Unmanned aerial vehicle small target image detection method based on YOLO11 improved network

Abstract

The invention provides an unmanned aerial vehicle small target image detection method based on a YOLO11 improved network. The method comprises the following steps: constructing an unmanned aerial vehicle target image data set; a precision and lightweight collaborative optimization network is constructed, the network comprises a backbone feature extraction network, a neck feature fusion network and a detection head which are connected in sequence, and the backbone feature extraction network comprises a plurality of feature extraction structures which are cascaded in sequence and are provided with SPDConv down-sampling modules. A dynamic up-sampling operator and an efficient feature enhancement module are adopted in a feature fusion path of a cross-scale feature aggregation network structure of the neck feature fusion network; and inputting a to-be-detected unmanned aerial vehicle image into the trained precision and lightweight collaborative optimization network, and outputting a detection result of the unmanned aerial vehicle target image, the trained precision and lightweight collaborative optimization network being a network obtained by training the unmanned aerial vehicle target image data set and the composite loss function. Therefore, high-precision detection of the small target of the unmanned aerial vehicle is realized, and the real-time reasoning efficiency is considered.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) target detection technology, and in particular to a method for detecting small targets in UAV images based on an improved YOLO11 network. Background Technology

[0002] The widespread application of drones in reconnaissance, logistics delivery, agricultural plant protection, and aerial photography has made drone detection and identification technology a crucial link in airspace safety supervision. Vision-based drone detection has become a research hotspot due to its low cost and ease of deployment. However, drone targets in images are typically small in size, have varied shapes, and fly at high speeds, posing significant challenges to detection tasks. Especially in complex natural backgrounds and urban environments, drones share similar features with distracting elements such as birds and building details, further increasing the difficulty of accurate detection. In recent years, deep learning-based target detection algorithms have achieved significant breakthroughs in both accuracy and speed. The YOLO series of single-stage detectors has become the preferred solution for real-time drone detection due to its good balance between speed and accuracy. However, existing YOLO-based detectors still suffer from insufficient feature extraction, low multi-scale fusion efficiency, and high model complexity when dealing with small drone targets, making it difficult to achieve high-precision real-time detection on edge computing devices.

[0003] Currently, existing UAV target detection methods typically employ bilinear interpolation or nearest-neighbor interpolation for upsampling during multi-scale feature fusion. However, these fixed interpolation methods do not consider the actual content distribution of the input features, which can easily lead to problems such as blurred target edges and feature misalignment when reconstructing high-resolution feature maps, severely affecting the detection accuracy of small UAV targets. Existing UAV target detection methods often improve detection accuracy by stacking attention mechanisms or introducing complex modules, but this significantly increases model complexity while improving performance, resulting in a decrease in inference speed and low real-time inference efficiency. In practical applications, existing UAV target detection methods struggle to balance small target detection accuracy with real-time inference efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting small targets in UAV images based on an improved YOLO11 network, which solves the problem that existing technologies face in practical applications where it is difficult to balance the accuracy of small target detection with the efficiency of real-time inference.

[0005] To address the aforementioned technical problems, the embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for small target image detection on unmanned aerial vehicles (UAVs) based on an improved YOLO11 network, comprising: Construct a dataset of drone target images; A precision and lightweight co-optimization network for small target detection of UAVs is constructed. The precision and lightweight co-optimization network includes a backbone feature extraction network, a neck feature fusion network and a detection head connected in sequence. The backbone feature extraction network includes multiple feature extraction structures cascaded in sequence. Each feature extraction structure includes an SPDConv downsampling module. The neck feature fusion network adopts a cross-scale feature aggregation network structure. In the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is used to extract features. The efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​in the feature fusion module C3k2. The drone image to be detected is input into the trained precision and lightweight co-optimization network to detect small drone targets and output the detection results of the drone target image. The trained precision and lightweight co-optimization network is a network obtained by training the precision and lightweight co-optimization network using the drone target image dataset and a composite loss function, which includes bounding box regression loss and classification loss.

[0006] A second aspect of the present invention provides a small target image detection device for unmanned aerial vehicles based on an improved YOLO11 network, comprising: The first building module is used to construct the UAV target image dataset; The second construction module is used to construct a precision and lightweight co-optimization network for small target detection of UAVs. The precision and lightweight co-optimization network includes a backbone feature extraction network, a neck feature fusion network and a detection head connected in sequence. The backbone feature extraction network includes multiple feature extraction structures cascaded in sequence. Each feature extraction structure includes an SPDConv downsampling module. The neck feature fusion network adopts a cross-scale feature aggregation network structure. In the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is used to extract features. The efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​in the feature fusion module C3k2. The detection module is used to input the drone image to be detected into the trained precision and lightweight co-optimization network to detect small targets on the drone and output the detection results of the drone target image. The trained precision and lightweight co-optimization network is a network obtained by training the precision and lightweight co-optimization network using the drone target image dataset and a composite loss function. The composite loss function includes bounding box regression loss and classification loss.

[0007] Compared to existing technologies, the UAV small target image detection method based on the improved YOLO11 network provided by this invention constructs a UAV target image dataset; it constructs a precision and lightweight co-optimization network for UAV small target detection, which includes a backbone feature extraction network, a neck feature fusion network, and a detection head connected in sequence. The backbone feature extraction network includes multiple cascaded feature extraction structures, each of which includes an SPDConv downsampling module. The neck feature fusion network adopts a cross-scale feature aggregation network structure, and dynamic features are used in the feature fusion path of the cross-scale feature aggregation network structure. An upsampling operator performs upsampling, and an efficient feature enhancement module is used to extract features. This efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​within the feature fusion module C3k2. The UAV image to be detected is input into a trained precision and lightweight co-optimization network to detect small UAV targets and output the detection results of the UAV target image. The trained precision and lightweight co-optimization network is obtained by training the precision and lightweight co-optimization network using the UAV target image dataset and a composite loss function, which includes bounding box regression loss and classification loss. In this way, by co-designing the efficient feature enhancement module, cascaded multiple feature extraction structures, SPDConv downsampling module, and composite loss function in the cross-scale feature aggregation network structure, high-precision detection of small UAV targets is achieved without relying on additional sensors or significantly increasing the computational burden, while also ensuring real-time inference efficiency. Attached Figure Description

[0008] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein: Figure 1 A flowchart illustrating a method for small target image detection on UAVs based on an improved YOLO11 network is shown schematically. Figure 2 A schematic diagram illustrating the size distribution of UAV targets is provided. Figure 3 A schematic diagram of the structure of the precision and lightweight co-optimization network is shown. Figure 4 A schematic diagram of the SPDConv downsampling module is shown. Figure 5 A schematic diagram of a cross-scale feature aggregation network structure is shown. Figure 6 A schematic diagram of the structure of the dynamic upsampling operator DySample is shown. Figure 7 A schematic diagram of the structure of the high-efficiency feature enhancement module C3k2-iREB is shown. Figure 8 A schematic diagram of the detection head structure is shown. Figure 9 A schematic diagram illustrating the relationship between the predicted bounding box and the ground truth bounding box is shown. Figure 10 The results of drone target detection are shown schematically. Figure 1 ; Figure 11 The results of drone target detection are shown schematically. Figure 2 ; Figure 12 A schematic diagram of a UAV small target image detection device based on an improved YOLO11 network is shown. Detailed Implementation

[0009] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0010] It should be noted that, unless otherwise stated, the technical or scientific terms used in this invention should have the ordinary meaning as understood by one of ordinary skill in the art.

