Unmanned aerial vehicle target detection method based on cross-spatial frequency domain and electronic equipment

Abstract

The invention discloses an unmanned aerial vehicle target detection method based on a cross-spatial frequency domain and electronic equipment, and belongs to the technical field of image detection. According to the method, an unmanned aerial vehicle target detection network based on a cross-space frequency domain is constructed to carry out unmanned aerial vehicle target real-time detection; the unmanned aerial vehicle target detection network is composed of a cross-space frequency domain feature extraction network, a detail information hybrid encoder and a position relation decoder; inputting the unmanned aerial vehicle target image into a cross-space frequency domain feature extraction network, and extracting multi-scale frequency domain features of the unmanned aerial vehicle target image; performing multi-scale feature fusion and feature enhancement to obtain a multi-scale feature map; and the position relation decoder generates a target category and bounding box coordinates according to the multi-scale feature map in combination with the relative position relation features between the targets. According to the method, complementary information of a spatial domain and a frequency domain and potential spatial position relation characteristics between targets can be fully utilized, and the target detection precision and robustness of the unmanned aerial vehicle under a complex background are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection, and particularly to a method for detecting unmanned aerial vehicle (UAV) targets based on cross-spatial frequency domain and an electronic device. Background Art

[0002] With the rapid development of UAV technology and the gradual reduction of costs, UAVs have been widely used in many fields. In these application scenarios, it is a key task to use the vision sensors carried by UAVs to detect specific targets in the field of view in real time and accurately, which is directly related to the success, failure and efficiency of the tasks. However, since UAVs usually fly at relatively high altitudes and the sizes of the observed targets themselves are small, in the acquired images, the targets (especially small targets) only occupy very few pixel areas, with low resolution, lack of rich texture and structural details, and are also affected by complex environments. This poses a huge challenge to the effective detection of targets.

[0003] Currently, the mainstream deep learning object detection models mainly include methods such as YOLO and Faster R-CNN based on convolutional neural network (CNN) and methods such as DETR series based on Transformer architecture. The former mainly relies on convolutional operations to extract local features, and the latter uses self-attention mechanism to model global dependencies. Although these deep learning object detection models perform excellently in many scenarios, when applied to the object detection task in the complex environment from the UAV perspective, there is still room for improvement in terms of accuracy, robustness and the ability to capture small targets in complex environments.

[0004] Most of the existing object detection methods mainly perform feature extraction and relationship modeling in the spatial domain of images, focusing on learning the local shape, texture, context information or global spatial dependencies of targets. However, for small targets with sparse pixels and unclear features, relying solely on spatial domain information may not be sufficient to effectively distinguish them from complex backgrounds or noises. The frequency domain information of an image contains the frequency characteristics of the gray level changes in the image, usually corresponding to the sudden change regions such as edges and details of the image and the overall contour and smooth regions of the image. Theoretically, the frequency domain information contains the key clues for distinguishing targets. However, the existing detection methods often ignore the effective use of frequency domain information and fail to design a suitable mechanism to efficiently fuse the complementary information in the spatial domain and the frequency domain to enhance the feature expression of small targets.

[0005] Therefore, although significant progress has been made in deep learning object detection technology, in the specific and challenging scenario of small object detection in drones, there is still room for improvement in the existing methods in terms of detection accuracy (especially for small objects), robustness to complex backgrounds, and model efficiency. How to design a method that can fully utilize the multi-dimensional information of images, especially effectively combine the spatial domain features and frequency domain features and the spatial position relationship between objects, to overcome problems such as insufficient object information and susceptibility to background interference, thereby improving the object detection performance under the drone platform. Summary of the Invention

[0006] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a drone object detection method and an electronic device based on cross-space frequency domain.

