An unmanned aerial vehicle image target detection method and system based on YOLOv11n improvement

Abstract

This invention belongs to the field of UAV image target detection technology, specifically disclosing a UAV image target detection method and system based on an improvement of YOLOv11n. The method acquires UAV-captured image data and improves the YOLOv11n model: the Bottleneck structure in the C3k2 module is replaced with a StripCGLU module to construct a C3k2_StripCGLU module; the FFN in the C2PSA module is replaced with an EDFFN module that fuses frequency domain information to construct a C2PSA_EDFFN module; a FEMD module is introduced into the shallow feature layers p1 and p2, and the Shape-IoU loss function is used to optimize bounding box regression. After training, the UAV image target detection model is obtained, and the detection results are output for the image to be tested. This invention can enhance the expression of small targets and weak texture features, improve detection accuracy and localization stability, and is suitable for lightweight deployment.

Description

Technical Field

[0001] This invention belongs to the field of UAV image target detection technology, specifically relating to a UAV image target detection method and system based on YOLOv11n improvement. Background Technology

[0002] Target detection in UAV images is widely used in low-altitude scenarios such as aerial surveying, power line inspection, traffic monitoring, and emergency search and rescue. Compared with ordinary ground-view images, UAV images typically have characteristics such as large variations in shooting altitude, significant differences in target scale, low target pixel ratio, complex background interference, and significant influence from environmental factors such as rain, fog, and changes in lighting. Therefore, higher requirements are placed on the detection accuracy, positioning accuracy, and real-time processing capabilities of target detection algorithms. Existing target detection algorithms mainly include two-stage target detection algorithms and single-stage target detection algorithms. Two-stage target detection algorithms typically generate candidate regions first, and then classify and regress bounding boxes for the candidate regions. Representative algorithms include R-CNN, Fast R-CNN, Faster R-CNN, and Mask R-CNN. This type of algorithm has good detection accuracy and robustness, but the computational complexity of candidate region generation and subsequent fine-tuning processes is high, and the inference latency is large, making it difficult to meet the requirements of real-time detection and lightweight deployment on UAVs.

[0003] Single-stage object detection algorithms omit the candidate region generation process, directly completing object category prediction and bounding box regression; the YOLO series of algorithms is a typical example. With the development of the YOLO series models, structures such as feature pyramids, anchorless bounding boxes, attention mechanisms, and lightweight detection heads have been continuously introduced, improving detection efficiency and multi-scale object detection capabilities. YOLOv11n, as a lightweight model, has certain advantages in deployment on low-resource devices. It employs modules such as C3k2 and C2PSA to improve feature extraction capabilities and reduces redundant computation through depthwise separable convolutions.

[0004] However, existing YOLOv11n and related improved algorithms still have shortcomings in UAV image detection: First, small target pixels account for a low percentage in UAV images, and detailed features are easily lost after multiple downsampling, resulting in a high false negative rate; second, complex backgrounds, occlusions, rain, fog, and changes in lighting weaken target texture and edge information, affecting the model's ability to extract weak features; third, existing loss functions are not sensitive enough to the scale, shape, and localization deviation of small target bounding boxes, easily leading to low localization accuracy; fourth, some high-precision models have a large number of parameters and computational load, making it difficult to meet the real-time and lightweight deployment requirements of UAV platforms. Therefore, there is an urgent need for a target detection method for UAV image scenarios that takes into account multi-scale feature extraction, small target enhancement, localization optimization, and lightweight computation. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical problems of existing UAV image detection technology based on YOLOv11n. This invention proposes an improved UAV image target detection method and system based on YOLOv11n, which aims to improve the detection performance of UAVs for multi-scale targets in scenarios such as aerial surveying and power line inspection, reduce the number of model parameters and computational complexity, and enhance the ability to capture weak textures and small target features.

[0006] To achieve the above objectives, this invention provides a UAV image target detection method based on an improved YOLOv11n, comprising: Acquire image data captured by drones, and obtain training image data from the image data for training the improved YOLOv11n model; The YOLOv11n model is improved to obtain an improved YOLOv11n model. The improvement process includes: replacing the Bottleneck structure in the C3k2 module of the model backbone network with a StripCGLU module based on strip convolution and gating mechanism to construct the C3k2_StripCGLU module; replacing the feedforward network FFN in the C2PSA module of the backbone network with an EDFFN module that integrates frequency domain information to construct the C2PSA_EDFFN module; introducing a FEMD feature enhancement module based on multi-branch convolution and dilated convolution into the p1 shallow feature layer and p2 shallow feature layer of the backbone network; and replacing the GIoU and DIoU loss functions used by the YOLOv11n model in bounding box regression training with a Shape-IoU loss function that introduces scale factor and shape factor. The improved YOLOv11n model is iteratively optimized and trained using the training image data to obtain a UAV image target detection model. The UAV image target detection model is then used to detect the image data to be tested, and the target detection results are output.

