Unmanned aerial vehicle target detection method based on improved RTDETR

Abstract

The invention discloses an improved RTDETR-based unmanned aerial vehicle target detection method, and the method comprises the steps: carrying out the feature extraction of four-scale feature maps of an image photographed by an unmanned aerial vehicle through a backbone network employing an LSKNet neural network architecture, improving the feature extraction performance, and reducing the model calculation amount; the feature fusion network is improved to form a new feature extraction and feature fusion network framework, the feature extraction and fusion efficiency is improved, the network calculation amount and the loss in the feature information transmission process are reduced, and the high detection real-time performance is kept while the target detection precision is improved.

Description

Background Art

[0002] UAV target detection technology is a key technology in the intelligent perception system of UAVs, and is widely used in many fields such as forest fire fighting, intelligent security, emergency rescue, anti-terrorism and explosion prevention. The targets under the aerial view of UAVs have complex background information, few pixels occupied by the targets, easy occlusion, and large changes in target scale. This technology is mainly based on the single-stage detection method of deep learning. The features of the drone aerial images are extracted through convolutional neural networks, and then the features are fused to densely predict the target category and regress the object bounding box on the feature map. This process is usually completed by a deep learning model, such as SSD and YOLO series algorithms. The single-stage target detection method is suitable for situations where the computing power of the drone airborne platform is low and real-time performance is required because of its fast speed and relatively small resource occupation.

[0003] The current single-stage target detection methods are represented by the YOLO series of algorithms, which have good real-time performance in some versions, but the accuracy is low when detecting targets in drone overhead images, especially when the scale of the target in the image is very small. With the development of single-stage target detection technology, single-stage detection methods based on Transformer have gradually emerged. Algorithms represented by DETR, Deformable-DETR and RT-DETR have surpassed the YOLO series in terms of accuracy and real-time performance. In previous detection algorithms, the post-processing operation of non-maximum suppression of the prediction box is required, which has a great impact on real-time performance, and the relevant parameter settings will also affect the model accuracy and real-time performance. In the DETR series of algorithms, this operation is removed to improve the overall performance and real-time performance of the model. The RT-DETR algorithm has broad application prospects in the field of drone target detection due to its high real-time performance and accuracy. RT-DETR has poor performance in detecting small targets in drone overhead images. Small target objects are usually detected on high-resolution images, while the feature map resolution used by RT-DETR is low, and the backbone has limited ability to extract small target features. In addition, the large number of model parameters is not conducive to deployment on drone airborne platforms. Summary of the invention

[0004] In view of the problems existing in the prior art, the purpose of the present disclosure is to provide a drone target detection method, an electronic device and a computer-readable storage medium based on improved RTDETR, thereby overcoming one or more problems caused by the limitations and defects of the relevant technology at least to a certain extent.

[0005] The first aspect of the present invention provides a method for detecting unmanned aerial vehicle targets based on improved RTDETR, comprising the following steps:

[0006] The LSKNet neural network architecture is used to extract four scale feature maps B2, B3, B4, and B5 of the drone-photographed images;

[0007] The feature maps B2, B3, B4, and B5 of the four scales are made consistent in size by using average pooling, maximum pooling, and bilinear interpolation. After being merged in the channel dimension, they are convolved by the DSConv convolution module to obtain the global feature map.

[0008] After convolution processing, the global feature map is split using the Split function and the feature maps inject_P3 and inject_P4 are output;

[0009] The feature map B3 and the feature map inject_P3 are fused to generate the feature map P3, and the feature map B4 and the feature map inject_P4 are fused to generate the feature map P4;

[0010] The feature map B5 is processed by HLD Encoder to generate the feature map P5;

[0011] After the feature maps P3 and P4 are processed by average pooling and maximum pooling, they are concatenated with the feature map P5 in the channel dimension, and then processed by the HLD-GD module for feature extraction. After that, they are split using the Split function and the feature maps are output as inject_N4 and inject_N5.

[0012] The feature map P4 and the feature map inject_N4 are fused to generate the feature map N4, and the feature map P5 and the feature map inject_N5 are fused to generate the feature map N5;

[0013] The feature maps P3, N4 and N5 are fused and merged and then output to the target detection head module for target detection.

