Dynamic adaptive feature balance method and system for small target detection of unmanned aerial vehicle

Abstract

The present application relates to the technical field of unmanned aerial vehicle aerial photography, and particularly relates to a dynamic adaptive feature balancing method and system for small target detection of unmanned aerial vehicles. The method comprises: acquiring an unmanned aerial vehicle aerial image; performing target detection according to the acquired unmanned aerial vehicle aerial image and a preset target detection model to obtain a target detection result, thereby completing small target detection and recognition of the unmanned aerial vehicle image; wherein the target detection model is improved based on a YOLOv11-nano model, a dynamic spatial semantic decoupling unit is embedded in a C3k2 unit of a backbone network of the YOLOv11-nano model, and an adaptive multi-scale perception fusion unit is used to replace a final down-sampling layer in a neck network model of the YOLOv11-nano. The present application realizes adaptive balancing of spatial-semantic feature flow through cooperation of the adaptive multi-scale perception fusion unit and the dynamic spatial semantic decoupling unit, thereby significantly improving the detection accuracy of aerial small targets.

Description

Technical Field

[0001] This invention relates to the field of drone aerial photography technology, and in particular to a dynamic adaptive feature balancing method and system for small target detection by drones. Background Technology

[0002] Aerial image target detection is a core technology of UAV vision systems, which faces severe challenges such as extremely small target size, complex background interference, and limited onboard computing power.

[0003] While general-purpose detectors, such as the YOLO series, perform excellently in natural scenes, their current network design presents a fundamental contradiction between spatial positioning accuracy and semantic discriminative ability in aerial photography. The shallow layers retain the high-resolution spatial features necessary for small target localization but lack discriminative semantic information; the deeper layers extract robust contextual semantic features, but at the cost of spatial granularity, failing to support accurate localization of small targets. Although existing technologies attempt to alleviate this contradiction through cross-scale fusion of feature pyramids, they still employ fixed feature fusion and channel allocation strategies throughout the detection process, leading to a misalignment between spatial and semantic features, ultimately resulting in missed detections and localization errors for small aerial targets. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a dynamic adaptive feature balancing method and system for small target detection on UAVs. Inspired by the dynamic resource allocation and functional differentiation mechanism of the mammalian visual cortex, it achieves adaptive balancing of spatial-semantic feature flows through the cooperation of an adaptive multi-scale perception fusion unit and a dynamic spatial semantic decoupling unit, significantly improving the detection accuracy of small targets in aerial photography while maintaining excellent real-time inference performance, and is suitable for deployment of UAV-borne edge devices.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a dynamic adaptive feature balancing method for small target detection using unmanned aerial vehicles (UAVs) is provided, including: Acquire drone aerial images; Target detection is performed based on the acquired drone aerial images and the preset target detection model to obtain the target detection results, thus completing the small target detection and recognition of drone images; The target detection model employs an improved YOLOv11-nano model. A dynamic spatial-semantic decoupling unit is embedded in the C3k2 unit of the YOLOv11-nano model's backbone network, and an adaptive multi-scale perceptual fusion unit replaces the final downsampling layer in the neck network of the YOLOv11-nano model. After inputting the UAV aerial image into the backbone network, the dynamic spatial-semantic decoupling unit adaptively allocates the channel ratio between spatial and semantic branches, outputting a spatially and semantically aligned feature map. This feature map is then input into the neck network, where the adaptive multi-scale perceptual fusion unit performs adaptive fusion of multi-scale features, ultimately inputting the detection head to generate the detection result.

[0006] Secondly, a dynamic adaptive feature balancing system for small target detection on UAVs is provided, including: The acquisition module is configured to acquire aerial images taken by the drone. The detection module is configured to perform target detection based on the acquired UAV aerial images and a preset target detection model, obtain the target detection results, and complete the small target detection and recognition of UAV images. The target detection model employs an improved YOLOv11-nano model. A dynamic spatial-semantic decoupling unit is embedded in the C3k2 unit of the YOLOv11-nano model's backbone network, and an adaptive multi-scale perceptual fusion unit replaces the final downsampling layer in the neck network of the YOLOv11-nano model. After inputting the UAV aerial image into the backbone network, the dynamic spatial-semantic decoupling unit adaptively allocates the channel ratio between spatial and semantic branches, outputting a spatially and semantically aligned feature map. This feature map is then input into the neck network, where the adaptive multi-scale perceptual fusion unit performs adaptive fusion of multi-scale features, ultimately inputting the detection head to generate the detection result.