[0011] The methods described in the embodiments of the present invention will be explained in detail below.

[0012] The UAV small target image detection method proposed in this invention, based on the improved YOLO11 network, is based on a redesigned feature fusion method. A cross-scale feature aggregation network structure is designed, and on this structure, a dynamic upsampling mechanism and a newly designed efficient special enhancement module are synergistically integrated. At the same time, a composite loss function is designed for the UAV small target localization task, thus constructing a UAV detection method that can maximize the preservation and enhancement of small target feature information under limited computing resources.

[0013] Figure 1 A flowchart illustrating a method for small target image detection on UAVs based on an improved YOLO11 network, as shown in this embodiment of the invention, is provided. Figure 1 As shown, this UAV small target image detection method based on the improved YOLO11 network can include: S101. Construct a dataset of drone target images.

[0014] To train and evaluate the accuracy-lightweight co-optimization network of this invention, a large-scale UAV-bird adversarial detection dataset, namely the UAV target image dataset, was constructed, focusing on small targets. This UAV target image dataset aims to simulate real-world complex airspace scenarios, where UAV targets often interfere with birds of similar appearance. The UAV target image dataset includes publicly available UAV small target images collected online and images taken from various complex backgrounds, ensuring that UAV targets are small in size, have varied postures, and are subject to different lighting conditions. To improve the accuracy-lightweight co-optimization network's ability to distinguish between UAVs and birds, the UAV target image dataset specifically includes a large number of bird images as interference samples.

[0015] The UAV target image dataset is divided into training, validation, and test sets in a ratio of approximately 7:1.5:1.5. The training set contains 12,751 images, the validation set contains 2,770 images, and the test set contains 2,625 images, for a total of 18,146 images. This partitioning strategy ensures that the model has sufficient training data and allows for reliable performance evaluation through independent validation and test sets, while also preventing overfitting. All UAV and bird targets in the images are accurately labeled using bounding boxes, and the annotation information is saved in standard YOLO format. Figure 2 The image schematically illustrates the size distribution of drone targets, showing the size proportions of targets within the image as follows: Figure 2 As shown, the targets are mainly distributed around the coordinates (0.0, 0.0), so the UAV target image dataset also meets the detection requirements of small targets on UAVs.

[0016] To enhance the model's generalization ability and robustness, various data augmentation techniques were employed during the training phase, including random rotation (±45°), random horizontal and vertical flipping, random adjustment of brightness and contrast, Gaussian noise injection, and Mosaic enhancement. These strategies effectively increased data diversity and improved the model's detection performance in unknown scenarios.

[0017] S102. Construct a network for the coordinated optimization of accuracy and lightweight design for small target detection by UAVs.

[0018] The precision and lightweight collaborative optimization network comprises a backbone feature extraction network, a neck feature fusion network, and a detection head, connected sequentially. The backbone feature extraction network consists of multiple cascaded feature extraction structures, each including an SPDConv downsampling module. The neck feature fusion network employs a cross-scale feature aggregation network structure. Within this structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is employed for feature extraction. This efficient feature enhancement module integrates the inverted residual efficient multi-scale block iREB ​​within the feature fusion module C3k2. Each feature extraction structure also includes a first feature fusion module C3k2.

[0019] Specifically, the precision-lightweight co-optimization network YoloDroneNet, used for small target detection in UAVs, aims to effectively balance accuracy and speed in detecting small targets in UAV images. This precision-lightweight co-optimization network first extracts a multi-scale feature pyramid from the input image through a backbone network. Then, it utilizes an improved path aggregation neck network, namely the neck feature fusion network, to deeply fuse semantic and spatial features at different levels, enhancing the network's ability to represent multi-scale targets (especially small targets in UAVs). Finally, the three enhanced feature maps output by the neck feature fusion network are fed into a decoupled detection head to predict the target's bounding box coordinates, confidence score, and class probability in parallel.

[0020] The backbone feature extraction network serves as the foundational feature extractor for a network that optimizes both accuracy and lightweight design. Its core objective is to efficiently generate a series of feature maps with different resolutions from the input image. These feature maps together constitute a multi-scale feature representation: shallow feature maps have high resolution and rich spatial details, while deep feature maps have low resolution and strong semantic information.

[0021] Figure 3 The schematic diagram illustrates the structure of the network that coordinates accuracy and lightweight optimization. See [link / reference]. Figure 3 As shown, the specific processing flow of the backbone feature extraction network includes: Step A1: Downsample the UAV target image dataset using a standard convolutional layer to obtain a shallow feature map.

[0022] The input image is 640×640×3 pixels. It is first processed by a standard convolutional layer, which uses a 3×3 convolutional kernel and downsampling with a stride of 2. The output is a shallow feature map of size 320×320 pixels. This stage is mainly used to capture basic low-level features such as edges and textures.

[0023] Step A2: Input the shallow feature map into multiple cascaded feature extraction structures so that the multiple feature extraction structures output feature maps with different downsampling rates.

[0024] The processing flow of multiple cascaded feature extraction structures is as follows: In each feature extraction structure, the input feature map is sequentially input into the SPDConv downsampling module and the first feature fusion module C3k2. The SPDConv downsampling module performs scale reduction and information preservation on the input feature map, and the first feature fusion module C3k2 performs receptive field expansion and feature extraction to generate feature maps with different downsampling rates step by step. The feature maps with different downsampling rates include the deepest feature map.

[0025] Specifically, steps A2 and A3 are the steps for constructing the multi-stage feature pyramid. The core of the backbone feature extraction network consists of four cascaded feature extraction structures, aiming to progressively construct multi-scale feature representations. Each stage first uses the SPDConv downsampling module to perform scale transformation and resolution reduction. Figure 4 A schematic diagram of the SPDConv downsampling module is shown, such as... Figure 4 As shown, the SPDConv downsampling module halves the feature map size through spatial-to-depth transformation while preserving all pixel information through channel recombination, minimizing the loss of details of small drone targets during downsampling. Subsequently, the feature map enters the feature fusion module C3k2. This C3k2 feature fusion module, based on a cross-stage partial network structure, integrates configurable convolutional operations, enabling flexible switching between standard convolutions and lightweight large-kernel convolutions. This allows for receptive field expansion and depth feature extraction at the current resolution level, enhancing the network's ability to model small targets in complex backgrounds.

[0026] After being processed step-by-step by four cascaded feature extraction structures, the network outputs four feature maps with different spatial resolutions, i.e., feature maps with different downsampling rates, namely 4x, 8x, 16x, and 32x. These feature maps constitute a multi-scale feature pyramid covering different receptive fields, providing a foundation for cross-scale fusion of the subsequent neck network.

[0027] Step A3: Input the deepest feature map into the SPPF module to perform multi-scale global context information aggregation on the deepest feature map, output an intermediate feature map, and input the intermediate feature map into the C2PSA semantic enhancement module to perform semantic enhancement on the intermediate feature map, outputting a first target scale feature map; wherein, the shallow feature maps output by at least the first two feature extraction structures among the multiple feature extraction structures constitute at least one second target scale feature map; the first target scale feature map and at least one second target scale feature map constitute multiple target scale feature maps for input into the neck feature fusion network.

