Lightweight small target detection method for unmanned aerial vehicle aerial image

Abstract

The application provides a lightweight small target detection method for unmanned aerial vehicle aerial image. A lightweight small target detection network based on improved YOLOv11s is constructed, a mixed architecture of multi-scale prediction head optimization, full-dimension dynamic convolution ODConv and lightweight deep convolution LDConv is fused, a shape-sensitive loss function is combined, hierarchical feature reservation, dynamic-static collaborative perception, geometric constraint enhancement and boundary box shape error quantization are realized, and target detection precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, specifically relating to a lightweight small target detection method for drone aerial images. It is applicable to the detection of small targets such as pedestrians and vehicles in drone aerial images and can be deployed on embedded platforms. Background Technology

[0002] Current drone aerial target detection technology faces significant challenges: traditional two-stage detection models (such as Faster R-CNN) have high accuracy, but their computational complexity is high (typical models have over 40M parameters and 187 GFLOPs of computation), making it difficult to meet the real-time requirements of drone embedded platforms; while single-stage lightweight models (such as the YOLO series) often sacrifice small target detection performance when compressing parameters, and their deep convolutions lead to severe loss of small target features. In the VisDrone2019 dataset, the baseline model YOLOv11s has an mAP@0.5 (mAP@0.5 represents the mean accuracy value at an IoU threshold of 0.5) of only 38.2%. Existing methods generally suffer from three main drawbacks: First, multi-scale prediction head designs are redundant; large target detection heads (20×20 resolution) are almost ineffective for small targets (<32 pixels), yet increase computation by 23.2%. Second, static convolutional kernels are ill-suited to the diverse target shapes and complex backgrounds in UAV scenarios, and the feature extraction module lacks dynamic adjustment capabilities. Third, traditional CIoU loss functions focus only on the geometric relationships of bounding boxes, ignoring target shape and scale information (e.g., aspect ratio deviations lead to IoU calculation errors of up to 15%), resulting in insufficient localization accuracy. Although existing improvement schemes attempt to optimize the model through channel pruning (e.g., YOLOv6) or attention mechanisms (e.g., RepVGG-YOLO), mAP@0.5 decreases by 5.2% when the number of parameters is compressed to 34.2M, and the shape sensitivity problem remains unresolved. Furthermore, the characteristics of UAV images, such as dense targets, large scale differences (coexistence of 0.5-30 meter resolutions), and strong background interference, further exacerbate the imbalance between accuracy and efficiency in existing models, and there is an urgent need for a detection scheme that takes into account both lightweight design and shape perception. Summary of the Invention

[0003] To overcome the technical bottlenecks in existing UAV aerial image small target detection, such as the difficulty in balancing accuracy and efficiency, the inability of traditional convolution to adapt to complex backgrounds, and large shape localization errors, this invention provides a lightweight small target detection method for UAV aerial images. A lightweight small target detection network based on an improved YOLOv11s is constructed, integrating a hybrid architecture of multi-scale prediction head optimization, full-dimensional dynamic convolution ODConv, and lightweight deep convolution LDConv. Combined with a shape-sensitive loss function, hierarchical feature preservation, dynamic-static collaborative perception, geometric constraint enhancement, and bounding box shape error quantification are achieved, thereby improving target detection accuracy.

[0004] A lightweight small target detection method for UAV aerial images, characterized by the following steps: Step 1: Perform size normalization, pixel normalization, and data augmentation on the images in the existing UAV aerial image dataset and the corresponding remote sensing image dataset, and convert the data annotation format; then, divide the dataset into training dataset, test dataset, and validation dataset respectively. Step 2: Construct an object detection network based on YOLOv11s. Specifically, this includes: replacing the C3K2 modules in layers 2, 4, and 6 of the backbone network of YOLOv11s with the full-dimensional dynamic convolution ODConv module; replacing the standard convolutions in layers 1 and 3 of the PANet upsampling path in the neck network of YOLOv11s with the lightweight deep convolution LDConv module; adding a detection head with a resolution of 160×160; deleting the original 20×20 resolution prediction head and its corresponding three C3 modules located in layers 6, 7, and 8 of the neck network; and deleting the associated upsampling layers and feature fusion pathways. Step 3: Train the object detection network constructed in Step 2 using the dataset obtained in Step 1. The total network loss is set as follows: (1) in, Indicates the total loss. Represents shape loss, and IoU represents fundamental loss; The shape loss The calculation formula is as follows: (2) in, ω The aspect ratio weighting coefficient is calculated according to... Calculations show that w g t This indicates a broad true target. h gt This indicates that the actual target is high. w pred This indicates a broad forecast target. h pred This indicates a high forecast target; The formula for calculating the basic loss IoU is as follows: (3) in, B pred Indicates the predicted bounding box. B g t Represents the true bounding box; Step 4: Compress and quantize the trained network, then input the drone aerial image to be detected into the network, and output the target detection result.

