Small target detection method and system based on unmanned aerial vehicle vision

Abstract

The application discloses a kind of small target detection method and system based on unmanned aerial vehicle vision, and the specific steps of the method include: obtaining unmanned aerial vehicle image dataset, and pre-processing, construct standardized dataset;Improved YOLO model is constructed, and the model includes main network, neck network and detection head in turn;Multi-scale dynamic gate fusion unit is integrated in the main network, and the branch specially used for small scale target detection is contained in the detection head;Based on the standardized dataset, the improved YOLO model is trained using multi-scale weighted comprehensive loss function;Unmanned aerial vehicle image is detected using the trained model.The method improves through double-layer structure, significantly improves the detection robustness and precision of network, can realize efficient, accurate identification to multiple categories of traffic targets under complex traffic intersection environment, and provides reliable technical support for unmanned aerial vehicle video traffic behavior analysis and traffic flow monitoring.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and computer vision technology, specifically relating to a small target detection method and system based on UAV vision. Background Technology

[0002] With the rapid development of intelligent transportation systems, drones have become an important tool for road traffic monitoring due to their advantages such as high mobility, wide coverage, and low-cost deployment. The top-down view of drones can acquire high spatiotemporal resolution video data, providing a rich information foundation for road traffic flow statistics, target detection, trajectory extraction, and safety conflict analysis. However, due to the high shooting angle and wide field of view of drones, the scale of targets in the images varies significantly, especially the features of small targets such as distant vehicles and pedestrians are not obvious, leading to a significant decrease in the detection accuracy of existing detection algorithms from the drone's perspective.

[0003] Currently, most mainstream traffic target detection methods are based on convolutional neural networks (CNNs), among which the YOLO series algorithms are widely used in traffic monitoring tasks due to their end-to-end detection capabilities and high detection speed. However, traditional YOLO algorithms are mainly designed for fixed ground-based perspectives, and their feature extraction networks have limited responsiveness to small-scale targets, resulting in missed detections and false detections of vehicles, non-motorized vehicles, and pedestrians from distant perspectives. Furthermore, drone photography is affected by factors such as clouds, shadows, reflections, and camera shake, blurring target edge information. The fixed receptive field of traditional convolutional kernels cannot fully capture cross-scale contextual features, further limiting detection performance.

[0004] To address these issues, some researchers have introduced Transformer architectures or multi-scale attention mechanisms to enhance the network's global modeling capabilities. However, these methods often suffer from excessively large model parameters and insufficient real-time performance, hindering real-time inference deployment on UAVs. Therefore, improving the accuracy and feature representation capabilities of small target detection from the UAV's perspective while maintaining detection speed has become a key technical challenge in the field of UAV traffic target detection.

[0005] Therefore, there is an urgent need to propose a lightweight detection model that can take into account multi-scale feature fusion and global context modeling while ensuring real-time performance, so as to improve the detection accuracy and robustness in UAV aerial photography environments. Summary of the Invention

[0006] To overcome the shortcomings of existing UAV traffic target detection methods in small-scale target recognition, global feature modeling, and robustness in complex environments, this invention proposes a small target detection method and system based on UAV vision.

[0007] The technical solution adopted in this invention is as follows:

[0008] A small target detection method based on UAV vision, comprising the following steps:

[0009] S1. Acquire drone imagery datasets, preprocess them, and construct standardized datasets;

[0010] S2. Construct an improved YOLO model (GSA-YOLO model), which consists of a backbone network, a neck network, and a detection head; the backbone network integrates multi-scale dynamic gating fusion units, and the detection head contains a branch specifically for small-scale target detection;

[0011] S3. Based on the standardized dataset, the improved YOLO model is trained using a multi-scale weighted comprehensive loss function;

[0012] S4. Use the trained model to perform target detection on UAV imagery.

[0013] Furthermore, the preprocessing in step S1 includes: obtaining the annotation file of the corresponding image data, unifying the annotation file format to a YOLO compatible format, and normalizing the target center coordinates and width and height parameters; dividing the training set, validation set and test set according to the proportion; and introducing data augmentation strategies including random flipping, scale transformation, brightness perturbation and color jitter.

[0014] Furthermore, the multi-scale dynamic gating fusion unit in the backbone network is a C3k2_MDGF unit, which is implemented by replacing the C3k2 unit in the original backbone network.

[0015] Furthermore, the C3k2_MDGF unit includes a parallel multi-scale dynamic attention (MDTA) branch and a gated depth feedforward (GDFN) branch, as well as a fusion layer that fuses the outputs of these two branches.