[0028] Specifically, after obtaining the deepest feature map through the aforementioned layers, i.e., the feature map with a downsampling rate of 32 times, the network uses an SPPF module to perform high-speed aggregation of multi-scale contextual information. This SPPF module uses three parallel max-pooling layers to capture receptive fields of different ranges and fuses their outputs with the original features, thereby giving the feature map a more powerful global perception capability without significantly increasing the computational load.

[0029] Finally, the SPPF module outputs features, i.e., intermediate feature maps, which flow through a C2PSA semantic enhancement module. This C2PSA semantic enhancement module innovatively combines the local modeling advantages of convolution with the global dependency capture capability of partial self-attention. By applying a lightweight self-attention mechanism on some channels, it further improves the semantic discriminativeness and robustness of high-level features.

[0030] After going through steps A1 to A3, the backbone feature extraction network finally outputs three key-size feature maps, namely three target-scale feature maps, with downsampling rates of 8x, 16x and 32x respectively. Together, they form a feature pyramid that takes into account both high-resolution details and strong semantic information, laying a solid foundation for subsequent neck fusion and head detection prediction.

[0031] As a key component connecting the backbone feature extraction network and the detection head, the neck feature fusion network has the core function of efficiently fusing multiple target-scale feature maps output by the backbone feature extraction network to construct a feature pyramid that combines rich spatial details and high-level semantic information.

[0032] To address the problems of poor feature perception and uneven allocation of computational resources in traditional feature fusion methods, this invention reconstructs the neck network. While maintaining the path aggregation network structure PANet, a cross-scale feature aggregation network structure CFAN is designed. The structure of the cross-scale feature aggregation network is described in [link to CFAN]. Figure 5 As shown, Figure 5A schematic diagram of the cross-scale feature aggregation network structure is shown. The core efficiency advantage of the cross-scale feature aggregation network is that, through a lightweight pre-channel alignment step, the channel dimension after feature concatenation is significantly reduced, thereby greatly reducing the computational overhead of subsequent fusion operations as a whole. In a traditional feature pyramid, feature maps of different scales have increasing channel numbers, such as the shallow layer C1, the middle layer C2, and the deep layer C3 (C1 < C2 < C3). When performing cross-scale fusion, the computational amounts of operations such as upsampling, feature concatenation, and feature extraction in feature aggregation and the convolutional operations of downsampling are proportional to this large channel number, resulting in low computational efficiency. In the CFAN structure of the present invention, first, a convolutional module with a lightweight 1×1 convolutional kernel is used to unify the channel numbers of the input features to a preset and moderate intermediate dimension C (usually the channel number of the small-scale feature map). This convolutional operation brings a small amount of computational amount, but when subsequent operations are performed on the feature map at this channel dimension, the computational amount of the overall network will be greatly reduced, and at the same time, the quality of the small-scale feature map can be ensured not to be reduced. Then, through feature aggregation, the channel numbers of the feature maps of the input detection head are unified again to C, ensuring the quality of the small-scale feature map and at the same time reducing the computational amount of the current and subsequent detection heads again, achieving the purpose of significantly improving computational efficiency. CFAN achieves a significant reduction in the computational amount in the entire fusion path by sacrificing the fusion quality of the secondary large-scale and medium-scale feature maps through a low-cost alignment operation, and can also ensure that the detection accuracy of the small-scale feature map for small targets of drones remains basically unchanged, thereby realizing the tilting of computing resources towards small targets of drones.

[0033] Meanwhile, based on the neck feature fusion network CFAN, by introducing a dynamic upsampling operator and an efficient feature enhancement module, a balance between the accuracy and speed of the overall network during drone detection is achieved.

[0034] See Figure 5 As shown, the processing flow of the cross-scale feature aggregation network includes: Step B1: Using the first convolutional module, compress the channel numbers of the first target scale feature map and at least one second target scale feature map to a preset dimension respectively, obtaining at least two intermediate features.

[0035] Specifically, the first convolutional module is a convolutional module with a 1×1 convolutional kernel. The input of the neck feature fusion network is feature maps of three key scales from the backbone feature extraction network. First, use a convolutional module with a 1×1 convolutional kernel to compress the channel numbers of all input features to a unified dimension, changing the inherent mode of increasing channel numbers in the traditional feature pyramid. The shallow and middle layer features closely related to the detection of small targets of drones can obtain sufficient parameter support, thereby realizing the balanced allocation of computing resources among tasks.

[0036] Step B2: In the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used to upsample the deepest feature among at least two intermediate features.

[0037] Step B3: The upsampled features are concatenated with the first shallow features from at least two intermediate features. The concatenated first feature map is input into the first feature fusion module C3k2 for feature extraction. The features extracted by the first feature fusion module C3k2 are sequentially aligned in channel dimension and upsampled using the second convolution module and the dynamic upsampling operator. The upsampled and aligned features are concatenated with the second shallow features from at least two intermediate features. The concatenated second feature map is input into the first efficient feature enhancement module for feature extraction to obtain a high-resolution feature map for detecting small targets.

[0038] Specifically, the second convolutional module is a 1×1 convolutional module. The fusion process still follows PANet's processing strategy. In the top-down semantic enhancement path, the network starts from the deepest features and uses the dynamic upsampling operator DySample to upsample them by a factor of 2. The structure of the dynamic upsampling operator DySample is as follows: Figure 6 As shown, Figure 6 The diagram illustrates the structure of the dynamic upsampling operator DySample. This operator achieves adaptive feature reconstruction by predicting dynamic convolution kernels related to the input content, effectively avoiding the problems of blurred edges and loss of details of small UAV targets caused by traditional interpolation methods. The upsampled and aligned features are concatenated with the second shallow feature from at least two intermediate features and then fed into the C3k2 module for further feature extraction. Next, a 1×1 convolution module is used again to align the channel dimensions and concatenate the upsampled features with adjacent shallow features. The concatenated feature map is then input into the first efficient feature enhancement module C3k2-iREB for feature extraction. The resulting high-resolution feature map is the small-scale feature map to be output for the detection head, used to detect small UAV targets.

[0039] The feature fusion mechanism of the Cross-Scale Feature Aggregation Network (CFAN) of this invention uses a lightweight 1×1 convolutional module to unify the number of channels of all input features to a preset dimension suitable for UAV small target detection before feature aggregation. Then, the number of channels is unified again in the feature extraction module before the feature map is input to the detection head. This breaks the traditional paradigm of increasing the number of channels and refocuses computing resources from the deep large target detection branch to the shallow and mid-level UAV small target feature extraction.

[0040] The high-efficiency feature enhancement module C3k2-iREB is an inverted residual high-efficiency multi-scale block iREB ​​that integrates into the standard C3k2 structure. Its structure is as follows: Figure 7 As shown, Figure 7 The schematic diagram illustrates the structure of the high-efficiency feature enhancement module C3k2-iREB. The core design element, iREB, is an improved module based on the inverted residual moving block iRMB. It generally follows an expansion-processing-shrinking structure and utilizes its core operators... This approach achieves an efficient fusion of the local modeling capabilities of convolution and the global modeling capabilities of attention. Input features are first enlarged to a 1×1 convolutional module using a convolutional kernel, then proceed to the core processing stage. This stage is handled by core operators. Completed, its specific implementation is defined as the following expression: ; in, For depthwise convolutional modules, For skip connections, For efficient multi-scale attention mechanisms.

