A Target Detection Method for UAV Aerial Images Based on Improved DEIM

Abstract

This invention discloses a target detection method for UAV aerial images based on an improved DEIM network, belonging to the field of computer vision. The method includes: acquiring and preprocessing a UAV image dataset to obtain a standardized image dataset; constructing an improved DEIM network, which includes a backbone network, an encoder, and a decoder; inputting the standardized image dataset into the improved DEIM network for training; wherein, the backbone network uses a multi-receptor domain adaptive fusion module to process the input features of the standardized image data to obtain multi-scale enhanced features; the encoder uses a multi-path gating state space modeling module to perform channel alignment and cross-scale fusion of the multi-scale enhanced features to obtain semantically enhanced features; and the decoder decodes based on the semantically enhanced features to output the UAV target detection result. The improved DEIM network provided by this invention achieves a significant improvement in target detection performance in UAV aerial images.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to a target detection method for UAV aerial images based on an improved DEIM. Background Technology

[0002] With the rapid development of drone technology, drones are increasingly widely used in urban management, traffic monitoring, disaster relief, agricultural inspection, and security monitoring. Drone-captured images typically exhibit characteristics such as a top-down perspective, large variations in shooting altitude, wide target scale spans, dense target distribution, and complex backgrounds. This places higher demands on the multi-scale modeling capabilities, long-range dependency modeling capabilities, and lightweight performance of target detection algorithms. In drone target detection tasks, small targets account for a high proportion, and there are occlusion and dense distribution phenomena among targets. Although traditional convolutional neural network-based target detection models can expand the receptive field by stacking convolutional layers, they still rely on local convolutional operations and struggle to effectively model long-distance spatial dependencies, especially in high-resolution scenes where they are prone to insufficient feature representation or loss of detail. In recent years, the development of attention mechanisms and state-space modeling methods has provided new ideas for global dependency modeling. However, the computational complexity of attention mechanisms is usually related to the square of the feature dimension, resulting in significant overhead in high-resolution scenes. Furthermore, existing state-space models mostly employ single-path modeling methods and lack path-level adaptive control mechanisms, limiting their expressive capabilities in complex scenes.

[0003] The DEIM network, as a target detection network combining convolutional structure and efficient modeling mechanism, has certain advantages in improving detection accuracy while maintaining low computational complexity. However, in complex UAV scenarios, the original DEIM network still has the following shortcomings: First, in the high-level semantic modeling stage, its feature modeling path is relatively simple, lacking a multi-path collaborative mechanism, making it difficult to fully explore the differences in feature expression under different spatial propagation modes; Second, in the cross-scale fusion process, the integration method of features from different receptive domains is relatively fixed, lacking adaptive weighting and dynamic adjustment mechanisms, and there is still room for improvement in the fusion effect in scenarios with dense small targets; Third, in the long-range dependency modeling process, it lacks a path-level gating adjustment structure, making it unable to effectively suppress redundant information.

[0004] Therefore, this invention proposes a target detection method for UAV aerial images based on improved DEIM. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a target detection method for UAV aerial images based on improved DEIM, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a target detection method for UAV aerial images based on improved DEIM, comprising: Acquire drone image datasets and preprocess them to obtain standardized image datasets; An improved DEIM network is constructed, comprising a backbone network, an encoder, and a decoder; The standardized image dataset is input into the improved DEIM network for training. The backbone network uses a multi-receptive-domain adaptive fusion module to process the input features of standardized image data to obtain multi-scale enhanced features. The encoder uses a multi-path gated state space modeling module to perform channel alignment and cross-scale fusion on the multi-scale enhancement features to obtain semantic enhancement features. The decoder performs decoding based on the semantic enhancement features and outputs the UAV target detection results.

[0007] Optionally, the UAV target detection method based on the improved DEIM network according to claim 1 is characterized in that the process of inputting the standardized image dataset into the improved DEIM network for training includes: The convolutional normalization processing module 1 in the backbone network is used to perform convolutional feature extraction and normalization operations on UAV images in the standardized image dataset to obtain the initial feature map; The initial feature map is sequentially input into multiple sets of staged hierarchical feature extraction modules consisting of a convolutional normalization processing module and a multi-receptive field adaptive fusion module for processing to obtain a multi-scale feature map. The multi-scale feature map is channel-aligned using the convolutional normalization processing module in the encoder to obtain an aligned feature map. A self-attention feature encoding module is used to perform global semantic modeling on the high-level feature map in the aligned feature map to obtain semantic encoded features; The semantic encoded features are upsampled using a feature upsampling module to obtain upsampled features. After the upsampled features are concatenated and fused with the aligned feature maps of the corresponding scale, they are input into the multi-path gated state space modeling module for processing to obtain semantically enhanced features. The semantic enhancement features are decoded using a decoder to output the UAV target detection results.

[0008] Optionally, the process of sequentially inputting the initial feature map into multiple sets of staged hierarchical feature extraction modules, consisting of a convolutional normalization processing module and a multi-receptive-field adaptive fusion module, to obtain a multi-scale feature map includes: The initial feature map is input into the staged hierarchical feature extraction module, and downsampling is performed based on the convolution normalization processing module to obtain the downsampled feature map. The downsampled feature map is then input into the multi-receptive-domain adaptive fusion module for multi-receptive-domain feature fusion processing to obtain the enhanced feature map of the current stage. The process of inputting the enhanced feature map of the current stage into the next stage-level feature extraction module and performing downsampling and multi-receptive-domain feature fusion processing in the next stage-level feature extraction module is repeated until a multi-scale feature map is obtained.

[0009] Optionally, the multi-receptive-domain adaptive fusion module includes: a channel expansion module, a feature generation module, a channel equalization module, a progressive enhancement sub-module group, a feature splicing module, and a channel compression fusion module; The process of inputting the downsampled feature map into the multi-receptive-domain adaptive fusion module for multi-receptive-domain feature fusion processing to obtain the enhanced feature map for the current stage includes: The channel expansion module is used to linearly map the downsampled feature map to obtain the first mapped feature; Based on the first mapping feature, convolution operation is performed using feature generation module 1 to obtain basic semantic features; Based on the first mapping feature, the feature generation module 2 performs depthwise separable convolution operation to obtain local texture features; The channel equalization module is used to divide the first mapping feature into channels to obtain the original semantic sub-features and the sub-features to be enhanced; Based on the sub-features to be enhanced, the progressive enhancement sub-module group is used to perform multi-level progressive feature enhancement to obtain enhanced features; The feature splicing module is used to splice and fuse the basic semantic features, the local texture features, the original semantic sub-features, and the enhanced features to obtain the fused features; Based on the fusion features, the channel compression fusion module is used to perform convolution operations and channel compression to obtain the output features of the current stage; The enhanced feature map for the current stage is obtained based on the output features of the current stage.