[0007] Thirdly, an electronic device is also provided, comprising: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0008] Fourthly, a computer-readable storage medium is provided having a program stored thereon that, when executed by a processor, implements the method described in the first aspect above.

[0009] Fifthly, a computer program product is provided, employing the following technical solution: A computer program product includes software code, wherein a program in the software code performs the steps of the method described in the first aspect of the present invention.

[0010] The above technical solution has the following advantages or beneficial effects: This invention designs a Dynamic Spatial Semantic Decoupling Unit (DSSD) to simulate the functional differentiation and dynamic resource allocation mechanism of the mammalian visual cortex, using end-to-end learnable channel segmentation parameters. This achieves a dynamic balance between spatial and semantic features at different network levels, fundamentally solving the problem of spatial-semantic feature misalignment caused by static network design, and significantly improving the positioning accuracy and classification accuracy of small aerial targets.

[0011] The adaptive multi-scale perception fusion unit designed in this invention simulates the multi-scale receptive field collaboration mechanism of the visual system. It replaces the serial multi-core structure with parallel dilated convolution branches, which reduces computational complexity while preserving the multi-scale receptive field. By using intra-branch channel attention and globally learnable fusion weights, it dynamically strengthens the scale features that are most discriminative for small targets, thereby improving the model's ability to perceive extremely small targets in aerial photography with extremely low computational overhead. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0013] Figure 1 This is a flowchart of a dynamic adaptive feature balancing method for small target detection by UAVs in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the improved YOLOv11-nano model in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the C3k2 unit of the backbone network embedded in the Dynamic Spatial Semantic Decoupling Unit (DSSD) in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the neck network structure after replacing the final downsampling layer in the neck network with an adaptive multi-scale sensing fusion unit in a specific embodiment of the present invention. Detailed Implementation

[0014] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0015] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0019] Example 1 This embodiment provides a dynamic adaptive feature balancing method for small target detection by UAVs.

[0020] like Figure 1 As shown, the dynamic adaptive feature balancing method for small target detection by UAVs according to an embodiment of the present invention includes the following steps: S1: Acquire aerial images of the object to be inspected taken by the drone.

[0021] S2: Build and train the object detection model.

[0022] S3: After the acquired UAV aerial images are input into the backbone network of the preset target detection model, the channel ratio of spatial and semantic branches is adaptively allocated through the dynamic spatial-semantic decoupling unit, and the spatial and semantic aligned feature map is output. Then the feature map is input into the neck network, and the adaptive multi-scale perception fusion unit completes the adaptive fusion of multi-scale features. Finally, the detection head is input to generate the detection result, thus completing the small target detection and recognition of UAV images.

[0023] In some embodiments, step S1 specifically involves: acquiring an RGB image taken by a drone, scaling the image to the network input resolution, normalizing the pixel values ​​to the [0,1] range, and simultaneously performing preprocessing operations such as image denoising and format conversion to generate a network input tensor. In this embodiment, the image is scaled to 640×640 or 1280×1280.

[0024] In this embodiment, the target detection model constructed in step S2 is an improvement on the YOLOv11-nano model. The original YOLOv11-nano model consists of three parts: a backbone network, a neck network, and a head network. Step S2 specifically includes: S2.1: Embedding the Dynamic Spatial Semantic Decoupling Unit (DSSD) into the four C3k2 units of the backbone network, corresponding to the four core stages of the network; S2.2: Replacing the final downsampling layer in the neck network with the Adaptive Multi-Scale Perceptual Fusion Unit (AMPF); S2.3: Training the improved model.

[0025] In this embodiment, the improved model is referred to as the DynBalance-YOLO network, and its overall structure is as follows: Figure 2 As shown.

[0026] In this embodiment, S2.1: The Dynamic Spatial Semantic Decoupling Unit (DSSD) is embedded into the four C3k2 units of the backbone network. Specifically, as follows: Figure 3 As shown, the DSSD unit is embedded in the main branch after the Split operation of each C3k2 unit, replacing the original residual module sequence. After the input features are convolved with 1×1 and then split, the main branch enters the DSSD unit for spatial-semantic dynamic decoupling; short-path edges maintain identity mapping, and short-path edge features are output. The DSSD unit uses a learnable channel segmentation ratio parameter. Features are dynamically segmented into semantic and spatial branches, which are then processed by 3×3 and 1×1 convolutions, respectively. A cross-path attention mechanism is then used to achieve bidirectional feature relabeling and fusion. The output of DSSD is concatenated with the short-path side features and finally output via a 1×1 convolution.