[0041] In step B3, the stitched second feature map is input into the first efficient feature enhancement module for feature extraction, resulting in a high-resolution feature map for detecting small targets, including: The stitched second feature map is input into at least one inverted residual high-efficiency multi-scale block (iREB) contained in the first high-efficiency feature enhancement module for processing. The processing flow includes: Step B31: Expand the channels of the stitched second feature map through the first 1×1 convolution module.

[0042] Step B32: Input the expanded features sequentially into the efficient multi-scale attention mechanism (EMA) and the deep convolution module for joint processing.

[0043] Step B33: Compress the processed features through the second 1×1 convolution module.

[0044] Step B34: Perform a residual connection between the compressed features and the stitched second feature map to obtain the enhanced features.

[0045] Step B35: Based on the enhanced features processed by at least one inverted residual efficient multi-scale block iREB, the first efficient feature enhancement module outputs a high-resolution feature map; wherein, the first efficient feature enhancement module completes feature extraction through at least one inverted residual efficient multi-scale block iREB ​​to output a high-resolution feature map.

[0046] Specifically, the features first pass through the efficient multi-scale attention mechanism EMA, and its output is then fed into a deep convolutional module containing skip connections for further processing. This processed feature is then passed through a 1×1 convolutional module to reduce the number of channels back to the original dimension. Finally, a residual connection is formed with the original input to obtain the module's final output. The inverted residual efficient multi-scale block iREB ​​retains the lightweight and structural advantages of the inverted residual mobile block iRMB as an efficient mobile infrastructure, while significantly enhancing the module's multi-scale representation capabilities and cross-dimensional interaction capabilities through the integration of EMA. This makes the inverted residual efficient multi-scale block iREB ​​a more powerful and expressive module. When used in the feature fusion module C3k2, it ensures that the features receive guidance and enhancement from the attention mechanism at each key transformation step of the module, ultimately outputting more informative and discriminative fused features. This provides higher-quality input for the downstream detection head, allowing the efficient feature enhancement module C3k2-iREB to achieve more efficient feature extraction capabilities without significantly increasing computational cost.

[0047] Step B4: Starting with the high-resolution feature map, the third convolution module is used to downsample the high-resolution feature map. The downsampled first feature is concatenated with the deepest feature and input into the second efficient feature enhancement module for feature extraction to obtain a medium-resolution feature map for detecting medium-sized targets.

[0048] Step B5: The fourth convolution module is used to downsample the medium-resolution feature map, and the downsampled second feature is concatenated with the first target scale feature map. The concatenation is then input into the second feature fusion module C3k2 for feature extraction to obtain a low-resolution feature map for detecting large targets.

[0049] Both the third and fourth convolutional modules are 3×3 convolutional modules.

[0050] Specifically, in the bottom-up fusion path, the network starts with the fused maximum feature map, i.e., the high-resolution feature map, and uses a 3×3 convolutional module with a stride of 2 for downsampling. After the downsampled features are concatenated with the deep features, they are also extracted by the efficient feature enhancement module C3k2-iREB. The resulting medium-resolution feature map is the medium-scale feature map to be output by the detection head, used to detect medium-sized UAV targets. Finally, the downsampled features from the 3×3 convolutional module are concatenated with adjacent deep features, and further feature extraction is performed by the feature fusion module C3k2. The resulting low-resolution feature map is the large-scale feature map to be output by the detection head, used to detect large targets. It is worth noting that when outputting the final feature maps for detection, this invention uses the efficient feature enhancement module C3k2-iREB for feature extraction of both the small-scale feature maps used for detecting small targets on UAVs and the medium-scale feature maps used for detecting medium-sized targets. However, the original feature fusion module C3k2 is used for feature processing of the large-scale feature maps used for detecting large targets. This is because the computational cost of the efficient feature enhancement module C3k2-iREB is greater than that of the original feature fusion module C3k2. However, the UAVs detected by this invention are mostly small targets. Using the efficient feature enhancement module C3k2-iREB for small-scale and medium-scale feature maps can better improve the detection accuracy of small targets on UAVs while minimizing the complexity of the model.

[0051] Finally, the neck feature fusion network outputs three deeply enhanced feature maps, corresponding to high, medium, and low resolutions, respectively. These feature maps are rich in complementary information from details to semantics and are fed in parallel to the detection head to jointly complete the accurate identification and localization of UAV targets.

[0052] The C3k2-iREB feature enhancement module of this invention integrates an efficient multi-scale attention mechanism. It incorporates an iREB ​​module into the standard C3k2 structure. This iREB ​​module combines an inverse residual structure with an efficient multi-scale attention mechanism to synergistically enhance local and global feature modeling capabilities in a lightweight manner, significantly improving the feature discrimination of small targets on UAVs. Simultaneously, the module is selectively deployed on the high- and medium-resolution feature paths of the neck network, specifically for detecting small and medium-sized targets on UAVs; while the low-resolution path still uses the standard C3k2 module to handle large targets, thereby achieving optimized allocation of computational resources.

[0053] The detection head is the final prediction module of the accuracy-lightweight collaborative optimization network. It is responsible for decoding the high, medium, and low resolution features fused by the neck feature fusion network and outputting the target's location and category information. This model adopts the decoupled detection head structure of YOLO11. Each prediction unit in the detection head adopts a decoupled design and contains two independent and parallel processing branches.

[0054] Figure 8 A schematic diagram of the detection head structure is shown below. Figure 8 As shown, the processing flow of the detection head includes: Step C1: Input the high-resolution feature map, medium-resolution feature map, and low-resolution feature map into the regression branch and the classification branch respectively. In the regression branch, encode the high-resolution feature map, medium-resolution feature map, and low-resolution feature map to output the coordinate distribution representation of the target bounding box. In the classification branch, encode the high-resolution feature map, medium-resolution feature map, and low-resolution feature map to output the category probability distribution of the target.

[0055] Specifically, the regression branch receives high-resolution, medium-resolution, and low-resolution feature maps from the neck feature fusion network. These maps are then sequentially enhanced using two 3×3 convolutional modules. Finally, a combination of a 1×1 convolutional kernel, a depthwise separable convolution (DWConv 3x3), and a pointwise convolution (Conv 1x1) outputs the coordinate distribution representation of the target bounding box. Simultaneously, the classification branch processes the same input features, efficiently extracting discriminative features from two sets of data. The original logistic values ​​for each category are then output by a 1×1 convolutional module, which, after sigmoid activation, yield the class probabilities.

[0056] Step C2: Based on the coordinate distribution of the target bounding box and the probability distribution of the target's category, generate the corresponding initial prediction box set at each scale.

[0057] Specifically, each prediction unit generates a predefined set of initial prediction boxes at each anchor point location on its corresponding feature map. For each initial prediction box in the set, its output vector contains the precise coordinates of the bounding box, the object confidence reflecting the possibility of a target within the box, and the class probability distribution of the drone relative to the bird.

[0058] Step C3: Summarize the initial set of predicted boxes, and then perform preliminary filtering based on confidence and non-maximum suppression based on intersection-union ratio on the summarized set of predicted boxes in sequence, outputting bounding boxes with class labels and corresponding confidence scores.

