An unmanned aerial vehicle aerial small target detection method, detection system and computer equipment based on improved RT-DETR

Abstract

An unmanned aerial vehicle aerial small target detection method and system based on an improved RT-DETR and a computer device, wherein a fine-grained feature enhancement convolution module is introduced into the backbone network of the detection model of the present application, and the fine-grained feature enhancement convolution module realizes differentiated enhancement of features through a multi-branch design: the first branch retains the original features to ensure that deep semantic information is not lost, and the second branch extracts edge features and texture features through an edge-texture parallel extraction module; this design branches the features and cross-fuses them at key levels, which not only maintains shallow feature information but also improves deep semantic expression, reduces repeated calculations and accelerates gradient flow, and is conducive to improving detection accuracy while maintaining lightweight.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) target detection technology, specifically to a method, system, and computer equipment for detecting small targets in UAV aerial photography based on an improved RT-DETR. Background Technology

[0002] In recent years, unmanned aerial vehicles (UAVs) have been widely used in numerous fields such as military reconnaissance, intelligent transportation, and agriculture due to their small size, low cost, ease of operation, and high maneuverability. With the continuous development of deep learning technology in UAV aerial image detection, deep learning-based image recognition algorithms are facing a series of challenges, including small target size, large number of targets, complex backgrounds, and high real-time requirements. Furthermore, in complex remote sensing contexts, images captured by UAVs are often obscured by dense buildings, vehicles, trees, and other objects, and changes in flight altitude and shooting angle further affect image clarity and accuracy. These problems significantly increase the difficulty of feature extraction and limit the applicability of existing target detection technologies.

[0003] Object detection technology has evolved from traditional methods to deep learning-based approaches, with mainstream algorithms currently categorized into two-stage and one-stage methods. Two-stage algorithms, exemplified by R-CNN, require generating candidate object regions in the first stage and then classifying and precisely locating them in the second. However, this two-stage approach consumes more computational resources and time, resulting in slower inference speeds. In contrast, single-stage algorithms such as YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector) do not require candidate region generation and can directly predict object categories and bounding box coordinates. This design typically allows for faster inference speeds, making them more suitable for scenarios with high speed requirements. The YOLO series of algorithms, in particular, has been widely used in various industrial scenarios due to its good balance between detection speed and accuracy. Since YOLOv3 and YOLOv4, the series has undergone multiple rounds of architectural iterations and optimizations, with current mainstream versions including YOLOv5, YOLOv8, and YOLOv11. With continuous improvements in network architecture design, loss function, and training strategy, the YOLO series has continuously improved its efficiency and accuracy in target detection tasks, providing an optional basic model for real-time detection of small targets on UAV platforms.

[0004] DETR (Detection Transformer) introduces the Transformer framework into end-to-end baseline networks for object detection, offering the advantage of global context modeling without relying on anchor generation or non-maximum suppression. However, DETR suffers from slow convergence and high computational cost during training. To address these challenges, Wang et al. proposed Deformable DETR, which alleviates the convergence limitations of traditional DETR through a deformable attention mechanism. However, due to its multi-scale feature interactions and deformable attention modules, it suffers from a large parameter size (42.86M) and high computational cost. Li et al. proposed DN-DETR, which employs a denoising training strategy to accelerate convergence through noise box reconstruction, achieving state-of-the-art performance within 50 training epochs. However, its uniform noise design is difficult to fully simulate dynamic interference in real-world scenarios, and it lacks explicit processing of high-frequency detail features, leading to an increased missing rate of small objects in complex backgrounds. At the same time, this method has high resource requirements. DINO proposed by Li et al. achieves a COCO detection performance of 50.7 AP by using a dynamic anchoring mechanism and contrastive denoising, but the complex multi-head self-attention architecture and dynamic anchoring design cause excessively high computational costs for resource-constrained UAV platforms. RT-DETR launched by Baidu, as a real-time end-to-end detection model, eliminates confidence thresholding and non-maximum suppression, and has advantages in speed and accuracy compared with the YOLO model of the same scale.

