Re-parameterization unmanned aerial vehicle target detection method based on multi-core fusion and omnidirectional connection

Abstract

The invention discloses a re-parameterization unmanned aerial vehicle target detection method based on multi-core fusion and omnidirectional connection, and the method comprises the following steps: obtaining an unmanned aerial vehicle aerial image data set, and constructing a detection model comprising a backbone network, a feature fusion network and a deformable task decoupling detection head; the backbone network extracts multi-scale and multi-direction features by six parallel paths in a training stage through a wide-branch re-parameterization convolution module, and re-parameterization is carried out in a reasoning stage to obtain single-path convolution; the feature fusion network performs down-sampling through a spatial deep convolution module and reduces spatial information loss, and realizes cross-stage omnidirectional feature interaction and double attention enhancement in combination with an omnidirectional kernel cross-stage partial connection module; and the deformable task decoupling detection head decouples the classification and regression features, optimizes feature expression, weights the classification features and then performs aggregation decoding. The model is trained to be used for a test set to output a detection result, detection precision and reasoning efficiency can be balanced, and the model adapts to a complex aerial photography scene of an unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to a reparameterized UAV target detection method based on multi-core fusion and omnidirectional connectivity, belonging to the field of computer vision. Background Technology

[0002] Intelligent video surveillance systems, as a key component of the public safety management system, play an irreplaceable role in traffic control, security patrols, and emergency rescue. With the rapid development of drone technology and aerial imaging technology, drones, with their advantages of flexibility, wide coverage, and convenient deployment, have become the core carrier for acquiring high-altitude scene data in intelligent video surveillance systems, enabling real-time dynamic monitoring of areas such as urban building complexes, transportation hubs, and large-scale event sites. However, drone aerial photography scenarios are characterized by large differences in target scale, complex background environments, and unstable target motion states, posing significant challenges to drone target detection. Simultaneously, the limited computing resources and power consumption of drone platforms place higher demands on the inference speed and lightweight nature of detection models. Traditional target detection methods relying on manual annotation and fixed-structure networks are insufficient to fully meet the dual requirements of high precision and high efficiency in real-time drone monitoring scenarios. Therefore, intelligent target detection technology for drone scenarios has become a research hotspot and challenge in the field of computer vision.

[0003] The core requirement of UAV target detection lies in achieving accurate identification and localization of multi-scale targets in complex aerial photography scenarios, while ensuring efficient inference capabilities of the model on embedded hardware. Existing UAV target detection methods are mostly based on traditional convolutional neural networks (CNNs) to construct feature extraction and detection frameworks, such as using single-stage or two-stage detection models. However, the fixed receptive field of traditional CNNs has limitations in adapting to the dynamic changes in target scale within UAV scenarios, leading to insufficient feature representation for small targets and a high risk of missed or false detections. Furthermore, traditional feature fusion modules often employ simple upsampling and concatenation operations, failing to fully exploit the spatial information and channel relationships of feature maps, resulting in poor semantic consistency of the fused features and further impacting detection accuracy. In addition, to improve model performance, existing methods often enhance feature representation capabilities by increasing network depth and width and introducing complex attention mechanisms. However, this leads to a sharp increase in the number of model parameters and computational load, making real-time inference on embedded UAV platforms difficult and severely limiting their deployment and application in practical UAV monitoring scenarios.

[0004] To address the core issue of balancing accuracy and efficiency in traditional UAV target detection methods, researchers have proposed structural reparameterization techniques. These techniques enhance feature extraction capabilities by constructing multi-branch parallel structures during training, and then reparameterize these structures into single-path convolutions during inference, achieving both strong training performance and high inference efficiency. For example, some methods design dual-branch reparameterized convolution modules, executing standard convolutions and depthwise convolutions in parallel during training, and merging them into single-path convolutions during inference to reduce computational cost. Other methods introduce cross-stage partial connection structures to improve feature fusion efficiency through multi-path feature interaction. However, existing reparameterization methods often limit branch design to single-scale or single-dimensional feature extraction, failing to fully utilize multi-scale and multi-directional feature information, resulting in insufficient adaptability of the model to multi-scale targets in UAV scenarios. Furthermore, the feature fusion stage lacks coordinated optimization of spatial and channel information, easily losing detailed features of small targets, and the cross-stage connection methods fail to achieve omnidirectional feature interaction, further limiting the expressive power of fused features. Furthermore, existing detection heads mostly adopt a design pattern of shared features and unified prediction, which fails to decouple the features of classification tasks and regression tasks. This leads to conflict between the optimization objectives of the two tasks, affecting the accuracy of detection box localization and category judgment, and making it difficult to meet the needs of accurate multi-target detection in UAV scenarios.