[0014] Furthermore, the feature maps B2, B3, B4, and B5 of the four scales are combined and then convolved by the DSConv convolution module, including:

[0015] The merged feature map is processed by the CBS module and the processed features are retained;

[0016] The features processed by the CBS module are then processed by the DBS module;

[0017] The features processed by the CBS module and the features processed by the DBS module are concatenated in the channel dimension, and then the channel features of the two parts are randomly mixed through the Shuffle module to output the global feature map.

[0018] Furthermore, the feature map B5 is processed by the HLD Encoder including:

[0019] Input feature map B5 into the first branch's multi-head deformable attention and then concatenate the maps;

[0020] The feature map B5 is average-pooled to use the downsampled feature map as the key and value of the multi-head deformable attention input to the second branch;

[0021] The downsampled feature map is input into the multi-head deformable attention post-concatenation map of the second branch;

[0022] Concatenate the output results of the multi-head deformable attention of the first branch and the multi-head deformable attention of the second branch in the channel dimension;

[0023] Add the original feature map and the corresponding position of the feature map after HLD Attention and normalize them;

[0024] The normalized feature map is added to the corresponding position of the feature map after the feed-forward neural network (FFN) and normalized to obtain the feature map P5.

[0025] Furthermore, feature maps P3 and P4 are processed by average pooling and maximum pooling, and then concatenated with feature map P5 in the channel dimension and subjected to feature extraction processing by the HLD-GD module, including:

[0026] After the feature maps P3 and P4 are processed by average pooling and maximum pooling respectively, a downsampled feature map aligned with the size of the feature map P5 is output;

[0027] The downsampled feature maps P3 and P4 are concatenated with the feature map P5 in the channel dimension;

[0028] The concatenated feature map is input into the HLD Attention module and then concatenated with the map and the feed-forward neural network FFN for feature extraction.

[0029] A second aspect of the present invention provides an electronic device, comprising

[0030] Processor; and

[0031] A memory having computer-readable instructions stored therein, wherein the computer-readable instructions implement the above method when executed by the processor.

[0032] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the above method when executed by a processor.

[0033] The backbone network of the present invention adopts the LSKNet neural network architecture to extract features of four-scale feature maps of images taken by drones, which improves the feature extraction performance while reducing the model calculation amount, and improves the feature fusion network to form a new feature extraction and feature fusion network framework, which improves the feature extraction and fusion efficiency, reduces the network calculation amount and the loss in the feature information transmission process, improves the target detection accuracy while maintaining high detection real-time performance.

[0034] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The above and other features and advantages of the present disclosure will become more apparent by describing in detail example embodiments thereof with reference to the attached drawings.

[0036] Figure 1 A block diagram of the working principle of a UAV target detection method based on improved RTDETR according to an embodiment of the present invention;

[0037] Figure 2 A block diagram of the working principle of the LSKNet module according to an embodiment of the present invention;

[0038] Figure 3 A block diagram of the working principle of the LSKNet neural network architecture according to an embodiment of the present invention;

[0039] Figure 4 A block diagram of the working principle of LSKNet Block according to an embodiment of the present invention;

[0040] Figure 5 A block diagram of the working principle of a DSConv convolution module according to an embodiment of the present invention;

[0041] Figure 6 A block diagram of the working principle of a DS-GD module according to an embodiment of the present invention;

[0042] Figure 7 A block diagram of the working principle of a CBS module according to an embodiment of the present invention;

[0043] Figure 8 A block diagram of the working principle of a DBS module according to an embodiment of the present invention;

[0044] Fig. 9 A block diagram of the working principle of the HLD Attention module according to an embodiment of the present invention;

[0045] Fig.10 A block diagram of the working principle of the HLD Encoder module according to an embodiment of the present invention;

[0046] Fig.11 A block diagram of the working principle of the FFN module according to an embodiment of the present invention;

[0047] Fig.12 The figure is a block diagram of the working principle of the HLD-GD module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be comprehensive and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted.

[0049] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of the specific details, or other methods, components, materials, devices, steps, etc. may be adopted. In other cases, known structures, methods, devices, implementations, materials or operations are not shown or described in detail to avoid blurring the various aspects of the present disclosure.

[0050] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or these functional entities or parts of functional entities may be implemented in one or more software hardened modules, or these functional entities may be implemented in different networks and / or processor devices and / or microcontroller devices.

