A detection method for establishing a small target detection model for drones based on improved YOLO11s

Abstract

The present invention relates to the field of computer vision, and in particular to a detection method for establishing an improved unmanned aerial vehicle (UAV) small target detection model based on YOLO11s. A YOLO11s-UAV model is established, where the YOLO11s-UAV model is an improved YOLO11s model. The last downsampling fusion layer in a backbone network of the YOLO11s model is deleted, and the downsampling fusion layer includes a Conv module and a C3k2 module. The remaining C3k2 modules in the backbone network of the YOLO11s model and the C3k2 modules in the neck network of the YOLO11s model are replaced with variable simple parameter-free attention mechanism modules. The remaining Conv modules in the backbone network of the YOLO11s model are replaced with space-to-depth residual convolution modules. The Upsample module in the neck network of the YOLO11s model is replaced with a content-aware feature reorganization upsampling module. The last downsampling fusion layer in the backbone network of the YOLO11s model is deleted, so that the entire network retains more fine feature information of small targets and reduces the complexity of the model.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a detection method for establishing a small target detection model for unmanned aerial vehicles (UAVs) based on an improved YOLO11s. Background Art

[0002] As mobile embedded platforms, drones face strict power consumption and computational constraints on their computing units. Existing lightweighting solutions, such as dynamic convolution and depthwise separable convolution, can reduce model parameters and computational complexity by over 20%, but they often result in significant degradation of various accuracy metrics, creating a vicious cycle of lightweighting and precision loss.

[0003] Existing models generally lack targeted design for retaining small target features. First, the downsampling process of traditional convolutional neural networks will continuously lose the fine-grained features of small targets. Second, mainstream target detection architectures often do not deploy dedicated small target attention mechanisms, resulting in the model ignoring the extraction of salient features of small targets in shallow networks.

[0004] Single-stage models, such as the YOLO series, leverage their streamlined architecture and fast inference speed to maintain high accuracy while meeting the real-time requirements of various application scenarios. However, in drone scenarios, these general-purpose single-stage models still struggle to detect small objects due to complex backgrounds, a small pixel ratio of objects, and low-resolution images. Summary of the Invention

[0005] The purpose of the present invention is to provide a detection method for establishing a small target detection model of unmanned aerial vehicle based on YOLO11s, which reduces the loss of feature information.

[0006] In order to achieve the above object, the present invention adopts such technical solution:

[0007] A method for establishing a small target detection model for unmanned aerial vehicles (UAVs) based on an improved YOLO11s algorithm is proposed. The YOLO11s-UAV model is an improved YOLO11s model. The YOLO11s-UAV model includes a backbone network for extracting features from input images, a neck network for extracting and fusing features from feature maps, and a head structure for detecting and classifying the fused feature maps output by the neck network.

[0008] Delete the last downsampling fusion layer in the backbone network of the YOLO11s model. The downsampling fusion layer includes the Conv module and the C3k2 module. Replace the remaining C3k2 modules in the backbone network of the YOLO11s model and the C3k2 modules in the neck network of the YOLO11s model with variable simple parameter-free attention mechanism modules. Replace the remaining Conv modules in the original backbone network with spatial-to-depth residual convolution modules.

[0009] Replace the Upsample module in the neck network of the YOLO11s model with a content-aware feature reorganization upsampling module;

[0010] Between the C2PSA module of the backbone network of the YOLO11s model and the first content-aware feature reconstruction upsampling module of the neck network, a first convolution module, a first splicing module and a variable simple parameter-free attention mechanism module are sequentially arranged along the data transmission direction.