[0010] Optionally, based on the sub-feature to be enhanced, the process of performing multi-level progressive feature enhancement using the progressive enhancement sub-module group to obtain enhanced features includes: The sub-feature to be enhanced is used as the initial input feature. The initial input feature is then input into the first lightweight grouping adaptive feature aggregation submodule for feature enhancement processing to obtain the first-level enhanced feature. Based on the first-level enhanced features, a second lightweight grouped adaptive feature aggregation submodule is used to perform feature enhancement processing to obtain the second-level enhanced features; Similarly, based on the enhanced features of the previous level, the next level of lightweight grouping adaptive feature aggregation submodule is used for feature enhancement processing, until the progressive processing of n lightweight grouping adaptive feature aggregation submodules is carried out to obtain the nth level enhanced features; The initial input features, the first-level enhancement features, the second-level enhancement features, and up to the nth-level enhancement features are collectively used as the enhancement features.

[0011] Optionally, the multi-path gating state space modeling module includes: a channel mapping module, a normalization module, a gating-enhanced state space feature reconstruction sub-module, a random path discarding module, a hierarchical scaling module, a residual scaling module, and a channel mixing module; The process of concatenating and fusing the upsampled features with the aligned feature map at the corresponding scale, and then inputting it into the multi-path gated state space modeling module for processing to obtain semantically enhanced features includes: The upsampled features and the aligned feature map of the corresponding scale are concatenated along the channel dimension using a feature concatenation module to obtain the concatenated features; The channel mapping module is used to linearly map the splicing features to obtain the first mapped features; The first mapping feature is normalized using the normalization module to obtain the first normalized feature; The first normalized feature is subjected to state-space modeling and gating fusion processing using the gated enhanced state-space feature reconstruction submodule to obtain the reconstructed feature. The reconstructed features are processed by the random path discarding module to obtain the first discarded features; The hierarchical scaling module is used to adjust the hierarchical scale of the first discarded feature to obtain the first scaled feature; The residual scaling module is used to adjust the residual ratio of the first mapping feature to obtain the second scaled feature; The first fused feature is obtained by residual fusion based on the first scaling feature and the second scaling feature; The first fused feature is normalized using the normalization module to obtain the second normalized feature; The channel mixing module is used to perform inter-channel information interaction and nonlinear mapping processing on the second normalized feature to obtain the mixed feature; The random path discarding module is used to perform random path discarding processing on the mixed features to obtain the second discarded feature; The hierarchical scaling module is used to adjust the hierarchical scale of the second discarded feature to obtain the third scaled feature; The residual scaling module is used to adjust the residual ratio of the first fused feature to obtain the fourth scaled feature; The semantic enhancement feature is obtained by residual fusion based on the third scaling feature and the fourth scaling feature.

[0012] Optionally, the process of the decoder using DFINETransformer to decode the semantically enhanced features includes: The DFINETransformer is used to perform unified dimensional mapping and feature integration on the semantic enhancement features to obtain integrated features; Based on the integrated features, a query selection mechanism is used to filter candidate target regions from the feature map to obtain candidate region features; A multi-layer decoding structure is used to progressively update and optimize the semantic and spatial location information of the candidate region features to obtain the target category prediction result and the target bounding box position; Based on the target category prediction result and the target bounding box position, the UAV target detection result is output.

[0013] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0014] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0015] The present invention also includes a computer program product comprising computer program instructions that, when executed by a processor, implement the steps of the method.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs a multi-receptive-domain adaptive fusion structure and a multi-path gated state-space modeling module based on the DEIM network structure. By structurally enhancing the original feature extraction and semantic modeling processes, it achieves collaborative modeling of multi-scale information and long-range dependency information. Specifically, the multi-receptive-domain adaptive fusion structure can perform domain-specific processing and progressive enhancement of features from different receptive domains, improving the expressive power of small-scale targets and complex background regions. The multi-path gated state-space modeling module, by setting a multi-path feature propagation mechanism and combining it with gating adjustment strategies, achieves adaptive control of features from different propagation paths and suppression of redundant information, enhancing the integrity and stability of high-level semantic information expression. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the algorithm of the UAV target detection method based on the improved DEIM network according to an embodiment of the present invention. Figure 2 This is a model architecture diagram of the improved DEIM network according to an embodiment of the present invention; Figure 3 This is a structural diagram of the multi-receptive-domain adaptive fusion module according to an embodiment of the present invention; Figure 4 This is a structural diagram of the progressive enhancement submodule group according to an embodiment of the present invention; Figure 5 This is a structural diagram of the lightweight grouping adaptive feature aggregation submodule according to an embodiment of the present invention; Figure 6 This is a structural diagram of the multi-path gating state space modeling module according to an embodiment of the present invention; Figure 7 This is a structural diagram of the gating-enhanced state-space feature reconstruction submodule in an embodiment of the present invention; Figure 8 This is a structural diagram of the adaptive gated channel coupling unit according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] In this invention, the numbered modules are only used to distinguish different structural units and do not represent a sequential, hierarchical or functional priority relationship.

[0021] like Figure 1 As shown, this embodiment provides a target detection method for UAV aerial images based on improved DEIM, including the following steps: Step 1. Obtain the drone image dataset and preprocess it to obtain a standardized drone image dataset. Divide the standardized drone image dataset into a training set, a validation set, and a test set.

[0022] As an optional implementation of this embodiment, in Step 1, the UAV image dataset is preprocessed to obtain a standardized UAV image dataset, which is then divided into a training set, a validation set, and a test set, specifically as follows: Step 101. Perform image resizing, normalization, and data augmentation on each image captured by the drone to obtain a standardized drone image dataset; Step 102. Divide the standardized UAV image dataset into a training set, a validation set, and a test set in a 7:2:1 ratio.

[0023] In this embodiment, the drone image dataset comes from the publicly available VisDrone2019 dataset.

[0024] Step 2. Construct an improved DEIM network, which consists of three main parts: a backbone network, an encoder, and a decoder. The backbone network is used for multi-scale feature extraction, the encoder is used for high-level semantic modeling and cross-scale fusion, and the decoder is used for final target prediction.

[0025] In this embodiment, the improved DEIM network is built based on the DEIM-Dfine-s version; the DEIM-Dfine-s network is a multi-scale target detection network structure, which includes a backbone network, an encoder, and a decoder.

[0026] The backbone network is used to perform hierarchical convolutional feature extraction on the input UAV images. It sequentially constructs multi-scale feature representations through an initial feature extraction module (StemBlock) and multiple staged hierarchical feature extraction modules (HG_Stage). As the network depth increases, the spatial resolution of the feature maps gradually decreases while the semantic expressive power gradually increases, ultimately forming... , , and Feature maps of different scales (such as reducing the spatial size of the image to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original size) are used to represent target information at different scales.

[0027] The encoder is used for channel alignment and cross-scale fusion of multi-scale features. First, a convolutional normalization module (ConvNormLayer_fuse) performs a unified mapping of features at different scales. Then, a self-attention feature encoding module (TransformerEncoderBlock) is used to perform global semantic modeling of high-level features, establishing long-range dependencies between spatial locations. Subsequently, a bidirectional feature fusion path combining top-down and bottom-up approaches is constructed through a feature upsampling module, a feature concatenation module, and a multi-branch feature enhancement module (RepNCSPELAN4) to achieve multi-scale feature interaction.

[0028] The decoder uses a Transformer decoding module (DFINETransformer) based on the DFINE mechanism to perform cross-scale interactive modeling of the fused multi-scale features and output the target class probability and bounding box regression parameters to achieve target detection in UAV images.