[0005] In the head design and target prediction stages, existing UAV target detection methods often ignore the inherent differences between classification and regression tasks, using the same feature mapping for both types of tasks. This leads to inconsistent expression preferences between classification and regression features. For example, classification tasks require strong semantic features to distinguish target categories, while regression tasks require precise spatial features for localization. This task confusion severely impacts detection performance. While some methods attempt to decouple features through task decomposition modules, they fail to specifically optimize for the spatial offset problem of regression features and the channel redundancy problem of classification features. For instance, the spatial alignment of regression features relies on fixed convolution kernels, making it difficult to adapt to target deformation and positional shifts; the channel weight allocation of classification features lacks a dynamic adjustment mechanism, failing to highlight key category information and resulting in low confidence in category prediction. Furthermore, existing methods often employ simple non-maximum suppression operations for the aggregation and decoding of multi-scale detection results, failing to consider the distribution differences of features at different scales. This easily leads to overlapping detection boxes or scale mismatches, further reducing the detection accuracy of multi-scale targets in UAV scenarios.

[0006] In summary, current UAV target detection methods still have significant shortcomings in the three core stages of feature extraction, feature fusion, and detection prediction: First, the reparameterization structure design in the feature extraction stage is simplistic, and multi-scale feature mining is insufficient, making it difficult to adapt to the dynamic changes in target scale in UAV scenarios; second, the feature fusion stage lacks coordinated optimization of spatial and channel information, and its omnidirectional feature interaction capability is insufficient, resulting in weak semantic expression of fused features; third, the detection prediction stage fails to achieve accurate alignment between classification and regression tasks, and the multi-scale result aggregation method is simplistic, affecting detection accuracy and stability. Therefore, a reparameterization UAV target detection method with multi-core fusion and omnidirectional connectivity is urgently needed to meet the application requirements of practical UAV monitoring scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a reparameterized UAV target detection method based on multi-core fusion and omnidirectional connectivity, comprising the following steps:

[0008] This invention is implemented using the following scheme:

[0009] S1: Obtain the drone aerial image dataset;

[0010] S2: Preprocess the drone dataset;

[0011] S3: Establish a UAV image target detection model, which includes a backbone network, a feature fusion network, and a deformable task decoupled detection head cascaded in sequence; wherein the backbone network outputs multi-scale features, the feature fusion network fuses the multi-scale features, and the deformable task decoupled detection head outputs detection boxes and class probabilities based on the fused features.

[0012] S3.1: In the feature extraction stage, the wide-branch reparameterized convolution module is used to dynamically expand the receptive field of the feature extraction network for feature representation through parallel multi-scale feature extraction and structural reparameterization mechanism; the wide-branch reparameterized convolution module extracts multi-scale information through parallel structure in the training stage; in the inference stage, the parallel structure is reparameterized into single-path convolution and the features are output.

[0013] S3.2: In the feature fusion stage, the spatial depth convolution module is used to downsample the feature map and retain complete spatial information; the omnidirectional kernel cross-stage partial connection module is used to perform cross-stage partial connection fusion on the downsampled features. The omnidirectional kernel cross-stage partial connection module includes a frequency domain channel attention module, a spatial channel attention module, a multi-directional depth convolution module and a feature fusion module, and outputs fused features.

[0014] S3.3: In the prediction stage, the deformable task decoupling detection head is used to feed multi-scale features into a shared convolution in sequence. After extracting the basic features, the task decomposition module is used to generate class features and regression features respectively. The dynamic deformable convolution is used to spatially align the regression features, and the spatial channel attention module is used to weight the class features. The fused regression results and class results are scaled by the scale scaling module, and then aggregated and decoded by the distribution aggregation and decoding module to output the final detection box and class probability.