[0011] Preferably, the YOLO11s-UAV model further includes a second convolution module, a third convolution module and a fourth convolution module;

[0012] The second convolution module is used to perform a convolution operation on the fusion features output by the first variable simple parameter-free attention mechanism module of the backbone network and then transmit the convolution operation to the splicing module after the first content-aware feature reorganization upsampling module of the neck network;

[0013] The third convolution module is used to perform a convolution operation on the fusion features output by the second variable simple parameter-free attention mechanism module of the backbone network and then transmit the convolution operation to the first splicing module of the neck network;

[0014] The fourth convolution module is used to perform a convolution operation on the fusion features output by the first variable simple parameter-free attention mechanism module after the content-aware reorganization of the neck network and the interactive feature pyramid network, and then transmit them to the last splicing module of the neck network.

[0015] Preferably, the space-to-depth residual convolution module includes a space-to-depth conversion module that divides the input feature map into multiple pixel blocks and rearranges them to the depth dimension, a non-stride convolution module that performs feature map channel number conversion operations on the pixel blocks, and a dilated residual module that extracts multi-scale context information from the feature map output by the non-stride convolution module, as well as merging and batch normalization processing of the feature map output by the non-stride convolution module and the feature map output by the dilated residual module, and applying an activation function to output the feature map.

[0016] Preferably, the variable simple parameter-free attention mechanism module performs channel dimension transformation operation and feature segmentation operation on the input feature map in sequence to obtain segmentation features. The segmentation feature selects a path according to the value of the Boolean parameter c3k. When the value of the Boolean parameter c3k is true, the segmentation feature is processed by a branch path composed of multiple C3kSimAM modules. When the value of the Boolean parameter c3k is false, the segmentation feature is processed by multiple Bottleneck modules, and the segmentation feature is residually connected and fused with the feature map output by the corresponding branch path. The fused feature map is then output after feature fusion and dimensionality restoration are completed through the convolution layer.

[0017] Preferably, the Bottleneck module in the C3k module is replaced with a BottleneckSimAM module to obtain the C3kSimAM module.

[0018] Preferably, the content-aware feature reconstruction upsampling module includes a kernel prediction module and a content-aware reconstruction module. The kernel prediction module performs channel compression, content encoding and generates a reassembled kernel on the input feature map in sequence, and applies an activation function to the kernel to output a predicted kernel. The content-aware reconstruction module uses a weighted sum operator to reconstruct the predicted kernel and outputs the reconstructed kernel.

[0019] A method for detecting small targets of drones, comprising the following steps performed in sequence:

[0020] S1: Obtain images taken by a drone, annotate them with normalized target frame coordinates and category labels, and divide the annotated images into training, validation, and test sets;

[0021] S2: Input the training set into the YOLO11s-UAV model established by any of the above-mentioned methods for establishing a UAV small target detection model based on the improved YOLO11s for training;

[0022] S3: Input the test set into the trained YOLO11s-UAV model for detection and output the detection results.

[0023] By adopting the above-mentioned design scheme, the beneficial effects of the present invention are:

[0024] In view of the characteristics of small target detection tasks, the last downsampling fusion layer in the original backbone network is deleted, so that the entire network retains more fine feature information of small targets while reducing the complexity of the model.

[0025] Replacing the C3k2 module in the network with a variable simple parameter-free attention mechanism module not only further reduces the model complexity, but also significantly expands the perception range of shallow networks for small object features by using real three-dimensional attention weights.

[0026] The remaining Conv modules in the original backbone network are replaced with spatial-to-depth residual convolution modules. This design eliminates the information loss caused by the traditional strided convolution operation that reduces the spatial dimension of the feature map, while greatly reducing the difficulty of the network in capturing multi-scale contextual information, providing a new technical solution for the effective retention of small target features. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a network architecture diagram of the YOLO11s-UAV model of the present invention;

[0028] Figure 2 It is a variable simple parameter-free attention mechanism module of the present invention;

[0029] Figure 3 This is a flowchart of the processing of an image by the spatial-to-depth residual convolution module of the present invention;

[0030] Figure 4 This is a flowchart of the processing of an image by the content-aware feature recombinant upsampling module of the present invention;