[0029] The DEIM-Dfine-s network possesses multi-scale feature extraction and global semantic modeling capabilities, but in UAV high-altitude shooting scenarios, it still suffers from insufficient representation of small targets and limited feature discrimination capabilities under complex background interference. To address these issues, this invention improves upon the existing structure by introducing a multi-receptive-domain adaptive fusion module and a multi-path gating state space modeling module to enhance local detail representation capabilities and cross-regional semantic modeling capabilities.

[0030] As an optional implementation of this embodiment, the backbone network part of Step 2 includes a convolutional normalization processing module and a multi-receptive field adaptive fusion module.

[0031] Among them, such as Figure 3 As shown, the multi-receptive-domain adaptive fusion module includes a channel expansion module, a feature generation module, a channel equalization module, a progressive enhancement sub-module group, a feature stitching module, and a channel compression fusion module; as shown... Figure 4 As shown, the progressive enhancement submodule group adopts a lightweight grouped adaptive feature aggregation submodule; as Figure 5 As shown, the lightweight grouped adaptive feature aggregation submodule includes a channel domain segmentation module, a point-level enhancement submodule, a local enhancement submodule, a mid-range enhancement submodule, a global enhancement submodule, a multi-domain fusion module, a channel feedforward mapping module, a random path discarding module, a normalization module, and a channel mapping module.

[0032] As an optional implementation of this embodiment, the data processing process of the multi-receptive-domain adaptive fusion module is as follows: The downsampled feature map of the UAV image input to the multi-receptive-domain adaptive fusion module is... The input channel expansion module processes the image features by linearly mapping the channel dimensions to enhance their expressive power, resulting in the first mapped features. ; map the first feature The input feature generation module 1 processes the data and extracts basic semantic features from the image through convolution operations to obtain the basic semantic features. ; map the first feature The input feature generation module 2 processes the data and extracts local spatial texture features from the UAV image through depthwise separable convolution, thus obtaining local texture features. ; map the first feature The input channel equalization module processes the image features, dividing them into two sub-feature spaces according to the channel dimension to obtain the original semantic sub-features. and features of the enhancer ,in Preserve the original semantic feature information. Used for further multi-layer feature enhancement; the sub-features to be enhanced The input is processed by a progressive enhancement submodule group to progressively enhance the features of the UAV image layer by layer, thereby strengthening the feature representation capability of the target region in the image and obtaining enhanced features. ; Features ,feature ,feature and characteristics The input feature concatenation module 5 processes the image features from different receptive fields to obtain fused features. ; to integrate features The input channel compression and fusion module processes the data, performing channel compression and information integration on the fused image features through convolution operations to obtain the output features of the multi-receptive-domain adaptive fusion module. The enhanced feature map for the current stage is obtained based on the output features of the current stage. This process is expressed by the following formula: ; ; ; ; ; ; ; in, This represents a 1×1 convolution operation. This represents depthwise separable convolution operations. This indicates a channel splitting operation. This represents the feature transformation process of a progressive enhancement submodule group. This indicates a feature splicing operation.

[0033] As an optional implementation of this embodiment, the data processing procedure for the progressive enhancement submodule group is as follows: The input sub-features to be enhanced to the progressive enhancement submodule group are processed... Let these be the initial input features. ; Initial input features The input is processed by a lightweight grouped adaptive feature aggregation submodule to enhance local region features in the UAV image, resulting in the first-level enhanced features. ; enhance the first-level features The image features are then processed by the next lightweight grouped adaptive feature aggregation submodule to further enhance them, resulting in the second-level enhanced features. ; progressive feature enhancement processing is performed sequentially to enhance the feature representation capability of the target region in the UAV image layer by layer; if the progressive enhancement submodule group includes A lightweight grouped adaptive feature aggregation submodule satisfies: ; in, This represents the feature transformation process of the lightweight grouping adaptive feature aggregation submodule; it transforms the features... ,feature ,feature ,...,feature Output features of the progressive enhancement submodule group .

[0034] As an optional implementation of this embodiment, the data processing procedure of the lightweight grouping adaptive feature aggregation submodule is as follows: The data processing of the first... indivual( The lightweight grouping adaptive feature aggregation submodule takes UAV image features as input. Characteristic ; Features The input channel segmentation module processes the data, dividing the UAV image features according to the channel dimension to form multiple feature subspaces, allowing for the extraction of image information at different scales to obtain features. , , and ; Features The input point-level enhancement submodule processes the data, using a point-level attention mechanism to enhance pixel-level fine-grained information in the UAV image, highlighting key pixel features of the target region, and obtaining the feature... ; Features The input is processed by the local enhancement submodule, which uses a local attention mechanism to model the features of local neighborhood regions in the UAV image to enhance local spatial structure information and obtain features. ; Features The input is processed by the mid-range enhancement submodule, which models the feature relationships between mid-scale regions in the UAV image through a mid-range attention mechanism to enhance the contextual information between the target region and its surrounding environment, thereby obtaining features. ; Features The input is processed by the global enhancement submodule, which models the global semantic information of the UAV image through a global attention mechanism to enhance the long-distance dependencies between different spatial regions and obtain features. ; Features , , and The input is processed by a multi-domain fusion module to fuse image features from different receptive domains, resulting in feature extraction. ; Features The input channel feedforward mapping module processes the data, performing further feature transformation on the fused image features through nonlinear mapping to obtain the desired features. ; Features The input is processed by the random path drop-off module 1. The random path drop-off strategy improves the model's generalization ability and yields features. ; Features The input is processed by normalization module 1 to normalize the image features, thereby stabilizing the network training process and obtaining the features. ; Features and characteristics Residual fusion is performed to preserve the original image feature information and enhance the deep semantic expression capability, resulting in feature extraction. ; Features The input channel mapping module 1 processes the image features, performing channel mapping and information integration to obtain the first... The output features of the lightweight grouped adaptive feature aggregation submodule are denoted as features. The process can be expressed by the following formula: ; ; ; ; ; ; ; ; in, Indicates channel partitioning operation; , , and These represent point-level attention operations, local attention operations, mid-range attention operations, and global attention operations, respectively. This represents the normalization operation. Indicates feature concatenation operation; This represents the channel feedforward mapping operation, implemented through convolution operations, normalization operations, and nonlinear activation functions; This indicates a random path discard operation. This represents a 1×1 convolution operation.

[0035] As an optional implementation of this embodiment, the encoder in Step 2 includes a convolutional normalization processing module, a self-attention feature encoding module, a feature upsampling module, a feature concatenation module, a multi-path gating state space modeling module, a feature downsampling module, and a multi-branch feature enhancement module.

[0036] Among them, such as Figure 6 As shown, the multi-path gated state space modeling module includes a channel mapping module, a normalization module, a gated enhanced state space feature reconstruction submodule, a random path discarding module, a hierarchical scaling module, a residual scaling module, and a channel mixing module; as shown... Figure 7 As shown, the gated enhanced state space feature reconstruction submodule includes a depthwise separable convolution module, a state space feature unit hybridization module, an adaptive gated channel coupling unit, and a channel mapping module; as shown... Figure 8 As shown, the adaptive gated channel coupling unit includes a channel mapping module, a depthwise separable convolution module, and a random deactivation module.