[0007] As a further improvement to the above technical solution, the improved YOLOv11n model includes a backbone network, a neck network, and a detection head along the image data processing direction; The backbone network receives the training image data and, through a stepwise downsampling feature extraction path including the C3k2_StripCGLU module, forms a p1 shallow feature layer, a p2 shallow feature layer, and a backbone network terminal feature layer. The shallow features output by the p1 and p2 shallow feature layers are processed by the FEMD feature enhancement module and input as shallow enhancement features into the neck network. The high-level features output by the backbone network terminal feature layer are processed by the C2PSA_EDFFN module and input as high-level enhancement features into the neck network. The neck network performs multi-scale feature fusion on the shallow and high-level enhancement features and inputs the fused feature map into the detection head. The detection head generates predicted bounding boxes based on the fused feature map. During training, the annotation information of the training image data provides the ground truth bounding boxes, and the Shape-IoU loss function calculates the bounding box regression loss based on the predicted bounding boxes and the ground truth bounding boxes.

[0008] As a further improvement to the above technical solution, the processing procedure of the C3k2_StripCGLU module includes: taking the feature map input to the C3k2_StripCGLU module as the C3k2 input feature map, and dividing the C3k2 input feature map into a first input sub-feature map and a second input sub-feature map; performing convolution processing on the first input sub-feature map to obtain a first convolutional feature map; performing depth feature extraction on the second input sub-feature map through multiple cascaded StripCGLU structures to obtain a second depth feature map; and fusing the first convolutional feature map and the second depth feature map to obtain the output feature map of the C3k2_StripCGLU module. The processing procedure for any of the StripCGLU structures includes: using the feature map input to the StripCGLU structure as the StripCGLU input feature map, and assigning the StripCGLU input feature map to a first strip convolution branch and a second strip convolution branch to form a first branch input feature map and a second branch input feature map; processing the first branch input feature map with a 1×1 strip convolution to obtain a first strip convolution feature map; processing the second branch input feature map with a 19×1 strip convolution to obtain a second strip convolution feature map; concatenating the first strip convolution feature map and the second strip convolution feature map in the channel dimension to obtain a concatenated feature map; performing batch normalization on the concatenated feature map to obtain a normalized feature map; and inputting the normalized feature map into the CGLU activation function to obtain the StripCGLU output feature map. The formula for calculating the CGLU activation function is: CGLU(x) = σ(Conv(x)) Conv(x), where x is the feature map input to the CGLU activation function, σ is the Sigmoid function, and Conv is the convolution operation. This indicates element-wise multiplication.

[0009] As a further improvement to the above technical solution, the processing procedure of the EDFFN module in the C2PSA_EDFFN module includes: using the feature map input to the EDFFN module as the EDFFN input feature map; inputting the EDFFN input feature map into a 1×1 pointwise convolution to expand the number of channels, obtaining a channel-expanded feature map; inputting the channel-expanded feature map into a 3×3 depthwise separable convolution to obtain a local neighborhood feature map; segmenting the local neighborhood feature map into a first segmentation feature x1 and a second segmentation feature x2; performing GELU activation processing on the first segmentation feature x1 to obtain a gated control feature GELU(x1); and combining the gated control feature GELU(x1) with the... The second segmentation feature x2 is multiplied element-wise to obtain a gated fusion feature map; the gated fusion feature map is segmented into multiple 8×8 local image blocks; a two-dimensional fast Fourier transform is performed on each of the 8×8 local image blocks to map the features from the spatial domain to the frequency domain, resulting in a frequency domain feature block; the frequency domain feature block is multiplied element-wise with the learnable complex tensor parameter W to obtain a filtered frequency domain feature block; an inverse fast Fourier transform is performed on the filtered frequency domain feature block to restore it back to the spatial domain, resulting in a spatial domain reconstructed feature block; the spatial domain reconstructed feature blocks are reassembled into the original resolution feature map, and the number of channels is restored to the target dimension through a 1×1 convolution to obtain the EDFFN output feature map.