[0015] S4: Train the model on the training set and perform iterative training;

[0016] S5: Apply the trained weights to the test set and output the detection results.

[0017] Furthermore, the wide-branch reparameterized convolution module includes six parallel processing paths during the training phase. The first path performs a standard k×k convolution followed by batch normalization. The second path performs a 1×1 convolution followed by batch normalization when the number of groups is less than the number of output channels. The third path selects its processing flow based on the number of groups: when the number of groups is less than the number of output channels, it sequentially performs 1×1 convolution, padding batch normalization, average pooling, and batch normalization; otherwise, it performs average pooling followed by batch normalization. The four paths select either identity 1×1 convolution or ordinary 1×1 convolution based on channel conditions, and then sequentially perform batch normalization, k×k convolution, and batch normalization operations; the fifth path performs vertical asymmetric convolution followed by batch normalization; the sixth path performs horizontal asymmetric convolution followed by batch normalization; the output features of the six paths are summed and fused, and then processed by a nonlinear activation function; during the inference stage, the operations of the six paths are merged into a single-path convolution using structural reparameterization technology, and then processed by a nonlinear activation function.

[0018] Furthermore, the spatial depth convolution module is composed of a spatial splitting module, a channel concatenation module, and a convolutional fusion module connected in sequence: the spatial splitting module extracts four sub-feature maps from the input feature map according to the row and column odd / even indices; the channel concatenation module concatenates the four sub-feature maps along the channel dimension to form an intermediate feature map with four times the number of channels and half the spatial size; the convolutional fusion module performs a 3×3 convolution on the intermediate feature map, outputs a feature map with a specified number of channels, completes downsampling, and retains complete spatial information.

[0019] Furthermore, the omnidirectional kernel cross-stage partial connection module consists of a feature splitting convolution module, an omnidirectional kernel feature extraction module, and a feature fusion convolution module; wherein, the omnidirectional kernel feature extraction module consists of an input 1×1 convolutional layer, a multi-directional depth convolution module, a frequency domain channel attention module, a spatial channel attention module, a feature gating module, and an output 1×1 convolutional layer.

[0020] Furthermore, the deformable task decoupling detection head consists of the following core modules: a shared convolution module, a task decomposition module, the task decomposition module including a classification task decomposition module and a regression task decomposition module; a spatial offset generation module, a dynamic deformable convolution module, a classification probability generation module, an output convolution module, a scale scaling module, and a distribution aggregation and decoding module.

[0021] Furthermore, the omnidirectional kernel cross-stage partial connection module divides the input into a main path and a shortcut path proportionally through the feature splitting convolution module. After the main path is processed by the input 1×1 convolution layer, multi-directional feature extraction is performed through the multi-directional depth convolution module. Then, dual attention enhancement is performed through the frequency domain channel attention module and the spatial channel attention module. After processing by the feature gating module, it is finally concatenated with the shortcut path features and then output as a fused feature through a 1×1 convolution.

[0022] Furthermore, the deformable task decoupled detection head first extracts multi-scale basic features through a shared convolution module; the task decomposition module decouples the features into classification features and regression features; the spatial offset generation module generates dynamic deformable convolution parameters; the dynamic deformable convolution module achieves spatial alignment of regression features; the spatial attention weight generation module generates spatial attention weights and weights the classification features; the output convolution module generates classification scores and regression offsets; the scale scaling module adaptively adjusts the feature scale; and the distribution aggregation and decoding module aggregates the feature distribution and decodes it into the final detection box coordinates and class probabilities.