[0051] like Figure 1 As shown, a UAV target detection method based on improved RTDETR includes the following steps:

[0052] Step S210: Using the LSKNet neural network architecture to extract four scale feature maps B2, B3, B4, and B5 of the drone image; wherein the sizes of the feature maps B2, B3, B4, and B5 decrease in sequence, the feature map B2 has the largest size and contains relatively rich feature information, and the feature map B5 has the smallest size and contains relatively less feature information.

[0053] Step S220: average pooling, maximum pooling and bilinear interpolation are used to make the feature maps B2, B3, B4 and B5 of the four scales consistent in size and then merged in the channel dimension and convolved by the DSConv convolution module to obtain a global feature map;

[0054] Step S230: Split the global feature map using the Split function after convolution processing and output feature maps inject_P3 and inject_P4;

[0055] Step S240: The feature map B3 and the feature map inject_P3 are fused to generate the feature map P3, and the feature map B4 and the feature map inject_P4 are fused to generate the feature map P4;

[0056] Step S250: Processing the feature map B5 through the HLD Encoder to generate the feature map P5;

[0057] Step S260: After the feature maps P3 and P4 are processed by average pooling and maximum pooling, they are concatenated with the feature map P5 in the channel dimension, and then feature extraction is performed by the HLD-GD module, and then the Split function is used to split and output the feature maps inject_N4 and inject_N5;

[0058] Step S270: The feature map P4 and the feature map inject_N4 are fused to generate the feature map N4, and the feature map P5 and the feature map inject_N5 are fused to generate the feature map N5;

[0059] Step S280: The feature map P3, the feature map N4 and the feature map N5 are merged and output to the target detection head module for target detection.

[0060] In one embodiment of the present invention, in step S210, the LSKNet neural network architecture is used to replace the traditional backbone neural network for extraction, which can dynamically adjust the spatial receptive field to better capture the background information characteristics of different objects in the remote sensing scene, thereby improving the model detection capability. LSKNet uses two large convolution kernels and expansion rates to increase the receptive field range, and then uses a spatial selection mechanism to enhance the network's ability to focus on the spatial context area most relevant to the target after fusing the convolved features. Compared with traditional target detection methods, LSKNet improves the detection accuracy of small targets by locating key points, utilizes multi-scale feature information to better adapt to targets of different sizes to improve adaptability, and enhances the ability to extract local features of the target through a spatial selection mechanism, which helps to capture the detailed information of small targets. At the same time, LSKNet improves the feature extraction performance while reducing the model calculation amount, and can achieve good performance under hardware conditions with different computing power.

[0061] like Figure 2As shown in Figure 2, the LSKNet module consists of a large kernel convolution sequence and a spatial kernel selection mechanism, which are embedded in the LK Selection sub-block of the LSK Block (see Figure 4 ). Figure 3 It is the LSKNET backbone network structure. After convolution and normalization, it is input to LSK Block, and then input to the next stage after normalization and sorting. LSKNet Block is a repeatable stacking block in the backbone network. Each LSK Block includes two residual sub-blocks, namely, Large Kernel Selection (LK Selection) and Feed-forward Network (FFN). Figure 4 . The LKSelection sub-block dynamically adjusts the network's receptive field as needed. The FFN sub-block is used for channel mixing and feature refinement, and consists of a fully connected layer, a deep convolution, a GELU activation, and a second fully connected layer. In order to make the model pay more attention to the key background information of the target in space, a spatial selection mechanism is used to spatially select feature maps from large convolution kernels of different scales. First, the features from convolution kernels with different receptive fields are concatenated, and then channel-level average pooling and maximum pooling are applied to extract spatial relationships to obtain the corresponding spatial feature descriptors SA. In order to realize the information interaction of different spatial descriptors, the spatial pooled features are spliced ​​using convolution layers, and the pooled features of 2 channels are converted into N spatial attention feature maps. Afterwards, the Sigmoid activation function is applied to each spatial attention feature map (SA)~ to obtain an independent spatial selection mask corresponding to each decoupled large convolution kernel. Then, the features of the decoupled large convolution kernel sequence are weighted with the corresponding spatial selection mask and fused through the convolution layer to obtain the attention feature S. Finally, the output of the LSK module can be obtained by the element-by-element point synthesis of the input feature X and the attention feature S.