[0007] The technical solution for the present invention to solve its technical problems is: a drone object detection method based on cross-space frequency domain, constructing a drone object detection network based on cross-space frequency domain and training it using a drone object image dataset, and using the trained drone object detection network for real-time drone object detection; the drone object detection network consists of a cross-space frequency domain feature extraction network, a detail information hybrid encoder, and a position relationship decoder; inputting the drone object image into the cross-space frequency domain feature extraction network to extract the multi-scale frequency domain features of the drone object image; inputting the multi-scale frequency domain features into the detail information hybrid encoder for multi-scale feature fusion and feature enhancement to obtain a multi-scale feature map; the position relationship decoder receives the multi-scale feature map generated by the detail information hybrid encoder and combines the relative position relationship features between objects to generate the object category and bounding box coordinates.

[0008] Further preferably, the cross-space frequency domain feature extraction network consists of a convolutional layer and four sequentially connected cross-space frequency domain modules, which can be respectively called the first cross-space frequency domain module, the second cross-space frequency domain module, the third cross-space frequency domain module, and the fourth cross-space frequency domain module. Among them, the frequency domain features output by the second cross-space frequency domain module, the third cross-space frequency domain module, and the fourth cross-space frequency domain module are used as the first-scale frequency domain feature, the second-scale frequency domain feature, and the third-scale frequency domain feature respectively.

[0009] Further preferably, the cross-space frequency domain module includes pointwise convolution and a spatial frequency domain module. The input features are processed by pointwise convolution, and then channel splitting is performed. The features of one branch after channel splitting are processed by three consecutive spatial frequency domain modules. The features of the other branch after channel splitting are channel concatenated with the features output by the three spatial frequency domain modules, and finally, pointwise convolution is performed again to obtain the output of the cross-space frequency domain module.

[0010] Further preferably, the process of the detail information hybrid encoder for multi-scale feature fusion and feature enhancement is as follows: Use pointwise convolution to adjust the number of channels of the first-scale frequency-domain features, the second-scale frequency-domain features, and the third-scale frequency-domain features respectively; after the third-scale frequency-domain features pass through pointwise convolution, they are input into the intra-scale feature interaction module based on self-attention. Calculate the attention weights between positions within the third-scale frequency-domain features through the intra-scale feature interaction module based on self-attention. The features output by the intra-scale feature interaction module based on self-attention are convolved to obtain feature P1. Feature P1 is upsampled and then concatenated with the second-scale frequency-domain features after pointwise convolution in the channel dimension, and then feature enhancement is performed through the detailed information enhancement module and convolution processing is carried out to obtain feature P2; feature P2 is upsampled and then concatenated with the first-scale frequency-domain features after pointwise convolution in the channel dimension, and feature enhancement is performed through the detailed information enhancement module to obtain feature P3; feature P3 is convolved and then concatenated with feature P2 in the channel dimension, and feature enhancement is performed through the detailed information enhancement module to obtain feature P4; feature P4 is convolved and then concatenated with feature P1 in the channel dimension, and feature enhancement is performed through the detailed information enhancement module to obtain feature P5; feature P3, feature P4, and feature P5 are concatenated in the channel dimension to obtain a multi-scale feature map.

[0011] Further preferably, the detailed information enhancement module first performs a convolution on the input features for feature transformation; subsequently, the transformed features are fed into three processing paths in parallel: the first path first performs channel splitting, and the features of some channels are input into three cascaded detailed information enhancement units. The features output by the three detailed information enhancement units and the features that have not been processed by the detailed information enhancement units are used as the output of the first path; the second path performs convolution, depth convolution, and pointwise convolution in sequence; the third path only performs a convolution once; finally, the features output by the three paths are concatenated in the channel dimension, and information is integrated through convolution to obtain the final output of the detailed information enhancement module.