[0031] Figure 5 This is a heat map of the shallow features of the network before and after using the variable simple parameter-free attention mechanism module of the present invention;

[0032] Figure 6 This is a diagram showing the detection effects of the present invention under night and exposure scenarios;

[0033] Figure 7 This is a scatter plot comparing the spatial-to-depth residual convolution module of the present invention with other downsampling modules. DETAILED DESCRIPTION

[0034] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It is apparent that the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0035] The terms "first," "second," "third," and so on, in the description and claims of the present invention and the accompanying drawings are used to distinguish between different items, not to describe a specific order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0036] A method for establishing a small target detection model for drones based on the improved YOLO11s is established. Figure 1 The YOLO11s-UAV model shown is an improved YOLO11s model. The YOLO11s-UAV model includes a backbone network for extracting features from an input image, a neck network for extracting and fusing features from a feature map, and a head structure for detecting and classifying the fused feature map output by the neck network.

[0037] The last downsampling fusion layer in the backbone network of the YOLO11s model is deleted. The downsampling fusion layer includes the Conv module and the C3k2 module. Based on the characteristics of the small target detection task, unnecessary downsampling fusion layers in the backbone network are removed, so that the entire network retains more fine feature information of small targets, while further reducing the model complexity.

[0038] The remaining C3k2 modules in the backbone network of the YOLO11s model and the C3k2 modules in the neck network of the YOLO11s model are replaced with a variable simple parameter-free attention mechanism module (FlexSimAM). The variable simple parameter-free attention mechanism module is obtained by combining the C3k2 module with a simple parameter-free attention module (SimAM), that is, replacing the Bottleneck module in the C3k module with the BottleneckSimAM module to obtain the C3kSimAM module. The specific structure is as follows Figure 2 shown.

[0039] The simple parameter-free attention mechanism module is a lightweight attention mechanism. Its core is to leverage the local self-similarity of feature maps to efficiently generate realistic three-dimensional attention weights without introducing additional training parameters. This makes it particularly suitable for small target detection tasks on drones with limited onboard resources. The core innovation of the variable simple parameter-free attention mechanism module lies in its unique dynamic module switching mechanism, which uses the c3k Boolean value to flexibly switch between the bottleneck module (Bottleneck) and the C3kSimAM module in the model.

[0040] The variable simple parameter-free attention mechanism module sequentially passes the input feature map through the 1×1 Conv module for channel dimension transformation operation and the feature segmentation module (Split) for feature segmentation operation to obtain segmentation features. The segmentation feature selects the path according to the value of the Boolean parameter c3k. When the value of the Boolean parameter c3k is true, the segmentation feature is processed through a branch path composed of multiple C3kSimAM modules. When the value of the Boolean parameter c3k is false, the segmentation feature is processed through multiple Bottleneck modules, and the segmentation feature is residually connected and fused with the feature map output by the corresponding branch path. The fused feature map is then output after feature fusion and dimension restoration are completed by the 1×1 Conv module.

[0041] The variable simple parameter-free attention mechanism module has the following three significant advantages:

[0042] First, the number of C3kSimAM modules can be manually adjusted based on the image features of the dataset being tested. This quantity-driven adaptive mechanism significantly enhances the adaptability and generalization capabilities of the YOLO11s-UAV model in this application in different scenarios.

[0043] Second, if Figure 5 As shown in the figure, this mechanism addresses the difficulty of detecting small targets in drone aerial images by generating real three-dimensional attention weights, enhancing the sensitivity of shallow networks to feature extraction of small targets, and significantly improving the recognition rate of small targets.

[0044] Third, the lightweight design of this mechanism enables the YOLO11s-UAV model to effectively control the complexity of the model while ensuring detection accuracy.

[0045] Table 1 shows that the VisDrone-DET2019 dataset is used to perform multiple tests on the method proposed in this application using a variable simple parameter-free attention mechanism module (FlexSimAM), which is used to demonstrate the influence of the c3k Boolean parameter in the variable simple parameter-free attention mechanism module on each layer.