[0059] Specifically, all predicted bounding boxes generated at the three scales are aggregated and further processed. First, the final score is calculated as the product of the object confidence score and the probability of the highest class, and all predicted bounding boxes are initially filtered. Then, a non-maximum suppression algorithm is used to remove highly overlapping redundant detection boxes based on a preset intersection-union ratio (IU) threshold, ensuring that only one optimal detection result is retained for the same target. Through this process, the model finally outputs a series of bounding boxes with class labels and their confidence scores, achieving accurate identification and localization of UAV targets in the image.

[0060] S103. Input the UAV image to be detected into the trained precision and lightweight co-optimization network to detect small targets on the UAV and output the detection results of the UAV target image.

[0061] The precision-lightweight co-optimization network, which has been trained, is obtained by training the network using a UAV target image dataset and a composite loss function. The composite loss function includes bounding box regression loss and classification loss.

[0062] Before inputting the drone image to be detected into the trained precision-lightweight co-optimization network to detect small drone targets and output the detection result of the drone target image, the method further includes: training the precision-lightweight co-optimization network using the drone target image dataset and a composite loss function to obtain the trained precision-lightweight co-optimization network. The trained precision-lightweight co-optimization network is then deployed.

[0063] Model deployment involves loading the weight file of the trained precision and lightweight co-optimized network into the inference engine and performing 100 idle inference operations on the GPU computing unit to warm up the model, eliminate performance fluctuations caused by cold starts, and ensure the stability and accuracy of the inference process.

[0064] The training steps are as follows: (1) Configure the network model training environment and set the training parameters. A training environment based on the PyTorch framework was built, with the batch size set to 16, the maximum number of iterations set to 100 training epochs, and the input image size set to 640. The Adam optimizer was used to update the network parameters, and a cosine annealing strategy with preheating was used to dynamically adjust the learning rate to improve model convergence efficiency and training stability. This strategy first adjusts the learning rate from the initial learning rate within the first 5 epochs. Linear warm-up to maximum learning rate It then decays to the minimum learning rate following a cosine function. The update formula for the learning rate during the decay phase is: ; in, This represents the current iteration number. This represents the total number of iterations. This represents the number of cycles in the preheating phase.

[0065] (2) Construct a composite loss function for a network that optimizes accuracy and lightweightness.

[0066] The loss function is crucial in guiding the learning direction of a neural network model. To simultaneously optimize the accuracy of UAV target localization and classification, the accuracy-lightweight collaborative optimization network of this invention employs a composite loss function. This function is composed of a weighted sum of bounding box regression loss and classification loss. The expression for the composite loss function is as follows: ; in, For composite loss function, For bounding box regression loss, For classification loss.

[0067] (21) Bounding box regression loss: To accurately locate small targets on UAVs and specifically address the core issues of uneven distribution of regression sample quality and sensitivity to bounding box geometry in small target detection, this invention designs a bounding box regression loss. This loss combines the sample focusing mechanism of Focaler-IoU with the shape and scale-aware mechanism of Shape-IoU, achieving complementary advantages through weighted summation. Specifically, the bounding box regression loss includes both the Focaler-IoU loss function and the Shape-IoU loss function.

[0068] The bounding box regression loss achieves two-level collaborative optimization: First, the Focaler-IoU loss function, through dynamic sample weight adjustment, filters out a large number of low-quality outliers in the early stages of training, making the network model's gradient more concentrated on potentially positive samples, effectively alleviating the problem of ambiguous optimization direction caused by the sparse positive samples of small UAV targets. Second, the Shape-IoU loss function, building on this, provides more precise gradient guidance for regression tasks sensitive to target geometry, especially strengthening the penalty for deviations in the short side direction of the target, thereby guiding the network to generate prediction boxes with higher shape fit. These two functions work together, optimizing sample selection and refining the regression target, ultimately significantly improving the model's localization accuracy for small UAV targets.

[0069] Specifically, the calculation process for bounding box regression loss includes: Step D1: Determine the Focaler-IoU loss function based on the original Intersection over Union (IoU) values ​​of the predicted bounding box and the ground truth bounding box, and the preset interval parameters.

[0070] Specifically, the core idea of ​​the Focaler-IoU loss function is to reconstruct the Intersection over Union (IoU) value through linear interval mapping, enabling the loss function to dynamically adjust the attention given to regression samples of different qualities (IoU values). The specific formula for the Focaler-IoU loss function is as follows: ; ; ; in, The predicted bounding box and the original bounding box are the original intersection-union ratio. To predict the area of ​​the bounding box, Let be the area of ​​the actual bounding box. This is the IoU value after dynamic focusing (Focaler) modulation. The lower limit parameter of the interval. The upper limit parameter of the interval. The Focaler-IoU loss function; The preset range parameter is an adjustable range, which can be set by... and This approach can proactively reduce focus on low-quality (low IoU) outliers and converged high-quality (high IoU) samples, allowing model training to concentrate more on medium-quality samples with improvement potential. Addressing the common characteristics of low-quality initial candidate bounding boxes for small targets on UAVs, which easily generate a large number of low-IoU negative samples, this invention sets... This effectively reduces gradient interference from low-quality samples and accelerates the optimization process of the model on valid positive samples.

[0071] Step D2: Determine the Shape-IoU loss function based on the center coordinates, width, and height of the predicted bounding box and the ground truth bounding box, as well as the orientation weights determined based on the target aspect ratio.

[0072] The core of the Shape-IoU loss function lies in recognizing that the shape and scale of the regression sample itself significantly affect localization accuracy. For non-square UAV targets, the same pixel deviation along the long and short sides has different effects on the IoU value. Therefore, this loss introduces adaptive orientation weights based on the target shape.

[0073] Specifically, step D2 includes: Step D21: Determine the normalized directional weights in the horizontal and vertical directions based on the width and height of the true bounding box; Step D22: Calculate the center point distance loss for shape perception based on the center point coordinates of the predicted box and the ground truth box, the normalized direction weights, and the diagonal length of the minimum bounding rectangle covering the predicted box and the ground truth box.

[0074] Specifically, the center point distance loss in shape perception The expression is: ; ; in, Figure 9 The diagram illustrates the relationship between the predicted bounding box and the ground truth bounding box, as shown below. Figure 9 As shown, The width of the actual bounding box. The actual height of the bounding box. The coordinates of the center point of the true bounding box. The coordinates of the center point of the prediction box. The diagonal length of the smallest bounding rectangle that covers both the predicted and ground truth boxes; The scale factor controls how sensitive the weights are to the target scale. It is usually set according to the average scale of the targets in the dataset. For UAV small target datasets, this value can be set to greater than 1 (e.g., 1.2 or 1.5) to enhance the sensitivity to small-scale targets. This represents the normalized directional weights in the horizontal direction. For the normalized direction weights in the vertical direction, satisfying When a target is a horizontal rectangle ( ),have and This means that horizontal deviations will be penalized more severely; conversely, for vertical rectangles, the penalty will be less severe. .

[0075] The center point distance loss for shape perception is no longer a simple Euclidean distance. For example, for a lateral UAV (ww>hh), its center point deviation in the vertical direction (y direction) is given a higher weight (multiplied by ww), prompting the network to prioritize center alignment in the shorter side direction, as this is more critical for improving the IoU of such targets.