[0005] However, RT-DETR still has significant shortcomings in small target detection on UAVs: it typically uses classic convolutional networks such as ResNet as its backbone, extracting features through multiple downsampling steps. However, small targets in UAV aerial images have extremely low resolution, and multiple downsampling steps lead to the loss of a large amount of crucial details. While high-level semantic features contain global information, they neglect details such as low-level edges and local textures that are crucial for small target localization. Furthermore, RT-DETR's Transformer encoder often only uses features from the last layer of the backbone network, further weakening the utilization of low-level fine-grained information, resulting in insufficient representation of small target features. Due to the lack of explicit attention to shallow details, RT-DETR easily incorporates background information into its feature representation, which is particularly noticeable in UAV images with complex backgrounds. In summary, the RT-DETR backbone network suffers from insufficient feature granularity, loss of spatial information, and significant impact from background noise in small target detection tasks, leading to decreased detection accuracy. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method, system, and computer device for detecting small targets in UAV aerial photography based on an improved RT-DETR. This method retains shallow feature information while enhancing deep semantic representation, reducing redundant calculations and accelerating gradient flow, which helps the network improve detection accuracy while maintaining a lightweight design.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a method for detecting small targets in UAV aerial photography based on improved RT-DETR, comprising the following steps: Acquire the original image and preprocess it. The preprocessed original image is input into a pre-trained detection model to obtain small object detection results. The detection model is improved and trained based on RT-DETR. A fine-grained feature enhancement convolution module is introduced into the backbone network of the detection model. The fine-grained feature enhancement convolution module includes a multi-branch stacking module and an edge-texture parallel extraction module. The edge-texture parallel extraction module includes an attention fusion submodule, a residual multilayer perception submodule, a multi-receptive field fusion submodule, and multiple feature extraction submodules. The detection model uses the following methods to detect the preprocessed original image: The preprocessed original image is processed by feature splitting using a multi-branch overlay module to obtain the first branch and the second branch. The second branch is then divided into the first component Y1 and the second component Y2 by spilt after a 1×1 convolution operation. The first feature extraction submodule uses the Scharr operator to perform edge filtering on the second component Y2 to obtain edge features, and the remaining feature extraction submodules sequentially use Gaussian convolution to perform smoothing filtering on the second component Y2 to obtain texture features. The attention fusion submodule is used to add edge features or texture features to the second component Y2 to obtain temporary features att. The temporary features att are then processed and residually connected with the second component Y2 to output the connection features. The connection features are processed using a residual multilayer sensing submodule to obtain fused and enhanced features; The receptive fields of the fusion enhancement features are extracted using the multi-receptive field fusion submodule, and after splicing and fusion, they are residually connected with the first component Y1 to obtain the fusion features. The first branch and the fused features are fused and classified, and the small target detection results are output.

[0008] As a further optimization of the above technical solution: the formula for dividing the second branch into the first component Y1 and the second component Y2 using split after a 1×1 convolution operation is as follows: Y1,Y2=split(Conv 1×1 (X),(C h C h )); in, C1 is the number of input channels, C2 is the number of output channels, and e is the expansion factor.

[0009] As a further optimization of the above technical solution: the formula for obtaining edge features by performing edge filtering on the second component Y2 using the Scharr operator through the first feature extraction submodule is as follows: Among them, F in Y2 is the second component of the input.

