An improved unmanned aerial vehicle aerial target detection method based on RT-DETR

Abstract

The application discloses an improved lightweight end-to-end unmanned aerial vehicle aerial image small target detection method based on RT-DETR, and steps are as follows: firstly, aerial images are acquired through image acquisition equipment carried by the unmanned aerial vehicle. Then, the aerial images are input into the improved RT-DETR target detection network for detection. The improvement of the model comprises a cross-level multi-path feature aggregation (CHMFA) module, a partial convolution and compression excitation (PCSE) module and a space-to-channel feature reuse (SCFR) module. Finally, a layer adaptive magnitude pruning (LAMP) strategy is adopted to prune redundant weights of the improved model, so that the model calculation overhead is further reduced. The application can reduce the model parameter quantity and the calculation complexity while maintaining the detection accuracy of the original model, and has high practical application value.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and UAV image intelligent perception technology, specifically involving a lightweight UAV aerial image small target detection method based on improved RT-DETR. Background Technology

[0002] With the rapid development of UAV technology, target detection methods based on UAV aerial photography have been widely applied in scenarios such as traffic monitoring, urban management, disaster relief, and military reconnaissance. UAV aerial images are characterized by high viewing angle, wide field of view, and complex backgrounds. Small targets (such as pedestrians, vehicles, and small objects) have low pixel ratios and limited texture information, posing a significant challenge to target detection.

[0003] In recent years, end-to-end object detection methods based on the Transformer architecture have demonstrated excellent detection performance in complex scenes due to their global context modeling capabilities. Among them, RT-DETR, as an efficient real-time detection Transformer framework, achieves rapid localization and recognition of multi-scale targets by decoupling the encoder-decoder structure.

[0004] However, existing RT-DETR methods still have the following shortcomings in practical applications:

[0005] High computational overhead: Due to its reliance on global attention mechanisms, the model has a high number of parameters and computational complexity, making it difficult to directly adapt to the edge deployment needs of drone platforms with limited computing power.

[0006] Limited small target detection capability: The downsampling operation in the RT-DETR backbone network is prone to losing detailed information of small targets, and the feature fusion process does not make sufficient use of shallow spatial features, resulting in a high false negative rate for small targets;

[0007] Multi-scale feature fusion is insufficient: Existing feature fusion methods need to improve their ability to integrate contextual information and local details, and it is difficult to balance receptive field expansion and detail preservation.

[0008] Therefore, how to reduce model complexity and improve small target detection performance while ensuring detection accuracy, so as to achieve efficient and accurate small target detection in UAV aerial images, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] The technical problem to be solved by this invention is to address the problems of high computational cost, high false negative rate of small targets, and insufficient multi-scale feature fusion in the existing RT-DETR method for small target detection in UAV aerial photography. This invention provides a lightweight UAV aerial image small target detection method based on RT-DETR improvement, which reduces the number of model parameters and computational cost while ensuring detection accuracy, thereby improving the small target detection performance and providing a more flexible and stable small target detection method for UAV aerial photography tasks.

[0010] This invention achieves the above objectives through the following technical solution: an improved RT-DETR target detection network is constructed based on the original RT-DETR network architecture. The improvements include the following three aspects:

[0011] i. Cross-level multi-path feature aggregation module

[0012] Because small target pixels account for a relatively low proportion in drone aerial images, some key pixels are easily discarded after max pooling, thus affecting the integrity of semantic features. To retain more important information and increase the receptive field of the network, a Cross Hierarchical Multi-path Feature Aggregation (CHMFA) module is proposed, with the structure as follows: Figure 2 As shown.

[0013] The CHMFA module first processes the input features X∈R. C×H×W The channels are reduced in dimensionality using 1×1 convolutional layers to prevent parameter inflation caused by the aggregation of multiple features. Then, the reduced-dimensional features X1∈R are... C / 2×H×W Input branch 1 (pooling and convolution branch) and branch 2 (convolution branch) respectively.