[0005] Specifically, the size normalization process in step 1 involves scaling the image to a fixed size of 640×640, with any insufficient areas filled with gray and RGB values ​​of [114, 114, 114]; the pixel normalization process involves normalizing the pixel values ​​in the image to [0,1]; the data augmentation process employs a Mosaic enhancement strategy, where random cropping ensures that at least one target larger than 4×4 pixels is included; and the data annotation format conversion involves converting the original annotations to YOLO format.

[0006] Specifically, the processing procedure of the full-dimensional dynamic convolution ODConv module in step 2 is as follows: First, the input feature map is compressed using full average pooling, as expressed by the formula: (4) in, GAP (·) indicates a full average pooling operation. X Indicates the input feature map, X gap This represents the compressed image; Then, four-dimensional attention weights are generated through a fully connected layer, expressed by the formula: (5) in, α Indicates the attention weights in the spatial dimension; β Indicates the attention weights along the input channel dimension; γ Indicates the attention weights for the output channel dimension; δ Indicates the weights fused from multiple convolutional kernels; Softmax (·) represents the normalized exponential function; FC Represents the dynamic convolution kernel weight matrix; ReLU (·) denotes the modified linear unit activation function; Next, the original convolutional kernel is adjusted using four-dimensional attention weights, as expressed by the formula: (6) in, W Represents the original convolution kernel. Represents element-wise multiplication. W od This represents the adjusted convolution; Finally, the output feature map is calculated, and the formula is expressed as: (7) Where Y represents the output feature map. Conv(·) represents the convolution operation. b This represents the activation amount of the bias term feature translation.

[0007] Specifically, the processing procedure of the lightweight deep convolutional LDConv module in step 2 is as follows: First, the input feature map is processed by a convolutional layer to generate an offset field, expressed by the formula: (8) Where, Δ P Represents the offset field. Conv2d (·) represents a 3×3 convolution operation. X Indicates the input feature map; Next, the coordinates are updated using the initial sampling grid, as expressed by the formula: (9) in, Indicates the updated sampling coordinates; Indicates the predefined coordinates of the initial sampling grid; initial sampling grid Distributed in a 5-point cross shape, N The number of sampling points. This represents the offset coordinates obtained from the learning process; Then, bilinear interpolation feature extraction is performed to obtain the interpolated feature values, expressed by the formula: (10) in, Represents the interpolated eigenvalues; i , j ) represents the offset of the nearest point, and iterates through the four nearest integer coordinate points; Represents the decimal part of the x-coordinate. Represents the decimal part of the y-coordinate; Indicates the interpolation weights; Finally, the interpolated features are input into subsequent convolutional layers to extract the final feature representation.

[0008] Specifically, in step 3, the network training uses the AdamW optimizer with an initial learning rate of 0.01, a weight decay of 0.05, and a learning rate scheduling period of 200 epochs. The backpropagation algorithm is used, and the training process is as follows: First, pre-trained YOLOv11s weights without the detector head are loaded. For the first 50 epochs, the backbone network parameters are frozen, and only the detector head and neck network are trained with a learning rate of 0.01. For the next 150 epochs, all network parameters are trained, and the learning rate is reduced to 0.001. If the evaluation metrics on the validation set do not improve for 10 consecutive epochs, the training is terminated early.

[0009] Specifically, the method for compressing and quantizing the trained network in step 4 is as follows: use FP16 mixed precision training data stream and remove channels with weights less than 1e-3 in the full-dimensional dynamic convolution ODConv module.