[0062] In one embodiment of the present invention, in step S220, the feature maps B2, B3, B4, and B5 of the four scales are merged and then convolved by the DSConv convolution module, including:

[0063] The merged feature map is processed by the CBS module and the processed features are retained;

[0064] The features processed by the CBS module are then processed by the DBS module;

[0065] The features processed by the CBS module and the features processed by the DBS module are concatenated in the channel dimension, and then the channel features of the two parts are randomly mixed through the Shuffle module to output the global feature map.

[0066] Specifically, if Figure 5 and 6 As shown in the figure, CBS is a convolution-batch normalization-activation module, and DBS is a dilated convolution-batch normalization-activation module. The input feature map first passes through the CBS module to retain the partial features and is spliced ​​with the features that pass through the DBS module in the channel dimension. The channel features of the two parts are randomly mixed through the Shuffle module to output the final feature map. Each compression of the spatial dimension (width and height) and the expansion of the channel dimension will cause the loss of some semantic information. DSConv can maximize the retention of hidden connections between each channel with low time complexity. Dilated convolution allows the network to capture information from a wider context without significantly increasing the number of parameters. By introducing gaps in the convolution kernel, the receptive field is expanded without increasing the computational load, enhancing the network's understanding of complex patterns in the data and improving the detection ability of small targets. The Shuffle module fully mixes the information of CBS into the output of DBS, promotes information exchange between channels, and effectively reduces computational costs.

[0067] like Figure 7 and Figure 8 The following are schematic diagrams of the working principles of the CBS and DBS modules, where the CBS module consists of a standard convolution layer, a batch normalization layer, and a SiLU activation function to extract image features. The DBS module consists of a hole convolution layer, a batch normalization layer, and a SiLU activation function. The hole convolution allows the network to capture information from a wider context without significantly increasing the number of parameters. By introducing gaps in the convolution kernel, the receptive field is expanded without increasing the computational load, enhancing the network's understanding of complex patterns in the data and improving the ability to detect small targets.

[0068] In one embodiment of the present invention, in step S250, processing the feature map B5 by the HLD Encoder includes:

[0069] Input feature map B5 into the first branch's multi-head deformable attention and then concatenate the maps;

[0070] The feature map B5 is average-pooled to use the downsampled feature map as the key and value of the multi-head deformable attention input to the second branch;

[0071] The downsampled feature map is input into the multi-head deformable attention post-concatenation map of the second branch;

[0072] Concatenate the output results of the multi-head deformable attention of the first branch and the multi-head deformable attention of the second branch in the channel dimension;

[0073] Add the original feature map and the corresponding position of the feature map after HLD Attention and normalize them;

[0074] The normalized feature map is added to the corresponding position of the feature map after the feed-forward neural network (FFN) and normalized to obtain the feature map P5.

[0075] Specifically, if Fig. 9 As shown in the figure, the HLD Attention mechanism consists of two branches. One branch uses the input feature map after multi-head deformable attention and then splices and maps it. The other branch first undergoes average pooling and uses the downsampled feature map as the key and value of the input multi-head deformable attention. The query comes from the original feature map. After splicing and mapping, it is spliced ​​with the output result of the first branch in the channel dimension and output. N is the total number of attention heads, and α is the segmentation ratio. HLD Attention uses a deformable attention mechanism to effectively reduce training time while helping the model better understand and utilize spatial relationships. By using two sizes of feature maps, the fusion of global and local features is achieved, thereby improving the ability to detect small targets. The input feature map B5 is as follows Fig.10 The final output of the HLD Encoder. Applying this module to the B5 feature map fully exploits the rich semantic properties of high-level features, greatly reducing computational requirements and improving processing speed without compromising performance robustness. This optimized hybrid encoder enables feature interactions within scales, converting multi-scale features into a sequence of serialized image features, and enhancing the model's ability to capture dense, high-resolution small object features.