[0023] The beneficial effects of this invention are:

[0024] This invention proposes a reparameterized UAV target detection method based on multi-core fusion and omnidirectional connectivity. This method constructs a UAV image target detection model comprising feature extraction, feature fusion, and prediction stages to detect targets in UAV aerial images. In the feature extraction stage, a wide-branch reparameterized convolutional module is designed. During training, six parallel paths are used to extract multi-scale and multi-directional features, expanding the network's receptive field to enhance feature representation. In the inference stage, the parallel structure is reparameterized into a single-path convolution to ensure detection accuracy and improve inference efficiency. In the feature fusion stage, a spatial depth convolutional module is used, achieving downsampling through spatial partitioning, channel concatenation, and convolutional fusion to preserve... This invention preserves complete spatial information of feature maps, avoiding detail loss caused by traditional downsampling. It also integrates multi-directional deep convolution and dual attention mechanisms with an omnidirectional kernel cross-stage partial connection module to achieve cross-stage omnidirectional feature interaction, improving the semantic consistency and discriminative ability of fused features. In the prediction stage, the invention designs a deformable task-decoupled detection head. A task decomposition module decouples classification and regression features, while dynamic deformable convolution optimizes the spatial alignment accuracy of regression features. A channel attention module strengthens key information of classification features, and a scale scaling and distribution aggregation decoding module integrates multi-scale detection results, improving the detection accuracy and localization accuracy of multi-scale targets in UAV scenarios. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating the overall implementation of this invention.

[0027] Figure 2 This is a framework diagram of the overall network model in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the structure of the wide-branch reparameterized convolution module in an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of the omnidirectional core cross-stage connection module structure in an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of the deformable task decoupling detection head structure in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be further described below with reference to the accompanying drawings, which are not intended to limit the present invention.

[0032] like Figure 1As shown, the present invention provides a reparameterized UAV target detection method based on multi-core fusion and omnidirectional connectivity, comprising the following steps:

[0033] Step 1: Obtain the UAV target detection dataset, preprocess it, and input it into the built model;

[0034] Step 2: Dynamically expand the receptive field through a wide-branch reparameterized convolution module to enhance the representation of target features;

[0035] Step 3: Decompose the spatial dimensions of the feature map using the spatial depth convolution module, concatenate the channels, and then perform convolutional fusion to preserve complete spatial details;

[0036] Step 4: Integrate multi-directional deep convolution and dual attention through the omnidirectional kernel cross-stage partial connection module to achieve cross-stage omnidirectional feature interaction and optimize semantic consistency and discriminative power;

[0037] Step 5: First, separate the classification and regression branches through the variable task decoupling module, decouple the features using channel attention and dynamic deformable convolution, and then integrate the multi-scale results through scale scaling and distribution aggregation to output the target category and location.

[0038] Furthermore, in step 1, the VisDrone and UAVDT dataset images are resized to a uniform size and their label formats are converted. Specifically, the VisDrone and UAVDT dataset images are scaled to a 640×640 pixel block size for the network input, while maintaining the original image aspect ratio. Sides with shorter sides less than 640 pixels are padded with a grayscale value of 114. Additionally, the original label files are converted to YOLO format .txt files.

[0039] Furthermore, step 1 establishes a UAV target detection model: this model consists of a cascaded backbone network, a feature fusion network, and a deformable task decoupling detection head. The backbone network utilizes a wide-branch reparameterized convolutional module to dynamically expand the receptive field of the feature extraction network for feature representation through parallel multi-scale feature extraction and structural reparameterization mechanisms. During training, the wide-branch reparameterized convolutional module extracts multi-scale information through parallel structures. During inference, it reparameterizes the parallel structures into single-path convolutions and outputs features. In the feature fusion stage, a spatial depth convolutional module downsamples the feature maps while preserving complete spatial information. An omnidirectional kernel cross-stage partial connection module performs cross-stage partial connection fusion on the downsampled features to enhance the representational ability of multi-scale features, thereby outputting fused features. In the prediction stage, a deformable task decoupling detection head sequentially feeds multi-scale features into a shared convolution. After extracting basic features, a task decomposition module generates categorical and regression features respectively. Dynamically deformable convolutions spatially align the regression features, and a spatial channel attention module weights the categorical features. The fused regression and categorical results are scaled by a scaling module, then aggregated and decoded by a distribution aggregation and decoding module to output the final detection box and category probability.