[0010] As a further optimization of the above technical solution: the remaining feature extraction submodules obtain texture features by smoothing the second component Y2 through Gaussian convolution as follows: The second component Y2 is smoothed by a 5×5 depth-separable Gaussian convolution to obtain the upper and lower texture information G(I); The upper and lower texture information G(I) is normalized by BatchNorm2d and then activated by SiLU to obtain texture features; The formula for obtaining the upper and lower texture information G(I) is: Where i,j∈{-2,-1,0,1,2},

[0011] As a further optimization of the above technical solution: the method of adding edge features or texture features to the second component Y2 using the attention fusion submodule to obtain temporary features att, processing the temporary features att and performing a residual connection with the second component Y2 to output the connection features is as follows: The attention fusion submodule is used to element-wise add edge features or texture features to the second component Y2 in the channel dimension; Channel attention weights w are generated through a channel attention mechanism, and temporary features att are adaptively weighted. The weighted temporary feature att is coupled with the second component Y2 through a residual connection to output the connection feature.

[0012] As a further optimization of the above technical solution: the specific method for adaptively weighting the temporary feature att using the channel attention mechanism to generate channel attention weights w is as follows: Channel attention weights are generated by 1×1 channel mapping, global average pooling, one-dimensional convolution and sigmoid activation function, and the temporary feature att is adaptively weighted. The formulas for 1×1 channel mapping and global average pooling are as follows: M = Conv 1×1 (att), Z = GAP(M); Where GAP stands for Global Average Pooling, Conv 1×1 It is a 1×1 convolution; The formula for generating channel attention weights using one-dimensional convolution and the sigmoid activation function is as follows: w = σ(Conv 1d(z)); Where, w∈R C×1 σ is the Sigmoid activation function, and Conv1d is a one-dimensional convolution.

[0013] The residual multilayer perceptron submodule includes two 1×1 convolutional layers and a DropPath mechanism, which is integrated into the residual connection path of the residual multilayer perceptron submodule.

[0014] The method for extracting the receptive fields of the fused enhancement features using the multi-receptive-field fusion submodule, and then concatenating and fusing them with the first component Y1 to obtain the enhanced features is as follows: The multi-receptive field fusion submodule extracts multi-scale receptive field features from at least three sub-branches with different convolutional kernel sizes; the receptive field features are concatenated in the channel dimension, fused by 1×1 convolution, and then residually connected with the first component Y1 to obtain the enhanced features.

[0015] A small target detection system for UAV aerial photography, comprising an improved RT-DETR-based method for small target detection in UAV aerial photography to achieve any of the above-mentioned methods. The acquisition module is used to acquire the raw image and preprocess the raw image; The model running module is used to run a pre-trained detection model to detect the pre-processed original image and generate small object detection results.

[0016] Computer equipment, including: Memory, used to store computer programs; A processor for reading and executing computer programs to achieve any of the above-mentioned small target detection methods for UAV aerial photography based on improved RT-DETR.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0018] 1) The backbone network of the detection model of this invention introduces a fine-grained feature enhancement convolution module. The fine-grained feature enhancement convolution module achieves differentiated feature enhancement through a multi-branch design: the first branch retains the original features to ensure that deep semantic information is not lost, and the second branch extracts edge features and texture features respectively through the edge-texture parallel extraction module. This design splits the features and cross-fused them at key levels, which not only maintains shallow feature information but also enhances deep semantic expression, reduces redundant calculations and accelerates gradient flow, which helps the network improve detection accuracy while maintaining lightweight design.

[0019] 2) This invention uses the Scharr operator to perform edge filtering on the second component Y2 to obtain edge features, effectively capturing the contours and texture edges of small targets; the remaining feature extraction sub-modules sequentially use Gaussian convolution to perform smooth filtering on the second component Y2 to obtain texture features, supplementing local context information and making up for the limitations of a single filter in extracting features of small targets in complex backgrounds.

[0020] 3) By generating channel attention weights through 1×1 channel mapping, global average pooling, one-dimensional convolution and sigmoid activation function, temporary features are adaptively weighted to dynamically highlight feature channels related to small targets, suppress background noise and improve the detector's ability to distinguish small targets. Attached Figure Description

[0021] Figure 1 This is a diagram of the RT-DETR network structure; Figure 2 This is a diagram of the CEG module structure; Figure 3 This is a diagram of the CEG Block module structure; Figure 4 This is the EG module structure diagram. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.