[0037] As an optional implementation of this embodiment, the data processing process of the multi-path gating state space modeling module is as follows: The stitching features of the UAV image input to the multi-path gating state space modeling module are processed... The input channel mapping module 2 processes the image features by performing a linear mapping on the channel dimensions through convolution operations. This adjusts the feature channel distribution and enhances the image feature representation capability, resulting in the first mapped feature. ; map the first feature The input is processed by normalization module 2 to normalize the image features, thereby stabilizing the feature distribution and improving the stability of network training, resulting in the first normalized features. ; the first normalized feature The input is processed by a gated enhanced state-space feature reconstruction submodule. A state-space modeling mechanism is used to model spatial dependencies in the UAV image, capturing contextual information between distant regions in the image to obtain the reconstructed features. ; Reconstruct features The input is processed by the random path discarding module 2. The random path discarding strategy is used to improve the model's generalization ability in complex scenarios, and the first discarding feature is obtained. ; discard the first feature The input is processed by the hierarchical scaling module 1, which adjusts the scale of image features using a hierarchical scaling function to adapt to the semantic information expression needs of different levels, thus obtaining the first scaled feature. ; map the first feature The input is processed by residual scaling module 1, which uses a residual scaling function to adjust the scale of the original image features to obtain the second scaled features. ; The first scaling feature Second scaling features Residual fusion is performed to preserve the original image feature information and enhance the semantic feature representation ability after state space modeling, resulting in the first fused feature. ; the first fusion feature The input is processed by normalization module 3, which normalizes the fused image features to obtain the second normalized features. ; the second normalized feature The input channel mixing module processes the data, further fusing and reconstructing the UAV image features through inter-channel information interaction and nonlinear mapping to obtain hybrid features. ; Mixed features The input random path drop-off module 3 is used for processing to improve the robustness of the model in different image scenes, and the second drop-off feature is processed. ; The second discard feature The input is processed by the hierarchical scaling module 2 to adjust the hierarchical scale of the image features, resulting in the third scaled feature. ; the first fusion feature The input residual scaling module 2 is processed to adjust the scale of the residual branch features, resulting in the fourth scaled feature. ; the third scaling feature and the fourth scaling feature Residual fusion is performed to fuse the original image features with the semantic information after channel mixing, resulting in semantically enhanced features output by the multi-path gated state space modeling module. The process can be expressed by the following formula: ; ; ; ; ; in, This represents a 1×1 convolution operation. This represents the normalization operation; This represents the feature transformation process of the gated enhanced state-space feature reconstruction submodule. This indicates a random path discard operation. and These represent the residual scaling function and the hierarchical scaling function, respectively. This represents a channel blending operation, implemented through convolution, non-linear activation functions, and random deactivation.

[0038] As an optional implementation of this embodiment, the data processing procedure of the gated enhanced state space feature reconstruction submodule is as follows: The UAV image features input to the gated enhanced state space feature reconstruction submodule are processed... Recorded as ;Multiple sets of learnable gating scaling parameters of the gating-enhanced state space feature reconstruction submodule ( )conduct Nonlinear mapping yields gated scaling coefficients. ( ), used for adaptive weight adjustment of different feature branches; to transfer features The input is processed by depthwise separable convolution module 1, which extracts local spatial texture information from the UAV image through depthwise separable convolution operations to obtain features. ; Features and Based on the first gating scaling factor Weighted fusion is performed to adaptively balance the contribution ratio between the original image features and the local convolutional features, thereby obtaining the features. ; Features The input state-space feature unit hybrid module processes the data and models the spatial sequence relationships in the UAV images through a state-space modeling mechanism, thereby capturing the long-distance dependencies between different spatial regions in the image and obtaining features. ; Features and Based on the second gating scaling factor Weighted fusion is performed to achieve adaptive fusion between local convolutional features and global contextual information, resulting in features. ; Features The input depthwise separable convolutional module 2 is used for processing to further model the local structure of the image features, thereby enhancing the spatial texture representation of the target region in the UAV image and obtaining the features. ; Features and Based on the third gating scaling factor Weighted fusion is performed to enhance local spatial detail information while preserving global semantic information, thereby obtaining features. ; Features The input is processed by an adaptive gated channel coupling unit. Through inter-channel information exchange and a gating mechanism, image features are reconstructed at the channel level to obtain the features. ; Features and Based on the fourth gating scaling factor Weighted fusion is performed to achieve adaptive integration of features between multiple layers of image features, resulting in feature extraction. ; Features The input channel mapping module 3 processes the image features through convolution operations, performing channel mapping and information integration to obtain the output features of the gated enhanced state space feature reconstruction submodule, denoted as... The process can be expressed by the following formula: ; ; ; ; ; ; ; in, for function, This represents depthwise separable convolution operations; The mixed operation of state-space feature units is represented by serialization expansion operation, normalization operation and state-space modeling operation; This represents the feature transformation process of the adaptive gated channel coupling unit. This represents a 1×1 convolution operation.

[0039] As an optional implementation of this embodiment, the data processing procedure of the adaptive gated channel coupling unit is as follows: The UAV image features input to the adaptive gated channel coupling unit are processed... Recorded as ; Features The input channel mapping module 4 processes the data by performing linear channel mapping on the UAV image features through pointwise convolution and dividing them along the channel dimension to obtain gated components. and characteristics The gated component V is used to perform gated modulation on subsequent image features; the features are... The input is processed by depthwise separable convolution module 3. This module models the spatial neighborhood information of the UAV image features through depthwise separable convolution to extract edge information and local texture features of the UAV target region, thus obtaining the feature... ; Gating components and characteristics Element-wise multiplication is performed to achieve gated modulation, which enhances the target-related feature responses in the UAV image while suppressing background features, thus obtaining the feature... ; Features The input is processed by random deactivation module 1. This random deactivation operation probabilistically masks some feature channels to improve the network's generalization ability to different UAV image scenes, thus obtaining the feature... ; Features The input channel mapping module 5 processes the data, remapping and fusing the feature channels through pointwise convolution to obtain the feature data. ; Features The input is processed by random deactivation module 2. The features are then regularized by performing another random deactivation operation to obtain the final features. ; Features and characteristics Residual fusion is performed, which allows the network to superimpose gated enhanced feature representations while preserving the original image feature information, resulting in the output features of the adaptive gated channel coupling unit, denoted as... The process can be expressed by the following formula: ; ; ; in, and All of these are learnable pointwise convolution kernel parameter sets. This indicates a channel partitioning operation. This represents depthwise separable convolution operations. This represents element-wise multiplication. This indicates a random deactivation operation.

[0040] As an optional implementation of this embodiment, Step 2 decoder includes a decoder module.

[0041] In this embodiment, the decoder module uses DFINETransformer to perform target detection decoding on the multi-scale features of the UAV image output by the aforementioned feature extraction network and feature fusion module. DFINETransformer can perform unified dimensional mapping and feature integration on the input UAV image features to obtain integrated features. Based on the integrated features, a query selection mechanism is used to select candidate target regions with high response from the feature map to obtain candidate region features. Then, through a multi-layer decoding structure, the semantic information and spatial location information of the candidate region features are gradually updated and optimized to obtain more accurate target category prediction results and target bounding box positions. Through the above decoding process, target objects can be effectively identified and located in UAV images with complex backgrounds, thereby completing the target detection task in UAV images.

