the backbone network is divided into five progressive feature extraction stages. Each stage sequentially performs convolutional downsampling and feature extraction on the input image, with the number of channels increasing layer by layer and the feature map size decreasing layer by layer. Each feature extraction stage is equipped with a CBS module as a basic module. Each CBS module contains a convolutional layer, a batch normalization layer (BN layer), and an activation layer. A 3×3 standard convolutional kernel is used for spatial feature extraction. The number of input channels of the convolutional layer is consistent with the number of output channels of the previous stage, and the number of output channels increases gradually according to the stage level. The batch normalization layer is used to normalize the convolutional output features, accelerate network training convergence, and avoid gradient vanishing. The activation layer uses the ReLU or SiLU activation function to achieve nonlinear transformation of features and enhance the feature representation capability of the network.

[0031] Furthermore, to improve the effectiveness of feature extraction, a lightweight feature enhancement LFEM module is embedded within each feature extraction stage starting from the second feature extraction stage. The residual structure adopts a cascaded combination of "convolution-normalization-activation". The LFEM module is used to further refine the features output by the residual structure, which not only ensures the high-resolution localization information of shallow features, but also realizes the strong semantic expression capability of deep features. At the same time, by embedding the lightweight module, the discriminativeness of features is improved while controlling the network complexity.

[0032] As an example, the backbone network includes a first CBS module, a second CBS module, a first LFEM module, a third CBS module, a second LFEM module, a fourth CBS module, a third LFEM module, a fifth CBS module, a fourth LFEM module, a Spatial Pyramid Pooling Module (SPPF), and a Cross-Stage Local Pyramid Compressed Attention (C2PSA) module, all cascaded in sequence. The feature image set includes the output image F3 of the second LFEM module, the output image F4 of the third LFEM module, and the output image F5 of the C2PSA module.

[0033] Please see Figure 2 and Figure 3 Furthermore, the LFEM module splits the input features into two branch paths through convolution operations. One branch path is fed into n consecutive MDFR modules for processing; the other branch path is retained and fused with the integrated feature representation output by the MDFR module. In this way, the lightweight feature extraction module LFEM enhances the fusion of multi-channel information, thereby improving the discriminative feature extraction capability for small targets.

[0034] Furthermore, the MDFR module processing procedure includes a first stage and a second stage.

[0035] Specifically, the first-stage residual structure begins with a partially convolutional (PConv) layer using a 3 × 3 kernel, which processes the feature map corresponding to one-quarter of the total channels. The processed channels are then concatenated with the remaining original channels to maintain consistency in the input and output channel dimensions. The concatenated features are then passed through two pointwise convolutional (PWConv) layers to generate the main branch features. Finally, the main branch features are element-wise added to the input feature map to generate the first-stage output features. This efficient design not only facilitates comprehensive integration of channel information but also has extremely low computational cost.

[0036] Specifically, the second stage designs a multi-attention coordination mechanism that complements parallel spatial attention and channel attention mechanisms. The output features of the first stage are divided into two independent branches along the channel dimension: a first branch and a second branch. Each branch receives an input feature F of size H×W×C / 2, where H, W, and C are the height, width, and number of channels, respectively. Each branch uses a multi-scale kernel to divide the feature map into k groups along the channel dimension, denoted as F1, F2, ..., F... k .

[0037] In the first branch, each subgroup F i Processed by the spatial attention module to capture contextual information of small targets, highlighting the importance weights of relevant spatial regions. Corresponding to sub-features. Spatial attention characteristics The following calculation was performed: ; Where σ is the sigmoid activation function; In terms of spatial dimension Perform global max pooling to extract spatial statistics; and These represent the convolution weight matrix and bias of size H×W×1, respectively, which help in the representation of the coding space.

[0038] The second branch employs a channel attention mechanism to model the inter-channel dependencies related to small targets, and obtains the sub-features through the following calculations. Channel attention features : ; in, It is in the channel dimension Perform global average pooling to aggregate channel-level statistics; and Let represent the convolution weight matrix and bias of size 1×1×C / 2k, respectively.

[0039] Then, the first and second branches generate k sets of spatial attention and channel attention feature maps, each with dimensions H × W × C / 2k. These feature maps are then shuffled and rearranged to ensure effective feature redistribution, and then reconstructed into a new integrated feature representation. Each MDFR module effectively associates cross-dimensional representation features through this parallel spatial-channel attention mechanism, which can effectively aggregate small target features while suppressing redundant background noise and interference.