[0010] As a further improvement to the above technical solution, the processing procedure of the FEMD feature enhancement module includes: taking the output features of the p1 shallow feature layer or p2 shallow feature layer input to the FEMD feature enhancement module as the FEMD input feature map, and allocating the FEMD input feature map to four feature branches; forming an equivalent mapping of the FEMD input feature map through the first feature branch using a residual structure to obtain a residual mapping feature map; performing a cascaded standard convolution operation with a kernel size of 1×3 on the FEMD input feature map through the second feature branch, and then... A 3×3 dilated convolutional layer is used to process the FEMD input feature map to obtain the first dilated convolutional feature map. A cascaded standard convolution operation with a kernel size of 3×1 is performed on the FEMD input feature map through the third feature branch, and then processed through a 3×3 dilated convolutional layer to obtain the second dilated convolutional feature map. A cascaded standard convolution operation with a kernel size of 3×3 is performed on the FEMD input feature map through the fourth feature branch to obtain the standard convolutional feature map. The residual mapping feature map, the first dilated convolutional feature map, the second dilated convolutional feature map, and the standard convolutional feature map are then fused to obtain the FEMD output feature map.

[0011] As a further improvement to the above technical solution, the calculation process of the Shape-IoU loss function includes: during the bounding box regression training of the improved YOLOv11n model, obtaining the predicted bounding boxes output by the improved YOLOv11n model and the ground truth bounding boxes corresponding to the training image data; determining a scale factor based on the size of the predicted bounding boxes and the ground truth bounding boxes relative to the image; determining a width shape factor based on the similarity between the predicted bounding boxes and the ground truth bounding boxes in the width direction; determining a height shape factor based on the similarity between the predicted bounding boxes and the ground truth bounding boxes in the height direction; calculating the bounding box regression loss based on the scale factor, width shape factor, height shape factor, and the overlap relationship between the predicted bounding boxes and the ground truth bounding boxes; and updating the parameters of the improved YOLOv11n model based on the bounding box regression loss.

[0012] As a further improvement to the above technical solution, the training process of the improved YOLOv11n model includes: During forward propagation, the input feature map in the downsampling feature extraction path of the backbone network is processed by the C3k2_StripCGLU module to obtain the C3k2 enhanced feature map. The C3k2 enhanced feature map is then passed along the backbone network to the end feature layer of the backbone network, and the high-level features output from the end feature layer of the backbone network are input into the C2PSA_EDFFN module to obtain the high-level enhanced feature map. The output features of the p1 shallow feature layer and the p2 shallow feature layer are input into the FEMD feature enhancement module to obtain the p1 shallow enhanced feature map and the p2 shallow enhanced feature map, respectively. The p1 shallow enhanced feature map, the p2 shallow enhanced feature map, and the high-level enhanced feature map are input into the neck network for multi-scale feature fusion to obtain the fused feature map. The fused feature map is then input into the detection head to generate the predicted bounding box. During backpropagation, the bounding box regression loss is calculated using the Shape-IoU loss function based on the predicted bounding box and the ground truth bounding box corresponding to the training image data, and the parameters of the improved YOLOv11n model are updated based on the bounding box regression loss.

[0013] This invention also provides a UAV image target detection system based on an improved YOLOv11n, comprising: The image acquisition module is used to acquire image data captured by the drone, and to obtain training image data for training the improved YOLOv11n model and test image data for detection from the image data. The model improvement unit is used to improve the YOLOv11n model to obtain an improved YOLOv11n model. The YOLOv11n model includes a backbone network, a neck network, and a detection head. The model improvement unit includes: a first module reconstruction unit, used to replace the Bottleneck structure within the C3k2 module of the model backbone network with a StripCGLU module based on strip convolution and gating mechanisms to construct the C3k2_StripCGLU module; and a second module replacement unit, used to replace the structure within the backbone network... The feedforward network FFN of the C2PSA module in the network is replaced with the EDFFN module that fuses frequency domain information to construct the C2PSA_EDFFN module; the feature enhancement unit is used to introduce the FEMD feature enhancement module based on multi-branch convolution and dilated convolution into the p1 shallow feature layer and p2 shallow feature layer in the backbone network; the loss function optimization unit is used to replace the GIoU and DIoU loss functions used by the YOLOv11n model in bounding box regression training with the Shape-IoU loss function that introduces scale factor and shape factor; The model training module is used to iteratively optimize and train the improved YOLOv11n model using the training image data to obtain the UAV image target detection model. The target detection module is used to detect the image data to be tested using the UAV image target detection model and output the target detection result.