[0027] The specific execution steps of the Dynamic Spatial Semantic Decoupling Unit (DSSD) are as follows: S2.1.1: For the input feature map Introduce independent, learnable scalar parameters for each stage of the backbone network. By mapping to the 0-1 interval using the sigmoid function, the learnable channel segmentation ratio parameter for the i-th stage can be obtained. This allows the network to automatically learn the optimal channel allocation ratio at different levels through gradient descent, simulating the dynamic resource allocation mechanism of a vision system.

[0028] S2.1.2: Based on Calculate the channel scaling coefficients for semantic and spatial branches to ensure that the feature magnitudes of the two branches remain consistent, thus avoiding gradient imbalance during training.

[0029] in, The number of channels allocated to semantic branches, The number of channels allocated to the spatial branches. This is the scaling factor for the semantic branch channel; Channel branch channel scaling factor Subsequently, soft segmentation of features is achieved through mask scaling along the channel dimension. This operation is fully differentiable, avoiding the gradient interruption problem caused by hard channel routing. (2); in, Semantic branch features; It is a spatial branching feature. and For the channel mask, the values ​​for the corresponding channel positions are as follows: and , ⊙ represents element-wise multiplication.

[0030] S2.1.3: Design hierarchy Association regularization terms constrain adjacent levels Differences are simulated to create a smooth transition mechanism between different levels of the visual cortex, avoiding discrepancies between different levels. Excessive differentiation ensures smooth feature propagation:

[0031] in, Let N be the channel segmentation ratio in the i-th stage after learning, N be the total number of network stages, and the regularization term weight be... The initial setting is 1e-4, when adjacent levels When the difference is greater than 0.5, the weight is increased to 1e-3 to avoid excessive constraints affecting the network's adaptive learning ability.

[0032] S2.1.4: Based on learned... Dynamic channel pruning is performed, adaptively adjusting the total number of channels at each level to simulate the dynamic suppression mechanism of the vision system for redundant resources and reduce computational redundancy.

[0033] Set channel retention rate , and The relationship is: ; in, This is the cutting factor, which is set to 0.3 by default. The maximum cutting ratio is 30%, but those skilled in the art can also set it according to their actual needs.

[0034] based on The total number of channels at this level has been adjusted as follows:

[0035] Total number of channels after adjustment at each level Based on this, the number of channels for semantic and spatial branches is reallocated. This operation is achieved through reparameterization and is performed only after training is complete, without affecting the end-to-end training process of the model. The purpose of this step is to first suppress redundant channels, reduce the overall computational cost, and then flexibly adjust the ratio of spatial to semantic distribution within the channels.

[0036] S2.1.5: The features of the segmented semantic and spatial branches are processed in parallel. The feature interaction between the two branches is achieved through a cross-path attention mechanism. Finally, the feature map of the DSSD unit is fused and output. The specific steps are as follows: S2.1.5.1: Semantic Branch Pairs Applying 3×3 convolution enhances feature interactions between channels, extracts high-level contextual semantic features, and outputs... This corresponds to the complex cellular functions of the visual cortex; spatial branching pairs Applying 1×1 pointwise convolution maximizes the preservation of fine-grained spatial location information and avoids the loss of spatial details caused by convolution operations, outputting... These correspond to simple cellular functions of the visual cortex; S2.1.5.2: Output from semantic branches Generate channel attention map This is achieved through global average pooling and fully connected layers with sigmoid activation, used to recalibrate features of spatial branches, allowing semantic information to guide the enhancement of spatial details.

[0037]

[0038] S2.1.5.3: Simultaneously, output from the spatial branch. Generate spatial attention map This is achieved through lightweight convolutional blocks and sigmoid activation, used to modulate features of semantic branches, allowing spatial location information to guide the focus of semantic features:

[0039]

[0040] S2.1.5.4: Add the features of the two branches that have been recalibrated through cross-attention element-wise to obtain the final output of the DSSD unit. + This enables bidirectional guidance between semantic context and spatial details, alleviating the problem of feature misalignment.