[0014] Branch 1 contains two sub-branches: Branch 1-1 (max pooling, convolution branch) and Branch 1-2 (average pooling, convolution branch). Each sub-branch first performs feature filtering (max pooling) and smoothing (average pooling) on ​​X1 through pooling operations, and then enters a 1×1 convolutional layer. The pooled feature X... 11 ∈R C / 2×H×W and X 21 ∈R C / 2×H×W Feature X participates in feature stacking via a skip connection method; the convolutional feature X... 12 ∈R C / 2×H×W and X 22 ∈R C / 2×H×W The feature stacking is performed via direct connection. This design not only expands the receptive field of the network but also enhances the nonlinear expressive power of pooling features. The calculation process is shown in equations (1)-(4):

[0015] (1)

[0016] (2)

[0017] (3)

[0018] (4)

[0019] In the formula, CBR() represents 2D convolution operation + batch normalization + ReLU activation function, MaxPool() and AvgPool() represent max pooling and average pooling operations respectively, and Concat[,] represents stacking operation.

[0020] Branch 2 employs a 3×3 convolutional layer for refined feature extraction and directly participates in feature stacking, aiming to preserve richer local detail information and ensure the model's high sensitivity to subtle features. Finally, the aggregated multiple features are used with a 1×1 convolutional layer to recover the number of input channels, obtaining the final feature X. out ∈R C×H×W The calculation process is shown in equations (5) and (6):

[0021] (5)

[0022] (6)

[0023] By utilizing the aforementioned idea of ​​cross-level multi-path feature aggregation, the model can make fuller use of the image's contextual information and local subtle features, thereby providing a more comprehensive feature representation for the detection task.

[0024] ii. Partial Convolution and Compression Excitation Module

[0025] The Basic Block module in the original RT-DETR backbone network utilizes consecutive ordinary convolutional layers to achieve hierarchical integration of features, thereby stabilizing the feature representation capability during deep network training. However, limited by the fixed kernel size and number of channels of ordinary convolution (Conv), this module incurs high computational overhead. To balance lightweight design with the representation capability of key features, this invention introduces partial convolution and squeezed excitation (SE) attention, constructing a Partial Convolution and Squeeze-Excitation (PCSE) module with a hybrid architecture of Conv, PConv, and SE attention, as shown in the figure. Figure 3 As shown.

[0026] Branch 1 incorporates 1×1 and 3×3 ordinary convolutional layers before some convolutional layers, where the 1×1 convolutional layer is used to process the input features X∈R. C×H×WThe channel dimension is controlled to ensure the consistency between network input and output. The 3×3 convolutional layer is responsible for extracting input features. The calculation process is shown in formula (7):

[0027] (7)

[0028] In the formula, CBR() represents 2D convolution operation + batch normalization + ReLU activation function, and PConv() represents 2D partial convolution operation. Partial convolution uses the first or last consecutive channels of the input as representatives of the feature map for calculation, dynamically adjusts the effective area of ​​the convolution kernel, selects a portion of channels for ordinary convolution operation, and keeps the remaining channels unchanged. This method can effectively reduce memory access. The formulas for calculating the memory access of partial convolution and ordinary convolution are as follows:

[0029] (8)

[0030] (9)

[0031] In the formula, H, W, and C are the height, width, and number of channels of the input feature map, k is the size of the convolution kernel, and cp is the number of channels used by the convolution kernel in the partial convolution. By adjusting c... p The / C ratio allows for a flexible trade-off between detection accuracy and computational overhead.

[0032] Branch 2 uses a 1×1 convolutional layer to output feature X1∈R from the main branch. C×H×W Residual fusion is performed, and the fused features can effectively alleviate the network degradation problem and enrich the feature information. The calculation process is shown in formula (10):

[0033] (10)

[0034] The fused feature X2∈R C×H×W This approach utilizes SE attention to adaptively calibrate the inter-channel dependencies, thereby enhancing the representation capability of key features. The scheme consists of three stages: squeezing, excitation, and scaling.