[0023] A method for detecting small targets in UAV aerial photography based on an improved RT-DETR, comprising S1 and S2.

[0024] S1. Acquire the original images and preprocess them; the original images used in this invention are from the VisDrone2019 dataset compiled by the AISKYEYE team at Tianjin University. This dataset covers outdoor aerial scenes taken in 14 cities in China under diverse environmental and weather conditions, including 400 video clips, 265,200 frames of images, and 8,629 still images, with approximately 2.6 million bounding boxes labeled for 10 types of targets such as pedestrians, vehicles, and bicycles.

[0025] To ensure that the evaluation results only reflect the performance improvements of the channel enhancement and fine-grained feature enhancement convolutional modules, this study performed targeted data processing: the 8,629 static images provided by the official team were divided into a 7:2:1 ratio into a training set of 6,040 images, a test set of 1,726 images, and a validation set of 863 images; at the same time, only size normalization (adjusted to 640×640) and pixel normalization operations were retained, and other data augmentation methods such as random cropping, flipping, rotation, color jittering, and noise injection were deliberately omitted.

[0026] This setup offers three advantages: First, it avoids data augmentation strategies interfering with detection performance, thus accurately attributing performance improvements to the fine-grained feature-enhanced convolutional module; second, it simplifies hyperparameter selection and experimental procedures, improving the reproducibility of the research; and third, it provides a fair benchmark for comparison with existing methods under default data conditions. Based on this baseline, the study further conducted comprehensive ablation experiments and control experiments to fully evaluate the effectiveness of the proposed improved algorithm in UAV aerial imagery for small target detection.

[0027] S2. The preprocessed original image is input into the pre-trained detection model to obtain the small target detection result. The detection model is improved and trained based on RT-DETR. A fine-grained feature enhancement convolution module is introduced into the backbone network of the detection model. The fine-grained feature enhancement convolution module is represented by CEG. The fine-grained feature enhancement convolution module includes a multi-branch stacking module and an edge-texture parallel extraction module. The edge-texture parallel extraction module is represented by EG. The edge-texture parallel extraction module includes an attention fusion submodule, a residual multilayer perception submodule, a multi-receptive field fusion submodule, and multiple feature extraction submodules.

[0028] The detection models use methods to perform detection on preprocessed original images, including:

[0029] S21. The preprocessed original image is processed by feature segmentation using a multi-branch overlay module to obtain a first branch and a second branch. The first branch retains the original information for deep propagation, while the second branch is used for efficient convolution. After a 1×1 convolution operation, the second branch is segmented into a first component Y1 and a second component Y2 using Split. The formula for segmenting the second branch into the first component Y1 and the second component Y2 using Split after a 1×1 convolution operation is as follows: Y1,Y2=split(Conv 1×1 (X),(C h C h )); in, C1 is the number of input channels, C2 is the number of output channels, and e is the expansion factor, which controls the proportion of hidden channels.

[0030] S22. Using the first feature extraction submodule, the second component Y2 is edge-filtered using the Scharr operator to obtain edge features, thereby accurately capturing fine-grained edge information of the target; the formula for obtaining the edge features is: Among them, F in Y2 is the second component of the input.

[0031] The Scharr operator performs Scharr convolution to calculate the horizontal and vertical gradient responses, respectively. The sum of squares and the square root are then used to obtain the edge intensity, and the edge features are obtained through a two-dimensional batch normalization layer and the SiLU activation function.