[0040] like Figure 2 As shown, in step S3.1, the wide-branch reparameterized convolution module extracts and fuses multi-scale and multi-morphological features of the input features through six parallel processing paths. These six paths are a standard k×k convolution path, a 1×1 convolution path, an average pooling path, 1×1 and k×k sequential convolution paths (conditional convolution), a vertical asymmetric convolution path, and a horizontal asymmetric convolution path. The output of each path is batch normalized. Subsequently, the output features of the six paths are summed and fused element-wise, and then processed by a nonlinear activation function to obtain the final output. During the inference stage, the above multi-branch parallel structure is equivalently converted into a single-path standard convolution operation through structural reparameterization technology, thereby significantly improving inference efficiency. The implementation definition process of this module is as follows:

[0041] Branch1 = BN(Conv_k×k(x))

[0042] Branch2 = BN(Conv_1×1(x))

[0043] Branch3 = IF(G <C,

[0044] BN(AvgPool(BN(Conv_1×1(x)))),

[0045] BN(AvgPool(x)) )

[0047] Branch4=IF(internal_channels_1x1_3x3==in_channels,

[0048] BN(Conv_k×k(Pad(BN(Conv_1×1_id(x))))),

[0049] BN(Conv_k×k(BN(Pad(Conv_1×1(x))))) )

[0051] Branch5=BN(Conv_vertical_asym(Crop(x)))

[0052] Branch6=BN(Conv_horizontal_asym(Crop(x)))

[0053] Fusion=Activation(Branch1+Branch2+Branch3+Branch4+Branch5+Branch6)

[0054] Output_infer=Activation(Repara_Conv(x))

[0055] Where: BN represents batch normalization, Conv_k×k represents standard k×k convolution, Conv_1×1 represents 1×1 convolution, Conv_1×1_id represents identity 1×1 convolution, C is the number of output channels, G is the number of convolution groups, Conv_vertical represents vertical asymmetric convolution, Conv_horizontal represents horizontal asymmetric convolution, AvgPool represents average pooling, Pad represents padding operation, Crop represents input pruning operation, Activation represents Silu nonlinear activation function, Repara_Conv represents single-path convolution after structure reparameterization, Output_train is the output of the training phase, and Output_infer is the output of the inference phase.

[0056] like Figure 3 As shown, in step S3.2, the spatial depthwise convolution module downsamples the input feature map through spatial recombination and convolutional fusion operations. Specifically, this module first uses a spatial splitting operation to evenly divide the input feature map into four sub-feature maps according to the row, column, parity, and evenness indices of the pixel positions; then, it uses a channel concatenation operation to concatenate the four sub-feature maps along the channel dimension to form the original feature map with the number of channels equal to the original feature map. Figure 4 Multiples, spatial dimensions are the original features Figure 1The intermediate feature map is represented in half; finally, a 3×3 convolution is performed on the intermediate feature map through a convolution fusion operation to fuse cross-channel information and output a feature map with a specified number of channels, thereby achieving a 2x downsampling while fully preserving the spatial structure information of the original feature map. The implementation definition process of this module is as follows:

[0057] Sub_feat1,Sub_feat2,Sub_feat3,Sub_feat4=Space_Split(x)

[0058] Cat_feat=Channel_Concat(Sub_feat1,Sub_feat2,Sub_feat3,Sub_feat4)

[0059] Output = Conv3 × 3(Cat_feat)

[0060] Sub_feat1 to Sub_feat4 represent the four sub-feature maps obtained after spatial splitting, respectively. Space_Split represents the spatial splitting operation based on row and column odd / even indices. Channel_Concat represents the channel dimension concatenation operation. Conv3×3 represents the convolution operation with a kernel size of 3×3. Output is the output feature map of this module.

[0061] like Figure 4 As shown, further, in step S3.2, the omnidirectional core cross-stage partial connection module processes the input feature map x according to the following process:

[0062] Split_feat=Conv1×1_split(x)

[0063] Main_path=OmniKernel(Split_feat[0:e])

[0064] Skip_path = Split_feat[e:]

[0065] Merge_feat=Concat(Main_path,Skip_path)

[0066] Output=Conv1×1_fuse(Merge_feat)