[0014] Compared with existing technologies, the UAV image target detection method and system provided by this invention have the following advantages: By expanding the long-range spatial receptive field and reducing redundant computation through C3k2_StripCGLU, fusing spatial and frequency domain information through C2PSA_EDFFN, enhancing the texture and context features of shallow small targets in p1 and p2 by combining FEMD, and optimizing bounding box regression with Shape-IoU, the model can reduce the false negative rate while maintaining its lightweight nature, and improve the detection accuracy and localization stability of multi-scale, small and blurry targets in UAV images. Attached Figure Description

[0015] Figure 1 This is a flowchart of the UAV image target detection method based on YOLOv11n in an embodiment of the present invention; Figure 2 This is a block diagram of a UAV image target detection system based on YOLOv11n improved in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall structure of the improved YOLOv11n model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the basic structure related to C3k2 in the YOLOv11n model; Figure 5 This is a schematic diagram of the C2PSA module structure in the YOLOv11n model; Figure 6a This is a schematic diagram of the splicing and fusion path between the ordinary convolutional branch and the concatenated StripCGLU branch in the C3k2_StripCGLU module; Figure 6b This diagram illustrates the strip convolution branch, channel splicing, batch normalization, and CGLU activation output path in the StripCGLU structure. Figure 7 This is a schematic diagram of the internal structure of C2PSA; Figure 8 This is a schematic diagram of the EDFFN module structure; Figure 9 This is a schematic diagram of the FEMD feature enhancement module structure. Detailed Implementation

[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings. The following embodiments are used to illustrate the implementation of the present invention, and are not intended to limit the scope of protection of the present invention; in the absence of conflict, the technical features of the various embodiments can be combined with each other.

[0017] Example 1 like Figure 1 As shown, this embodiment provides a UAV image target detection method based on an improved YOLOv11n. First, image data captured by the UAV is acquired. Training image data for training the improved YOLOv11n model is obtained from this image data, and validation image data and test image data are generated according to training needs. For image data with annotation information, the target category, ground truth bounding box position, and size information are organized according to the format required by the YOLOv11n model. Then, the YOLOv11n model is improved to obtain an improved YOLOv11n model. The improved YOLOv11n model is iteratively optimized and trained using the training image data to obtain a UAV image target detection model. Finally, the test image data is input into the trained UAV image target detection model, and the target detection result is output.

[0018] Combination Figure 1The S1-S4 process shown in this embodiment can be further represented as follows: S1, acquire image data captured by the UAV, and filter, label, and organize the training image data for training the improved YOLOv11n model from the image data, forming verification image data and test image data when necessary; S2, improve the structure of the YOLOv11n model to obtain an improved YOLOv11n model containing the C3k2_StripCGLU module, C2PSA_EDFFN module, FEMD feature enhancement module, and Shape-IoU loss function; S3, use the training image data to iteratively optimize and train the improved YOLOv11n model until the UAV image target detection model is obtained; S4, use the trained UAV image target detection model to detect the test image data and output the target detection result.

[0019] In the above model improvement process, to address the issues of low pixel ratio of small targets, complex backgrounds, and easy attenuation of detail information during downsampling in UAV-captured images, this embodiment makes the following improvements to the YOLOv11n model: The Bottleneck structure within the C3k2 module of the model's backbone network is replaced with a StripCGLU module based on strip convolution and gating mechanisms to construct the C3k2_StripCGLU module; the feedforward network FFN within the C2PSA module of the backbone network is replaced with an EDFFN module that fuses frequency domain information to construct the C2PSA_EDFFN module; a FEMD feature enhancement module based on multi-branch convolution and dilated convolution is introduced into the p1 and p2 shallow feature layers of the backbone network; and the GIoU and DIoU loss functions used in the bounding box regression training of the YOLOv11n model are replaced with a Shape-IoU loss function that incorporates scale and shape factors.