[0042] Step 3. Input the training set and validation set obtained in Step 1 into the improved DEIM network constructed in Step 2 for training and validation evaluation; the network transforms the input UAV image into a multi-scale feature representation step by step through layer-by-layer feature extraction and feature fusion operations, and finally outputs the target detection result.

[0043] As an optional implementation method of this embodiment, such as Figure 2 As shown, Step 3, which involves inputting the training set obtained in Step 1 into the improved DEIM network constructed in Step 2, includes the following steps: Step 301. Transfer the drone images from the training set. I The input is processed by the convolutional normalization processing module 1, which extracts the initial feature map through convolutional feature extraction and normalization operations. The convolution normalization processing module 1 has 3 input channels, 16 output channels, and a kernel size of 3. 3, step size is 2.

[0044] Suppose the input is a drone image. ,in and These represent the height and width of the image, respectively, and 3 represents the three RGB color channels; the operation process of convolution normalization module 1 can be represented as follows: ; in, This represents the normalization operation. This represents a convolution operation with a kernel size of 3×3 and a stride of 2.

[0045] This operation allows for the extraction of initial low-level visual features from UAV images, including target edge information, texture structure, and local shape features. This provides a foundational feature representation for subsequent feature enhancement in the multi-receptive-domain adaptive fusion module and the multi-path gating state space modeling module.

[0046] Step 302. Transfer the feature map obtained in Step 301. The input is processed by the convolutional normalization module 2; a new feature map is obtained through convolutional feature extraction and normalization operations. .

[0047] In this module, the convolution normalization processing module 2 has 16 input channels, 32 output channels, and a convolution kernel size of 3. 3. The stride is 2, and its operation process is the same as that of the convolution normalization module 1. Through this operation, the feature map spatial size is further reduced, while the channel dimension is increased, so that the network can extract richer semantic information.

[0048] Step 303. Transfer the feature map obtained in Step 302. The input is processed by the multi-receptive-domain adaptive fusion module 1, which obtains enhanced feature maps through multi-scale receptive-domain feature modeling and adaptive fusion operations. .

[0049] The multi-sensory-domain adaptive fusion module 1 has 32 input channels and 64 output channels; the operation process of the multi-sensory-domain adaptive fusion module 1 can be expressed as follows: ; in, This represents the feature transformation process of the multi-receptive-domain adaptive fusion module.

[0050] Through this operation, the network performs feature modeling on target regions in UAV images within different spatial receptive domains, enabling the feature maps to simultaneously contain local texture information and broader contextual semantic information. This enhances the network's ability to represent features of small-scale targets, dense targets, and complex background regions in UAV scenes, thereby improving the network's ability to distinguish between target and background regions.

[0051] Step 304. Transfer the feature map obtained in Step 303. The input is processed by the convolutional normalization module 3, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 3 has 64 input channels, 64 output channels, and a kernel size of 3. 3. The step size is 2, and its operation process is the same as that of convolution normalization module 1.

[0052] This operation further spatially downsamples the feature map, reducing its spatial resolution while gradually enhancing its semantic information. This expands the receptive field of the features, enabling the network to perceive a wider range of target structure information in UAV images.

[0053] Step 305. Transfer the feature map obtained in Step 304. The input is processed by the multi-receptive-domain adaptive fusion module 2, which obtains enhanced feature maps through multi-scale receptive-domain feature modeling and adaptive fusion operations. The multi-sensory-domain adaptive fusion module 2 has 64 input channels and 128 output channels, and its operation process is the same as that of the multi-sensory-domain adaptive fusion module 1.

[0054] Through this operation, the network performs multi-receptive-domain feature fusion on target regions in UAV images at a higher semantic level, enabling the network to simultaneously capture feature information of targets at different scales, thereby improving the detection capability of multi-scale targets in complex scenes.

[0055] Step 306. Transfer the feature map obtained in Step 305. The input is processed by the convolutional normalization module 4, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 4 has 128 input channels, 256 output channels, and a kernel size of 3. 3. The step size is 2, and its operation process is the same as that of the above convolution normalization processing module 1.

[0056] This operation further reduces the spatial size of the feature map while increasing the channel dimension, enabling the network to extract more abstract semantic features, thereby enhancing its ability to express the overall structural information of the target in UAV images.

[0057] Step 307. Transfer the feature map obtained in Step 306 The input is processed by the multi-receptive-domain adaptive fusion module 3, which obtains enhanced feature maps through multi-scale receptive-domain feature modeling and adaptive fusion operations. The multi-sensory-domain adaptive fusion module 3 has 256 input channels and 256 output channels, and its operation process is the same as that of the multi-sensory-domain adaptive fusion module 1.

[0058] Through this operation, the network further integrates feature information from different spatial receptive domains, enabling the feature map to simultaneously contain local detail information and global contextual information, thereby enhancing the network's ability to semantically represent target regions in UAV images.

[0059] Step 308. Transfer the feature map obtained in Step 307. The input is processed by the convolutional normalization module 5, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 5 has 256 input channels, 512 output channels, and a kernel size of 3. 3, with a step size of 2; its operation process is the same as that of convolution normalization module 1.

[0060] Through this operation, the network obtains higher-level semantic features, enabling the feature map to describe the overall semantic information of the target in the UAV image, thereby providing a basic feature representation for subsequent high-level semantic modeling and cross-scale feature fusion.

[0061] Step 309. Transfer the feature map obtained in Step 308. The input is processed by the multi-receptive-domain adaptive fusion module 4, which obtains enhanced feature maps through multi-scale receptive-domain feature modeling and adaptive fusion operations. The multi-sensory-domain adaptive fusion module 4 has 512 input channels and 512 output channels, and its operation process is the same as that of the multi-sensory-domain adaptive fusion module 1.

[0062] Through this operation, the network further integrates feature information from different receptive domains at a high semantic level, enabling the feature map to simultaneously represent the local structural features and global semantic information of the target in the UAV image. This improves the network's ability to express multi-scale targets in complex scenes and provides high-quality feature input for cross-scale feature fusion in the subsequent encoder stage.

[0063] Step 310. Transfer the feature map obtained in Step 305. The input is processed by the convolutional normalization module 6, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 6 has 128 input channels, 256 output channels, and a kernel size of 1. 1. Step size is 1.

[0064] The operation process of convolution normalization module 6 can be represented as follows: ; in, This represents the normalization operation. This represents a convolution operation with a kernel size of 1×1 and a stride of 1.

[0065] This operation transforms the channel dimension and remaps the feature map, enhancing the feature representation capability while maintaining the spatial resolution. This allows features from different network layers to be mapped to a unified channel dimension, thus providing a consistent feature representation for subsequent multi-scale feature fusion and attention feature encoding.

[0066] Step 311. Transfer the feature map obtained in Step 307. The input is processed by the convolutional normalization module 7, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 7 has 256 input channels, 256 output channels, and a kernel size of 1. 1. The step size is 1, and its operation process is the same as that of the above convolution normalization processing module 6.

[0067] This operation, while keeping the feature map spatial size unchanged, performs channel transformation and feature reorganization on the mid-level semantic features, making the feature representation more compact and further enhancing the semantic information expression of the target region in the UAV image, thereby providing a stable feature input for subsequent cross-level feature fusion.