[0012] Further preferably, the position relationship decoder first uniformly adjusts the number of channels of the multi-scale feature map to a fixed dimension and flattens it into a one-dimensional feature sequence; generates dense reference anchor boxes according to the shape of the multi-scale feature map to cover the potential target areas of the image; uses a prediction head to process the one-dimensional feature sequence to generate initial class scores and bounding box offset predictions; selects the top N candidate anchor points and their corresponding features as the initial input for the decoding process; the decoder is stacked by three identical decoding layers and gradually refines the query and reference bounding boxes in an iterative manner; each decoding layer includes the following steps: performs multi-head self-attention calculation on the query embedding of the current layer, and adjusts the interaction intensity between different queries in the query embedding by an attention mask; performs multi-head deformable cross-attention calculation on the updated query embedding and the multi-scale feature map to extract features from dynamically sampled feature point positions; the output of the cross-attention is subjected to non-linear feature transformation through a feed-forward network, and all operation results are integrated through residual connection and layer normalization; uses a classification head and a regression head to predict the class score and bounding box offset of the target based on the output features of the current layer respectively, and applies the predicted bounding box offset to the input reference bounding box of the current layer for bounding box refinement to obtain a normalized refined bounding box.

[0013] Further preferably, the position relationship decoder explicitly calculates the relative position relationship features between targets. For the attention mask of the i-th layer, according to the refined bounding box output by this layer and the reference bounding box input by the previous layer, the relative position relationship weight matrix is calculated through a dedicated position relationship encoding module; the relative position relationship weight matrix is encoded by calculating the relative position relationship features between the refined bounding box output by the current layer and the reference bounding box of the previous layer and using a set of linear layer-Gaussian error linear unit activation function-linear layer.

[0014] The present invention also provides an electronic device, including a memory and a processor. The memory stores computer-readable instructions. When the instructions are executed by the processor, the processor is enabled to implement the above-mentioned UAV target detection method based on cross-space frequency domain.

[0015] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned UAV target detection method based on cross-space frequency domain is implemented.

[0016] The beneficial effects of the present invention: (1) Existing methods such as YOLOv8 only use spatial convolution, while the present invention adopts a cross - spatial - frequency - domain feature extraction network. By jointly optimizing the frequency - domain response through the cross - spatial - frequency - domain module, it can efficiently fuse the spatial details and frequency - domain characteristics of images simultaneously, deeply integrate global and local context information, significantly enhance the ability to distinguish targets from the background, overcome the problem of easy confusion between the two, and significantly improve the edge feature extraction ability of small targets.

[0017] (2) Given that targets often appear as detailed information in images, this detail - information hybrid encoder strengthens the detailed information of the features after multi - scale fusion, thereby focusing on improving the network's target detection performance.

[0018] (3) The present invention proposes a position - relationship decoder. By explicitly calculating the spatial relative - position features between targets and integrating this relationship information into the bounding - box prediction process, it provides crucial context assistance for decoding, and thus significantly improves the target detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a structural diagram of the UAV target detection network of the present invention.

[0020] Figure 2 It is a structural diagram of the cross - spatial - frequency - domain module of the present invention.

[0021] Figure 3 It is a structural diagram of the detail - information enhancement module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0022] The present invention will be further described in detail below in conjunction with the embodiments and the drawings.

[0023] In image processing, detailed information mainly includes fine textures of objects, fine details such as hair, etc. The detection of small targets highly depends on whether the algorithm can effectively extract and utilize the detailed information in the image. If the detailed information is insufficient or lost during the processing, it is very difficult to detect small targets. Therefore, when designing a target - detection algorithm, how to better retain and enhance the detailed information in the features is crucial for improving the small - target detection performance.

[0024] To improve the small - target detection performance, this embodiment provides a UAV target detection method based on cross - spatial - frequency domain, including: constructing a UAV target detection network based on cross - spatial - frequency domain and training it using a UAV target - image data set, and using the trained UAV target detection network for real - time UAV target detection. As Figure 1As shown, the UAV target detection network consists of a cross - spatial - frequency - domain feature extraction network, a detailed information hybrid encoder, and a position - relationship decoder. The UAV target image is input into the cross - spatial - frequency - domain feature extraction network to extract the multi - scale frequency - domain features of the UAV target image. The multi - scale frequency - domain features are input into the detailed information hybrid encoder for multi - scale feature fusion and feature enhancement to obtain a multi - scale feature map. The position - relationship decoder receives the multi - scale feature map generated by the detailed information hybrid encoder and combines the relative position - relationship features between targets to generate the target category and bounding - box coordinates.