[0076] In one embodiment of the present invention, in step S260, the feature maps P3 and P4 are processed by average pooling and maximum pooling, and then spliced ​​with the feature map P5 in the channel dimension and subjected to feature extraction processing by the HLD-GD module, including:

[0077] After the feature maps P3 and P4 are processed by average pooling and maximum pooling respectively, a downsampled feature map aligned with the size of the feature map P5 is output;

[0078] The downsampled feature maps P3 and P4 are concatenated with the feature map P5 in the channel dimension;

[0079] The concatenated feature map is input into the HLD Attention module and then concatenated with the feed-forward neural network FFN for feature extraction. Specifically, Fig.11As shown in the figure, FFN stands for Feed-Forward Neural Network, which is a basic neural network structure used for feature transformation in deep learning models. FFN usually contains two linear transformations with a nonlinear activation function ReLU between them. The first linear layer, called the hidden layer, maps the input features to a larger dimension, and the activation function ReLU is applied to the hidden layer to introduce nonlinearity, which helps the network capture more complex features. The second linear layer maps the output of the hidden layer back to the original feature dimension. FFN is a way to independently transform the features at each position in the input sequence, without considering other positions in the sequence. It increases the expressive power of the model and allows the network to learn more complex feature representations.

[0080] like Fig.12 As shown in the figure, the HLD-GD module concatenates the P3 and P4 feature maps with the P5 layer features in the channel dimension after average pooling and maximum pooling to obtain the feature map f0. The concatenated feature map f0 is added to the corresponding position of the feature map of the feature map f0 after HLDAttention to obtain the feature map f1, and then the feature map f1 is added to the corresponding position of the feature map of the feature map f1 after the FFN module to obtain the feature map f2. Split is to split the feature map f2 into feature maps inject_N4 and inject_N5 according to the channel, which are used to fuse with the P4 and P5 layers to generate feature maps N4 and N5. The hybrid pooling structure used in this model helps to reduce the spatial dimension of the features, achieve translation invariance, and enhance the robustness of the network to spatial changes and translations of the input image. This module downsamples all feature maps to the smallest P5 feature map, obtains high-level semantic information, and reduces the amount of calculation in subsequent steps. The use of residual structure can effectively improve the efficiency of information use and enhance network stability.

[0081] In one embodiment of the present invention, the target detection head module in step S280 includes feature selection, decoder and detection head. This part is IoU-aware Query Selection, Decoder and Head. The function of IoU-aware Query Selection is to select a fixed number of features from the feature sequence output by the feature fusion part as object queries. The feature selection method based on DETR adds an IoU judgment selection method to generate high classification scores for features with high IoU and low classification scores for features with low IoU, so that the prediction box corresponding to the features selected by the model according to the classification score has both high classification and high IoU scores. The Decoder uses a multi-scale deformable attention mechanism to replace the original cross attention mechanism, so that the attention mechanism samples in a small range and only considers the most valuable points to reduce the amount of calculation and speed up the convergence speed. The detection head consists of a classification prediction network and a border prediction network, wherein the classification prediction network is composed of a fully connected network layer, and the border prediction network is composed of a multi-layer perceptron. The input of the detection module is the feature sequence after feature fusion, and a fixed number of feature sequences are extracted through feature selection, and then the final prediction result is output through the Decoder and Head.

[0082] It will be appreciated by those skilled in the art that various aspects of the present invention may be implemented as systems, methods or program products. Therefore, various aspects of the present invention may be specifically implemented in the following forms, namely: complete hardware embodiments, complete software embodiments (including firmware, microcode, etc.), or embodiments combining hardware and software aspects, which may be collectively referred to herein as "circuits", "modules" or "systems".

[0083] The electronic device is presented in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit, at least one storage unit, a bus connecting different system components (including the storage unit and the processing unit), and a display unit.

[0084] The storage unit stores program codes, which can be executed by the processing unit, so that the processing unit performs the steps according to various exemplary embodiments of the present invention described in the above “Exemplary Method” section of this specification. For example, the processing unit can perform the following steps: Figure 2 Steps S210 to S280 shown in FIG.

[0085] The storage unit may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) and / or a cache memory unit, and may further include a read-only memory unit (ROM).

[0086] The storage unit may also include a program / utility having a set (at least one) of program modules, such program modules including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0087] The bus may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0088] The electronic device can also communicate with one or more external devices (such as keyboards, pointing devices, Bluetooth devices, etc.), and can also communicate with one or more devices that enable users to interact with the electronic device, and / or communicate with any device (such as routers, modems, etc.) that enables the electronic device to communicate with one or more other computing devices. This communication can be carried out through an input / output (I / O) interface. In addition, the electronic device can also communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs) and / or public networks, such as the Internet) through a network adapter. As shown in the figure, the network adapter communicates with other modules of the electronic device through a bus. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0089] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiment of the present disclosure.