[0076] Step D23: Based on the width and height of the predicted bounding box and the ground truth bounding box, calculate the shape-aware width and height loss by applying the normalized direction weights in the vertical direction to the width error and the normalized direction weights in the horizontal direction to the height error.

[0077] Shape-perceived width and height loss The expression is: ; ; in, The width of the prediction box. The height of the predicted bounding box, Weighted shape difference in the width direction. To weight the shape difference in the height direction, similar to the center point distance loss, the width and height loss also introduces cross weights. For lateral targets, the width deviation is weighted by a smaller weight. Adjustment, while the height deviation is weighted more heavily. Adjustment. This means that during optimization, the network prioritizes reducing the size error of small UAV targets along their shorter sides, as the accuracy of the shorter side dimensions is more critical for improving the IoU value of such non-square targets. This design makes the loss function more sensitive to the actual physical shape of the small UAV targets.

[0078] Step D24: Determine the Shape-IoU loss function based on the center point distance loss of shape perception, the width and height loss of shape perception, and the original intersection-union ratio of the predicted box and the ground truth box.

[0079] Shape-IoU loss function The expression is: ; Among them, among them, This is the original intersection-over-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. A constant of 0.5 is used to balance the contributions of the center point loss and the width and height loss.

[0080] Step D3: The Focaler-IoU loss function and the Shape-IoU loss function are weighted and summed to form the bounding box regression loss.

[0081] The formula for calculating the bounding box regression loss is as follows: ; in, For bounding box regression loss, As the first weighting coefficient, The Focaler-IoU loss function is used. This is the second weighting coefficient. The first weight coefficient is the Shape-IoU loss function. Second weighting coefficient To balance the weighting coefficients of these two loss functions, they can be set to 1:1.

[0082] (22) Classification loss: To address the potential class imbalance between drones and bird targets, the classification task employs a combination of Variable Focal Loss (VFL), the primary classification loss, and Focal Loss (FL), the secondary classification loss. If a target is present in the current batch, the primary classification loss is used. Otherwise, use the second classification loss. This classification loss, by dynamically down-adjusting the weights of easily classified samples, allows the model training to focus more on samples that are difficult to distinguish, effectively mitigating the performance limitations imposed by class imbalance. The formula for calculating the classification loss is as follows: ; ; in, The predicted probability of the target. The target quality score is typically the intersection-over-union (IoU) ratio between the predicted bounding box and the ground truth bounding box. As a class balance factor, This is a focusing parameter used to adjust the weights of difficult samples. The target is the true label (0 or 1). The classification loss is either the first classification loss or the second classification loss. During training, if there is a target in the current batch, the first classification loss is used; if there is no target in the current batch, the second classification loss is used. For the first category of loss, This is the second category of loss.

[0083] (3) Data preprocessing.

[0084] First, images from the UAV target image dataset are loaded, and all input images are uniformly scaled to a fixed size (640×640) to fit the network's input requirements. Then, various data augmentation strategies are employed to enhance the model's generalization ability, including random rotation (angle range ±45°), random horizontal and vertical flipping (flip probability 0.5), Gaussian noise injection, and contrast adjustment. Finally, the augmented images are normalized to accelerate network model training convergence and improve stability.

[0085] (4) Online training.

[0086] The preprocessed UAV target data is input into the YOLODroneNet network, a precision and lightweight co-optimization network for UAV small target detection. The network parameters are iteratively optimized using the backpropagation algorithm. During training, an early stopping mechanism is employed, evaluating model performance on the validation set every 20 epochs to monitor training effectiveness and prevent overfitting. Finally, the weight file of the network model that performs best on the validation set is saved as the optimal model parameters for subsequent inference and deployment.

[0087] The image of the drone to be detected is input into the trained precision and lightweight co-optimization network. This network is then used to detect small targets on the drone and outputs the detection results for the drone target image. The detection results are as follows: Figure 10 and Figure 11 As shown, Figure 10 The results of drone target detection are shown schematically. Figure 1 , Figure 11 The results of drone target detection are shown schematically. Figure 2 The drone target detection results are all presented in the selected form.

[0088] Compared to existing technologies, this invention precisely solves the fundamental contradiction between accuracy and speed in small target detection using UAVs, and brings the following significant technical effects and advantages: Optimizing computational resource allocation to ensure the ability to extract small targets from UAVs: The CFAN module avoids the traditional incremental feature pyramid design and ensures sufficient parameter support for shallow and mid-level features most relevant to UAV detection through balanced channel allocation. This addresses the resource misallocation problem in existing technologies, where computational resources are excessively tilted towards large target detection, allowing the network model's computational power to be more focused on the core task—UAV small target detection. This results in improved performance for this critical task while reducing the overall computational load.

[0089] This invention systematically enhances the representation, differentiation, and localization capabilities of small targets on UAVs, thereby improving UAV small target detection accuracy. Based on the optimized structure provided by CFAN, this invention systematically improves the detection accuracy of small targets on UAVs through a collaborative design across three levels: feature reconstruction, feature enhancement, and training optimization. Specifically, DySample is used to replace traditional fixed interpolation upsampling, achieving adaptive feature alignment and detail recovery, effectively alleviating edge blurring and information distortion of small targets on UAVs during feature fusion. Simultaneously, the C3k2-iREB feature enhancement module, selectively deployed on the critical path, is integrated into the neck network. This module combines local convolution and efficient multi-scale attention, significantly enhancing the feature differentiation capability for small UAVs and similar bird-like interference. Furthermore, addressing the challenge of UAV small target localization, a composite bounding box regression loss function fusing Focaler-IoU and Shape-IoU is designed. This function, through sample focusing and shape awareness mechanisms, accelerates training convergence while significantly improving the geometric fit of the bounding box. Experiments show that the synergistic effect of the above improvements significantly improves the model's mean accuracy (mAP@0.5) by 6.72% on the UAV-bird dataset, and improves precision and recall by 5.52% and 5.91% respectively, effectively suppressing false negatives and false positives.

[0090] Achieving model lightweighting and significantly improving inference speed: By employing a lightweight downsampling module based on space-to-depth transformation to achieve efficient feature utilization through downsampling, and performing channel alignment before and after feature fusion, this invention also precisely deploys the computationally optimized and efficient feature enhancement module C3k2-iREB on the critical path. This collaborative approach reduces model complexity at both the operator and structural levels. The number of model parameters is reduced by 42.31%, and the computational load is reduced by 26.56%. This enables the model to run efficiently on resource-constrained edge computing devices, achieving an end-to-end inference speed of up to 153 FPS on GPUs, fully meeting the deployment requirements of real-time monitoring systems and overcoming the drawback of existing lightweight designs that often result in a sharp performance decline.

[0091] Experimental results show that the proposed YoloDroneNet network significantly improves the detection accuracy and robustness of small drone targets in complex scenarios compared with mainstream drone detection models in the prior art. It effectively achieves a balance between high detection accuracy, high recall rate and high inference speed, providing a reliable technical means for drone supervision.