[0025] The UAV target image dataset used to train the UAV target detection network can be constructed in the following way: S1.1. Patrol and photograph the target area through the image sensor carried by the UAV to obtain the original UAV target image dataset. Alternatively, a publicly available UAV target image dataset that meets the application - scenario requirements can be used as the basic data source.

[0026] S1.2. Manually or using an automated script, screen the collected original UAV target image dataset, remove images with poor quality to ensure the basic quality of the UAV target image dataset. Use annotation tools to accurately label the bounding boxes of the targets, especially small targets, in the screened UAV target image dataset and assign corresponding category labels.

[0027] S1.3. For categories with a small number of samples or specific types of small targets in the UAV target image dataset, apply a series of data - augmentation techniques, such as random rotation, cropping, scaling, color jitter, Mosaic, or MixUp, etc., to expand the scale of the UAV target image dataset and improve the generalization ability of the model and its robustness to small targets.

[0028] S1.4. Divide the pre - processed and augmented UAV target image dataset into a training set and a validation set according to a predetermined ratio for the training and performance evaluation of the UAV target detection network.

[0029] Set the corresponding parameters before training, including: the number of training epochs, the training batch size, the learning rate, the input image size, and the AdamW optimizer. Use the constructed training set to train the UAV target detection network and verify the detection effect through the validation set.

[0030] The core task of the cross - spatial frequency - domain feature extraction network is to capture the frequency - domain information of an image at different scales and generate a set of multi - scale frequency - domain features. Subsequently, these multi - scale frequency - domain features are passed to the detailed information mixing encoder; the detailed information mixing encoder is responsible for effectively fusing and interacting with the multi - scale frequency - domain features to integrate global and local information. The detailed information enhancement module in it processes the fused features and specifically enhances the detailed information in them, aiming to sharpen the details and highlight the unique features of small targets. Finally, the enhanced multi - scale feature maps are fed into the positional relationship decoder. This positional relationship decoder not only utilizes rich feature information but also explicitly combines the potential positional relationships between targets, and finally performs accurate bounding - box localization and class prediction on the targets.

[0031] As Figure 1 shown, the cross - spatial frequency - domain feature extraction network consists of a convolutional layer and four sequentially connected cross - spatial frequency - domain modules. The four cross - spatial frequency - domain modules can be respectively called the first cross - spatial frequency - domain module, the second cross - spatial frequency - domain module, the third cross - spatial frequency - domain module, and the fourth cross - spatial frequency - domain module. Among them, the frequency - domain features output by the second cross - spatial frequency - domain module, the third cross - spatial frequency - domain module, and the fourth cross - spatial frequency - domain module are used as the first - scale frequency - domain feature, the second - scale frequency - domain feature, and the third - scale frequency - domain feature respectively.

[0032] As Figure 2 shown, the cross - spatial frequency - domain module includes point - wise convolution and a spatial frequency - domain module. The input features are processed by point - wise convolution and then channel - split. The features of one branch after channel - split are processed by three consecutive spatial frequency - domain modules. The features of the other branch after channel - split are channel - concatenated with the features output by the three spatial frequency - domain modules, and finally, another point - wise convolution is performed to obtain the output of the cross - spatial frequency - domain module. The cross - spatial frequency - domain module designed in this way can reduce the computational amount while maintaining the flow of gradient information, enhance the learning ability of the model, and can effectively fuse the spatial frequency - domain features.