[0020] Example 2 In this embodiment, the overall connectivity of the improved YOLOv11n model is described in detail. For example... Figure 3As shown, the improved YOLOv11n model comprises a backbone network, a neck network, and a detection head along the image data processing direction. The backbone network receives training image data and, through a stepwise downsampling feature extraction path including the C3k2_StripCGLU module, forms a p1 shallow feature layer, a p2 shallow feature layer, and a backbone network terminal feature layer. The shallow features output from the p1 and p2 shallow feature layers are processed by the FEMD feature enhancement module and then input as shallow enhancement features into the neck network. The high-level features output from the backbone network terminal feature layer are processed by the C2PSA_EDFFN module and then input as high-level enhancement features into the neck network. The neck network performs multi-scale feature fusion on the shallow and high-level enhancement features and inputs the fused feature map into the detection head; the detection head generates predicted bounding boxes based on the fused feature map.

[0021] During training, the labeled information of the training image data provides the true bounding boxes. The Shape-IoU loss function calculates the bounding box regression loss based on the predicted bounding boxes and the true bounding boxes, and updates the parameters of the improved YOLOv11n model through backpropagation. Figure 4 This shows the basic structure related to C3k2 in the original YOLOv11n. Figure 5 The overall structure of the original C2PSA module is shown in this embodiment. Figure 4 and Figure 5 The corresponding modules shown in the diagram are used as the basis for replacement and reconstruction.

[0022] Further integration Figure 2 and Figure 3 , Figure 2 The image acquisition module, model improvement unit, model training module, and object detection module in the model are respectively... Figure 1 The data acquisition, model structure improvement, iterative optimization training, and detection output stages are all part of the process. Figure 3 This demonstrates the feature transfer path between the backbone network, neck network, and detection head in the improved YOLOv11n model. In the backbone network, the C3k2_StripCGLU module replaces the C3k2 feature extraction unit in the stepwise downsampling path; the C2PSA_EDFFN module is positioned at a critical location before high-level semantic features enter the neck network; the FEMD feature enhancement module is positioned at the output of the p1 and p2 shallow feature layers, first enhancing the high-resolution shallow texture before inputting it into the neck network for multi-scale fusion. Figure 4 and Figure 5 The replaced or reconstructed C3k2 and C2PSA infrastructures are shown respectively, making the replacement locations, input-output relationships and inter-module connection boundaries clearer in this embodiment.

[0023] In one optional implementation, after the input image enters the backbone network at a resolution of 640×640, the shallow to high-level features can sequentially form feature maps of scales such as 320×320, 160×160, 80×80, and 40×40; where the p1 shallow feature layer corresponds to 320×320 scale features, the p2 shallow feature layer corresponds to 160×160 scale features, and the feature layer at the end of the backbone network corresponds to 40×40 high-level semantic features. The C3k2_StripCGLU module maintains the output feature map size and channel connection relationship unchanged in each downsampling stage, the C2PSA_EDFFN module performs frequency domain enhancement on the high-level semantic features, and the FEMD feature enhancement module retains more edge, texture, and local contextual information before the shallow features enter the neck network, thereby improving the small object detection capability without disrupting the overall inference path of YOLOv11n.

[0024] Example 3 This embodiment further illustrates the specific operation process of the C3k2_StripCGLU module. For example... Figure 6a As shown, in the C3k2_StripCGLU module, the feature map input to this module is used as the C3k2 input feature map, which is then divided into a first input sub-feature map and a second input sub-feature map. The first input sub-feature map is processed by convolution to obtain a first convolutional feature map; the second input sub-feature map undergoes depth feature extraction through multiple cascaded StripCGLU structures to obtain a second depth feature map. Subsequently, the first convolutional feature map and the second depth feature map are fused to obtain the output feature map of the C3k2_StripCGLU module. Through this structure, while retaining the channel segmentation and fusion path of the C3k2 module, the original Bottleneck structure is replaced with a StripCGLU structure that can capture long-range spatial information.

[0025] like Figure 6b As shown, the processing steps for any StripCGLU structure include: using the feature map of the input StripCGLU structure as the input feature map of StripCGLU, and assigning the StripCGLU input feature map to the first strip convolution branch and the second strip convolution branch to form the first branch input feature map and the second branch input feature map; processing the first branch input feature map with a 1×1 strip convolution to obtain the first strip convolution feature map; processing the second branch input feature map with a 19×1 strip convolution to obtain the second strip convolution feature map; concatenating the first strip convolution feature map and the second strip convolution feature map in the channel dimension to obtain the concatenated feature map; performing batch normalization on the concatenated feature map to obtain the normalized feature map; and inputting the normalized feature map into the CGLU activation function to obtain the StripCGLU output feature map.