[0046] Table 1. The impact of the c3k Boolean parameter on each layer in the FlexSimAM (VisDrone-DET2019 dataset)

[0047]

[0048] As shown in the table above, three comparative experiments were conducted on the VisDrone-DET2019 dataset, comparing the Boolean values (true: √; false: ×) of the c3k parameters in the FlexSimAM modules at different layers in our proposed method. The baseline architecture retains all original C3k2 modules in the corresponding layers. Experiment 3 achieved the best detection accuracy. Therefore, for drone datasets under different scenarios, dynamically adjusting the Boolean values of the c3k parameters in FlexSimAM can achieve better detection performance while also reducing the model's weight.

[0049] The remaining Conv modules in the backbone network of the YOLO11s model are replaced with the spatial-to-depth residual convolution module (S2DResConv). The spatial-to-depth residual convolution module of this application is based on the spatial-to-depth conversion module (SPD) and the dilated residual module (DWR). Figure 3 As shown in the figure, the spatial-to-depth residual convolution module includes a spatial-to-depth conversion module that divides the input feature map into multiple pixel blocks and rearranges them to the depth dimension, a non-stride convolution module that converts the number of feature map channels on the pixel blocks, and a dilated residual module that extracts multi-scale context information from the feature map output by the non-stride convolution module, as well as merging the feature maps output by the non-stride convolution module and the feature maps output by the dilated residual module for batch normalization, and applying the Gaussian error linear unit (GELU) activation function to output the feature map; the normalization process makes the activation value smoother, ensures the output stability and is conducive to gradient flow, and the smooth nonlinear transformation provided by the GELU activation function avoids the neuron death problem related to the ReLU activation function and enhances the expressive power of the entire model.

[0050] In this embodiment, it is assumed that the feature map M∈(S,S,C1) processed by the spatial depth conversion module is divided into several pixel blocks with a step size of 2, and these feature blocks are combined into four sub-blocks along the spatial dimension 、 、 、 , as shown in formula (1):

[0051] (1);

[0052] The size of each sub-block is (S / 2, S / 2, C1). After the space-to-depth conversion, the number of channels increases to 4C1, the length and width are halved, and finally a feature map of size (S / 2, S / 2, 4C1) is obtained.

[0053] The non-stride convolution module applies the feature map channel number conversion operation to the feature map output by the spatial depth conversion module, and adjusts the channel dimension of the feature map to the specified value C2. The non-stride convolution module does not perform any operation to reduce the spatial dimension of the feature map, thereby avoiding the loss of target feature information.

[0054] The dilated residual module uses a two-step approach to extract multi-scale contextual information: the first is regional residualization (RR) to generate residual features from the input features, and the second is semantic residualization (SR) to apply multi-rate dilated depth convolution to perform morphological filtering on features of regions of different sizes. Finally, the features are merged into a residual through point-by-point convolution and added to the input feature map to create a more comprehensive feature representation.

[0055] like Figure 7 As shown in the figure, the VisDrone-DET2019 dataset was used to test the proposed method using the spatial-to-depth residual convolution module, and other conventional general-purpose convolution modules were used for comparative experiments. The scatter plot intuitively shows that the S2DResConv downsampling module proposed in the proposed method exhibits significant advantages in the drone small target detection task on the VisDrone-DET2019 dataset: it comprehensively outperforms other compared downsampling modules in terms of detection accuracy indicators, while only introducing a small number of model parameters, which reflects its excellent balance between detection accuracy and model efficiency.

[0056] The Upsample module in the neck network of the YOLO11s model is replaced by the Content-Aware Feature Restructuring Upsampling module (CARAFE); this content-aware feature restructuring upsampling module shows significant advantages in small object detection tasks through large receptive field context aggregation and dynamic kernel generation mechanism.