[0033] The processing process of the cross - spatial frequency - domain module can be described as: ; ; ; ; ; where, represents the features input to the cross - spatial frequency - domain module; represents the point - wise convolution for channel adjustment; represents channel - split, which evenly divides the feature channels into two parts, is the feature for the main - path processing, For the characteristics of the bypass connection part; respectively represent three consecutive spatial frequency domain modules, represents the output characteristics of the first spatial frequency domain module, represents the output characteristics of the second spatial frequency domain module, represents the output characteristics of the third spatial frequency domain module; represents the channel concatenation operation; represents the pointwise convolution for integrating features; is the output characteristics of the cross-spatial frequency domain module.

[0034] The process of the detailed information hybrid encoder for multi-scale feature fusion and feature enhancement is as follows: The number of channels of the first-scale frequency domain features, the second-scale frequency domain features, and the third-scale frequency domain features are respectively adjusted using pointwise convolution; after the third-scale frequency domain features with lower resolution and richer semantic information are processed by pointwise convolution, they are input into the in-scale feature interaction module based on self-attention, and the attention weights between positions within the third-scale frequency domain features are calculated through the in-scale feature interaction module based on self-attention, enabling the third-scale frequency domain features to adaptively aggregate context information according to their own content, thereby achieving a more efficient and flexible in-scale feature representation. The features output by the in-scale feature interaction module based on self-attention are convolved to obtain feature P1, and after feature P1 is upsampled, it is concatenated with the second-scale frequency domain features after pointwise convolution, and then feature enhancement is performed through the detailed information enhancement module and convolution processing to obtain feature P2; after feature P2 is upsampled, it is concatenated with the first-scale frequency domain features after pointwise convolution, and feature enhancement is performed through the detailed information enhancement module to obtain feature P3; after feature P3 is convolved, it is concatenated with feature P2 in channels, and feature enhancement is performed through the detailed information enhancement module to obtain feature P4; after feature P4 is convolved, it is concatenated with feature P1 in channels, and feature enhancement is performed through the detailed information enhancement module to obtain feature P5; feature P3, feature P4, and feature P5 are concatenated in channels to obtain a multi-scale feature map.

[0035] The detailed information enhancement module focuses on deeply extracting and strengthening high-frequency information. Such as Figure 3As shown, the detailed information enhancement module first performs a convolution on the input features for feature transformation. Subsequently, the transformed features are fed into three processing paths in parallel: in the first path, channel splitting is first performed, and the features of some channels are input into three cascaded detailed information enhancement units. The features output by the three detailed information enhancement units and the features that have not been processed by the detailed information enhancement units are used as the output of the first path, aiming to focus on enhancing the detailed information related to small targets; the second path performs convolution, depth convolution, and pointwise convolution in sequence; the third path only performs a convolution once. Finally, the features output by the three paths are concatenated in the channel dimension, and the information is integrated through convolution to generate the final output of this detailed information enhancement module.

[0036] Organize top-down high-resolution to low-resolution and bottom-up low-resolution to high-resolution bidirectional paths to achieve dense interaction of cross-scale features. Each channel concatenation is responsible for processing a pair of adjacent-scale features. The concatenated features are then fed into a detailed information enhancement module for fine-grained depth extraction of detailed information, especially enhancing the details related to small targets, and outputting the features at this level. By stacking and iterating multiple such channel concatenations according to the above bidirectional paths, it is ensured that the feature information between different scales can flow and fuse fully and bidirectionally, and finally global and deep cross-scale feature integration is achieved.

[0037] The multi-scale feature maps simultaneously contain rich semantic and detailed information. Finally, the multi-scale feature maps are passed to the subsequent position relationship decoder for performing the final object detection and precise localization tasks.

[0038] The position relationship decoder receives the multi-scale feature maps generated by the detailed information hybrid encoder and combines the relative position relationship features between objects to generate the object category and bounding box coordinates. Specifically, the number of channels of the multi-scale feature maps is first uniformly adjusted to a fixed dimension and flattened into a one-dimensional feature sequence. Dense reference anchor boxes are generated according to the shape of the multi-scale feature maps to cover the potential target regions of the image. A lightweight prediction head is used to preliminarily process the one-dimensional feature sequence to generate initial class scores and bounding box offset predictions. The 300 candidate anchor points with the highest scores and their corresponding features are selected as the initial inputs for the decoding process.