[0016] Furthermore, the MDTA branch for input features Output Calculated using the following formula:

[0017]

[0018] in, These are the mapping tensors generated by 1×1 and 3×3 convolutions, respectively. This represents the scaling factor for the key vector dimension. Tensor The transpose operation, This represents the tensor reconstruction process, used to restore the product result to the spatial dimension of the original input. This indicates normalization on the appropriate dimension.

[0019] Furthermore, the GDFN branch for input features Output Calculated using the following formula:

[0020] ,

[0021] in, The convolution mapping matrix, represents element-wise multiplication, and GELU represents a non-linear activation function.

[0022] Furthermore, the fusion layer employs either splicing compression or element-wise addition to achieve feature integration:

[0023] ,

[0024] And introduce a channel adaptive gating function:

[0025] ,

[0026] Final output:

[0027] ,

[0028] Where g represents the gating parameter, The activation function is Sigmoid, and GAP represents global average pooling. This is the channel gating weight matrix.

[0029] Furthermore, the detection head includes four scale detection branches based on feature maps of four different resolutions, with the first scale branch having the highest spatial resolution and being used specifically for detecting small-scale targets.

[0030] Furthermore, the multi-scale weighted comprehensive loss function is defined as follows:

[0031] ,

[0032] Here, P1, P2, P3, and P4 represent four detection branches at different scales, with the resolution decreasing sequentially. Let be the scale weight coefficient of the s-th branch, and assign the maximum weight to the small target branch P1; The classification loss is Focal Loss; The bounding box regression loss uses Complete IoU; The confidence loss is calculated using binary cross-entropy; This is the loss weighting coefficient.

[0033] A small target detection system based on UAV vision, used in the above method, includes:

[0034] Data processing module: used to acquire UAV imagery datasets, perform preprocessing, and construct standardized datasets;

[0035] Model building module: used to build an improved YOLO model, which includes a backbone network, a neck network and a detection head in sequence; the backbone network integrates a multi-scale dynamic gating fusion module, and the detection head contains a branch specifically for small-scale target detection;

[0036] Model training module: used to train the improved YOLO model based on the standardized dataset using a multi-scale weighted comprehensive loss function;

[0037] Target detection module: Used to perform target detection on UAV imagery using a trained model.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This invention combines the imaging characteristics of UAV aerial video with the complex and ever-changing visual features of traffic scenes, proposing an improved multi-scale dynamic gating fusion detection model, GSA-YOLO. Through deep optimization of the YOLOv11n network structure, this model enables more accurate identification and localization of distant vehicles, small traffic targets, and complex background areas from the UAV's perspective. Compared to existing lightweight detection algorithms, this invention significantly improves the hierarchical nature of feature representation and small target detection performance while maintaining high detection speed.

[0040] 2. This invention employs a multi-scale dynamic gating fusion structure, fusing multi-scale dynamic attention with a gated deep feedforward network in parallel to form the C3k2_MDGF feature extraction unit. This unit achieves interactive learning and effective fusion of cross-scale features through dynamic weight allocation and adaptive channel adjustment mechanisms, thereby improving the network's feature response strength for small targets at distant viewpoints. Compared to traditional convolutional stacking extraction methods, this invention maintains stable detection accuracy under complex lighting, occlusion, and high-density traffic conditions.

[0041] 3. This invention is based on the concept of multi-scale structure improvement, adding a small target detection branch to the original YOLOv11n detection head. This branch fully utilizes the complementarity between low-level detail features and high-level semantic features through upsampling and shallow feature fusion, achieving coordinated detection of multi-level features. This improvement effectively alleviates the problem of missed detection of small targets due to reduced resolution in UAV top-down scenes, and significantly improves the model's detection accuracy for small-sized traffic targets (such as distant vehicles, pedestrians, bicycles, etc.).

[0042] 4. This invention employs a multi-scale weighted loss function and an adaptive learning rate optimization strategy, combined with the feature distribution of the UAV perspective dataset, to design a training mechanism suitable for high-altitude, low-angle, and multi-scene conditions. Through techniques such as Mosaic data augmentation, brightness perturbation, and scale transformation, this invention effectively enhances the model's robustness to environmental changes while maintaining its lightweight characteristics. Experiments demonstrate that the average accuracy (mAP@0.5) of this invention on the VisDrone2019 dataset is approximately 3.9% higher than the original YOLOv11n model, while maintaining a high real-time detection speed.