[0026] The formula for calculating the CGLU activation function is: CGLU(x) = σ(Conv(x)) Conv(x), where x is the feature map of the input CGLU activation function, σ is the sigmoid function, and Conv is the convolution operation. This represents element-wise multiplication. This gating mechanism uses the Sigmoid function to generate a weight map and dynamically weights the feature map through element-wise multiplication, enabling the model to enhance the expression of important feature channels and long-range spatial dependencies while reducing the number of parameters.

[0027] from Figure 6a and Figure 6b As shown in the flowchart, the C3k2_StripCGLU module retains the splitting and fusion ideas of C3k2: one part of the features are quickly passed through ordinary convolution to maintain lightweight computation; the other part of the features are extracted through a concatenated StripCGLU structure to extract deeper directional spatial information. Compared with the local convolution path of "compression-convolution-expansion" in the original Bottleneck, StripCGLU expands the effective receptive field through strip convolution in the long direction, which can better capture the continuous structure of slender or low-pixel-ratio targets such as vehicles and pedestrians in UAV top-down images; at the same time, CGLU gating adaptively generates a weight map according to the input features, strengthens important channels and suppresses redundant responses, thereby reducing the computational overhead caused by repeated convolution and reducing the missed detection of small targets.

[0028] The C3k2_StripCGLU module serves as the core feature extraction unit in the backbone network. Its output enhanced features continue to propagate along the backbone network and provide a more sufficient spatial structure foundation for the frequency domain analysis of the subsequent C2PSA_EDFFN module. In other words, strip convolution first strengthens long-distance dependencies in the spatial domain, and EDFFN then performs frequency component filtering on high-level semantic features. The two form a continuous enhancement relationship in the forward feature flow.

[0029] Example 4 In this embodiment, the operation process of the C2PSA_EDFFN module is described in detail. Figure 5 and Figure 7 As shown, the original C2PSA module includes processing paths such as channel segmentation, PSABlock stacking, concatenation, and output convolution. The PSABlock internally includes an attention module and a traditional feedforward network (FFN). This embodiment, while maintaining the basic input-output paths of the C2PSA module, replaces the traditional feedforward network (FFN) in the C2PSA module with an EDFFN module, forming the C2PSA_EDFFN module.

[0030] like Figure 8As shown, the processing steps of the EDFFN module include: using the feature map input to the EDFFN module as the EDFFN input feature map; inputting the EDFFN input feature map into a 1×1 pointwise convolution to expand the number of channels, resulting in a channel-expanded feature map; inputting the channel-expanded feature map into a 3×3 depthwise separable convolution to obtain a local neighborhood feature map; dividing the local neighborhood feature map into a first segmentation feature x1 and a second segmentation feature x2; performing GELU activation on the first segmentation feature x1 to obtain a gated control feature GELU(x1); and multiplying the gated control feature GELU(x1) element-wise with the second segmentation feature x2 to obtain a gated fusion feature map.

[0031] Subsequently, the gated fusion feature map is segmented into multiple 8×8 local image patches. A two-dimensional Fast Fourier Transform (FFT) is performed on each 8×8 local image patch to map the features from the spatial domain to the frequency domain, resulting in a frequency domain feature block. The frequency domain feature block is then element-wise multiplied with a learnable complex tensor parameter W to obtain a filtered frequency domain feature block. An Inverse Fast Fourier Transform (IFT) is performed on the filtered frequency domain feature block to restore it to the spatial domain, resulting in a spatial domain reconstructed feature block. The spatial domain reconstructed feature blocks are then reassembled into the original resolution feature map, and a 1×1 convolution is used to restore the number of channels to the target dimension, yielding the EDFFN output feature map. Through this processing, the C2PSA_EDFFN module can introduce frequency domain filtering based on local neighborhood information in the spatial domain, enabling the model to better express the edges and textures of small targets and the global correlations of blurred targets.