[0039] The decoder is composed of three identical decoding layers stacked together, gradually refining the query and reference bounding boxes in an iterative manner. Each layer of the decoder includes the following steps: The query embedding of the current layer undergoes multi-head self-attention calculation, and the interaction intensity between different queries in the query embedding is regulated by an attention mask; the updated query embedding performs multi-head deformable cross-attention calculation with the multi-scale feature maps to extract features from the dynamically sampled feature point positions. The output of the cross-attention undergoes non-linear feature transformation through a feed-forward network, and all operation results are integrated through residual connections and layer normalization. A classification head and a regression head are used to predict the class scores and bounding box offsets of the targets respectively based on the output features of the current layer, and the predicted bounding box offsets are applied to the input reference bounding boxes of the current layer for bounding box refinement to obtain normalized refined bounding boxes.

[0040] Specifically, the position relationship decoder significantly improves the accuracy of object detection by explicitly calculating the relative position relationship features between objects. For the attention mask of the i-th layer, based on the refined bounding boxes output by this layer and the reference bounding boxes input by the previous layer, a relative position relationship weight matrix is calculated through a dedicated position relationship encoding module. The relative position relationship weight matrix is obtained by calculating the relative position relationship features between 300 refined bounding boxes output by the current layer and 300 reference bounding boxes of the previous layer, and is encoded using a set of linear layer - Gaussian error linear unit activation function - linear layer; the relative position relationship weight matrix is incorporated into the attention mask of the self-attention mechanism of the current i-th layer to guide and assist the current layer in generating more accurate bounding boxes. This design enables the decoder to effectively decode the accurate object positions and classes from the global image features.

[0041] The processing process of the relative position relationship weight matrix can be described as: ; ; , , ; ; ; Among them, , is the serial number index of the bounding box, N is the number of bounding boxes, , respectively represent the abscissa and ordinate of the center point of the bounding box ; , respectively represent the width of the bounding box and the width of the bounding box; , respectively represent the bounding box The abscissa and ordinate of the center point of , respectively represent the width of the bounding box and the width of the bounding box; represents the bounding box and the bounding box relative position relationship feature; represents the bounding box and the bounding box area of the intersection part, and respectively represent the bounding box and the bounding box area; represents the bounding box and the bounding box relative direction; are 300 refined bounding boxes output by the current layer, are 300 reference bounding boxes input by the previous layer, is the i-th refined bounding box of the current layer, is the j-th reference bounding box of the previous layer, is the relative position relationship feature between the i-th refined bounding box of the current layer and the j-th reference bounding box of the previous layer, is the relative position relationship feature set, represents the linear layer, represents the Gaussian error linear unit activation function, is the relative position relationship weight matrix.

[0042] In the training stage, the class scores and the list of refined bounding boxes of all decoding layers are used for loss calculation and model optimization; during inference, only the final refined bounding boxes of the last decoding layer and the corresponding class confidence are concerned. By gradually refining, feature point sampling of deformable cross-attention, and introducing the relative position relationship weight matrix between bounding boxes, the position relationship decoder can significantly improve the accuracy and robustness of object detection.