[0067] like Figure 4As shown, in step S3.2, the omnidirectional kernel cross-stage partial connection module performs omnidirectional perception and fusion of input features through a cross-stage partial connection structure. The module first uses a 1×1 convolution to divide the input features in the channel dimension into a main path and a shortcut path by a scaling factor e; the shortcut path is directly passed to the subsequent fusion stage, while the main path is processed by the omnidirectional kernel feature extraction module. The omnidirectional kernel feature extraction module consists of an input 1×1 convolutional layer, a multi-directional depth convolution module, a frequency domain channel attention module, a spatial channel attention module, and a feature gating module. Then, the processed main path features and shortcut path features are concatenated. Finally, the features are fused through a 1×1 convolution and the number of channels is adjusted to obtain the output. The implementation definition process of this module is as follows:

[0068] Main_path,Skip_path=Split(Conv1×1(x),ratio=e)

[0069] Main_path_processed=OmniKernel(Main_path)

[0070] Output=Conv1×1(Concat(Main_path_processed,Skip_path))

[0071] In the formula, Main_path and Skip_path represent the main path and shortcut path features after splitting, respectively; Split represents the channel splitting operation performed according to the ratio e; Conv1×1 represents the 1×1 convolution operation; Concat represents the channel dimension concatenation operation; and OmniKernel represents the omnidirectional kernel feature extraction unit. Its internal calculation process can be further expanded as follows:

[0072] X_proj = GELU(Conv1×1(X))

[0073] X_fca = FCA(X_proj)

[0074] X_sca = SCA(X_fca)

[0075] X_gated = FGM(X_sca)

[0076] X_conv=X+DW_13(X_proj)+DW_31(X_proj)+DW_33(X_proj)+DW_11(X_proj)

[0077] Output_Omni=ReLU(X_conv+X_gated)

[0078] ReturnConv1×1(Output_Omni)

[0079] The frequency domain channel attention FCA operation is defined as follows:

[0080] Att_Weight = Conv1 × 1(GAP(X))

[0081] X_fft = FFT2D(X)

[0082] X_fft_att=Att_Weight⊙X_fft

[0083] X_fca = IFFT2D(X_fft_att)

[0084] In the formula, GAP represents the global average pooling operation, which averages the spatial feature values ​​of each channel of the input feature map and outputs a channel-wide global description vector. FFT2D and IFFT2D represent the two-dimensional fast Fourier transform and its inverse transform, respectively, used to transform the feature map between the spatial and frequency domains. Their transformation and inverse transform follow the standard Cooley-Tukey algorithm. ⊙ represents element-wise multiplication. This submodule, through frequency domain transformation and weight modulation, can effectively capture the global context and periodic patterns that are difficult to detect in the spatial domain.

[0085] The Spatial Channel Attention (SCA) operation is defined as follows:

[0086] Att_Weight=Sigmoid(Conv1×1(GAP(X)))

[0087] X_sca=Att_Weight⊙X

[0088] In the formula, Sigmoid represents the S-shaped growth curve activation function, used to compress input values ​​into the (0,1) interval to generate attention weights. This submodule enables the network to adaptively emphasize important feature channels and suppress redundant channels.

[0089] The Feature Gating (FGM) module is used to adaptively filter and finely modulate the features optimized by the aforementioned attention mechanism. Its core lies in learning the importance weights of each position and channel in the feature map through a parameterized gating structure, thereby suppressing noise or redundant information and enhancing the expression of key features. The implementation of this module includes the following operations performed sequentially:

[0090] Gate_Weight=Sigmoid(Conv1×1(X_in))

[0091] X_gated=Gate_Weight⊙X_in

[0092] In the formula, X_in represents the input feature of the feature gating module, Conv1×1 represents a 1×1 convolutional layer, which is used to fuse cross-channel information to generate complex gating signals. The Sigmoid function compresses the numerical range of the gating signal to the (0,1) interval, thus obtaining a gating weight matrix Gate_Weight that can be regarded as an "importance score". ⊙ represents element-wise multiplication.

[0093] The multi-directional depthwise convolution is achieved by summing the results of parallel operations on horizontal, vertical, diagonal, and point-to-point depthwise separable convolutions:

[0094] DW_13=DepthwiseConv(Kernel=(1,K))

[0095] DW_31=DepthwiseConv(Kernel=(K,1))

[0096] DW_33=DepthwiseConv(Kernel=(K,K))

[0097] DW_11=DepthwiseConv(Kernel=(1,1))

[0098] In the formula, DepthwiseConv represents the depthwise convolution operation in depthwise separable convolution, where each input channel is independently convolved by a convolution kernel, and the groups parameter is equal to the number of input channels. K is the depthwise convolution kernel size, which is an odd number greater than 1.