[0057] The content-aware feature reconstruction upsampling module includes a kernel prediction module and a content-aware reconstruction module. The kernel prediction module performs channel compression, content encoding and generates a reassembled kernel on the input feature map in sequence, and applies an activation function to the kernel to output a predicted kernel. The content-aware reconstruction module uses a weighted sum operator to reconstruct the predicted kernel and outputs the reconstructed kernel.

[0058] In this embodiment, Figure 4 As shown in the figure, the workflow of the content-aware feature reorganization upsampling module is as follows: given the input feature map X∈(C×H×W) and the upsampling factor σ, by implementing the kernel prediction module and the content-aware reorganization module, the output feature map The generation process of ∈ (C×σH×σW) includes:

[0059] The kernel prediction module consists of three submodules, namely channel compressor, content encoder and kernel normalizer. They respectively perform channel compression, content encoding and generate reassembled kernels on the input feature map, and apply softmax function to this kernel. Central feature k×k neighborhood of This kernel prediction module predicts the location of each target Predict its position-dependent kernel :

[0060] ;

[0061] in, represents the kernel prediction module, represents the kernel size of the content encoder,

[0062] The content-aware reorganization module will be predicted kernel Reassemble features in the local area to obtain the output feature map :

[0063] ;

[0064] in, represents the operator of the content-aware reorganization module, represents the reorganized kernel size,

[0065] The content-aware reconstruction module uses a simple weighted sum operator By reorganizing, more attention can be paid to the information provided by the relevant points in the local area, and the semantics of the reassembled feature map may be stronger than the original feature map.

[0066] Between the C2PSA module of the backbone network of the YOLO11s model and the first content-aware feature reconstruction and upsampling module of the neck network of the YOLO11s model, a first convolution module, a first splicing module, and a variable simple parameter-free attention mechanism module are sequentially provided along the data transmission direction. In this embodiment, the first convolution module is a 1×1 Conv module.

[0067] The YOLO11s-UAV model also includes a second convolution module, a third convolution module, and a fourth convolution module; and the second convolution module, the third convolution module, and the fourth convolution module are all 3×3 Conv modules;

[0068] The second convolution module is used to perform a convolution operation on the fusion features output by the first variable simple parameter-free attention mechanism module of the backbone network and then transmit the convolution operation to the splicing module after the first content-aware feature reorganization upsampling module of the neck network;

[0069] The third convolution module is used to perform a convolution operation on the fusion features output by the second variable simple parameter-free attention mechanism module of the backbone network and then transmit the convolution operation to the first splicing module of the neck network;

[0070] The fourth convolution module is used to perform a convolution operation on the fusion features output by the first variable simple parameter-free attention mechanism module after the content-aware reorganization of the neck network and the interactive feature pyramid network, and then transmit it to the last splicing module of the neck network;

[0071] The fused features output by the first variable simple parameter-free attention mechanism module of the backbone network are delivered to the splicing module after the second content-aware feature reorganization and upsampling module (CARAFE) of the neck network. The fused features output by the second variable simple parameter-free attention mechanism module of the backbone network are delivered to the splicing module after the first content-aware feature reorganization and upsampling module (CARAFE) of the neck network.

[0072] The fused features output by the variable simple parameter-free attention mechanism module between the C2PSA module and the first content-aware feature reorganization upsampling module of the neck network of the YOLO11s model are delivered to the last splicing module of the neck network, and the fused features output by the variable simple parameter-free attention mechanism module before the second content-aware feature reorganization upsampling module (CARAFE) of the neck network are delivered to the penultimate splicing module of the neck network.

[0073] The introduction of the first convolution module, the first splicing module and the variable simple parameter-free attention mechanism module between the backbone network and the neck network, as well as the setting of the second convolution module, the third convolution module and the fourth convolution module is equivalent to introducing the skip connection layer in the YOLO11s-UAV model and adaptively aligning the channel dimension differences of the multi-scale feature maps, which significantly improves the fusion efficiency of feature information of different scales in the same deep network layer.