[0035] During the compression phase, a global average pooling operation is used to compress the spatial dimension (H×W) of each channel into a single scalar. This step aggregates global spatial information to generate a descriptor z∈R that contains the global statistical features of each channel. 1 ×1×C :

[0036] (11)

[0037] In the formula, z c For Fsq (X 2c The c-th element after z, X 2c (i,j) represents the element in the i-th row and j-th column of feature map X2.

[0038] The activation process then proceeds, generating channel weights through two fully connected layers and an activation function. First, a fully connected layer with a reduction ratio of r compresses the number of channels to C / r. After ReLU activation, another fully connected layer restores the original number of channels C. Finally, a sigmoid function outputs a weight vector s∈R to quantify the importance of each channel. 1×1×C :

[0039] (12)

[0040] In the formula, W1∈R C / r×C and W2∈R C×C / r These are the reduced-dimensional matrix and the increased-dimensional matrix, respectively. and These are the Sigmoid and ReLU functions, respectively.

[0041] Finally, a reweighting operation is performed, multiplying the generated weight vector s with the feature map X2 channel by channel to obtain the channel-recalibrated feature X. out ∈R C×H×W :

[0042] (13)

[0043] iii. Spatial-to-channel feature multiplexing

[0044] In UAV aerial image detection tasks, images contain numerous small targets. The original RT-DETR network, due to its excessive focus on high-level semantic information and neglect of shallow spatial features, is prone to false positives and false negatives. To address these issues, this paper maps the spatial information of the shallow spatial features (S2) extracted by the backbone network to channel information and fuses them with the small target detection layer through stacking, constructing a Spatial to Channel Feature Reuse (SCFR) module to compensate for the loss of small target information during extraction. The feature fusion idea and structure are as follows: Figure 4 As shown.

[0045] SCFR can be divided into two stages: non-parametric dimensionality transformation and parametric feature extraction. The non-parametric dimensionality transformation stage rearranges and combines spatially adjacent pixels through a sampling factor (Scale), mapping them into multiple sub-features to achieve feature downsampling. This process does not generate additional learnable parameters. The choice of Scale mainly considers the following aspects: In local regions, pixels have strong spatial similarity, and a low sampling factor better ensures that the rearranged sub-feature pixels maintain a continuous structural relationship. An excessively high Scale will map out more low-resolution sub-features, leading to attenuation of local texture and detail information, and the number of channels will become Scale × Scale, increasing the computational burden of subsequent feature fusion. Based on the above analysis, this invention sets Scale to 2 to maintain a balance between preserving spatial pixel information, ensuring reasonable sub-feature scale, and maximizing channel expansion.

[0046] When the sampling factor Scale is 2, the original input features S2∈R C×H×W Adjacent pixels are transformed and divided into four feature segments of the same size: S2(0,0), S2(1,0), S2(0,1), and S2(1,1), each with dimensions C×H / 2×W / 2. The transformation process is shown in formulas (14)-(18):

[0047] (14)

[0048] (15)

[0049] (16)

[0050] (17)

[0051] (18)

[0052] In the formula, [:] represents the segmentation operation, and Concat[,] represents the stacking operation. Then, all feature maps are stacked along the channel dimension to form a feature S with half the spatial dimension and four times the channel dimension. 21 ∈R 4C×H / 2×W / 2 .

[0053] The parameterized feature extraction stage uses a 1×1 convolutional layer for S 21 Channel reduction is performed to prevent parameter inflation caused by increased channels. The resulting features are then stacked with the output features of a 3×3 dilated convolutional layer using a skipping method to obtain feature S. 22 ∈R 4C×H / 2×W / 2 This design effectively controls computational complexity while expanding the receptive field and enhancing nonlinear expression. The computation process is shown in formula (19):

[0054] (19)

[0055] In the formula, CBR() represents a 2D convolution operation + batch normalization + ReLU activation function, and DCBR() represents a 2D dilated convolution operation with a dilation rate of 2 + batch normalization + ReLU activation function.