[0032] The remaining feature extraction sub-modules sequentially apply Gaussian convolution to smooth the second component Y2 to obtain texture features; the remaining feature extraction modules uniformly use Gaussian convolution to stabilize local texture and contextual information; the method for obtaining texture features is as follows: S221. A 5×5 depth-separable Gaussian convolution is used to smooth the second component Y2 to obtain the upper and lower texture information G(I). S222. The upper and lower texture information G(I) is normalized using BatchNorm2d and then activated by the SiLU function to obtain texture features; wherein, the formula for obtaining the upper and lower texture information G(I) is: Where i,j∈{-2,-1,0,1,2},

[0033] S23. Using the attention fusion submodule, the edge features or texture features are added to the second component Y2 in the channel dimension to obtain a temporary feature att. The temporary feature att is then processed and residually connected with the second component Y2 to output the connection feature. The method is as follows: S231. Use the attention fusion submodule to element-wise add the edge features or texture features to the second component Y2 in the channel dimension; S232. Generate channel attention weights w through a channel attention mechanism, and adaptively weight the temporary feature att. The method is as follows: S2321. Channel attention weights are generated through 1×1 channel mapping, global average pooling, one-dimensional convolution and sigmoid activation function, and the temporary feature att is adaptively weighted. S233. The weighted temporary feature att is coupled with the second component Y2 through residual connection, and the connection feature is output.

[0034] The formulas for 1×1 channel mapping and global average pooling are as follows: M = Conv 1×1 (att), Z = GAP(M); Where GAP represents global average pooling, Conv 1×1 Represents a 1×1 convolution; The formula for generating channel attention weights using one-dimensional convolution and the sigmoid activation function is as follows: w = σ(Conv 1d(z)); Where, w∈RC×1 σ is the Sigmoid activation function; The formula for adaptively weighting the temporary feature att is: att′=att⊙w; Where att′ is the intermediate feature, and ⊙ represents element-wise multiplication; The formula for outputting the connection feature after residual connection between the intermediate feature and the second component Y2 is: out = x + att'; Where out represents the connection feature.

[0035] The formula for obtaining the temporary feature att by adding the edge features or texture features to the second component Y2 in the channel dimension using the attention fusion submodule is as follows: Y ori This is the second component, Y2.

[0036] S24. A residual multilayer perceptron submodule is used to process connection features to obtain fused and enhanced features. The residual multilayer perceptron submodule includes two 1×1 convolutional layers and a DropPath mechanism. The first 1×1 convolutional layer compresses the number of channels in the input connection features, and the second 1×1 convolutional layer restores the number of channels. The DropPath mechanism is integrated into the residual connection paths of the residual multilayer perceptron submodule. During the detection model training phase, some feature propagation paths are randomly selected and discarded according to a preset probability, and the feature values ​​of the discarded paths are set to zero. During the detection model inference phase, this mechanism does not perform the discarding operation, and all feature paths remain propagated intact. Through the above random discarding operation, overfitting during the feature enhancement process is suppressed, and the network's ability to generalize and extract features from small targets is enhanced.

[0037] S25. The receptive fields of the fused enhancement features are extracted using the multi-receptive-field fusion submodule, and after splicing and fusion, residual connections are performed with the first component Y1 to obtain the enhancement features. The method is as follows: The multi-receptive-field fusion submodule extracts multi-scale receptive-field features from at least three sub-branches with different convolutional kernel sizes; the receptive-field features are concatenated along the channel dimension, fused by a 1×1 convolution, and then residually connected to the first component Y1 to obtain the fused features, as shown in the following formula: Z1 = EG 3×3 (Y2),Z i =EG 3×3 (Z i-1 ); Z final =Conv 1×1 (Concat(Y1,Z1,Z2,…,Z n )).

[0038] S26. Fuse and classify the first branch and the fusion features, and output the small target detection results.

[0039] The detection model of this invention introduces a fine-grained feature enhancement convolutional module into its backbone network. This module achieves differentiated feature enhancement through a multi-branch design: the first branch retains the original features to ensure that deep semantic information is not lost, while the second branch extracts edge features and texture features respectively through an edge-texture parallel extraction module. This design splits the features and cross-fused them at key levels, which not only preserves shallow feature information but also enhances deep semantic expression, reduces redundant computation, and accelerates gradient flow, thus improving detection accuracy while maintaining a lightweight architecture.