[0074] Table 2 shows the results of multiple tests using the VisDrone-DET2019 dataset using different upsampling modules for the proposed method, comparing the Content-Aware Feature Reconstruction Upsampling Module (CARAFE) with other upsampling modules. Nearest and Bilinear represent different upsampling methods in the Upsample module of the original YOLO11s model.

[0075] Table 2 Content-Aware Feature Recombination Upsampling Module (CARAFE) introduced in this application method

[0076] Comparative experiments with other upsampling modules (VisDrone-DET2019 dataset)

[0077]

[0078] As can be seen from the above table, six groups of replacement tests were conducted on the upsampling module in the method of this application. The results show that compared with other upsampling modules, the content-aware feature reconstruction upsampling module achieves better detection performance improvement under the premise of introducing a small amount of model parameters and calculation amount.

[0079] This embodiment also provides a detection method for performing detection in the YOLO11s-UAV model established by the above method.

[0080] A method for detecting small targets of drones, comprising the following steps performed in sequence:

[0081] S1: Obtain images taken by a drone, annotate them with normalized target frame coordinates and category labels, and divide the annotated images into training, validation, and test sets;

[0082] S2: Input the training set into the YOLO11s-UAV model established by any of the above-mentioned methods for establishing a UAV small target detection model based on the improved YOLO11s for training;

[0083] S3: Input the test set into the trained YOLO11s-UAV model for detection and output the detection results, such as Figure 6 shown.

[0084] Table 3 shows the ablation experiments on the three methods proposed in this application using the VisDrone-DET2019 dataset, where the newly introduced skip connection layer and the replacement content-aware feature reorganization upsampling module (CARAFE) together represent the improvement of the feature pyramid network structure of the original YOLO11s model to the content-aware reorganization and interactive feature pyramid network structure (CARIFPN).

[0085] Table 3 Ablation experiments of this application method (VisDrone-DET2019 dataset)

[0086]

[0087] As can be seen from the above table, the ablation experiment of the method of this application verifies the effectiveness of the improved superposition by gradually integrating three improvements on the baseline model YOLO11s: content-aware reorganization and interactive feature pyramid network structure (CARIFPN), variable simple parameter-free attention mechanism module (FlexSimAM) and spatial-to-depth residual convolution module (S2DResConv).

[0088] Table 4 Comparison experiment of mAP@0.5 index between the proposed method and the baseline method (VisDrone-DET2019 dataset)

[0089]

[0090] As can be seen from the above table, compared with the baseline model YOLO11s, this method achieved significant improvements on both the validation set and the test set of the VisDrone-DET2019 dataset. The key indicator mAP@0.5 improved by 7.8% on the validation set and 5.9% on the test set. This result fully demonstrates that this method is not only effective on the validation set, but also exhibits excellent detection performance on unknown test set data.

[0091] Table 5 Generalization experiments on other UAV aerial photography datasets

[0092]