[0056] iv. Layer-adaptive Magnitude-based Pruning (LAMP) Strategy

[0057] LAMP is a weight-based pruning method. Its core idea is to dynamically calculate the importance score of each weight (LAMP Score) to determine the priority of weight pruning and maintain the original performance of the model as much as possible.

[0058] like Figure 5 As shown, the LAMP strategy process can be summarized as follows: First, select the layers of the model to be pruned, and assign weight matrices W∈R to each layer. C×S×S Flattened into a one-dimensional vector W∈R N Arranged in ascending order of absolute weight, we obtain an ordered sequence {W1, W2, ..., W...} N Next, the LAMP Score for each weight is calculated, and low-scoring weights are pruned. Then, the weight matrix is ​​updated, and it is determined whether the desired sparsity is met. Finally, the remaining weight parameters that meet the sparsity are fine-tuned so that these parameters can adapt to the new network.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention introduces a cross-level multi-path feature aggregation module to fuse features from different levels through multiple paths, effectively enhancing the continuity of multi-scale semantic information and expanding the network's receptive field. Simultaneously, the spatial-to-channel feature reuse module maps shallow spatial features to channel features and supplements them to the small target detection layer, effectively compensating for the loss of detailed information in small targets and significantly improving small target detection capabilities.

[0061] 2. This invention replaces the basic convolutional modules in the original RT-DETR backbone network with partial convolutional and compression excitation modules. By performing convolution operations on only some channels and directly passing the rest, memory access and computational overhead are significantly reduced. The compression excitation mechanism enhances the feature representation capability of key regions while maintaining a lightweight design, achieving a balance between accuracy and efficiency.

[0062] 3. This invention employs a layer-adaptive amplitude pruning strategy, dynamically determining the pruning ratio based on the contribution of each network layer to the detection performance. Redundant weights are pruned while retaining key channels, further reducing the number of model parameters and computational complexity, and improving inference speed.

[0063] 4. Experimental results show that on the VisDrone2019 dataset, the improved RT-DETR outperforms the original model in terms of mAP. 50 A 0.6% improvement was achieved, while parameter count and GFLOPs decreased by 39.2% and 27.5%, respectively, with FPS reaching 97.1. On the HIT-UAV infrared dataset, precision, recall, and mAP were also improved. 50 The improvements were 2.2%, 0.8%, and 0.6%, respectively. Visualization results show that the improved model significantly reduced both the false negative and false positive rates under challenging conditions such as dense scenes, false target interference, and infrared imaging, demonstrating good generalization performance.

[0064] 5. This invention can maintain detection accuracy while reducing model complexity, and is especially suitable for edge deployment scenarios where the computing power of UAV platforms is limited. It can be widely used in tasks such as traffic monitoring, urban management, disaster relief and military reconnaissance, and has high practical application value. Attached Figure Description

[0065] Figure 1 This is a diagram of the overall architecture of the improved RT-DETR network of this invention.

[0066] Figure 2 This refers to the CHMFA module in this invention.

[0067] Figure 3 This refers to the PCSE module in this invention.

[0068] Figure 4 This is a diagram illustrating the SCFR module feature fusion concept and its structure in this invention.

[0069] Figure 5 This is a flowchart of the LAMP strategy in this invention.

[0070] Figure 6 This is a schematic diagram of the research scheme of the present invention.