[0040] Ablation experiments were conducted using this invention on the VisDrone2019 dataset, and the experimental data are shown in Table 1.

[0041] Table 1 Ablation Experiment RT-DETR 19.9M 57.3G 54.3％ 38.2％ 37.1％ 21.2％ RT-DETR+CEG 14.9M 53.0G 56.2％ 39.4％ 38.7％ 22.9％

[0042] Experiments were conducted on the VisDrone2019 dataset using the present invention and other algorithms. The experimental data comparison is shown in Table 2.

[0043] Table 2 Comparison of Experimental Data StarNet 20.1M 58.5G 53.0％ 37.2％ 35.9％ 20.9％ FasterNet 16.8M 49.8G 55.8％ 38.9％ 38.1％ 22.4％ HVI-CIDNET 13.8M 46.9G 55.0％ 38.5％ 37.6％ 21.5％ YOLOv5m 25.1M 64.2G 46.8％ 35.8％ 34.3％ 20.3％ YOLOv11m 25.3M 86.6G 48.3％ 37.0％ 35.8％ 20.3％ Our 14.9M 53.0G 56.2％ 39.4％ 38.7％ 22.9％

[0044] On the VisDrone2019 dataset, all evaluation metrics were improved compared to RT-DETR. Accuracy improved by 1.98%, recall by 1.2%, mAP50 by 1.64%, and mAP50:95 by 1.74%, fully validating the effectiveness and superiority of this invention in UAV aerial photography for small target detection.

[0045] A small target detection system for UAV aerial photography, comprising an improved RT-DETR-based method for small target detection in UAV aerial photography to achieve any of the above-mentioned methods. The acquisition module is used to acquire the raw image and preprocess the raw image; The model running module is used to run a pre-trained detection model to detect the pre-processed original image and generate small object detection results.

[0046] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the embodiments provided by this invention, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or modules may be electrical, mechanical, or other forms.

[0047] Computer equipment, including: Memory, used to store computer programs;

[0048] A processor for reading and executing computer programs to achieve any of the above-mentioned small target detection methods for UAV aerial photography based on improved RT-DETR.

[0049] The memory, serving as a carrier for resource storage, can be a read-only memory, random access memory, disk, or optical disk, etc. The resources stored on it can include the operating system, computer programs, etc., and the storage method can be temporary or permanent storage. The operating system is used to manage and control the various hardware devices and computer programs on the electronic device; it can be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of performing the adaptive emotion regulation method based on personalized reconfigurable music disclosed in the foregoing embodiments, the computer programs may further include computer programs capable of performing other specific tasks. The processor can be a general-purpose processor product based on architectures such as x86, IA64, RISC, MIPS, and ARM.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting small targets in UAV aerial photography based on an improved RT-DETR, characterized in that, Includes the following steps: Acquire the original image and preprocess it. The preprocessed original image is input into a pre-trained detection model to obtain small object detection results. The detection model is improved and trained based on RT-DETR. A fine-grained feature enhancement convolution module is introduced into the backbone network of the detection model. The fine-grained feature enhancement convolution module includes a multi-branch stacking module and an edge-texture parallel extraction module. The edge-texture parallel extraction module includes an attention fusion submodule, a residual multilayer perception submodule, a multi-receptive field fusion submodule, and multiple feature extraction submodules. The detection model uses the following methods to detect the preprocessed original image: The preprocessed original image is processed by feature splitting using a multi-branch overlay module to obtain the first branch and the second branch. The second branch is then divided into the first component Y1 and the second component Y2 by spilt after a 1×1 convolution operation. The first feature extraction submodule uses the Scharr operator to perform edge filtering on the second component Y2 to obtain edge features, and the remaining feature extraction submodules sequentially use Gaussian convolution to perform smoothing filtering on the second component Y2 to obtain texture features. The attention fusion submodule is used to add edge features or texture features to the second component Y2 to obtain temporary features att. The temporary features att are then processed and residually connected with the second component Y2 to output the connection features. The connection features are processed using a residual multilayer sensing submodule to obtain fused and enhanced features; The receptive fields of the fusion enhancement features are extracted using the multi-receptive field fusion submodule, and after splicing and fusion, they are residually connected with the first component Y1 to obtain the fusion features. The first branch and the fused features are fused and classified, and the small target detection results are output.

2. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 1, characterized in that, The second branch, after undergoing a 1×1 convolution operation, is split into a first component Y1 and a second component Y2 using the following formula: Y1,Y2=split(Conv 1×1 (X),(C h C h )); in, C1 is the number of input channels, C2 is the number of output channels, and e is the expansion factor.

3. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 1, characterized in that, The formula for obtaining edge features by performing edge filtering on the second component Y2 using the Scharr operator through the first feature extraction submodule is as follows: Among them, F in Y2 is the second component of the input.

4. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 1, characterized in that, The remaining feature extraction submodules obtain texture features by smoothing the second component Y2 through Gaussian convolution as follows: a 5×5 depth-separable Gaussian convolution is used to smooth the second component Y2 to obtain the upper and lower texture information G(I). The upper and lower texture information G(I) is normalized by BatchNorm2d and then activated by SiLU to obtain texture features; The formula for obtaining the upper and lower texture information G(I) is: Where i,j∈{-2,-1,0,1,2}, 5. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 1, characterized in that, The method for adding edge features or texture features to the second component Y2 using the attention fusion submodule to obtain temporary features att, and then processing the temporary features att and performing a residual connection with the second component Y2 to output the connection features is as follows: The attention fusion submodule is used to element-wise add edge features or texture features to the second component Y2 in the channel dimension; Channel attention weights w are generated through a channel attention mechanism, and temporary features att are adaptively weighted. The weighted temporary feature att is coupled with the second component Y2 through a residual connection to output the connection feature.

6. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 5, characterized in that, The specific method for adaptively weighting the temporary feature att using the channel attention mechanism to generate channel attention weights w is as follows: Channel attention weights are generated by 1×1 channel mapping, global average pooling, one-dimensional convolution and sigmoid activation function, and the temporary feature att is adaptively weighted. The formulas for 1×1 channel mapping and global average pooling are as follows: M=Conv 1×1 (that), Z=GAP(M); Where GAP stands for Global Average Pooling, Conv 1×1 It is a 1×1 convolution; The formula for generating channel attention weights using one-dimensional convolution and the sigmoid activation function is as follows: w = σ(Conv 1d(z)); Where, w∈R C×1 σ is the Sigmoid activation function, and Conv1d is a one-dimensional convolution.

7. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 1, characterized in that, The residual multilayer sensing submodule includes two layers of 1×1 convolution and a DropPath mechanism, which is integrated into the residual connection path of the residual multilayer sensing submodule.

8. The method for detecting small targets in UAV aerial photography based on improved RT-DETR as described in claim 1, characterized in that, The method for extracting the receptive fields of the fused enhancement features using the multi-receptive-field fusion submodule, and then concatenating and fusing them with the first component Y1 to obtain the enhanced features is as follows: The multi-receptive field fusion submodule extracts multi-scale receptive field features from at least three sub-branches with different convolutional kernel sizes; the receptive field features are concatenated in the channel dimension, fused by 1×1 convolution, and then residually connected with the first component Y1 to obtain the enhanced features.

9. A small target detection system for UAV aerial photography, used to implement the UAV aerial target detection method based on the improved RT-DETR as described in any one of claims 1-8, characterized in that, The system includes: The acquisition module is used to acquire the raw image and preprocess the raw image; The model running module is used to run a pre-trained detection model to detect the pre-processed original image and generate small object detection results.

10. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for reading and executing computer programs to implement the UAV aerial small target detection method based on the improved RT-DETR as described in any one of claims 1-8.

Images