[0040] Understandably, the LFEM module divides the input feature map into two equal branches. The first branch is used to efficiently capture the spatial distribution information of the target, while the second branch focuses on modeling the feature variations between channels. By fusing features from two dimensions at each pixel location, the LFEM module not only enhances the discriminative representation of small target features but also effectively improves computational efficiency.

[0041] Please see Figure 1 and Figure 4-6 Furthermore, in some embodiments, a Divide-and-Conquer Propagation Path (DCPP) strategy is proposed in the neck network. This DCPP strategy aims to efficiently perform multi-scale feature fusion and enhance the complementary advantages of the two types of information flows without introducing additional computational overhead by decoupling the processing paths of semantic information and spatial details.

[0042] Specifically, the DCPP strategy guides the progressive aggregation of contextual semantics and spatial details through dedicated attention mechanisms. In the top-down feature propagation path, a channel attention-guided semantic aggregation module (CASA) is introduced, focusing on the transmission and enhancement of semantic information to ensure high semantic consistency across multi-scale features, thereby improving the discriminative power of the fused features. In the bottom-up feature propagation path, a spatial attention-guided detail aggregation module (SADA) is used, focusing on the refinement and preservation of local spatial features, achieving layer-by-layer aggregation and enhancement of detailed information. Through this divide-and-conquer, dual-path design, the DCPP strategy can efficiently propagate contextual semantic information while accurately preserving spatial detail features, significantly enhancing the robustness and expressive power of multi-scale feature fusion. Ultimately, this strategy helps the model to more comprehensively understand complex scenes and achieve superior performance in object detection tasks at different scales.

[0043] In a specific implementation, in the top-down feature propagation path, deep features are enhanced by the CASA module, extracted by the LFEM module, and then progressively propagated to shallower layers through the upsampling module. These features are then fused across layers with the enhanced features from the corresponding shallow CASA module to output perceptual channel features. In the bottom-up feature propagation path, shallow features are enhanced by the SADA module and fused with the perceptual channel features output from the corresponding shallow CASA module. These features are then extracted by the LFEM module and output as aggregated features. Finally, they are progressively propagated to deeper layers through the downsampling module and fused across layers with the enhanced features from the corresponding deep SADA module and perceptual channel features.

[0044] Furthermore, the CASA module's core mechanism dynamically recalibrates channel-level feature responses to highlight semantically rich information and suppress noise. Simultaneously, through a stepwise aggregation strategy, this module enhances cross-scale semantic alignment and improves the discriminative power of fused features. The processing procedure of the CASA module includes: First, a dual-pooling channel attention module is used to process the input features by applying average pooling and max pooling operations in the spatial dimension. To minimize computational overhead, the pooled feature maps are first compressed through a shared 1×1 convolutional layer, reducing the channel dimension to 1 / 16 of the original feature map. Then, the channel dimension is restored through another 1×1 shared convolutional layer. The two resulting feature maps are combined by element-wise summation, and the sigmoid function is applied to generate the channel attention weights. : ; in, This indicates max pooling. Indicates average pooling; It consists of two 1 × 1 convolutional layers and a ReLU activation function; σ is the sigmoid activation function.

[0045] Then the CASA module obtains the attention weights. Input features Element-wise multiplication yields channel refinement features ; by complementary weights Deep feature mapping Modulation is performed to generate residual features. ;at last, and The channel-aware features are obtained by addition and calculation as follows. : .

[0046] Furthermore, the core mechanism of the SADA module computes spatial attention weights to model positional importance, thereby suppressing background noise while highlighting discriminative features crucial for target localization. In addition, a stepwise ensemble strategy is employed to aggregate these spatially refined features, improving the efficiency and fidelity of spatial detail propagation within the network. To preserve fine-grained spatial details, the input features of this SADA module are directly derived from the multi-level outputs of the backbone network. The processing steps of this SADA module include: First, a dual-pooling spatial attention module is used to aggregate spatial information from the input features through max pooling and average pooling. Then, the pooling graphs are fused by element-wise summation, and a spatial attention weight graph is generated by applying the sigmoid function. : ; in, This indicates max pooling. Indicates average pooling; σ represents a 3 × 3 convolution and the ReLU activation function; σ represents the sigmoid activation function. The SADA module then uses the obtained attention weights to spatially refine the input features, generating spatially refined features. Meanwhile, features from shallow layers Through complementary weights Forming residual characteristics This preserves the detailed features that spatial attention might suppress. The final aggregated features... These spatial refinement features are calculated using the following methods. residual characteristics and channel sensing features Enhanced integration: .