[0071] Figure 7 This is a visualization of the detection performance of the method of the present invention on the VisDrone2019 and HIT-UAV datasets compared to the original RT-DETR. Detailed Implementation

[0072] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0073] Example 1: Implementation of an improved UAV aerial target detection method based on RT-DETR

[0074] Step 1: Data Preparation

[0075] Aerial images are acquired using high-definition image acquisition equipment mounted on a drone. These images are then uniformly adjusted to the input size required by the model (e.g., 640×640 pixels) and normalized, scaling pixel values ​​to the [0,1] range. During the training phase, data augmentation strategies are employed, including random flipping, random cropping, and color dithering, to improve the model's generalization ability.

[0076] Step 2: Construct an improved RT-DETR target detection network

[0077] This invention improves upon the original RT-DETR network architecture, constructing an improved RT-DETR target detection network. The specific improvements are as follows:

[0078] (1) Introduce a cross-layer multi-path feature aggregation module into the backbone network

[0079] A cross-level multi-path feature aggregation module is introduced between different layers of the backbone network to fuse multi-scale features. For example... Figure 2 As shown, this module first performs 1×1 convolution to reduce the dimensionality of the input features, and then inputs the dimensionality-reduced features into branch 1 and branch 2 respectively. Branch 1 contains max pooling sub-branch and average pooling sub-branch, and each sub-branch performs pooling operation first and then convolution operation; branch 2 directly performs 3×3 convolution operation. Finally, the features output from each branch are stacked and fused along the channel dimension, and the number of channels is restored by 1×1 convolution.

[0080] (2) Introduce some convolution and compression excitation modules into the backbone network.

[0081] The basic convolutional modules in the original RT-DETR backbone network are replaced with some convolutional and compressed excitation modules. For example... Figure 3 As shown, this module first performs 1×1 convolution and 3×3 convolution on the input features sequentially, then performs partial convolution operations, only performing convolution operations on a portion of continuous channels, while the remaining channels are directly passed. The output of the partial convolution is residually fused with the output of a 1×1 convolution branch, and finally, the fused features are compressed and activated to generate channel weights, which are then weighted channel by channel to obtain the channel-calibrated features.

[0082] (3) Introducing a spatial-to-channel feature multiplexing module into the feature fusion network

[0083] A spatial-to-channel feature reuse module is introduced into the feature fusion network to supplement the shallow spatial features extracted by the backbone network into the small target detection layer. For example... Figure 4As shown, this module first performs a spatial-to-channel dimensionality transformation on the input shallow features, rearranging spatially adjacent pixels to the channel dimension by a sampling factor of 2, resulting in features with half the spatial dimension and four times the channel dimension. Then, the transformed features are subjected to 1×1 convolutional dimensionality reduction. The dimensionality-reduced features are input into a dilated convolution branch (with a dilation rate of 2) and a skip connection branch, respectively. The output features of the two branches are then stacked and fused along the channel dimension. Finally, the fused features are fused with the features from the small object detection layer.

[0084] Step 3: Model Training

[0085] The preprocessed aerial images are input into an improved RT-DETR object detection network. The network first extracts features through a backbone network, then performs multi-scale feature fusion through a cross-layer multi-path feature aggregation module, followed by shallow feature enhancement through a spatial-to-channel feature multiplexing module, and finally further fuses deep and shallow feature information through an FPN+PAN structure. The network employs an intersection-union-ratio (IUU) aware query mechanism, selecting image features from the feature sequence output by the encoder as the initial object query, and generating the final predicted bounding box and confidence score through subsequent iterative optimization.

[0086] The AdamW optimizer was used during training, with an initial learning rate of 1×10⁻⁶. -4 The weight decay is set to 1×10. -4 The batch size was set to 16, and the number of training epochs was 150. The loss function included classification loss, bounding box regression loss, and confidence loss, which were jointly optimized in an end-to-end manner.