[0043] To verify the effectiveness of the proposed UAV target detection method based on cross - spatial frequency domain in UAV target detection, a series of comparative experiments were conducted using the publicly available VisDrone UAV small target dataset. The experimental results were compared with YOLOv8, YOLOv9, YOLOv10, YOLOv11, GFL, RT - DETR, and RT - DETR+CSFHF algorithms respectively. GFL represents the Generalized Focal Loss algorithm, RT - DETR represents the Real - Time Detection Transformer, and RT - DETR+CSFHF represents RT - DETR plus the designed Cross - Spatial Frequency Domain Feature Extraction Network (CSFHF). Accuracy, recall, number of parameters, computational cost, and mean average precision evaluation metrics were selected to evaluate the advantages and disadvantages of the algorithms. Accuracy reflects the overall correctness of the algorithm's predictions, while recall focuses on its ability to find all relevant targets. For tasks such as object detection, mean average precision is a core metric that comprehensively evaluates the accuracy of the model in localizing and identifying objects at a 0.5 intersection - over - union threshold. At the same time, the efficiency and resource consumption of the model were also concerned: the number of parameters reveals the size and complexity of the model, and the computational cost measures the computational resources and speed required for its operation. By comprehensively analyzing these metrics, the performance, efficiency, and practicality of an algorithm can be judged more comprehensively. The specific experiments are shown in Table 1. The experimental results show that the UAV target detection method based on cross - spatial frequency domain achieves better results than other algorithms with lower numbers of parameters and computational costs.

[0044] Table 1 Comparative Experimental Results

[0045] The second embodiment of the present invention provides an electronic device, including a memory and a processor. The memory stores computer - readable instructions. When the instructions are executed by the processor, the processor is caused to implement the above - mentioned UAV target detection method based on cross - spatial frequency domain.

[0046] The third embodiment of the present invention provides a computer - readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the above - mentioned UAV target detection method based on cross - spatial frequency domain.

[0047] The above only expresses the preferred embodiments of the present invention and does not limit the present invention in other forms. Any person skilled in the relevant art may use the disclosed content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting UAV targets based on cross - spatial frequency domain, characterized in that, Construct a drone target detection network based on cross - space frequency domain and train it using a drone target image dataset, and use the trained drone target detection network for real - time drone target detection; the drone target detection network consists of a cross - space frequency domain feature extraction network, a detailed information hybrid encoder, and a position relationship decoder; input the drone target image into the cross - space frequency domain feature extraction network to extract multi - scale frequency domain features of the drone target image; Input the multi - scale frequency domain features into the detailed information hybrid encoder for multi - scale feature fusion and feature enhancement to obtain a multi - scale feature map; The position relationship decoder receives the multi - scale feature map generated by the detailed information hybrid encoder and combines the relative position relationship features between targets to generate the target category and bounding box coordinates.

2. The method for detecting an unmanned aerial vehicle target based on cross-space frequency domain according to claim 1, characterized in that, The cross - space frequency domain feature extraction network consists of a convolutional layer and four sequentially connected cross - space frequency domain modules. The four cross - space frequency domain modules are respectively called the first cross - space frequency domain module, the second cross - space frequency domain module, the third cross - space frequency domain module, and the fourth cross - space frequency domain module. Among them, the frequency domain features output by the second cross - space frequency domain module, the third cross - space frequency domain module, and the fourth cross - space frequency domain module are respectively used as the first - scale frequency domain feature, the second - scale frequency domain feature, and the third - scale frequency domain feature.

3. The method for detecting an unmanned aerial vehicle target based on cross-space frequency domain according to claim 2, wherein The cross - space frequency domain module includes point - wise convolution and a space - frequency domain module. The input features are processed by point - wise convolution, and then channel splitting is performed. The features of one branch after channel splitting are processed by three consecutive space - frequency domain modules. The features of the other branch after channel splitting are channel - concatenated with the features output by the three space - frequency domain modules, and finally, point - wise convolution is performed again to obtain the output of the cross - space frequency domain module.