[0010] The beneficial effects of this invention are as follows: By adding a small target detection head and removing redundant large target detection heads, hierarchical feature preservation is facilitated, effectively improving the network model's detection accuracy for small targets and reducing the number of model parameters; by introducing full-dimensional dynamic convolution and lightweight depthwise convolution, dynamic-static collaborative perception and geometric constraint enhancement are achieved, enabling the network model to adapt to scenarios such as changes in lighting conditions, local occlusion, background interference, and irregularly shaped small targets; by employing an innovatively designed shape loss function, the problem of large localization errors for asymmetric targets is solved; this invention can achieve high-precision detection of small targets, and experiments show that it can accurately detect 4×4 pixel targets. The recall rate is improved by 67.4%, and the localization error of asymmetric targets such as vehicles and pedestrians is reduced by 18.7%. This invention enables lightweight deployment, compressing the network model size to less than 20MB. Through dynamic feature learning and shape-aware optimization, the network model of this invention shows significant advantages in complex scenes and has strong adaptability to multi-scale and multi-form targets. Experiments on existing datasets show that the detection accuracy of small targets such as ships and wind turbines is improved by 23.5%. This invention provides a "high-precision, low-energy-consumption, and robust" embedded detection solution for resource-constrained UAV platforms, significantly breaking through the performance bottleneck of traditional models in complex aerial photography scenarios. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the target detection network structure of the present invention; Figure 2 This is a schematic diagram of the full-dimensional dynamic convolution ODConv module structure of the present invention; Figure 3 This is a schematic diagram of the workflow of the lightweight depthwise convolution LDConv module of the present invention; Figure 4 This is an example diagram of the Shape-IoU loss function of the present invention; Figure 5 This is a schematic diagram of the network training process of the present invention; Figure 6 This is a schematic diagram of the network inference process of the present invention; Figure 7 This is a comparison chart of the detection performance of the network model of this invention and the YOLOv11s model on the VisDrone dataset. Detailed Implementation

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. The present invention includes, but is not limited to, the following embodiments.

[0013] This invention provides a lightweight small target detection method for UAV aerial images, the specific implementation process of which is as follows: 1. Data Preprocessing and Augmentation The images in the existing UAV aerial image dataset (VisDrone2019-DET dataset in this embodiment) and the corresponding remote sensing image dataset (DIOR dataset in this embodiment) are subjected to size normalization and pixel normalization. Size normalization is to scale the image to a fixed size of 640×640 while maintaining the aspect ratio, and fill any insufficient parts with gray with RGB values ​​of [114, 114, 114]. Pixel normalization is to normalize the pixel values ​​in the image to [0,1].

[0014] Then, data augmentation was performed using a Mosaic augmentation strategy: four images were randomly selected and stitched together, with random scaling ratios ranging from [0.5, 1.5] and rotation angles ranging from [-45°, 45°], and then proportionally... Beta(1.0, 1.0) blends two images with linearly superimposed labels, expressed by the formula: (11) (12) in, X mix This represents the new image data generated by hybrid enhancement. X A This represents the first original image data to be mixed. X B This represents the second original image data that is being mixed. Y mix This represents the label or target value corresponding to the blended image. Y A Representing an image X A The corresponding original label or target value, Y B Representing an image X B The corresponding original label or target value; This represents the weighting coefficient, which is between 0 and 1 and is determined by the Beta(1.0,1.0) distribution. It is used to control the mixing ratio.

[0015] When cropping randomly, ensure that at least one target larger than 4×4 pixels is included to avoid full cropping and protect small targets.

[0016] Data annotation format conversion is also required, converting the original annotations to YOLO format to adapt to the web.

[0017] Finally, the dataset is divided into training, testing, and validation datasets. The specific number of samples for each dataset division in this embodiment is shown in Table 1.

[0018] Table 1

[0019] 2. Construct an object detection network model like Figure 1 As shown, this invention constructs an object detection network based on YOLOv11s, and the main improvements include: (1) The C3K2 modules in layers 2, 4, and 6 of the YOLOv11s backbone network are replaced with the full-dimensional dynamic convolution ODConv module.