[0087] Step 4: Model pruning (optional)

[0088] For scenarios requiring further reduction in computational overhead, a layer-adaptive magnitude pruning strategy is employed to remove redundant weights from the trained improved network. For example... Figure 5 As shown, the network layers to be pruned are first identified, their weight matrices are flattened into one-dimensional vectors and sorted in ascending order of absolute weight values. Then, the LAMP score for each weight is calculated, and low-scoring weights are pruned according to their priority. The weight matrix is ​​updated, and it is determined whether the preset sparsity has been achieved. Finally, the remaining weights are fine-tuned to adapt the model to the pruned network structure. The same optimizer as the initial training is used during fine-tuning, with a learning rate set to 1 / 10 of the initial learning rate, and 30 fine-tuning epochs.

[0089] Step 5: Target Detection and Output

[0090] The aerial image to be detected is input into the improved RT-DETR object detection network that has been trained (or fine-tuned after pruning). The network output layer outputs the class labels, confidence scores and bounding box location information of all detected objects.

[0091] Example 2: Model Performance Validation

[0092] To verify the technical effectiveness of the present invention, comparative experiments were conducted on the VisDrone2019 public dataset and the HIT-UAV infrared thermal imaging dataset, respectively.

[0093] This invention employs an improved RT-DETR target detection algorithm for small target detection in UAV aerial images, effectively enhancing the model's target perception capability in complex aerial photography environments. Table 1 shows the comparative experimental results; on the VisDrone2019 dataset, the improved RT-DETR model significantly outperforms the original model in terms of mAP. 50 The performance is improved by 0.6%, while the number of parameters and GFLOPs are reduced by 39.2% and 27.5% respectively, achieving an FPS of 97.1, meeting the real-time requirements of UAV aerial photography missions. Compared with classic object detection networks such as Faster R-CNN and SSD, the model of this invention maintains high detection accuracy while exhibiting better inference speed and computational efficiency, demonstrating a good balance between performance and efficiency.

[0094] As shown in Table 2, the improved RT-DETR achieved higher precision, recall, and mAP on the HIT-UAV infrared thermal imaging dataset. 50 The accuracy was improved by 2.2%, 0.8%, and 0.6% respectively, significantly outperforming other comparative models. Experiments verified that the improved model possesses excellent generalization performance and adaptability when dealing with cross-modal data.

[0095] from Figure 7 The visualization of the detection results shows that in scenes with relatively dense aerial drone photography ( Figure 7 In (a), RT-DETR shows significant missed detections of small targets. In contrast, the improved RT-DETR covers more targets that correspond to real-world scenarios. When faced with false vehicle reflections in glass (…), Figure 7 In (b), RT-DETR misidentified the shadow of a vehicle as a real vehicle, while the improved RT-DETR correctly predicted it as background. Figure 7 (c) and Figure 7 As shown in (d), in the highly challenging scenario of UAV infrared photography, the improved RT-DETR still demonstrates robust generalization performance, and the issues of missed detections and false detections are also improved compared to the original model.

[0096] Table 1. Performance comparison of different models on the VisDrone2019 dataset.

[0097]

[0098] Table 2. Performance comparison of different models on the HIT-UAV dataset

[0099]

Claims

1. A UAV aerial target detection method based on RT-DETR improvement, characterized in that, Includes the following steps: (1) Aerial images are acquired using image acquisition equipment mounted on a drone; (2) Construct an improved RT-DETR target detection network, wherein the improved RT-DETR target detection network includes: The backbone network integrates a cross-level multi-path feature aggregation module and a partial convolution and compression activation module; wherein, the cross-level multi-path feature aggregation module is used to fuse multi-scale features extracted by the backbone network, and the partial convolution and compression activation module is used to replace the basic convolution module in the backbone network to reduce computational overhead and enhance the feature representation of key regions. The feature fusion network integrates a spatial-to-channel feature multiplexing module, which is used to convert the shallow spatial features extracted by the backbone network into channel features and supplement them to the small target detection layer to improve the saliency of the small target features. (3) The aerial image is input into the improved RT-DETR target detection network constructed in step (2). After feature extraction by the backbone network, multi-scale feature fusion by the cross-level multi-path feature aggregation module, and shallow feature enhancement by the space-to-channel feature reuse module, the target category and bounding box information are output by the network output layer.