[0032] exist Figure 7 In the internal path of C2PSA shown, EDFFN replaces the feedforward network FFN part in the original PSABlock. Therefore, the original channel segmentation, attention modeling, concatenation and output convolution paths of the C2PSA module can still remain stable. Figure 8 The frequency domain enhancement process of EDFFN is further demonstrated: 1×1 pointwise convolution is used for channel expansion, 3×3 depthwise separable convolution is used to supplement local neighborhood context, GELU(x1)×x2 is used to form gated fusion features, and FFT and inverse FFT of 8×8 local image patches are used for learnable filtering and spatial domain reconstruction in the frequency domain. Figure 8 In this context, DConv represents depthwise convolution, PConv represents pointwise convolution, FFT represents Fourier transform, IFFT represents inverse Fourier transform, W represents the quantization matrix, and ⊙ represents element-wise multiplication. This represents the GELU function.

[0033] Because the 2D Fast Fourier Transform (FFN) can utilize the overall pixel distribution of each 8×8 local image patch simultaneously, the learnable complex tensor parameter W can adaptively preserve high-frequency components such as edges and textures based on training feedback, while suppressing redundant frequency responses related to complex backgrounds. For small, occluded, or blurred targets commonly found in UAV images, this frequency-spatial domain fusion mechanism can overcome the shortcomings of traditional FFN, which only performs spatial domain channel mixing, has a limited receptive field, and suffers from easy attenuation of high-frequency information, thereby enhancing the global receptive field and long-range dependency representation.

[0034] Example 5 In this embodiment, the working mode of the FEMD feature enhancement module is further explained. For example... Figure 9 As shown, the FEMD feature enhancement module is located at the output of the p1 shallow feature layer and the p2 shallow feature layer, and is used to enhance features before the shallow features are input into the neck network. Specifically, the output features of the p1 shallow feature layer or the p2 shallow feature layer input to the FEMD feature enhancement module are used as the FEMD input feature map, and the FEMD input feature map is assigned to four feature branches.

[0035] The first feature branch uses a residual structure to form an equivalent mapping on the FEMD input feature map, resulting in a residual mapping feature map. The second feature branch performs a cascaded standard convolution operation with a kernel size of 1×3 on the FEMD input feature map, followed by a 3×3 dilated convolution layer, to obtain the first dilated convolution feature map. The third feature branch performs a cascaded standard convolution operation with a kernel size of 3×1 on the FEMD input feature map, followed by a 3×3 dilated convolution layer, to obtain the second dilated convolution feature map. The fourth feature branch performs a cascaded standard convolution operation with a kernel size of 3×3 on the FEMD input feature map, resulting in a standard convolution feature map. The residual mapping feature map, the first dilated convolution feature map, the second dilated convolution feature map, and the standard convolution feature map are then fused to obtain the FEMD output feature map. Since the FEMD module does not change the feature map size, it can preserve shallow texture and local contextual information before inputting to the neck network, enhancing the feature representation capability of small targets in UAV images.

[0036] exist Figure 9 In the FEMD structure shown, the p1 shallow feature layer and p2 shallow feature layer can each correspond to a higher resolution shallow output. For example, when the p1 output is 320×320×64 and the p2 output is 160×160×128, the output after FEMD processing still maintains the same spatial size and channel connectivity, thus allowing it to be directly fed into the neck network. This setting enhances the texture, edge, and local context information of small targets before multi-scale fusion, preventing shallow details from being prematurely weakened during subsequent downsampling and fusion processes.

[0037] Specifically, the residual branch preserves the original key information in the shallow layer, the 1×3 and 3×1 branches enhance the horizontal and vertical local structures respectively, the 3×3 branch supplements the regular neighborhood semantics, and the two middle branches are superimposed with a 3×3 dilated convolution to expand the receptive field without changing the feature map size. Thus, the FEMD module balances shallow detail preservation and local context expansion, making it particularly suitable for solving the problems of low pixel ratio of small targets, weak shallow semantics, and target edges being easily buried by the background in UAV images.

[0038] Example 6 This embodiment uses the Shape-IoU loss function instead of the GIoU and DIoU loss functions used in the bounding box regression training of the YOLOv11n model. The calculation method for the Shape-IoU loss function is as follows: In the above formula (1), The loss is the Shape-IoU bounding box regression loss. The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. Indicates overlap loss; For shape perception center distance loss, For the loss of width and height shape, These are the weighting coefficients for the shape loss term.

[0039] In the above formulas (2) and (3), Represents a real bounding box; and These are the width and height of the actual bounding box, respectively; Scale factor; and These are the width and height shape weights determined by the actual frame width and height, which are used to apply scale weights to the subsequent center distance and width-height difference terms.