[0041] In this embodiment, S2.2: The execution step of replacing the final downsampling layer in the neck network with an adaptive multi-scale perceptual fusion unit AMPF is as follows: Figure 4 As shown, the Adaptive Multi-Scale Perceptual Fusion Unit (AMPF) receives feature input from the previous layer of the neck network and extracts multi-scale features through three parallel dilated convolutional branches (dilation rates d=1,2,3). Each branch is equipped with channel attention (CA) for channel recalibration. Subsequently, it undergoes global average pooling, descriptor concatenation, and MLP+Softmax to generate normalized fusion weights. Finally, the multi-branch features are weighted and summed for output. The specific execution steps are as follows: S2.2.1: Construct multiple parallel convolutional branches. In this embodiment, three parallel convolutional branches are constructed, employing 3×3 dilated convolutions with dilation rates d=1, d=2, and d=3 respectively. These correspond to equivalent receptive fields of 3×3, 5×5, and 7×7, covering the full scale range from extremely small targets to large targets in aerial photography scenes. This applies to the feature inputs of the previous layer in the neck network. Parallel processing is performed to generate multi-scale features corresponding to the receptive field:

[0042] in, The dilated convolution operator representing the i-th branch, compared to the standard large kernel convolution, can significantly reduce the number of parameters and computational complexity while keeping the receptive field unchanged, and avoids the continuous loss of spatial details caused by serial convolution.

[0043] S2.2.2: Equip each parallel branch with a lightweight channel attention (CA) submodule to simulate the selective enhancement mechanism of the vision system for key features, strengthen the channel features in each branch that are discriminative for small targets, and suppress the interference of background noise; for the output features of the i-th branch The channel attention module generates a weight vector along the channel dimension: :

[0044] in, and For trainable fully connected layer weights, It is the sigmoid activation function.

[0045] Subsequently, the branch features are recalibrated channel by channel using the generated weight vector:

[0046] S2.2.3: Based on global feature description, adaptive attention-weighted fusion is implemented, simulating the dynamic weighting mechanism of the visual system on the outputs of neurons in different receptive fields. This allows the network to automatically adjust the fusion weights of different receptive field branches according to the scale distribution of the target in the input image. First, the recalibrated features of each branch are processed. Perform global average pooling to extract global feature descriptors:

[0047] Subsequently, the global descriptors of all branches are concatenated, and a set of normalized fusion weights is generated through two layers of MLP and softmax activation. The sum of all weights is 1; finally, the output of the AMPF unit is obtained by weighted summation of the multi-branch features according to the fusion weights:

[0048] The specific steps for training the improved model in S2.3 are as follows: During the network training phase, end-to-end joint optimization is used to learn the channel segmentation parameters of the DSSD unit and the multi-scale fusion weights of the AMPF unit, thereby achieving network convergence. The objective function of the end-to-end joint optimization is:

[0049] in, Represents the network's regular trainable parameters. The channel segmentation parameters represent the network's learning parameters after N stages. Let D be the detection loss function, and D be the training dataset. For the inter-level Association regular terms, This is the weight coefficient for the regularization term.

[0050] In some embodiments, the specific parameters during the training phase are set as follows: the backbone network is initialized with ImageNet pre-trained weights, the basic input resolution is 640×640, and high-resolution training uses 1280×1280 input; the SGD optimizer is used, with momentum set to 0.937, weight decay to 5×10^-4, initial learning rate of 0.01, cosine annealing strategy is used to adjust the learning rate, batch size is 32, and training epochs are 300; standard data augmentation strategies such as Mosaic, MixUp, random flipping, scaling, and color jitter are used during training; the learnable parameters of the DSSD unit are... Initialized to 0.5, hierarchical association regularization weight. The initial setting is 1e-4, and the maximum clipping ratio for dynamic channel clipping is set to 30%.

[0051] The specific implementation steps of S3 are as follows: The preprocessed image to be detected is input into the trained DynBalance-YOLO network. The DSSD unit dynamically balances the spatial and semantic features across different network layers, while the AMPF unit performs adaptive perceptual fusion of multi-scale features. Specifically, the preprocessed image tensor is input into the network and sequentially passes through four stages of the backbone network. At each stage, the DSSD unit adaptively allocates the channel ratio between spatial and semantic branches, outputting spatially and semantically aligned feature maps. These feature maps are then input into the Neck structure, where the AMPF unit performs adaptive fusion of multi-scale features, finally inputting the detection head to generate the initial detection results.