[0092] This invention addresses the challenge of balancing accuracy and real-time performance in UAV small target detection by proposing a UAV small target image detection method based on an improved YOLO11 network. This method significantly improves the feature perception and localization capabilities for UAV small targets by reconstructing the feature extraction and fusion methods, while maintaining a lightweight model. Specifically, this invention designs a cross-scale feature aggregation network structure (CFAN) to replace the traditional feature pyramid design with its incrementally increasing channel number. This structure first uses a 1×1 convolutional module to unify the channel number of all input features before feature fusion. Then, it uses the final feature extraction module to maintain a consistent channel number in the final output feature map. This allows the model to maintain detection accuracy while focusing computational resources on the shallow and mid-level features most relevant to UAV detection, thus optimizing the allocation efficiency of computational resources from the source. Furthermore, to improve feature reconstruction quality, this invention uses the dynamic upsampling operator DySample instead of traditional interpolation. This operator predicts adaptive convolutional kernels based on the input feature content, achieving more accurate multi-scale feature alignment and detail restoration, effectively alleviating the edge blurring problem of UAV small targets. Furthermore, at the feature extraction level, this invention proposes a high-efficiency feature enhancement module, C3k2-iREB, which integrates an efficient multi-scale attention mechanism and embeds it into the key feature extraction path of the neck network. This module combines the lightweight mechanism of the inverse residual moving block with the global dependency capture capability of the efficient multi-scale dynamic attention mechanism, enhancing the network's ability to distinguish small targets and similar interference objects without significantly increasing the number of parameters. To further reduce model complexity, this invention introduces a lightweight convolution module, SPDConv, based on space-to-depth transformation, in the downsampling stage of the backbone network, achieving downsampling that retains input information to the maximum extent. At the model training and optimization level, this invention designs a composite bounding box regression loss function for UAV small target localization tasks. This loss function integrates the sample focusing mechanism of Focaler-IoU and the shape and scale awareness mechanism of Shape-IoU. Through a collaborative optimization strategy, it effectively solves the problems of uneven quality distribution and geometric sensitivity of UAV small target regression samples, thereby significantly improving the bounding box localization accuracy and model convergence efficiency.

[0093] Based on the above Figure 1As can be seen from the implementation method, the embodiments of the present invention construct a UAV target image dataset; construct a precision and lightweight co-optimization network for UAV small target detection, the precision and lightweight co-optimization network includes a backbone feature extraction network, a neck feature fusion network and a detection head connected in sequence, the backbone feature extraction network includes multiple feature extraction structures cascaded in sequence, each feature extraction structure includes an SPDConv downsampling module, the neck feature fusion network adopts a cross-scale feature aggregation network structure, in the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is used to extract features, the efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​in the feature fusion module C3k2; input the UAV image to be detected into the trained precision and lightweight co-optimization network to detect UAV small targets, and output the detection result of UAV target image, the trained precision and lightweight co-optimization network is the network obtained by training the precision and lightweight co-optimization network using the UAV target image dataset and a composite loss function, the composite loss function includes bounding box regression loss and classification loss. In this way, by coordinating the design of efficient feature enhancement modules, cascaded multiple feature extraction structures, SPDConv downsampling modules, and composite loss functions in the cross-scale feature aggregation network structure, high-precision detection of small targets on UAVs can be achieved without relying on additional sensors or significantly increasing the computational burden, while also taking into account real-time inference efficiency.

[0094] Based on the same inventive concept, as an implementation of the above-mentioned method for detecting small targets in UAV images based on an improved YOLO11 network, this embodiment of the invention also provides a device for detecting small targets in UAV images based on an improved YOLO11 network. Figure 12 This is a structural diagram of the device in an embodiment of the present invention. See also: Figure 12 As shown, the UAV small target image detection device based on the YOLO11 improved network may include: The first construction module 1201 is used to construct the UAV target image dataset; The second construction module 1202 is used to construct a precision and lightweight co-optimization network for small target detection of UAVs. The precision and lightweight co-optimization network includes a backbone feature extraction network, a neck feature fusion network and a detection head connected in sequence. The backbone feature extraction network includes multiple feature extraction structures cascaded in sequence. Each feature extraction structure includes an SPDConv downsampling module. The neck feature fusion network adopts a cross-scale feature aggregation network structure. In the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is used to extract features. The efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​in the feature fusion module C3k2. The detection module 1203 is used to input the UAV image to be detected into the trained precision and lightweight co-optimization network to detect small targets on the UAV and output the detection result of the UAV target image. The trained precision and lightweight co-optimization network is a network obtained by training the precision and lightweight co-optimization network using the UAV target image dataset and a composite loss function. The composite loss function includes bounding box regression loss and classification loss.

[0095] In detection module 1203, the expression for the composite loss function is: ; ; ; ; in, For composite loss function, For bounding box regression loss, For classifying losses, As the first weighting coefficient, The Focaler-IoU loss function is used. This is the second weighting coefficient. The Shape-IoU loss function is used. The predicted probability of the target. For the target quality score, As a class balance factor, To focus parameters, The target is the true label. The classification loss can be either the first-class classification loss or the second-class classification loss. During training, if the target is present in the current batch, the first-class classification loss is used; otherwise, the second-class classification loss is used. For the first category of loss, This is the second category of loss.

[0096] It should be noted that the above description of the UAV small target image detection device embodiment based on the YOLO11 improved network is similar to the description of the UAV small target image detection method embodiment based on the YOLO11 improved network, and has similar beneficial effects. For technical details not disclosed in the embodiments of the UAV small target image detection device based on the YOLO11 improved network of this invention, please refer to the description of the UAV small target image detection method embodiment based on the YOLO11 improved network of this invention for understanding.

[0097] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for small target image detection in unmanned aerial vehicles (UAVs) based on an improved YOLO11 network, characterized in that, include: Construct a dataset of drone target images; A precision and lightweight co-optimization network for small target detection of UAVs is constructed. The precision and lightweight co-optimization network includes a backbone feature extraction network, a neck feature fusion network, and a detection head connected in sequence. The backbone feature extraction network includes multiple feature extraction structures cascaded in sequence. Each feature extraction structure includes an SPDConv downsampling module. The neck feature fusion network adopts a cross-scale feature aggregation network structure. In the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is used to extract features. The efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​in the feature fusion module C3k2. The drone image to be detected is input into the trained precision and lightweight co-optimization network to detect small drone targets and output the detection result of the drone target image. The trained precision and lightweight co-optimization network is a network obtained by training the precision and lightweight co-optimization network using the drone target image dataset and a composite loss function. The composite loss function includes bounding box regression loss and classification loss.

2. The method for detecting small targets in UAV images based on an improved YOLO11 network according to claim 1, characterized in that, Each feature extraction structure further includes a first feature fusion module C3k2, and the processing flow of the backbone feature extraction network includes: The UAV target image dataset is downsampled using standard convolutional layers to obtain shallow feature maps; The shallow feature map is input into the multiple cascaded feature extraction structures, so that the multiple feature extraction structures output feature maps with different downsampling rates. The processing flow of the multiple cascaded feature extraction structures is as follows: in each feature extraction structure, the input feature map is sequentially input into the SPDConv downsampling module and the first feature fusion module C3k2. The SPDConv downsampling module performs scale reduction and information preservation on the input feature map, and the first feature fusion module C3k2 performs receptive field expansion and feature extraction to generate feature maps with different downsampling rates step by step. The feature maps with different downsampling rates include the deepest feature map. The deepest feature map is input into the SPPF module to perform multi-scale global context information aggregation on the deepest feature map, outputting an intermediate feature map. The intermediate feature map is then input into the C2PSA semantic enhancement module to perform semantic enhancement on the intermediate feature map, outputting a first target scale feature map. The shallow feature maps output by at least the first two feature extraction structures constitute at least one second target scale feature map. The first target scale feature map and the at least one second target scale feature map constitute multiple target scale feature maps for input into the neck feature fusion network.