[0020] like Figure 2 As shown, the processing procedure of the full-dimensional dynamic convolution ODConv module is as follows: First, the input feature map... ( C Indicates the dimension of the feature map. H Indicates the height of the feature map, W (representing the width of the feature map), which is compressed using full average pooling, as shown in the formula: (13) in, GAP (·) indicates a full average pooling operation. X Indicates the input feature map, X gap This represents the compressed image; Then, four-dimensional attention weights are generated through a fully connected layer, expressed by the formula: (14) in, This represents the spatial dimension attention weight, focusing on the target spatial location; This represents the attention weights for the input channel dimension, used to suppress invalid input channels; This represents the attention weights in the output channel dimension, enhancing target-related features; Indicates the weights of multi-convolution kernel fusion ( N (number of convolutional kernels), adapting to multi-scale target morphology; Softmax (·) represents the normalization exponential function, used for calculating the confidence score of the target classification; FC This represents the dynamic convolution kernel weight matrix, which is dynamically adjusted in the ODConv module through a four-dimensional attention mechanism; ReLU (·) denotes the modified linear unit activation function, used for nonlinear transformation of features after convolutional layers to enhance the model's robustness to changes in illumination. Next, the original convolutional kernel is adjusted using four-dimensional attention weights, as expressed by the formula: (15) in, W Represents the original convolution kernel. Representing element-wise multiplication, enabling four-dimensional dynamic modulation. W od This represents the adjusted convolution; Finally, the output feature map is calculated, and the formula is expressed as: (16) Where Y represents the output feature map. Conv (·) represents the convolution operation. b This represents the activation amount of the bias term feature translation, which improves the expressive power of the model.

[0021] The full-dimensional dynamic convolution ODConv can effectively enhance the ability to extract local features by introducing a four-dimensional attention mechanism into the backbone network.

[0022] (2) The standard convolutions in layers 1 and 3 of the PANet upsampling path in the neck network of YOLOv11s are replaced with the lightweight depthwise convolution LDConv module.

[0023] like Figure 3 As shown, the processing procedure of the lightweight depthwise convolutional LDConv module is as follows: First, the input feature map... The offset field is generated through convolutional layers, as expressed by the formula: (17) Where, Δ P Represents the offset field. Conv2d (·) represents a 3×3 convolution operation. X Indicates the input feature map; Next, the coordinates are updated using the initial sampling grid, as expressed by the formula: (18) in, This represents the updated sampling coordinates, enabling dynamic deformation alignment of the feature map; The initial sampling grid represents the predefined coordinates of the initial sampling grid, the fixed sampling positions of the standard convolution; the initial sampling grid. Distributed in a 5-point cross shape, N The number of sampling points. The learned offset coordinates allow the sampling position to adapt to target deformation. Then, bilinear interpolation feature extraction, which preserves the features of small targets, is performed to obtain the interpolated feature values, expressed by the formula: (19) in, This represents the interpolated feature values, addressing the problem of feature loss in small targets; i , j ) represents the offset of the nearest point, and iterates through the four nearest integer coordinate points; Represents the decimal part of the x-coordinate, precisely controlling the fusion of features in the horizontal direction. Represents the decimal part of the y-coordinate, precisely controlling the fusion of features in the vertical direction; This represents the interpolation weight, ensuring a smooth transition of features and eliminating jagged edges.

[0024] Finally, these interpolated (i.e., resampled) features are fed into subsequent convolutional layers to extract the final feature representation.

[0025] Lightweight deep convolution LDConv is a linear deformable convolution with a dynamic sampling mechanism. In the neck network, the sampling position is dynamically adjusted through offset learning, which can improve the perception of small targets with irregular shapes.

[0026] (3) Add a detection head with a resolution of 160×160.

[0027] The formula for the target detection head in YOLOv11s is: (20) in, This is the output feature map of the second layer of the backbone network. Output TinyHead This represents the final output of the lightweight object detection head. f decoder This represents the decoder function for the detection results. SPFF This represents the Spatial Pyramid Pooling Fusion (SPPF) module. C 3 K 2 indicates a cross-stage local convolution block.