[0052] The network outputs the detection results, and after non-maximum suppression, outputs the target category, confidence score, and bounding box coordinates in the aerial image. In this embodiment, the initial detection boxes output by the detection head are subjected to non-maximum suppression with an intersection-over-union (IoU) threshold of 0.5 to remove redundant duplicate detection boxes. Finally, the class label, classification confidence score, and precise bounding box coordinates of each target in the aerial image are output, completing the detection of small targets in aerial photography.

[0053] Experimental verification To verify the effectiveness of this invention, this embodiment conducted experimental verification on three major aerial target detection benchmark datasets: VisDrone2019-DET, UAVDT, and AI-TOD. All experiments were completed under the same hardware and software environment. The desktop testing environment consisted of an NVIDIA GeForce RTX4090 GPU, PyTorch 2.0.1, and CUDA 11.8, while the edge device testing environment consisted of NVIDIA Jetson Xavier NX and Jetson Orin NX. TensorRT 8.5 was used for quantization optimization.

[0054] Experimental results show that the proposed DynBalance-YOLO, using the YOLOv11-nano model as a baseline, achieves 35.8% mAP@0.5 on the VisDrone dataset at a 640×640 input resolution, an improvement of 8.2 percentage points over the baseline, with small target AP_S increasing from 8.7% to 14.2%, while maintaining a real-time inference speed of 162 FPS. On the UAVDT dataset, mAP@0.5 reaches 38.7%, an improvement of 7.5 percentage points over the baseline; on the AI-TOD dataset, mAP@0.5 reaches 36.1%, an improvement of 7.7 percentage points over the baseline, significantly outperforming all lightweight state-of-the-art (SOTA) models of the same scale. At a high resolution of 1280×1280, this method still maintains 38.7% mAP@0.5 and an inference speed of 78 FPS on the VisDrone dataset; on Jetson OrinNX edge devices, the inference speed after INT8 quantization can reach 68 FPS, fully meeting the real-time detection requirements of UAVs. In cross-dataset generalization experiments, this method was tested directly on UAVDT and AI-TOD after training on VisDrone. The mAP@0.5 improved by 6.8 and 6.1 percentage points respectively compared with the baseline, demonstrating strong generalization ability and robustness.

[0055] Example 2 This embodiment provides a dynamic adaptive feature balancing system for small target detection using unmanned aerial vehicles (UAVs), including: The acquisition module is configured to acquire aerial images taken by the drone. The detection module is configured to perform target detection based on the acquired UAV aerial images and a preset target detection model, obtain the target detection results, and complete the small target detection and recognition of UAV images. The target detection model employs an improved YOLOv11-nano model. A dynamic spatial-semantic decoupling unit is embedded in the C3k2 unit of the YOLOv11-nano model's backbone network, and an adaptive multi-scale perceptual fusion unit replaces the final downsampling layer in the neck network of the YOLOv11-nano model. After inputting the UAV aerial image into the backbone network, the dynamic spatial-semantic decoupling unit adaptively allocates the channel ratio between spatial and semantic branches, outputting a feature map aligned with both spatial and semantic aspects. This feature map is then input into the neck network, where the adaptive multi-scale perceptual fusion unit performs adaptive fusion of multi-scale features, ultimately inputting the detection head to generate the detection result.

[0056] The detailed steps are the same as those of the dynamic adaptive feature balancing method for small target detection of UAVs provided in Example 1, and will not be repeated here.

[0057] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0058] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0059] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0060] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0061] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0062] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0063] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0064] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0065] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0066] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the method as described in Embodiment 1 of the present invention.

[0067] The detailed steps are the same as those provided in Example 1, and will not be repeated here.

[0068] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0069] A computer program product includes software code, wherein the program in the software code performs the steps described in Embodiment 1 of the present invention.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic adaptive feature balancing method for small target detection using unmanned aerial vehicles (UAVs), characterized in that, include: Acquire drone aerial images; Target detection is performed based on the acquired drone aerial images and the preset target detection model to obtain the target detection results, thus completing the small target detection and recognition of drone images; The target detection model employs an improved YOLOv11-nano model. A dynamic spatial semantic decoupling unit is embedded in the C3k2 unit of the YOLOv11-nano backbone network, and an adaptive multi-scale perceptual fusion unit replaces the final downsampling layer in the neck network model of YOLOv11-nano. After inputting the UAV aerial image into the backbone network, the dynamic spatial semantic decoupling unit adaptively allocates the channel ratio between spatial and semantic branches, outputting a feature map aligned with both spatial and semantic aspects. This feature map is then input into the neck network, where the adaptive multi-scale perceptual fusion unit performs adaptive fusion of multi-scale features, ultimately inputting the detection head to generate the detection result.