[0043] 5. This invention is not only applicable to UAV traffic target detection, but can also be extended to other UAV visual inspection scenarios, including but not limited to: (1) abnormal detection and status assessment of infrastructure defects, cracks, corrosion, obstacles and other abnormalities in high-altitude inspection scenarios. The structure can enhance the ability to identify small structural targets in high-resolution images; (2) traffic density detection, vehicle trajectory extraction and congestion pattern recognition in urban traffic flow analysis scenarios. The structure can achieve stable detection and real-time response in multi-angle, complex background and multi-target interference environments; (3) dynamic target identification and location positioning of personnel, vehicles, equipment and other dynamic targets in emergency disaster monitoring and rescue scenarios. The structure can maintain high detection accuracy and response speed under strong noise, uneven lighting and occlusion conditions. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall method in an embodiment of the present invention.

[0045] Figure 2 This is a diagram of the improved YOLO model network structure in an embodiment of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be further described clearly and in detail below with reference to the accompanying drawings and specific examples.

[0047] like Figure 1 As shown, a small target detection method based on UAV vision includes the following steps:

[0048] S1. Acquire drone imagery datasets, preprocess them, and construct standardized datasets;

[0049] S2. Construct an improved YOLO model (GSA-YOLO model), which consists of a backbone network, a neck network, and a detection head; the backbone network integrates multi-scale dynamic gating fusion units, and the detection head contains a branch specifically for small-scale target detection;

[0050] S3. Based on the standardized dataset, the improved YOLO model is trained using a multi-scale weighted comprehensive loss function;

[0051] S4. Use the trained model to perform target detection on UAV imagery.

[0052] In one specific embodiment of the present invention, the drone image dataset in step S1 uses the VisDrone2019 drone perspective dataset downloaded from a public data source. This dataset includes drone image data and corresponding annotation files, covering drone aerial photography scenes from multiple angles, at different altitudes, and under various lighting conditions. The dataset comprises three subsets: a training set, a validation set, and a test set. The steps for processing the dataset are as follows:

[0053] Perform annotation format conversion: Convert the original annotation files of the VisDrone2019 dataset (rectangular bounding box annotations stored in .txt or .xml format) into a unified annotation format compatible with the YOLO series detection framework. During this conversion, the target center coordinates in the annotation data are... and target width and height parameters Normalization is performed separately to map the labeled values ​​to the range [0,1], thereby improving the input consistency and stability of the model. The calculation formula is as follows:

[0054]

[0055] in, and These are the original width and height of the image, respectively. These are the normalized coordinates and dimensions.

[0056] After completing the format conversion and parameter normalization, the training set, validation set and test set are divided into three sets in an 8:1:1 ratio. The converted annotation files are matched one-to-one with the corresponding image data to generate data list files that can be directly called by the model training module (training list file train.txt / validation list file val.txt / test list file test.txt).

[0057] Deploy a data preprocessing pipeline: Develop and configure batch processing scripts to perform batch format conversion and parameter normalization of raw data; synchronously build a multi-threaded data loader with asynchronous data prefetching capabilities to ensure efficient reading, caching, and transmission of data streams during model training, guaranteeing the real-time performance and stability of the training process.

[0058] In the data preprocessing stage, a variety of enhancement strategies can be introduced, including random flipping, scale transformation, brightness perturbation and color jitter, to enhance the model’s robustness to complex environments and images of different resolutions.

[0059] In step S2, this embodiment first designs a multi-scale dynamic gated fusion unit (MDGF) to improve the model's feature extraction capability under multi-scale traffic targets. The internal structure of the MDGF unit includes two parallel branches: multi-scale dynamic attention (MDTA) and gated deep feedforward network (GDFN). Its input features are convolved and mapped into the above two branches respectively, and finally output according to the fusion strategy.

[0060] Let the input features be... The MDTA branch generates query Q, key K, and value V through convolutional mapping:

[0061]

[0062] in, Depend on and The mapping tensor generated by convolution.

[0063] Expand the tensor into a sequence, flatten it along the spatial dimension, and aggregate features based on attention weights. The output format is as follows:

[0064]

[0065] in, The scaling factor for the key vector dimension. Tensor The transpose operation, This represents the tensor reconstruction process, used to restore the product result to the spatial dimension of the original input. This indicates normalization along appropriate dimensions, followed by summing of the residuals. (Input feature tensor). This structure can dynamically capture multi-scale contextual information.