2. The UAV aerial target detection method based on RT-DETR improvement according to claim 1, characterized in that, The processing flow of the cross-level multi-path feature aggregation module includes: (2.1) Perform channel dimensionality reduction on the input features from different network layers to obtain the dimensionality-reduced features; (2.2) Input the dimensionality-reduced features into the first branch and the second branch respectively; (2.3) The first branch includes a first sub-branch and a second sub-branch. The first sub-branch performs max pooling and convolution operations on the input features in sequence. The second sub-branch performs average pooling and convolution operations on the input features in sequence. The features after pooling operations participate in subsequent stacking through skip connections, and the features after convolution operations participate in subsequent stacking through direct connections. (2.4) The second branch performs a convolution operation on the input features; (2.5) Stack and fuse the features output by the first and second branches along the channel dimension; (2.6) Perform convolution operation on the stacked and fused features to restore the number of input channels and obtain the aggregated multi-scale features.

3. The UAV aerial target detection method based on RT-DETR improvement according to claim 1, characterized in that, The processing flow of the partial convolution and compression excitation module includes: (3.1) Perform 1×1 convolution and 3×3 convolution operations on the input features in sequence; (3.2) Perform partial convolution operation on the output features of step (3.1), only perform convolution operation on some continuous channels of the input features, and pass the remaining channels directly; (3.3) Perform residual fusion between the output features of step (3.2) and the output features of a 1×1 convolution branch; (3.4) Perform compression excitation operation on the residual fusion features to generate channel weights and weight them channel by channel to obtain the channel-calibrated features.

4. The UAV aerial target detection method based on RT-DETR improvement according to claim 3, characterized in that, The compression excitation operation includes: (4.1) Compression stage: The spatial dimension of each channel is compressed into a single scalar through global average pooling to generate channel descriptors; (4.2) Activation phase: The channel descriptors are processed by a learnable parameterized mapping network to generate channel weight vectors; (4.3) Reweighting stage: The generated channel weight vector is multiplied with the input feature channel by channel to achieve feature recalibration.

5. The UAV aerial target detection method based on RT-DETR improvement according to claim 1, characterized in that, The processing flow of the space-to-channel feature multiplexing module includes: (5.1) Perform spatial-to-channel dimension transformation on the shallow features extracted by the backbone network, rearrange spatially adjacent pixels to the channel dimension according to a preset sampling factor, and obtain features with halved spatial dimension and increased channel dimension; (5.2) Perform 1×1 convolution dimensionality reduction on the features output in step (5.1); (5.3) Input the dimensionality-reduced features into the dilated convolution branch and the skip connection branch respectively, and stack and fuse the output features of the two branches in the channel dimension; (5.4) The fused features are fused with the features of the small target detection layer.

6. The UAV aerial target detection method based on RT-DETR improvement according to claim 5, characterized in that, The preset sampling factor is a positive integer.

7. The UAV aerial target detection method based on RT-DETR improvement according to claim 5, characterized in that, The preset sampling factor is set to 2.

8. The UAV aerial target detection method based on RT-DETR improvement according to claim 1, characterized in that, Following step (3), the process also includes performing layer-adaptive magnitude pruning on the trained improved RT-DETR network: (8.1) Determine the network layer to be pruned, flatten its weight matrix into a one-dimensional vector, and arrange them in ascending order of absolute weight value; (8.2) Calculate the LAMP score for each weight and prune the low-scoring weights according to the score priority; (8.3) Update the weight matrix and determine whether the preset sparsity has been reached; (8.4) Fine-tune the remaining weights to fit the pruned network structure.

9. The UAV aerial target detection method based on RT-DETR improvement according to claim 8, characterized in that, The LAMP score is dynamically calculated based on the contribution of each weight in the network, and weights with lower scores are pruned first.

Images