[0047] Understandably, in the top-down propagation path, the CASA module ensures semantic consistency among multi-scale features, thereby fully leveraging semantic modeling capabilities. Correspondingly, in the bottom-up path, the SADA module focuses on the spatial dimension of the feature map, preserving and enhancing detailed information while highlighting fine-grained features crucial for accurate localization. Furthermore, both paths employ a progressively focusing fusion strategy to aggregate the refined features, maintaining semantic consistency among multi-scale features in the top-down path while ensuring the integrity of spatial details during bottom-up propagation, thus improving the efficiency and fidelity of spatial information flow. Through this separate yet complementary bidirectional propagation mechanism, the detection method of this application embodiment achieves significant accuracy improvement in small target detection tasks in UAV images with complex backgrounds.

[0048] As an example, the top-down feature propagation path includes a first CASA module, a fifth LFEM module, a first upsampling module, a second CASA module, a sixth LFEM module, a second upsampling module, and a third CASA module cascaded in sequence; the first CASA module receives the output features of the C2PSA module; the second CASA module receives the output features of the third LFEM module and the first upsampling module; and the third CASA module receives the output features of the second LFEM module and the second upsampling module.

[0049] The bottom-up feature propagation path includes a cascaded first SADA module, a seventh LFEM module, a first downsampling module, a second SADA module, an eighth LFEM module, a second downsampling module, a third SADA module, and a ninth LFEM module. The first SADA module receives the output features from the second LFEM module and the third CASA module, fuses them, and outputs an aggregated feature, which is then extracted by the seventh LFEM module and output as P3. The second SADA module receives the output features from the first downsampling module, the third LFEM module, and the second CASA module, fuses them, and outputs an aggregated feature, which is then extracted by the eighth LFEM module and output as P4. The third SADA module receives the output features from the second downsampling module, the C2PSA module, and the first CASA module, fuses them, outputs an aggregated feature, which is then extracted by the ninth LFEM module and output as P5.

[0050] Furthermore, the P3, P4, and P5 multi-scale fused feature maps output by the neck network are input into the detection network. The detection network then performs classification prediction and bounding box regression on these fused feature maps, outputting the target's category, confidence level, and location information to complete the target detection.

[0051] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows: (1) A multi-scale feature aggregation network for efficient real-time detection of small targets guided by multi-attention and divide-and-conquer is proposed. It integrates a novel feature extraction module and a fusion strategy to improve the accuracy and real-time stability of small target detection in UAV applications. (2) Design a lightweight feature extraction module and adopt a parallel spatial-channel attention mechanism to effectively associate cross-dimensional representations, thereby extracting discriminative features of small targets with higher accuracy and minimal computational overhead; (3) A divide-and-conquer propagation path strategy is proposed, which decouples the processing of semantic information and spatial information into a dedicated path to resolve their inherent conflict and enhance the differentiation and localization of small targets without increasing computational cost. (4) Both paths of the divide-and-conquer propagation path strategy adopt a progressively focused fusion strategy to aggregate the refined features, which maintains the semantic consistency of multi-scale features in the top-down path and ensures the integrity of spatial details in the bottom-up propagation process, thereby improving the efficiency and fidelity of spatial information flow.

[0052] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A divide-and-conquer method for efficient real-time detection of small targets guided by multiple attention, characterized in that, Includes the following steps: S1, Input the image of the UAV under test into the backbone network to obtain the feature image group; S2, input the feature image group into the neck network and use the DCPP strategy to perform multi-scale feature fusion; The DCPP strategy includes a top-down feature propagation path enhanced by the CASA module and a bottom-up feature propagation path enhanced by the SADA module. S3 inputs the fused feature map into the detection network and outputs the target detection result.

2. A divide-and-conquer method for efficient real-time detection of small targets guided by multiple attention, as described in claim 1, characterized in that: The backbone network includes multiple feature extraction stages, some of which embed LFEM modules to improve feature representation capabilities while controlling computational load.

3. A divide-and-conquer multi-attention-guided efficient real-time detection method for small targets according to claim 2, characterized in that: The backbone network includes a first CBS module, a second CBS module, a first LFEM module, a third CBS module, a second LFEM module, a fourth CBS module, a third LFEM module, a fifth CBS module, a fourth LFEM module, an SPPF module, and a C2PSA module, which are cascaded in sequence. The feature image group consists of the feature images output by the second LFEM module, the third LFEM module, and the C2PSA module.

4. The efficient real-time detection method for small targets guided by multi-attention using divide-and-conquer as described in claim 2, characterized in that: The LFEM module splits the input features into two branch paths through convolution operations. One branch path is fed into n consecutive MDFR modules for processing; the other branch path is retained and fused with the integrated feature representation output by the MDFR module.