[0066] The GDFN branch employs a gating mechanism to enhance the nonlinear representation of inter-channel features. It first obtains two activations through channel expansion mapping, and then merges them using gating. The calculation method is as follows:

[0067]

[0068] in, Let be the convolution mapping matrix, ⊙ denotes element-wise multiplication, and GELU be the non-linear activation function. This module achieves dynamic adjustment within channels while maintaining feature stability.

[0069] For the same input features Two outputs are obtained The fusion layer achieves feature integration using either splicing compression or element-wise addition.

[0070]

[0071] And introduce a channel adaptive gating function:

[0072] ,

[0073] Final output:

[0074]

[0075] By adaptively adjusting the gating parameter g, the model can dynamically filter multi-scale features, achieving robust feature modeling in complex scenarios.

[0076] The fusion layer should ensure the following during implementation: , Output channels are consistent or pass Convolutional alignment; the implementation of MDGF must balance parameter count and real-time performance, which can be reduced by depthwise separable convolutions or channel compression. This invention constructs an improved YOLO model (GSA-YOLO model) by building an improved backbone network to replace the traditional YOLOv11n backbone. Specifically, all C3k2 modules in the original backbone network are uniformly replaced with C3k2_MDGF modules, retaining the original input / output channels and tensor topology to ensure that the overall channel structure and input / output dimensions are consistent with the original model. After the replacement, the network remains compatible with subsequent feature pyramid modules (SPPF, C2PSA).

[0077] This structure significantly enhances the backbone's ability to distinguish features in small-scale targets and complex backgrounds, and improves the global modeling capability of deep semantic features.

[0078] Then, a multi-scale detection head extension structure was designed to improve the small target detection performance from the UAV's perspective. Based on the original YOLO detection head, a small target branch P1 was added, expanding the output layer from a three-scale to a four-scale structure, corresponding to multi-scale feature maps P1, P2, P3, and P4, which are used to detect traffic targets of different sizes. The output includes class confidence and bounding box regression results. Details are as follows:

[0079] Based on the given head structure, this invention adopts a four-scale input Detect layer: from node numbers or corresponding semantics P1, P2, P3, P4.

[0080] The Detect layer accepts four inputs. After convolution adjustment and fusion, features at each scale output predicted bounding box positions, target confidence, and class probabilities for small, medium, and large targets respectively, achieving accurate localization of targets at multiple scales.

[0081] In step S3, a multi-scale weighted loss function and training strategy are designed. The specific details are as follows:

[0082] The total loss function consists of classification loss, bounding box regression loss, and confidence loss, and its combined expression is:

[0083]

[0084] Here, P1, P2, P3, and P4 represent four detection branches at different scales, with the resolution decreasing sequentially. is the scale weight coefficient of the s-th branch. Larger weights can be set for branches with small targets to improve detection performance.

[0085] Classification loss Focal Loss or weighted cross-entropy can be used to mitigate class imbalance.

[0086] The regression loss is expressed in the form of CIoU / DIoU to improve regression accuracy, and is defined as follows:

[0087]

[0088] Where ρ is the distance from the center point, c is the length of the diagonal of the circumscribed rectangle, and v is the aspect ratio.

[0089] The confidence loss is calculated using binary cross-entropy or binary cross-entropy with a balance term (BCE).

[0090] During the training phase, the AdamW optimizer was used in combination with a cosine annealing learning rate strategy. The batch size was 8, the training epochs were 300, and an early stopping mechanism was introduced to prevent overfitting. The input image size was 640×640. Mosaic data augmentation, brightness perturbation, and scale transformation operations were used to enhance the model's generalization ability.

[0091] The model training and performance validation process is as follows:

[0092] The model was trained using the preprocessed VisDrone2019 dataset, and the training parameters are shown in Table 1. During the training phase, the model recorded the loss curve and saved the optimal weight file.

[0093] Table 1 Experimental Parameter Settings

[0094] Parameter Value Training rounds 300 Input image size 640×640 Batch 8 Optimizer SGD Initial learning rate 0.01 Weight decay 0. 0005 Momentum 0.9

[0095] Performance verification includes accuracy metrics, ablation experiments, and comparative experiments:

[0096] Accuracy metrics were calculated by taking the average mAP at IoU thresholds of 0.50:0.05:0.95, and further subdividing the mAP by target size (small target <32×32, medium target 32×32-96×96, large target >96×96). Ablation experiments were conducted to verify the independent contributions of MDTA, GDFN, and the newly added small target detector. Comparative experiments were conducted using YOLOv11n, YOLOv10n, YOLOv8n, and YOLOv5n models as baselines to compare the performance of mAP50 and mAP50–95. The results are shown in Tables 2 and 3.