[0068] Step 312. Transfer the feature map obtained in Step 309. The input is processed by the convolutional normalization module 8, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 8 has 512 input channels, 256 output channels, and a kernel size of 1. 1. The step size is 1; its operation process is the same as that of the convolution normalization processing module 6.

[0069] This operation performs channel compression and feature integration on high-level semantic features, reducing channel dimensions while maintaining spatial structure information. This allows the feature map to retain rich semantic information while reducing the computational complexity in subsequent feature modeling, thereby improving the overall computational efficiency of the network.

[0070] Step 313. Transfer the feature map obtained in Step 312. The input is processed by the self-attention feature encoding module, and a global semantically enhanced feature map is obtained through positional encoding, self-attention operation, and feedforward network transformation. The self-attention feature encoding module has a feature embedding dimension of 256, 8 attention heads, a hidden layer dimension of 1024, a random deactivation probability of 0.1, uses the ReLU activation function, and has a position encoding temperature parameter of 10000.

[0071] Let the input feature map ,in and These represent the height and width of the feature map, respectively. This represents the number of feature channels; the operation process of the self-attention feature encoding module is as follows: first, the two-dimensional feature map is expanded into sequence features: ; in, This represents the number of pixel positions in the feature map; then, two-dimensional positional encoding is added to the sequence features. , obtain the query vector With key vector : ; in, This represents the positional encoding of two-dimensional sine and cosine.

[0072] The self-attention operation process is represented as follows: ; in, This represents the normalization operation. The feature dimension is then represented; subsequently, the features are transformed through a feedforward network: ; in, , Represents a linear transformation matrix. This represents the ReLU activation function.

[0073] The final encoded feature map is obtained as follows: ; in, This means restoring the sequence features to a two-dimensional feature map structure.

[0074] Through this operation, the network can establish global dependencies between pixels across the entire feature map, enabling the feature representation to simultaneously contain local structural information and global contextual semantic information. This enhances the network's ability to semantically represent target regions in complex scenes in UAV images and improves its ability to recognize small-scale targets, dense targets, and complex background regions.

[0075] Step 314. Transfer the feature map obtained in Step 313. The input is processed by the convolutional normalization module 9, which extracts new feature maps through convolutional feature extraction and normalization operations. The convolution normalization processing module 9 has 256 input channels, 256 output channels, and a kernel size of 1. 1. The step size is 1; its operation process is the same as that of the convolution normalization processing module 6.

[0076] This operation performs channel transformation and feature reshaping on the features after self-attention encoding, further enhancing the semantic information expression while keeping the feature map space size unchanged, thereby providing a stable feature representation for subsequent feature upsampling and cross-scale feature fusion.

[0077] Step 315. Transfer the feature map obtained in Step 314. The input feature upsampling module 1 processes the data and obtains a new feature map through spatial interpolation. Among them, the upsampling factor of feature upsampling module 1 is 2, and the nearest neighbor interpolation method is used.

[0078] The operation process of feature upsampling module 1 can be represented as follows: ; in, This indicates an upsampling operation.

[0079] This operation improves the spatial resolution of the feature map, enabling high-level semantic features to be expressed at a finer spatial scale, thus providing a spatially aligned feature foundation for subsequent cross-scale fusion with shallow features.

[0080] Step 316. Transfer the feature map obtained in Step 311. and the feature map obtained in step 315 The input feature concatenation module 1 processes the data and obtains a fused feature map through channel-dimensional concatenation. The operation process of feature splicing module 1 can be represented as follows: ; This represents a feature concatenation operation along the channel dimension.

[0081] This operation fuses high-level semantic features with mid-level structural features, enabling the feature map to contain both rich semantic information and more refined spatial structural information, thereby enhancing the network's ability to represent features of multi-scale targets in UAV images.

[0082] Step 317. Transfer the feature map obtained in Step 316 The input to the multi-path gating state space modeling module 1 is processed, and a new feature map is obtained through multi-path feature modeling and gating state update operations. Among them, the multi-path gating state space modeling module 1 has 512 input channels and 256 output channels.

[0083] The computation process of the multi-path gating state-space modeling module 1 can be represented as follows: ; in, This represents the feature transformation process of the multi-path gating state-space modeling module.

[0084] Through this operation, the network can dynamically model input features on multiple feature paths and adaptively select and fuse feature information from different paths through a gating mechanism, thereby enhancing the ability of the feature map to express the spatial structure information of the target region in the UAV image and improving the network's detection performance for multi-scale targets in complex background environments.

[0085] Step 318. Transfer the feature map obtained in Step 317. The input is processed by the convolutional normalization module 10, which obtains a new feature map through convolutional feature extraction and normalization operations. The convolution normalization processing module 10 has 256 input channels, 256 output channels, and a kernel size of 1. 1. The step size is 1; its operation process is the same as that of the convolution normalization processing module 6.

[0086] This operation reshapes the channel dimension and remaps the features after multi-path gating state space modeling, further enhancing the feature representation capability while maintaining the spatial structure information, thus providing a unified feature representation for subsequent feature upsampling and cross-scale feature fusion.

[0087] Step 319. Transfer the feature map obtained in Step 318 The input feature upsampling module 2 processes the data and obtains a new feature map through spatial interpolation. The feature upsampling module 2 has an upsampling factor of 2 and uses the nearest neighbor interpolation method; its operation process is the same as that of the feature upsampling module 1.

[0088] This operation improves the spatial resolution of the feature map, enabling high-level semantic features to be expressed at a finer spatial scale, thereby enhancing the network's ability to model the spatial structure of small-scale target regions in UAV images.

[0089] Step 320. Transfer the feature map obtained in Step 310. and the feature map obtained in Step 319 The input feature concatenation module 2 processes the data and performs channel-dimensional concatenation to obtain the fused feature map. Its operation process is the same as that of feature splicing module 1.

[0090] This operation fuses shallow structural features with high-level semantic features, enabling the feature map to contain both rich detail information and high-level semantic information, thereby enhancing the network's ability to represent features of small-scale targets and complex background regions in UAV images.

[0091] Step 321. Transfer the feature map obtained in Step 320. The input is processed by the multi-path gating state space modeling module 2, and a new feature map is obtained through multi-path feature modeling and gating state update operations. The multi-path gating state space modeling module 2 has 512 input channels and 128 output channels; its operation process is the same as that of the multi-path gating state space modeling module 1.

[0092] Through this operation, the network can dynamically model fused features on multiple feature paths and adaptively adjust the information transmission of different feature paths through a gating mechanism. This enhances the global semantic expression ability while preserving detailed structural information, thereby improving the network's detection performance for small-scale targets, dense targets, and complex background regions in UAV images.

[0093] Step 322. Transfer the feature map obtained in Step 321. The input feature downsampling module 1 processes the data and obtains a new feature map through convolutional downsampling operations. The feature downsampling module 1 has 128 input channels, 256 output channels, and a convolutional kernel size of 3. 3. The step size is 2, and the SiLU activation function is used.

[0094] The operation process of feature downsampling module 1 can be represented as follows: ; in, This indicates a convolution operation with a kernel size of 3×3 and a stride of 2, used for spatial downsampling; This represents a convolution operation with a kernel size of 1×1 and a stride of 1, used for channel mapping.