[0028] To avoid feature misalignment and ensure resolution matching with anchors, this invention adds a small target detection head. Based on the existing three-scale prediction head (80×80, 40×40, 20×20) in YOLOv11s, it takes the downsampled values ​​from the second layer of the backbone network and inputs a feature map size of 160×160 as the resolution prediction head, specifically for detecting small targets larger than 4×4 pixels. At 160×160 resolution, each grid corresponds to the original... Figure 4The ×4 pixel region (640 / 160=4) is strictly matched with the minimum size of the Anchor (based on k-means clustering analysis of small target sizes in the VisDrone dataset, the Anchor is configured to cover targets from 4×4 to 16×16 pixels. The detection layer is configured with 3 Anchor Boxes with scales of [4,6], [8,12], and [12,16]).

[0029] (4) Remove redundant large target detection heads and associated layers, specifically: delete the original 20×20 resolution prediction head and its corresponding 3 C3 modules (located in the 6th, 7th and 8th layers of the neck network); delete the associated upsampling layer (2× NearestUpsample) and feature fusion pathway.

[0030] The formula for calculating parameter reduction is as follows: (twenty one) in, P original This represents the total number of parameters after parameter reduction. C 3params This indicates the number of parameters for a single C3 module. U psample_params This indicates the number of parameters associated with the upsampling layer.

[0031] The number of parameters decreased from 9.4M to 2.9M (↓69.8%). Computational cost reduced: from 21.3 GFLOPs to 17.6 GFLOPs (↓17.4%).

[0032] The target detection network constructed in this invention adds the following advantages to the original YOLOv11s model: 1) Dynamic-static hybrid architecture: combining full-dimensional dynamic convolution ODConv with deformable sampling lightweight depth convolution LDConvLDConv, improving the accuracy of small target detection under the premise of lightweight (5.0M model parameters); 2) Hierarchical feature preservation mechanism: By adding a 160×160 high-resolution prediction head, the details of small targets are preserved, thus solving the problem of feature loss in deep networks.

[0033] 3. Network training The object detection network constructed in step 2 is trained using the dataset obtained in step 1. Specifically, this invention designs a Shape-IoU loss function, which adds a shape loss to the basic IoU loss to form the total network loss: (twenty two) in, Indicates the total loss. The shape loss is represented by IoU, and the basic loss is represented by IoU.

[0034] Shape loss The calculation formula is as follows: (twenty three) in, ω The aspect ratio weighting coefficient is calculated according to... The calculations show that the optimized bounding box regression... w g t This indicates a broad true target. h gt This indicates that the actual target is high. w pred This indicates a broad forecast target. h pred This indicates a high forecast target; The formula for calculating the basic loss IoU is as follows: (twenty four) in, B pred This represents the predicted bounding box, the rectangular region of the predicted target location output by the model, in pixel coordinates. B g t Represents the true bounding box, the rectangular area of ​​the target's true location, in pixel coordinates.

[0035] By introducing a Shape-IoU loss function with aspect ratio weighting coefficients, the bounding box shape error can be quantified, which can improve the localization accuracy of asymmetric targets such as vehicles and pedestrians. Figure 4 An example of the Shape-IoU loss function is given. Shape-IoU emphasizes the shape and size of the bounding box, thus making the bounding box regression more accurate.

[0036] During training, the optimizer AdamW was set with an initial learning rate of 0.01, a weight decay of 0.05, and a learning rate scheduling period of 200 epochs. The backpropagation algorithm was used, and the training process was as follows: First, pre-trained YOLOv11s weights without the detector head were loaded. For the first 50 epochs, the backbone network parameters were frozen, and only the detector head and neck network were trained with a learning rate of 0.01. For the next 150 epochs, all network parameters were updated, and the learning rate was reduced to 0.001. If the evaluation metrics on the validation set did not improve for 10 consecutive epochs, training was terminated early.

[0037] The core indicators for evaluation include: (25) (26) Efficiency metrics include: number of parameters (M), computational cost (GFLOPs), and inference latency (ms / frame).

[0038] Where 'c' represents the category index and 'AP' represents the average precision. mAP This indicates the average precision of the mean. 0.5 represents the average precision across all categories when the IoU threshold is 0.5. This indicates the average accuracy calculated when the IoU threshold is between 0.5 and 0.95.