[0097] The results show that on the VisDrone2019 dataset, the proposed GSA-YOLO model achieves the highest mAP50-95. Compared with the YOLOv11n baseline model, it improves the accuracy metric mAP@0.5 by about 3.9%, the mAP@0.5:0.95 ratio by about 3.2%, and the recall rate by about 3.8%. It significantly improves the small target detection capability and feature representation accuracy while maintaining lightweight and real-time performance.

[0098] Table 2 Ablation Experiment

[0099] YOLOv11-N MDTA GDFN SmallHead Box(P) R mAP50 mAP50-95 √ 0.465 0.337 0.349 0.203 √ √ 0.458 0.342 0.346 0.203 √ √ 0.465 0.36 0.36 0.215 √ √ √ 0.477 0.351 0.361 0.216 √ √ 0.462 0.359 0.364 0.219 √ √ √ √ 0.5 0.375 0.388 0.235

[0100] Table 3 Comparative Experiments

[0101] Model Box(P) R mAP50 mAP50-95 yolov11-n 0.465 0.337 0.349 0.203 yolov10-n 0.458 0.341 0.349 0.204 yolov9-t 0.486 0.346 0.356 0.21 yolov8-n 0.474 0.341 0.35 0.205 yolov5-n 0.449 0.344 0.341 0.198 GSA-YOLO 0.5 0.375 0.388 0.235

[0102] This invention's multi-scale dynamic gating fusion structure achieves dynamic fusion of global features and local details by establishing an adaptive weight allocation mechanism among features at different scales. While maintaining low computational complexity and fast inference speed, it significantly improves the detection performance of small-sized targets, distant targets, and targets with weak textures. This structure can be directly embedded into lightweight convolutional neural networks or multi-branch deep detection models, exhibiting good versatility and portability. It can be adapted to embedded inference chips, edge computing devices, and UAV onboard computing platforms to achieve high-precision, low-latency intelligent visual detection, thus having broad practical value and promising prospects in various UAV missions and real-time monitoring applications.

[0103] The improved backbone network module C3k2_MDGF of this invention is fully compatible with the YOLOv11 framework structure at the system interface layer and data flow transmission layer. It can directly replace the original YOLOv11 backbone network module without modifying the number of model input / output channels, feature mapping size, and detector head configuration. The replacement process maintains the consistency of the overall network topology and parameter mapping relationship, and can achieve seamless migration in the model configuration file by replacing modules, thereby ensuring complete consistency of the training process, feature transmission path, and inference results.

[0104] The improved backbone network module C3k2_MDGF introduces a multi-scale dynamic gating fusion mechanism in its structure, which enhances the global context modeling capability and multi-scale feature perception capability while maintaining the lightweight characteristics of the original YOLOv11n model. It can significantly improve the detection accuracy and stability of the model in complex traffic scenes, dense areas of small targets, and long-distance view conditions. It can also be combined with end-to-end optimized training strategies to achieve rapid transfer learning and adaptive parameter adjustment, thereby enabling efficient inference and real-time deployment on resource-constrained embedded devices.

[0105] Those skilled in the art should understand that embodiments of the present invention can be implemented in various forms, including but not limited to hardware implementations, software implementations, or hybrid implementations combining hardware and software. For example, the improved YOLO model based on multi-scale dynamic gating fusion (GSA-YOLO) can be implemented programmatically on a deep learning platform equipped with a graphics processing unit (GPU), or deployed in edge computing devices, intelligent transportation terminals, or unmanned aerial vehicle (UAV) computing modules to achieve real-time traffic target detection and behavior analysis. The technical solution of the present invention achieves consistent technical effects and performance advantages under different computing architectures and operating environments.

[0106] This invention is described in conjunction with specific embodiments of the structure, method, and process. However, it should be understood that these embodiments are only used to illustrate the core ideas of the invention and not to limit the technical solutions of the invention. For specific modules involved in this invention (such as the MDTA module, GDFN module, C3k2_MDGF unit, etc.), their internal structure, connection method, parameter configuration, and calculation strategy can be flexibly adjusted according to different application scenarios. For example, the attention window size, channel compression ratio, gating weight parameters, or multi-scale loss weights can be appropriately modified based on the UAV's flight altitude, lens resolution, target density, or scene complexity to achieve optimal detection performance under different environmental conditions.