3. The UAV small target image detection method based on the improved YOLO11 network according to claim 2, characterized in that, The processing flow of the cross-scale feature aggregation network includes: Using the first convolution module, the number of channels of the first target scale feature map and the at least one second target scale feature map are compressed to a preset dimension to obtain at least two intermediate features; In the feature fusion path of the cross-scale feature aggregation network structure, the dynamic upsampling operator is used to upsample the deepest feature among the at least two intermediate features. The upsampled features are concatenated with the first shallow feature among the at least two intermediate features. The concatenated first feature map is input into the first feature fusion module C3k2 for feature extraction. The features extracted by the first feature fusion module C3k2 are sequentially aligned in channel dimension and upsampled using the second convolution module and the dynamic upsampling operator. The upsampled and aligned features are concatenated with the second shallow feature among the at least two intermediate features. The concatenated second feature map is input into the first efficient feature enhancement module for feature extraction, resulting in a high-resolution feature map for detecting small targets. Starting with the high-resolution feature map, the third convolution module is used to downsample the high-resolution feature map. The downsampled first feature is concatenated with the deepest feature and then input into the second efficient feature enhancement module for feature extraction to obtain a medium-resolution feature map for detecting medium-sized targets. The fourth convolution module is used to downsample the medium-resolution feature map, and the downsampled second feature is concatenated with the first target scale feature map. The concatenation is then input into the second feature fusion module C3k2 for feature extraction to obtain a low-resolution feature map for detecting large targets.

4. The UAV small target image detection method based on the improved YOLO11 network according to claim 3, characterized in that, The step of inputting the stitched second feature map into the first high-efficiency feature enhancement module for feature extraction to obtain a high-resolution feature map for detecting small targets includes: The stitched second feature map is input into at least one inverted residual high-efficiency multi-scale block (iREB) contained in the first high-efficiency feature enhancement module for processing. The processing flow includes: The stitched second feature map is then subjected to channel expansion via a first 1×1 convolution module; The expanded features are sequentially fed into the efficient multi-scale attention mechanism EMA and the deep convolution module for joint processing. The processed features are then subjected to channel compression via a second 1×1 convolutional module. The compressed features are residually joined with the stitched second feature map to obtain the enhanced features. Based on the enhanced features processed by the at least one inverted residual efficient multi-scale block iREB, the first efficient feature enhancement module outputs the high-resolution feature map; wherein, the first efficient feature enhancement module completes feature extraction through at least one inverted residual efficient multi-scale block iREB ​​to output the high-resolution feature map.

5. The UAV small target image detection method based on the YOLO11 improved network according to claim 4, characterized in that, The processing flow of the detection head includes: The high-resolution feature map, the medium-resolution feature map, and the low-resolution feature map are respectively input into the regression branch and the classification branch. In the regression branch, the high-resolution feature map, the medium-resolution feature map, and the low-resolution feature map are encoded to output the coordinate distribution representation of the target bounding box. In the classification branch, the high-resolution feature map, the medium-resolution feature map, and the low-resolution feature map are encoded to output the category probability distribution of the target. Based on the coordinate distribution of the target bounding box and the category probability distribution of the target, generate the corresponding initial prediction box set at each scale; The initial set of predicted boxes is aggregated, and the aggregated set of predicted boxes is then subjected to preliminary filtering based on confidence level and non-maximum suppression based on intersection-union ratio in sequence, outputting bounding boxes with class labels and corresponding confidence scores.

6. The method for detecting small targets in UAV images based on an improved YOLO11 network according to claim 1, characterized in that, The bounding box regression loss includes the Focaler-IoU loss function and the Shape-IoU loss function.

7. The method for detecting small targets in UAV images based on an improved YOLO11 network according to claim 6, characterized in that, The calculation process for the bounding box regression loss includes: The Focaler-IoU loss function is determined based on the original intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box, and the preset interval parameters. The Shape-IoU loss function is determined based on the center point coordinates, width, and height of the predicted bounding box and the ground truth bounding box, as well as the orientation weights determined based on the target aspect ratio. The Focaler-IoU loss function and the Shape-IoU loss function are weighted and summed to form the bounding box regression loss.

8. The method for detecting small targets in UAV images based on an improved YOLO11 network according to claim 7, characterized in that, The step of determining the Shape-IoU loss function based on the center coordinates, width, and height of the predicted bounding box and the ground truth bounding box, and the orientation weights determined based on the target aspect ratio, includes: Based on the width and height of the real bounding box, determine the normalized directional weights in the horizontal and vertical directions; The shape-aware center point distance loss is calculated based on the center point coordinates of the predicted bounding box and the ground truth bounding box, the normalized direction weight, and the diagonal length of the minimum bounding rectangle covering the predicted bounding box and the ground truth bounding box. Based on the width and height of the predicted bounding box and the ground truth bounding box, and by applying the normalized direction weight in the vertical direction to the width error and the normalized direction weight in the horizontal direction to the height error, the shape-aware width and height loss is calculated. The Shape-IoU loss function is determined based on the center point distance loss of the shape perception, the width and height loss of the shape perception, and the original intersection-union ratio of the predicted box and the ground truth box.

9. The method for detecting small targets in UAV images based on an improved YOLO11 network according to claim 7, characterized in that, The expression for the composite loss function is: ； ； ； ； in, The composite loss function, The bounding box regression loss is... For the classification loss, As the first weighting coefficient, The Focaler-IoU loss function is... This is the second weighting coefficient. The Shape-IoU loss function is... The predicted probability of the target. For the target quality score, As a category balance factor, To focus parameters, For the true label, the classification loss is either a first classification loss or a second classification loss. During training, if the target is present in the current batch, the first classification loss is used; if the target is not present in the current batch, the second classification loss is used. For the first category loss, This is the loss for the second category.

10. A UAV small target image detection device based on an improved YOLO11 network, characterized in that, include: The first building module is used to construct the UAV target image dataset; The second construction module is used to construct a precision and lightweight collaborative optimization network for small target detection of UAVs. The precision and lightweight collaborative optimization network includes a backbone feature extraction network, a neck feature fusion network, and a detection head connected in sequence. The backbone feature extraction network includes multiple feature extraction structures cascaded in sequence. Each feature extraction structure includes an SPDConv downsampling module. The neck feature fusion network adopts a cross-scale feature aggregation network structure. In the feature fusion path of the cross-scale feature aggregation network structure, a dynamic upsampling operator is used for upsampling, and an efficient feature enhancement module is used to extract features. The efficient feature enhancement module is a module that integrates the inverted residual efficient multi-scale block iREB ​​in the feature fusion module C3k2. The detection module is used to input the drone image to be detected into the trained precision and lightweight co-optimization network to detect small targets on the drone and output the detection result of the drone target image. The trained precision and lightweight co-optimization network is a network obtained by training the precision and lightweight co-optimization network using the drone target image dataset and a composite loss function. The composite loss function includes bounding box regression loss and classification loss.

Images