[0039] Figure 5 A schematic diagram of the network training process is given, wherein the inference network is the target detection network constructed in this invention, and the training and evaluation network represents the training and evaluation of the network.

[0040] 4. Model compression and quantization and target detection The model compression and quantization uses the FP16 mixed-precision training data stream, which can be converted to INT8 format (TensorRT) upon export. Based on channel importance scoring, channels with weights less than 1e-3 in the full-dimensional dynamic convolution ODConv module are pruned. Afterwards, it can be optimized and adapted for embedded deployment on NVIDIA Jetson Nano.

[0041] The aerial images of the drone to be detected are input into the final deployed network, and the output is the target detection result. For example... Figure 6 As shown.

[0042] To verify the effectiveness of this invention, data experiments were conducted in the following environment: Training environment: NVIDIA GTX 1080Ti GPU (11GB VRAM), Intel Xeon E5-2678 v3 CPU, 64GB RAM; Deployment environment: NVIDIA Jetson Nano (4GB RAM, 128 CUDA cores), Ubuntu 18.04 LTS; Deep learning framework: PyTorch 1.10.0 + TorchVision 0.11.1; Acceleration libraries: CUDA 11.2, cuDNN 8.1.0; Deployment tool: TensorRT 8.2.1.8.

[0043] Experiments show that, compared to the baseline model YOLOv11s, this invention achieves a 9.6% and 6.9% improvement in mAP@0.5 on the VisDrone2019 and DIOR datasets, respectively, with a 46.8% reduction in parameters (9.4M → 5.0M) and a 17.4% reduction in computation (21.3 → 17.6 GFLOPs). Real-time detection at 30 FPS is also achieved on embedded platforms such as the NVIDIA Jetson Nano. Through dynamic feature learning and shape-aware optimization, the model demonstrates significant advantages in complex scenes. 1) High-precision detection: The recall rate for 4×4 pixel targets was improved by 67.4% (34.7%→58.1%), and the positioning error for asymmetric targets such as vehicles and pedestrians was reduced by 18.7%; 2) Lightweight deployment: The model size is compressed to less than 20MB, supports FP16 / INT8 quantization, and maintains mAP@0.5 above 45% under conditions of ±30% illumination change and 40% target occlusion. 3) Cross-scene generalization: The detection accuracy of small targets (<8×8 pixels) such as ships and wind turbines in the DIOR remote sensing dataset is improved by 23.5%, which verifies the strong adaptability of the algorithm to targets of multiple scales and shapes.

[0044] The experimental data comparison is shown in Table 2.

[0045] Table 2

[0046] Figure 7 Examples of detection results on the VisDrone dataset using the network model of this invention and the YOLOv11s model are given respectively. It can be seen that this invention can better detect small targets.

[0047] This invention provides a "high-precision, low-energy-consumption, and robust" embedded detection solution for resource-constrained UAV platforms, significantly breaking through the performance bottleneck of traditional models in complex aerial photography scenarios.

Claims

1. A lightweight small target detection method for UAV aerial images, characterized in that... The steps are as follows: Step 1: Perform size normalization, pixel normalization, and data augmentation on the images in the existing UAV aerial image dataset and the corresponding remote sensing image dataset, and convert the data annotation format. Then, the dataset is divided into training dataset, test dataset and validation dataset respectively; Step 2: Construct an object detection network based on YOLOv11s. Specifically, this includes: replacing the C3K2 modules in layers 2, 4, and 6 of the backbone network of YOLOv11s with the full-dimensional dynamic convolution ODConv module; replacing the standard convolutions in layers 1 and 3 of the PANet upsampling path in the neck network of YOLOv11s with the lightweight deep convolution LDConv module; adding a detection head with a resolution of 160×160; deleting the original 20×20 resolution prediction head and its corresponding three C3 modules located in layers 6, 7, and 8 of the neck network; and deleting the associated upsampling layers and feature fusion pathways. Step 3: Train the object detection network constructed in Step 2 using the dataset obtained in Step 1. The total network loss is set as follows: （1） in, Indicates the total loss. Represents shape loss, and IoU represents fundamental loss; The shape loss The calculation formula is as follows: （2） in, ω The aspect ratio weighting coefficient is calculated according to... Calculations show that w g t This indicates that the actual target is broad. h gt This indicates that the actual target is high. w pred This indicates a broad forecast target. h pred This indicates a high forecast target; The formula for calculating the basic loss IoU is as follows: （3） in, B pred Indicates the predicted bounding box. B g t Represents the true bounding box; Step 4: Compress and quantize the trained network, then input the drone aerial image to be detected into the network, and output the target detection result.