[0093] As can be seen from the above table, the method of the present application shows significant advantages over the baseline model YOLO11s in two typical drone photography datasets, TinyPerson and UAVDT-DET. In the validation set, the mAP@0.5 of all categories increased by 7.2% and 3.0% respectively. This result shows that the present method can achieve significant improvements in the accuracy of small target detection in different drone scenarios, showing excellent dataset generalization performance. The specific implementation methods described above further explain the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for establishing a small target detection model for drones based on an improved YOLO11s, characterized by: Establish a YOLO11s-UAV model, which is an improved YOLO11s model. The YOLO11s-UAV model includes a backbone network for extracting features from input images, a neck network for extracting and fusing features from feature maps, and a head structure for detecting and classifying the fused feature maps output by the neck network. Delete the last downsampling fusion layer in the backbone network of the YOLO11s model. The downsampling fusion layer includes the Conv module and the C3k2 module. Replace the remaining C3k2 modules in the backbone network of the YOLO11s model and the C3k2 modules in the neck network of the YOLO11s model with variable simple parameter-free attention mechanism modules. Replace the remaining Conv modules in the original backbone network with spatial-to-depth residual convolution modules. Replace the Upsample module in the neck network of the YOLO11s model with a content-aware feature reorganization upsampling module; Between the C2PSA module of the backbone network of the YOLO11s model and the first content-aware feature reorganization upsampling module of the neck network, a first convolution module, a first splicing module, and a variable simple parameter-free attention mechanism module are sequentially arranged along the data transmission direction; The second convolution module is used to perform a convolution operation on the fusion features output by the first variable simple parameter-free attention mechanism module of the backbone network and then transmit it to the splicing module after the first content-aware feature reorganization upsampling module of the neck network; The third convolution module is used to perform a convolution operation on the fusion features output by the second variable simple parameter-free attention mechanism module of the backbone network and then transmit it to the first splicing module of the neck network; The fourth convolution module is used to perform a convolution operation on the fusion features output by the first variable simple parameter-free attention mechanism module after the content-aware reorganization of the neck network and the interactive feature pyramid network, and then transmit it to the last splicing module of the neck network; The spatial-to-depth residual convolution module includes a spatial-to-depth conversion module that divides the input feature map into multiple pixel blocks and rearranges them into the depth dimension, a non-strided convolution module that performs a feature map channel number conversion operation on the pixel blocks, and a dilated residual module that extracts multi-scale context information from the feature map output by the non-strided convolution module, and merges and batch normalizes the feature map output by the non-strided convolution module and the feature map output by the dilated residual module, and applies an activation function to output the feature map; The variable simple parameter-free attention mechanism module performs channel dimension transformation operation and feature segmentation operation on the input feature map in sequence to obtain segmentation features. The segmentation features select paths according to the value of the Boolean parameter c3k, and select branch paths composed of multiple C3kSimAM modules or multiple Bottleneck modules to process the segmentation features.

2. The method for establishing a small target detection model for unmanned aerial vehicles based on the improved YOLO11s as claimed in claim 1, characterized in that: The variable simple parameter-free attention mechanism module performs channel dimension transformation operation and feature segmentation operation on the input feature map in sequence to obtain segmentation features. The segmentation feature selects a path according to the value of the Boolean parameter c3k. When the value of the Boolean parameter c3k is true, the segmentation feature is processed by a branch path composed of multiple C3kSimAM modules. When the value of the Boolean parameter c3k is false, the segmentation feature is processed by multiple Bottleneck modules, and the segmentation feature is residually connected and fused with the feature map output by the corresponding branch path. The fused feature map is then output after feature fusion and dimensionality restoration are completed through the convolution layer.

3. The method for establishing a small target detection model for unmanned aerial vehicles based on the improved YOLO11s as claimed in claim 2, characterized in that: The Bottleneck module in the C3k module is replaced with the BottleneckSimAM module to obtain the C3kSimAM module.

4. The method for establishing a small target detection model for unmanned aerial vehicles based on the improved YOLO11s as claimed in claim 3 is characterized in that: The content-aware feature reorganization upsampling module includes a kernel prediction module and a content-aware reorganization module. The kernel prediction module sequentially performs channel compression and content encoding on the input feature map and generates a reassembled kernel, and applies an activation function to the kernel to output a predicted kernel. The content-aware reorganization module uses a weighted sum operator to reorganize the predicted kernel and outputs the reorganized kernel.

5. A method for detecting small targets on drones, characterized by: The process includes the following steps: S1: Obtain images taken by a drone, annotate them with normalized target frame coordinates and category labels, and divide the annotated images into training, validation, and test sets; S2: Inputting the training set into the YOLO11s-UAV model established by the improved UAV small target detection model establishment method based on YOLO11s according to any one of claims 1 to 4 for training; S3: Input the test set into the trained YOLO11s-UAV model for detection and output the detection results.

Images