5. The divide-and-conquer multi-attention guided efficient real-time detection method for small targets according to claim 4, characterized in that: The MDFR module processing procedure includes a first stage; The first stage uses a 3×3 partial convolutional layer to process the feature map corresponding to one-quarter of the total number of channels. The processed channels are then connected to the remaining original channels. The connected features are then passed through two pointwise convolutional layers to generate the main branch features. Then, the main branch features are added element-wise to the input feature map to generate the output features of the first stage.

6. The efficient real-time detection method for small targets guided by multi-attention using a divide-and-conquer approach, as described in claim 5, is characterized in that: The MDFR module processing also includes a second stage; The second stage divides the output features of the first stage into independent first and second branches along the channel dimension. The first and second branches respectively receive input features F of size H×W×C / 2, where H, W, and C are the height, width, and number of channels, respectively. Each branch uses a multi-scale kernel to divide the feature map into k groups along the channel dimension. In the first branch, sub-features Spatial attention module is used for processing, and the sub-features are obtained through the following calculations. Spatial attention characteristics : ； Where σ is the sigmoid activation function; In terms of spatial dimension Perform a global max pooling operation; and Let H and W represent the convolution weight matrix and bias of size H×W×1, respectively. In the second branch, sub-features Channel attention module is used for processing, and the sub-features are obtained through the following calculations. Channel attention features : ； in, It is in the channel dimension Perform global average pooling; and These represent the convolution weight matrix and bias, respectively, which are of size 1×1×C / 2k; Then, the first branch and the second branch generate k sets of spatial attention and channel attention feature maps, which are subsequently reconstructed into a new integrated feature representation.

7. The efficient real-time detection method for small targets guided by multi-attention using a divide-and-conquer approach according to claim 1, characterized in that: In the top-down feature propagation path, the deep features are enhanced by the CASA module and then extracted by the LFEM module. They are then gradually passed to the shallow layers through the upsampling module and fused with the CASA module enhancement features of the corresponding shallow layers to output the perceptual channel features. In the bottom-up feature propagation path, shallow features are enhanced by the SADA module and fused with the perceptual channel features output by the corresponding shallow CASA module. Then, the aggregated features are extracted and output by the LFEM module. Finally, the features are gradually passed to deeper layers through the downsampling module and fused with the enhanced features of the corresponding deep SADA module and the perceptual channel features across layers.

8. The efficient real-time detection method for small targets guided by multi-attention using a divide-and-conquer approach, as described in claim 7, is characterized in that: The top-down feature propagation path includes a first CASA module, a fifth LFEM module, a first upsampling module, a second CASA module, a sixth LFEM module, a second upsampling module, and a third CASA module, which are cascaded in sequence. The first CASA module receives the output features of the C2PSA module; the second CASA module receives the output features of the third LFEM module and the first upsampling module; the third CASA module receives the output features of the second LFEM module and the second upsampling module. The bottom-up feature propagation path includes a first SADA module, a seventh LFEM module, a first downsampling module, a second SADA module, an eighth LFEM module, a second downsampling module, a third SADA module, and a ninth LFEM module, which are cascaded in sequence. The first SADA module receives the output features of the second LFEM module and the third CASA module; the second SADA module receives the output features of the first downsampling module, the third LFEM module and the second CASA module; the third SADA module receives the output features of the second downsampling module, the C2PSA module and the first CASA module.

9. The divide-and-conquer multi-attention guided efficient real-time detection method for small targets according to claim 1, characterized in that, The processing procedure of the CASA module includes: First, a dual-pooling channel attention module is used to process the input features by applying average pooling and max pooling operations in the spatial dimension. The pooled feature maps are first compressed through a 1×1 shared convolutional layer, and then the channel dimension is restored through another 1×1 shared convolutional layer. The two feature maps are then combined by element-wise summation, and the sigmoid function is applied to generate the channel attention weights. : ； in, This indicates max pooling. Indicates average pooling; It consists of two 1 × 1 convolutional layers and a ReLU activation function; σ is the sigmoid activation function; Then the input features and deep feature mapping The channel sensing features are obtained through the following calculations. : 。 10. The divide-and-conquer multi-attention guided efficient real-time detection method for small targets according to claim 9, characterized in that: The processing procedure of the SADA module includes: First, a dual-pooling spatial attention module is used to aggregate spatial information from the input features. Then, the pooling graph is fused by element-wise summation, and a spatial attention weight map is generated by applying the sigmoid function. : ； in, This indicates max pooling. Indicates average pooling; σ represents a 3 × 3 convolution and the ReLU activation function; σ represents the sigmoid activation function. Then the input features Shallow features and channel sensing features The final aggregated feature is obtained through enhanced fusion using the following calculations. : 。

Images