[0095] This operation reduces the spatial resolution of the feature map while increasing the channel dimension, enabling the network to model the features of target regions in UAV images within a larger spatial receptive field, thereby enhancing its ability to express target structural information.

[0096] Step 323. Transfer the feature map obtained in Step 318. and the feature map obtained in step 322 The input feature concatenation module 3 processes the data and obtains the fused feature map through channel-dimensional concatenation. Its operation process is the same as that of feature splicing module 1.

[0097] This operation integrates feature information from different network layers, enabling the feature map to simultaneously contain high-resolution spatial structure information and higher-level semantic information, thereby enhancing the network's ability to represent features of target regions in complex scenes in UAV images.

[0098] Step 324. Transfer the feature map obtained in Step 323. The input is processed by the multi-branch feature enhancement module 1, which generates an enhanced feature map through multi-branch feature extraction and feature fusion operations. The multi-branch feature enhancement module 1 has 512 input channels, 256 output channels, 512 intermediate expansion channels, and 64 internal branch channels, and uses the SiLU activation function.

[0099] The operation process of the multi-branch feature enhancement module 1 can be represented as follows: First, perform channel mapping using a 1×1 convolution: ; Then the feature map is divided along the channel dimension: ; Next, feature extraction is performed on a subset of the features: ; ; The features from each branch are then concatenated: ; Finally, channel fusion is performed using 1×1 convolution to obtain the output feature map: ; in, This represents a convolution operation with a kernel size of 1×1. , This represents the feature extraction function composed of convolutional layers. This indicates that feature splicing operations are performed along the channel dimension.

[0100] Through this operation, the network can perform feature modeling on target regions in UAV images on multiple feature branches, so that the feature map contains feature information within different scale receptive fields at the same time, thereby enhancing the network's ability to represent small-scale targets, dense targets and complex background regions, and improving target detection performance.

[0101] Step 325. Transfer the feature map obtained in Step 324. The input feature downsampling module 2 processes the data and obtains a new feature map through convolutional downsampling operations. The feature downsampling module 2 has 256 input channels, 256 output channels, and a convolutional kernel size of 3. 3. The step size is 2, and the SiLU activation function is used; its operation process is the same as that of feature downsampling module 1.

[0102] This operation reduces the spatial resolution of the feature map while maintaining high semantic expressiveness, enabling the network to model the features of target regions in UAV images within a larger spatial receptive domain, thereby enhancing its ability to express the overall structural information of targets in complex scenes.

[0103] Step 326. Transfer the feature map obtained in Step 314. and the feature map obtained in step 325 The input feature concatenation module 4 processes the data and obtains the fused feature map through channel-dimensional concatenation. Its operation process is the same as that of feature splicing module 1.

[0104] This operation fuses feature information from different network layers, enabling the feature map to simultaneously contain high-resolution spatial structure information and deep-level semantic information, thereby enhancing the network's ability to represent features of complex background regions and multi-scale targets in UAV images.

[0105] Step 327. Transfer the feature map obtained in Step 326 The input is processed by the multi-branch feature enhancement module 2, which generates enhanced feature maps through multi-branch feature extraction and feature fusion operations. The multi-branch feature enhancement module 2 has 512 input channels, 256 output channels, 512 intermediate expansion channels, and 64 internal branch channels, and uses the SiLU activation function. Its operation process is the same as that of the multi-branch feature enhancement module 1.

[0106] Through this operation, the network can perform feature modeling on target regions in UAV images on multiple feature branches, so that the feature map contains feature information within different scale receptive fields at the same time, thereby enhancing the network's ability to express small-scale targets, dense targets and complex background regions, and improving target detection performance.

[0107] Step 328. Transfer the feature map obtained in Step 321 The feature map obtained in Step 324 and the feature map obtained in step 327 The input is processed by the decoder module to obtain the final detection output result. The decoder module performs joint modeling on the input multi-scale feature maps, fuses feature information of different spatial resolutions and semantic levels, and classifies and regresses the target region in the UAV image to output the target detection result.

[0108] Through this operation, the network can comprehensively utilize the spatial structure information and semantic information in feature maps of different scales, thereby improving the detection accuracy and localization accuracy of small-scale targets, dense targets, and targets with complex backgrounds in UAV images.

[0109] The above parameters are merely exemplary settings, and those skilled in the art can adjust them according to actual needs.

[0110] Step 4. Input the test set obtained in Step 1 into the improved DEIM network trained in Step 3 to obtain the UAV target detection results.

[0111] To further verify the performance of the method proposed in this invention, experiments were conducted on the VisDrone2019 dataset using the UAV target detection method based on the improved DEIM network proposed in this invention. The results were compared with existing methods to objectively evaluate their performance and effectiveness. Tables 1 and 2 show the comparison of YOLO and COCO metrics between the improved model proposed in this invention and existing models on the VisDrone2019 dataset, respectively.

[0112] Table 1 Table 2 The UAV target detection method provided by this invention includes the following steps: acquiring a target dataset; constructing and improving a DEIM (DETR with Improved Matching for Fast Convergence) network: designing and introducing a multi-receptive-domain adaptive fusion module and a multi-path gated state-space modeling module to obtain an improved DEIM network; training and evaluating the improved DEIM network based on the training and validation sets in the target dataset; inputting the test set data from the target dataset into the trained model to obtain detection results; this invention improves the existing DEIM network to obtain an improved DEIM network, achieving a significant improvement in UAV target detection performance.

[0113] This invention designs a multi-receptive-domain adaptive fusion module and a multi-path gated state-space modeling module based on the original network structure. The multi-receptive-domain adaptive fusion module achieves dynamic integration of features from different receptive domains through channel segmentation, progressive feature enhancement, and multi-domain fusion mechanisms, thereby improving the ability to express multi-scale features. The multi-path gated state-space modeling module constructs a multi-path state-space modeling structure and introduces a gating adjustment mechanism to achieve adaptive weighting and control of features from different propagation paths, thereby enhancing long-range dependency modeling capabilities and suppressing the propagation of redundant information.

[0114] This invention is based on the DEIM network and introduces a multi-receptive-domain adaptive fusion module into the backbone network and a multi-path gated state space modeling module into the feature modeling stage, which structurally improves the original feature extraction and high-level semantic modeling structure. This improvement enhances the cross-scale information fusion capability and long-range dependency modeling capability without destroying the original network framework, thereby improving the feature representation quality in complex scenarios.

[0115] To verify the effectiveness of the method of this invention, experimental comparison and analysis were conducted on the VisDrone2019 dataset. The experimental results show that the improved model proposed in this invention outperforms the YOLO series models and the original DEIM-Dfine-s model in multiple detection metrics.

[0116] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0117] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0118] The present invention also includes a computer program product comprising computer program instructions that, when executed by a processor, implement the steps of the method.

[0119] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A target detection method for UAV aerial images based on improved DEIM, characterized in that, Includes the following steps: Acquire drone image datasets and preprocess them to obtain standardized image datasets; An improved DEIM network is constructed, comprising a backbone network, an encoder, and a decoder; The standardized image dataset is input into the improved DEIM network for training. The backbone network uses a multi-receptive-domain adaptive fusion module to process the input features of standardized image data to obtain multi-scale enhanced features. The encoder uses a multi-path gated state space modeling module to perform channel alignment and cross-scale fusion on the multi-scale enhancement features to obtain semantic enhancement features. The decoder performs decoding based on the semantic enhancement features and outputs the UAV target detection results.