[0107] The multi-scale dynamic gating fusion structure proposed in this invention is not only applicable to target detection tasks on the VisDrone2019 UAV traffic dataset, but can also be widely applied to other multi-scale target detection and intelligent perception fields, such as urban road monitoring, traffic flow statistics, pedestrian density estimation, intelligent security monitoring, aerial infrastructure inspection, and port and airport operation monitoring. In these applications, the feature fusion and attention gating mechanism proposed in this invention can significantly improve the robustness of the model in small target recognition, long-distance detection, and complex backgrounds, thereby providing efficient and reliable visual detection support for intelligent transportation systems, smart cities, and autonomous perception of UAVs.

[0108] In summary, the above embodiments are merely preferred embodiments of the present invention. Although the present invention has been described through preferred embodiments, this does not limit the scope of protection of the present invention. For those skilled in the art, any equivalent substitutions, modifications, or improvements made to its structural form, network configuration, algorithm flow, or parameter settings without departing from the spirit and essence of the present invention should be considered to fall within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the appended claims.

Claims

1. A small target detection method based on UAV vision, characterized in that, The specific steps include: S1. Acquire drone imagery datasets, preprocess them, and construct standardized datasets; S2. Construct an improved YOLO model, which includes a backbone network, a neck network, and a detection head in sequence; the backbone network integrates multi-scale dynamic gating fusion units, and the detection head contains a branch specifically for small-scale target detection; S3. Based on the standardized dataset, the improved YOLO model is trained using a multi-scale weighted comprehensive loss function; S4. Use the trained model to perform target detection on UAV imagery; The multi-scale dynamic gating fusion unit in the backbone network is the C3k2_MDGF unit, which is implemented by replacing the C3k2 unit in the original backbone network. The C3k2_MDGF unit includes a parallel multi-scale dynamic attention (MDTA) branch and a gated depth feedforward (GDFN) branch, as well as a fusion layer that fuses the outputs of these two branches. The fusion layer achieves feature integration using either splicing compression or element-by-element addition. ， And introduce a channel adaptive gating function: ， Final output: ， Where g represents the gating parameter, The activation function is Sigmoid, and GAP represents global average pooling. This is the channel gating weight matrix; The detection head includes four scale detection branches based on feature maps of four different resolutions. The first scale branch has the highest spatial resolution and is used specifically for detecting small-scale targets.

2. The small target detection method based on UAV vision according to claim 1, characterized in that, The preprocessing in step S1 includes: obtaining the annotation file of the corresponding image data, unifying the annotation file format to a YOLO compatible format, and normalizing the target center coordinates and width and height parameters; dividing the dataset into training set, validation set and test set according to the proportion; and performing data augmentation operations on the dataset, including random flipping, scale transformation, brightness perturbation and color jitter.

3. The small target detection method based on UAV vision according to claim 1, characterized in that, The MDTA branch is for input features Output Calculated using the following formula: ， in, These are the mapping tensors generated by 1×1 and 3×3 convolutions, respectively. This represents the scaling factor for the key vector dimension. Tensor The transpose operation, This represents the tensor reconstruction process, used to restore the product result to the spatial dimension of the original input. This indicates normalization on the appropriate dimension.

4. The small target detection method based on UAV vision according to claim 1, characterized in that, The GDFN branch is for input features Output Calculated using the following formula: ， in, The convolution mapping matrix, represents element-wise multiplication, and GELU represents a non-linear activation function.

5. The small target detection method based on UAV vision according to claim 1, characterized in that, The multi-scale weighted comprehensive loss function is defined as follows: ， Here, P1, P2, P3, and P4 represent four detection branches at different scales, with the resolution decreasing sequentially. Let be the scale weight coefficient of the s-th branch, and assign the maximum weight to the small target branch P1; The classification loss is Focal Loss; The bounding box regression loss uses Complete IoU; The confidence loss is calculated using binary cross-entropy; This is the loss weighting coefficient.

6. A small target detection system based on UAV vision, characterized in that, To implement the method of any one of claims 1-5, comprising: Data processing module: used to acquire UAV imagery datasets, perform preprocessing, and construct standardized datasets; Model building module: used to build an improved YOLO model, which includes a backbone network, a neck network and a detection head in sequence; the backbone network integrates multi-scale dynamic gating fusion units, and the detection head contains a branch specifically for small-scale target detection; Model training module: used to train the improved YOLO model based on the standardized dataset using a multi-scale weighted comprehensive loss function; Target detection module: Used to perform target detection on UAV imagery using a trained model.

Images