[0099] like Figure 5 As shown, further, in step S3.3, the deformable task decoupling detection head module processes the fused feature F according to the following process:

[0100] Feat_shared = Shared_Conv(F)

[0101] Feat_cls,Feat_reg=Task_Decomp(Feat_shared)

[0102] Offset,Mask=Offset_Gen(Feat_shared)

[0103] Feat_reg_aligned=Deform_Conv(Feat_reg,Offset,Mask)

[0104] Cls_Weight=Cls_Prob_Gen(Feat_shared)

[0105] Feat_cls_weighted=Feat_cls⊙Cls_Weight

[0106] Reg_pred=Scale(Conv_reg(Feat_reg_aligned))

[0107] Cls_pred=Conv_cls(Feat_cls_weighted)

[0108] BBox,Cls=Decode(Reg_pred,Cls_pred)

[0109] Wherein: Shared_Conv represents the shared convolutional module, consisting of two Conv_GN layers connected in series, with the outputs of the two layers concatenated along the channel dimension, used to extract the shared basic feature Feat_shared from the input feature F for multi-path fusion. Task_Decomp represents the task decomposition module, which first performs adaptive average pooling on the shared feature Feat_shared to obtain global features, and then decouples Feat_shared and global features into task-specific features Feat_cls and Feat_reg through a two-branch structure. Offset_Gen represents the spatial offset generation module, which predicts the offset and modulation mask required for dynamic deformable convolution from the shared feature Feat_shared through a convolutional layer. Deform_Conv represents the dynamic deformable convolution module, which performs adaptive spatial sampling and alignment on the regression feature Feat_reg based on the offset and mask, and outputs the aligned feature Feat_reg_aligned. Cls_Prob_Gen represents the classification probability generation module, which generates a spatial attention weight map Cls_Weight from the shared feature Feat_shared through a two-level convolutional sub-network. ⊙ represents element-wise multiplication, used to spatially reweight the classification feature Feat_cls with the attention weight Cls_Weight to obtain the modulated feature Feat_cls_weighted. Conv_reg and Conv_cls represent independent output convolutional modules, used to map the aligned regression features and weighted classification features to regression offset prediction Reg_pred and classification score prediction Cls_pred, respectively. Scale represents the scaling module, used to adaptively adjust the scale of the regression features. Decode represents the distribution aggregation and decoding module, whose function is implemented by the DFL layer and the decode_bboxes method. It aggregates the predicted distribution parameters Reg_pred into the final bounding box coordinates BBox and applies Sigmoid activation to the classification score Cls_pred to obtain the class probability Cls.

[0110] The above description represents specific embodiments of the present invention. It should be noted that the present invention is not limited to the specific embodiments described above. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting a target of a UAV based on multi-core fusion and omni-directional connection reparameterization, characterized in that, The method comprises the following steps: S1: obtaining a UAV aerial image dataset; S2: preprocessing the UAV dataset; S3: establishing a UAV image target detection model, wherein the model comprises a backbone network, a feature fusion network and a deformable task decoupling detection head connected in sequence; wherein the backbone network outputs multi-scale features, the feature fusion network fuses the multi-scale features, and the deformable task decoupling detection head outputs a detection box and a class probability according to the fused features; S3.1: a feature extraction stage, wherein a wide branch reparameterization convolution module is used to dynamically expand the receptive field of the feature extraction network through parallel multi-scale feature extraction and structure reparameterization mechanism for feature expression; the wide branch reparameterization convolution module extracts multi-scale information through parallel structure in the training stage; in the inference stage, the parallel structure is reparameterized into a single convolution and the features are outputted; S3.2: a feature fusion stage, wherein a spatial depth convolution module is used to downsample the feature map and reduce the spatial information loss; an omnidirectional kernel cross-stage partial connection module is used to perform cross-stage partial connection fusion on the downsampled features to enhance the representation ability of the multi-scale features, and then output the fused features; S3.3: a prediction stage, wherein a deformable task decoupling detection head is used to sequentially input the multi-scale features into a shared convolution, and after extracting the basic features, a task decomposition module is used to generate class features and regression features respectively; a dynamic deformable convolution is used to align the regression features in space, and a spatial channel attention module is used to weight the class features; the fused regression results and class results are adjusted in scale by a scale adjustment module, and then the distribution is aggregated and decoded by a distribution aggregation and decoding module, and finally the final detection box and class probability are outputted; S4: training the model on the training set and performing iterative training; S5: using the trained weight for the test set to output the detection results.