2. The target detection method for UAV aerial images based on improved DEIM according to claim 1, characterized in that, The process of inputting the standardized image dataset into the improved DEIM network for training includes: The convolutional normalization processing module 1 in the backbone network is used to perform convolutional feature extraction and normalization operations on UAV images in the standardized image dataset to obtain the initial feature map; The initial feature map is sequentially input into multiple sets of staged hierarchical feature extraction modules consisting of a convolutional normalization processing module and a multi-receptive field adaptive fusion module for processing to obtain a multi-scale feature map. The multi-scale feature map is channel-aligned using the convolutional normalization processing module in the encoder to obtain an aligned feature map. A self-attention feature encoding module is used to perform global semantic modeling on the high-level feature map in the aligned feature map to obtain semantic encoded features; The semantic encoded features are upsampled using a feature upsampling module to obtain upsampled features. After the upsampled features are concatenated and fused with the aligned feature maps of the corresponding scale, they are input into the multi-path gated state space modeling module for processing to obtain semantically enhanced features. The semantic enhancement features are decoded using a decoder to output the UAV target detection results.

3. The target detection method for UAV aerial images based on improved DEIM according to claim 2, characterized in that, The process of sequentially inputting the initial feature map into multiple sets of staged hierarchical feature extraction modules, consisting of a convolutional normalization processing module and a multi-receptive-field adaptive fusion module, to obtain a multi-scale feature map includes: The initial feature map is input into the staged hierarchical feature extraction module, and downsampling is performed based on the convolution normalization processing module to obtain the downsampled feature map. The downsampled feature map is then input into the multi-receptive-domain adaptive fusion module for multi-receptive-domain feature fusion processing to obtain the enhanced feature map of the current stage. The process of inputting the enhanced feature map of the current stage into the next stage-level feature extraction module and performing downsampling and multi-receptive-domain feature fusion processing in the next stage-level feature extraction module is repeated until a multi-scale feature map is obtained.

4. The target detection method for UAV aerial images based on improved DEIM according to claim 3, characterized in that, The multi-receptive-domain adaptive fusion module includes: a channel expansion module, a feature generation module, a channel equalization module, a progressive enhancement sub-module group, a feature splicing module, and a channel compression fusion module; The process of inputting the downsampled feature map into the multi-receptive-domain adaptive fusion module for multi-receptive-domain feature fusion processing to obtain the enhanced feature map for the current stage includes: The channel expansion module is used to linearly map the downsampled feature map to obtain the first mapped feature; Based on the first mapping feature, convolution operation is performed using feature generation module 1 to obtain basic semantic features; Based on the first mapping feature, the feature generation module 2 performs depthwise separable convolution operation to obtain local texture features; The channel equalization module is used to divide the first mapping feature into channels to obtain the original semantic sub-features and the sub-features to be enhanced; Based on the sub-features to be enhanced, the progressive enhancement sub-module group is used to perform multi-level progressive feature enhancement to obtain enhanced features; The feature splicing module is used to splice and fuse the basic semantic features, the local texture features, the original semantic sub-features, and the enhanced features to obtain the fused features; Based on the fusion features, the channel compression fusion module is used to perform convolution operations and channel compression to obtain the output features of the current stage; The enhanced feature map for the current stage is obtained based on the output features of the current stage.

5. The target detection method for UAV aerial images based on improved DEIM according to claim 4, characterized in that, Based on the sub-features to be enhanced, the process of performing multi-level progressive feature enhancement using the progressive enhancement sub-module group to obtain enhanced features includes: The sub-feature to be enhanced is used as the initial input feature. The initial input feature is then input into the first lightweight grouping adaptive feature aggregation submodule for feature enhancement processing to obtain the first-level enhanced feature. Based on the first-level enhanced features, a second lightweight grouped adaptive feature aggregation submodule is used to perform feature enhancement processing to obtain the second-level enhanced features; Similarly, based on the enhanced features of the previous level, the next level of lightweight grouping adaptive feature aggregation submodule is used for feature enhancement processing, until the progressive processing of n lightweight grouping adaptive feature aggregation submodules is carried out to obtain the nth level enhanced features; The initial input features, the first-level enhancement features, the second-level enhancement features, and up to the nth-level enhancement features are collectively used as the enhancement features.

6. The target detection method for UAV aerial images based on improved DEIM according to claim 2, characterized in that, The multi-path gated state space modeling module includes: a channel mapping module, a normalization module, a gated enhanced state space feature reconstruction sub-module, a random path discarding module, a hierarchical scaling module, a residual scaling module, and a channel mixing module; The process of concatenating and fusing the upsampled features with the aligned feature map at the corresponding scale, and then inputting it into the multi-path gated state space modeling module for processing to obtain semantically enhanced features includes: The upsampled features and the aligned feature map of the corresponding scale are concatenated along the channel dimension using a feature concatenation module to obtain the concatenated features; The channel mapping module is used to linearly map the splicing features to obtain the first mapped features; The first mapping feature is normalized using the normalization module to obtain the first normalized feature; The first normalized feature is subjected to state-space modeling and gating fusion processing using the gated enhanced state-space feature reconstruction submodule to obtain the reconstructed feature. The reconstructed features are processed by the random path discarding module to obtain the first discarded features; The hierarchical scaling module is used to adjust the hierarchical scale of the first discarded feature to obtain the first scaled feature; The residual scaling module is used to adjust the residual ratio of the first mapping feature to obtain the second scaled feature; The first fused feature is obtained by residual fusion based on the first scaling feature and the second scaling feature; The first fused feature is normalized using the normalization module to obtain the second normalized feature; The channel mixing module is used to perform inter-channel information interaction and nonlinear mapping processing on the second normalized feature to obtain the mixed feature; The random path discarding module is used to perform random path discarding processing on the mixed features to obtain the second discarded feature; The hierarchical scaling module is used to adjust the hierarchical scale of the second discarded feature to obtain the third scaled feature; The residual scaling module is used to adjust the residual ratio of the first fused feature to obtain the fourth scaled feature; The semantic enhancement feature is obtained by residual fusion based on the third scaling feature and the fourth scaling feature.

7. The target detection method for UAV aerial images based on improved DEIM according to claim 1, characterized in that, The process of the decoder using DFINETransformer to decode the semantically enhanced features includes: The DFINETransformer is used to perform unified dimensional mapping and feature integration on the semantic enhancement features to obtain integrated features; Based on the integrated features, a query selection mechanism is used to filter candidate target regions from the feature map to obtain candidate region features; A multi-layer decoding structure is used to progressively update and optimize the semantic and spatial location information of the candidate region features to obtain the target category prediction result and the target bounding box position; Based on the target category prediction result and the target bounding box position, the UAV target detection result is output.

8. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in claim 1.

9. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 1.

10. A computer program product comprising computer program instructions, characterized in that, When the computer program instructions are executed by the processor, they implement the steps of the method of claim 1.

Images