[0090] In an exemplary embodiment of the present disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the above method of the present specification is stored. In some possible embodiments, various aspects of the present invention can also be implemented in the form of a program product, which includes a program code, and when the program product is run on a terminal device, the program code is used to enable the terminal device to perform the steps according to various exemplary embodiments of the present invention described in the above "Exemplary Method" section of the present specification.

[0091] The program product may be a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto, and in this document, a readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, apparatus, or device.

[0092] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0093] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0094] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the foregoing.

[0095] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0096] In addition, the above-mentioned figures are only schematic illustrations of the processes included in the method according to an exemplary embodiment of the present invention, and are not intended to be limiting. It is easy to understand that the processes shown in the above-mentioned figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0097] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure. The specification and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0098] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A UAV target detection method based on improved RTDETR, characterized in that: The steps include: The LSKNet neural network architecture is used to extract four scale feature maps B2, B3, B4, and B5 of the drone-photographed images; The feature maps B2, B3, B4, and B5 of the four scales are made consistent in size by using average pooling, maximum pooling, and bilinear interpolation. After being merged in the channel dimension, they are convolved by the DSConv convolution module to obtain the global feature map. After convolution processing, the global feature map is split using the Split function and the feature maps inject_P3 and inject_P4 are output; The feature map B3 and the feature map inject_P3 are fused to generate the feature map P3, and the feature map B4 and the feature map inject_P4 are fused to generate the feature map P4; The feature map B5 is processed by HLD Encoder to generate the feature map P5; After the feature maps P3 and P4 are processed by average pooling and maximum pooling, they are concatenated with the feature map P5 in the channel dimension, downsampled by the HLD-GD module, and then split using the Split function to output the feature maps inject_N4 and inject_N5; The feature map P4 and the feature map inject_N4 are fused to generate the feature map N4, and the feature map P5 and the feature map inject_N5 are fused to generate the feature map N5; The feature maps P3, N4 and N5 are fused and merged and then output to the target detection head module for target detection.

2. The drone target detection method according to claim 1, characterized in that: The feature maps B2, B3, B4, and B5 of the four scales are combined and then convolved through the DSConv convolution module, including: The merged feature map is processed by the CBS module and the processed features are retained; The features processed by the CBS module are then processed by the DBS module; After the features processed by the CBS module and the features processed by the DBS module are spliced ​​in the channel dimension, the channel features of the two parts are randomly mixed by the Shuffle module and then the global feature map is output.

3. The drone target detection method according to claim 1, characterized in that: The feature map B5 is processed by HLDEncoder including: Input feature map B5 into the first branch's multi-head deformable attention and then concatenate the maps; The feature map B5 is average-pooled to use the downsampled feature map as the key and value of the multi-head deformable attention input to the second branch; The downsampled feature map is input into the multi-head deformable attention post-concatenation map of the second branch; Concatenate the output results of the multi-head deformable attention of the first branch and the multi-head deformable attention of the second branch in the channel dimension; Add the original feature map and the corresponding position of the feature map after HLD Attention and normalize them; The normalized feature map is added to the corresponding position of the feature map after the feed-forward neural network (FFN) and normalized to obtain the feature map P5.

4. The drone target detection method according to claim 1, characterized in that: After the feature maps P3 and P4 are processed by average pooling and maximum pooling, they are concatenated with the feature map P5 in the channel dimension and then subjected to feature extraction processing by the HLD-GD module, including: After the feature maps P3 and P4 are processed by average pooling and maximum pooling respectively, a downsampled feature map aligned with the size of the feature map P5 is output; The downsampled feature maps P3 and P4 are concatenated with the feature map P5 in the channel dimension; The concatenated feature map is input into the HLD Attention module and then concatenated with the map and the feed-forward neural network FFN for feature extraction.

5. An electronic device, characterized in that: include Processor; and A memory having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by the processor, implement the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the method according to any one of claims 1 to 4 when executed by a processor.

Images