2. The multi-core fusion and omni-directional connection based reparameterization UAV target detection method according to claim 1, characterized in that: The wide branch reparameterization convolution module comprises six parallel processing paths in the training stage, wherein the first path performs a standard k×k convolution operation followed by a batch normalization operation; the second path performs a 1×1 convolution operation followed by a batch normalization operation when the number of groups is less than the number of output channels; the third path selects a processing flow according to the number of groups: when the number of groups is less than the number of output channels, a 1×1 convolution, padding batch normalization, average pooling and batch normalization operations are performed in sequence, otherwise an average pooling followed by a batch normalization operation is performed; the fourth path selects an identity 1×1 convolution or a normal 1×1 convolution according to the channel condition, and then performs padding batch normalization, k×k convolution and batch normalization operations in sequence; the fifth path performs a vertical asymmetric convolution followed by a batch normalization operation; the sixth path performs a horizontal asymmetric convolution followed by a batch normalization operation; the output features of the six paths are fused and processed by a nonlinear activation function; in the inference stage, the operations of the six paths are combined into a single convolution through structure reparameterization technology, and a nonlinear activation function is applied for processing. 3.The method of claim 1, wherein: The spatial depth convolution module is sequentially connected by a spatial splitting module, a channel splicing module and a convolution fusion module; the spatial splitting module extracts four sub-feature maps according to the row and column odd indexes of the input feature map; the channel splicing module splices the four sub-feature maps in the channel dimension to form an intermediate feature map with four times the channel number and half the spatial size; and the convolution fusion module performs one 3*3 convolution on the intermediate feature map to output a feature map with a specified channel number, completes down-sampling and retains complete spatial information.

4. The method of claim 1, wherein the method is based on multi-core fusion and omni-directional connection. The omnidirectional kernel cross-stage partial connection module is composed of a feature splitting convolution module, an omnidirectional kernel feature extraction module and a feature fusion convolution module; wherein the omnidirectional kernel feature extraction module is composed of an input 1*1 convolution layer, a multi-directional depth convolution module, a frequency domain channel attention module, a spatial channel attention module, a feature gating module and an output 1*1 convolution layer.

5. The method of claim 1, wherein the method is based on multi-core fusion and omni-directional connection. The deformable task decoupling detection head is composed of the following core modules: a shared convolution module, a task decomposition module, the task decomposition module including a classification task decomposition module and a regression task decomposition module, a spatial offset generation module, a dynamic deformable convolution module, a classification probability generation module, an output convolution module, a scale scaling module and a distribution aggregation and decoding module.

6. The method according to claim 4, wherein, The omnidirectional kernel cross-stage partial connection module divides the input into a main path and a shortcut path by the feature splitting convolution module, processes the main path through the input 1*1 convolution layer, performs multi-directional feature extraction through the multi-directional depth convolution module, executes double attention enhancement through the frequency domain channel attention module and the spatial channel attention module, processes through the feature gating module, finally splices the features of the shortcut path, and then outputs the fused features through the 1*1 convolution.

7. The multi-core fusion and omni-directional connection based reparameterization UAV target detection method according to claim 5, characterized in that, The deformable task decoupling detection head first extracts multi-scale basic features through the shared convolution module; The task decomposition module decouples the features into classification features and regression features; the spatial offset generation module generates dynamic deformable convolution parameters; the dynamic deformable convolution module realizes spatial alignment of the regression features; The classification probability generation module generates spatial attention weights and weights the classification features; the output convolution module generates classification scores and regression offsets; The scale scaling module adaptively adjusts the feature scale; The distribution aggregation and decoding module is used to aggregate the feature distribution and decode it into the final detection frame coordinates and class probability.

Images