2. The lightweight small target detection method for UAV aerial images as described in claim 1, characterized in that: The size normalization process described in step 1 involves scaling the image to a fixed size of 640×640, with any insufficient areas filled with gray, and RGB values ​​of [114, 114, 114]. The pixel normalization process normalizes the pixel values ​​in the image to [0,1]. The data augmentation process employs a Mosaic enhancement strategy, where random cropping ensures that at least one target larger than 4×4 pixels is included. The data annotation format conversion converts the original annotations to YOLO format.

3. The lightweight small target detection method for UAV aerial images as described in claim 1, characterized in that: The processing procedure of the full-dimensional dynamic convolution ODConv module in step 2 is as follows: First, the input feature map is compressed using full average pooling, as expressed by the formula: （4） in, GAP (·) indicates a full average pooling operation. X Indicates the input feature map, X gap This represents the compressed image; Then, four-dimensional attention weights are generated through a fully connected layer, expressed by the formula: （5） in, α Indicates the attention weights in the spatial dimension; β Indicates the attention weights along the input channel dimension; γ Indicates the attention weights for the output channel dimension; δ Indicates the weights fused from multiple convolutional kernels; Softmax (·) represents the normalized exponential function; FC Represents the dynamic convolution kernel weight matrix; ReLU (·) denotes the modified linear unit activation function; Next, the original convolutional kernel is adjusted using four-dimensional attention weights, as expressed by the formula: （6） in, W Represents the original convolution kernel. Represents element-wise multiplication. W od This represents the adjusted convolution; Finally, the output feature map is calculated, and the formula is expressed as: （7） Where Y represents the output feature map. Conv (·) represents the convolution operation. b This represents the activation amount of the bias term feature translation.

4. The lightweight small target detection method for UAV aerial images as described in claim 1, characterized in that: The lightweight depthwise convolutional LDConv module described in step 2 processes the following: First, the input feature map is processed by a convolutional layer to generate an offset field, expressed by the formula: （8） Where, Δ P Represents the offset field. Conv2d (·) represents a 3×3 convolution operation. X Indicates the input feature map; Next, the coordinates are updated using the initial sampling grid, as expressed by the formula: （9） in, Indicates the updated sampling coordinates; Indicates the predefined coordinates of the initial sampling grid; initial sampling grid Distributed in a 5-point cross shape, N The number of sampling points. This represents the offset coordinates obtained from the learning process; Then, bilinear interpolation feature extraction is performed to obtain the interpolated feature values, expressed by the formula: （10） in, Represents the interpolated eigenvalues; i , j ) represents the offset of the nearest point, and iterates through the four nearest integer coordinate points; Represents the decimal part of the x-coordinate. Represents the decimal part of the y-coordinate; Indicates the interpolation weights; Finally, the interpolated features are input into subsequent convolutional layers to extract the final feature representation.

5. A lightweight small target detection method for UAV aerial images as described in claim 1, characterized in that: The network training described in step 3 uses the AdamW optimizer with an initial learning rate of 0.01, a weight decay of 0.05, and a learning rate scheduling period of 200 epochs. The backpropagation algorithm is employed, and the training process is as follows: First, pre-trained YOLOv11s weights without the detector head are loaded. For the first 50 epochs, the backbone network parameters are frozen, and only the detector head and neck network are trained with a learning rate of 0.

01. For the next 150 epochs, all network parameters are trained, and the learning rate is reduced to 0.

001. If the evaluation metrics on the validation set do not improve for 10 consecutive epochs, training is terminated early.

6. The lightweight small target detection method for UAV aerial images as described in claim 1, characterized in that: The specific method for compressing and quantizing the trained network described in step 4 is as follows: use FP16 mixed precision training data stream and remove channels with weights less than 1e-3 in the full-dimensional dynamic convolution ODConv module.

Images