2. The dynamic adaptive feature balancing method for small target detection of UAVs as described in claim 1, characterized in that, The embedding of a dynamic spatial semantic decoupling unit into the C3k2 unit of the backbone network of the YOLOv11-nano model specifically involves: Embed the DSSD unit on the main branch after the Split operation of each C3k2 unit; After the input features are processed by 1×1 convolution and Split operation, the main branch enters the DSSD unit for spatial-semantic dynamic decoupling; the short-path edges maintain the identity mapping and output the short-path edge features. The DSSD unit uses a learnable channel segmentation ratio parameter. Features are dynamically segmented into semantic branches and spatial branches, which are then processed by 3×3 convolution and 1×1 convolution respectively. Finally, bidirectional feature recalibration and fusion are achieved through a cross-path attention mechanism. The output of DSSD is concatenated with the short-path features, and finally the feature map of the DSSD unit is output by 1×1 convolution.

3. The dynamic adaptive feature balancing method for small target detection in UAVs as described in claim 2, characterized in that, The learnable channel segmentation ratio parameter The steps for obtaining the input feature map are as follows: Introduce independent, learnable scalar parameters for each stage of the backbone network. The channel segmentation ratio for the i-th stage is obtained by mapping using the sigmoid function. .

4. The dynamic adaptive feature balancing method for small target detection by UAVs as described in claim 1, characterized in that, The method of replacing the final downsampling layer in the neck network model of YOLOv11-nano with an adaptive multi-scale perceptual fusion unit is as follows: Multiple parallel convolutional branches are constructed to process the input feature map in parallel and generate multi-scale features; Each parallel branch is equipped with a lightweight channel attention submodule to recalibrate the branch features channel by channel. Global average pooling is performed on the recalibrated features of each branch to extract global feature descriptors; Concatenate the global descriptors of all branches to generate the fusion weights; The multi-branch features are weighted and summed according to the fusion weights to output the final feature map of the adaptive multi-scale perceptual fusion unit.

5. The dynamic adaptive feature balancing method for small target detection in UAVs as described in claim 4, characterized in that, The construction of multiple parallel convolutional branches to process the input feature map in parallel and generate multi-scale features specifically involves: constructing three parallel convolutional branches, which respectively employ 3×3 dilated convolutions with dilation rates d=1, d=2, and d=3 to generate multi-scale features corresponding to the receptive field.

6. The dynamic adaptive feature balancing method for small target detection by UAVs as described in claim 1, characterized in that, The loss function of the object detection model during training is defined as follows: ； in, Represents the network's regular trainable parameters. The channel segmentation parameters represent the network's learning parameters after N stages. Let D be the detection loss function, and D be the training dataset. For the inter-level Association regular terms, This is the weight coefficient for the regularization term.

7. A dynamic adaptive feature balancing system for small target detection using unmanned aerial vehicles (UAVs), characterized in that, include: The acquisition module is configured to acquire aerial images taken by the drone. The detection module is configured to perform target detection based on the acquired UAV aerial images and a preset target detection model, obtain the target detection results, and complete the small target detection and recognition of UAV images. The target detection model employs an improved YOLOv11-nano model. A dynamic spatial semantic decoupling unit is embedded in the C3k2 unit of the YOLOv11-nano backbone network, and an adaptive multi-scale perceptual fusion unit replaces the final downsampling layer in the neck network model of YOLOv11-nano. After inputting the UAV aerial image into the backbone network, the dynamic spatial semantic decoupling unit adaptively allocates the channel ratio between spatial and semantic branches, outputting a feature map aligned with both spatial and semantic aspects. This feature map is then input into the neck network, where the adaptive multi-scale perceptual fusion unit performs adaptive fusion of multi-scale features, ultimately inputting the detection head to generate the detection result.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the dynamic adaptive feature balancing method for small target detection of unmanned aerial vehicles as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the dynamic adaptive feature balancing method for small target detection of UAVs as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps in the dynamic adaptive feature balancing method for small target detection of UAVs as described in any one of claims 1-6.

Images