4. The method for detecting UAV targets based on cross-space frequency domain according to claim 2, characterized in that, The process of the detailed information hybrid encoder for multi - scale feature fusion and feature enhancement is as follows: Use point - wise convolution to adjust the number of channels of the first - scale frequency domain feature, the second - scale frequency domain feature, and the third - scale frequency domain feature respectively; after the third - scale frequency domain feature is processed by point - wise convolution, it is input into the intra - scale feature interaction module based on self - attention. Calculate the attention weights between positions inside the third - scale frequency domain feature through the intra - scale feature interaction module based on self - attention. The features output by the intra - scale feature interaction module based on self - attention are convolved to obtain feature P1. Feature P1 is upsampled and then channel - concatenated with the second - scale frequency domain feature after point - wise convolution, and then feature enhancement is performed through the detailed information enhancement module and convolution processing to obtain feature P2; feature P2 is upsampled and then channel - concatenated with the first - scale frequency domain feature after point - wise convolution, and feature enhancement is performed through the detailed information enhancement module to obtain feature P3; feature P3 is convolved and then channel - concatenated with feature P2, and feature enhancement is performed through the detailed information enhancement module to obtain feature P4; feature P4 is convolved and then channel - concatenated with feature P1, and feature enhancement is performed through the detailed information enhancement module to obtain feature P5; channel - concatenate feature P3, feature P4, and feature P5 to obtain a multi - scale feature map.

5. The method for detecting an unmanned aerial vehicle target based on cross-space frequency domain according to claim 4, characterized in that, The detailed information enhancement module first performs a convolution on the input features for feature transformation; Subsequently, the transformed features are fed into three processing paths in parallel: in the first path, channel splitting is first performed, and the features of some channels are input into three cascaded detailed information enhancement units. The features output by the three detailed information enhancement units and the features not processed by the detailed information enhancement units are used as the output of the first path; in the second path, convolution, depth convolution, and pointwise convolution are performed in sequence; in the third path, only one convolution is performed; finally, the features output by the three paths are concatenated in the channel dimension, and the information is integrated through convolution to obtain the final output of the detailed information enhancement module.

6. The method for detecting an unmanned aerial vehicle target based on cross-spatial frequency domain according to claim 1, characterized in that, The position relationship decoder first uniformly adjusts the number of channels of the multi-scale feature map to a fixed dimension and flattens it into a one-dimensional feature sequence; dense reference anchor boxes are generated according to the shape of the multi-scale feature map to cover the potential target areas of the image; a prediction head is used to process the one-dimensional feature sequence to generate initial class scores and bounding box offset predictions; the N candidate anchor points with the highest scores and their corresponding features are selected as the initial inputs for the decoding process; the decoder is stacked by three identical decoding layers and gradually refines the query and reference bounding boxes in an iterative manner; each decoding layer includes the following steps: multi-head self-attention calculation is performed on the query embedding of the current layer, and the interaction strength between different queries in the query embedding is adjusted by an attention mask; the updated query embedding performs multi-head deformable cross-attention calculation with the multi-scale feature map to extract features from the dynamically sampled feature point positions; the output of the cross-attention is subjected to non-linear feature transformation through a feed-forward network, and all operation results are integrated through residual connection and layer normalization; a classification head and a regression head are used to predict the class scores and bounding box offsets of the target based on the output features of the current layer respectively, and the predicted bounding box offsets are applied to the input reference bounding box of the current layer for bounding box refinement to obtain the normalized refined bounding box.

7. The method for detecting an unmanned aerial vehicle target based on cross-space frequency domain according to claim 6, wherein The position relationship decoder explicitly calculates the relative position relationship features between targets. For the attention mask of the i-th layer, according to the refined bounding box output by the current layer and the reference bounding box input by the previous layer, a relative position relationship weight matrix is calculated through a dedicated position relationship encoding module; the relative position relationship weight matrix is encoded by calculating the relative position relationship features between the refined bounding box output by the current layer and the reference bounding box of the previous layer and using a set of linear layer-Gaussian error linear unit activation function-linear layer.

8. An electronic device, comprising a memory and a processor, wherein computer-readable instructions are stored in the memory, characterized in that, When the instructions are executed by the processor, the processor is caused to implement the drone target detection method based on cross-space frequency domain according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the drone target detection method based on cross-space frequency domain according to any one of claims 1-7.

Images