[0040] In the above formulas (4) to (6), , The coordinates of the center of the prediction box. , The coordinates of the center of the true bounding box. The diagonal length of the minimum bounding box between the predicted box and the ground truth box is given in formula (4). Normalization center shift; , To predict the width and height of the bounding box, , For the actual frame width and height, and Used to normalize the difference between width and height; , These represent the normalized shape differences in the width and height directions, respectively. Indicates direction Differences in shape on; The shape loss adjustment parameter is set to 4. This represents the summation of shape losses over the width and height directions, respectively.

[0041] By introducing a scale factor and shape factor , , It can perceive the size of the bounding box relative to the image and dynamically adjust the loss weights on small target samples, amplifying the effective gradient signal and guiding the model to focus more on the localization of tiny targets; simultaneously, Explicitly calculating the shape similarity between the predicted bounding box and the ground truth bounding box in the width and height directions provides a more accurate gradient direction for bounding box regression.

[0042] Compared to GIoU and DIoU losses, which are based solely on overlap area or center distance constraints, Shape-IoU incorporates target scale and width / height shape differences into the regression optimization process. When there are size discrepancies between the predicted and ground truth bounding boxes, target occlusion, or unclear labeled boundaries, this loss function can redistribute training attention through scale and shape factors, alleviating the problem of insufficient gradients for small target localization and improving the stability of bounding box fitting in complex aerial photography contexts.

[0043] During bounding box regression training, the predicted bounding boxes output by the improved YOLOv11n model and the ground truth bounding boxes corresponding to the training image data are obtained. The scale factor is determined based on the size of the predicted and ground truth bounding boxes relative to the image. The width shape factor is determined based on the similarity between the predicted and ground truth bounding boxes in the width direction, and the height shape factor is determined based on the similarity between the predicted and ground truth bounding boxes in the height direction. The bounding box regression loss is calculated based on the scale factor, width shape factor, height shape factor, and the overlap relationship between the predicted and ground truth bounding boxes, and the parameters of the improved YOLOv11n model are updated based on this bounding box regression loss.

[0044] During forward propagation, the C3k2_StripCGLU module processes the input feature map in the downsampling feature extraction path of the backbone network to obtain the C3k2 enhanced feature map. The C3k2 enhanced feature map is then passed along the backbone network to the end feature layer, and the high-level features output from the end feature layer are input into the C2PSA_EDFFN module to obtain the high-level enhanced feature map. The output features of the p1 and p2 shallow feature layers are input into the FEMD feature enhancement module to obtain the p1 and p2 shallow enhanced feature maps, respectively. The p1, p2, and high-level enhanced feature maps are then input into the neck network for multi-scale feature fusion to obtain the fused feature map. The fused feature map is then input into the detection head to generate the predicted bounding box. During backpropagation, the Shape-IoU loss function calculates the bounding box regression loss based on the predicted and ground truth bounding boxes, and feeds the corresponding gradients back to the aforementioned feature extraction, frequency domain enhancement, and shallow enhancement paths, thereby enabling each module to jointly optimize around small target feature representation and bounding box localization.

[0045] Furthermore, during training, the spatial augmentation features output by C3k2_StripCGLU provide a richer long-range structural basis for the frequency domain filtering of C2PSA_EDFFN; the high-level augmentation features obtained by C2PSA_EDFFN after frequency domain filtering are fused with the shallow augmentation features output by FEMD in the neck network, making the high-level semantic information and shallow texture details complementary; the Shape-IoU loss function allocates gradients according to the bounding box scale and shape error during the backpropagation stage, so that the above spatial domain augmentation, frequency domain augmentation and shallow augmentation paths all update parameters around small target recall and localization accuracy.

[0046] Example 7 like Figure 2 As shown, this embodiment also provides a UAV image target detection system based on improved YOLOv11n, including an image acquisition module, a model improvement unit, a model training module, and a target detection module. The image acquisition module is used to acquire image data captured by the UAV, and obtain training image data for training the improved YOLOv11n model and test image data for detection from the image data.

[0047] The model improvement unit is used to improve the YOLOv11n model, resulting in an improved YOLOv11n model. The model improvement unit includes a first module reconstruction unit, a second module replacement unit, a feature enhancement unit, and a loss function optimization unit. Specifically, the first module reconstruction unit replaces the Bottleneck structure within the C3k2 module of the model's backbone network with a StripCGLU module based on strip convolution and gating mechanisms to construct the C3k2_StripCGLU module; the second module replacement unit replaces the feedforward network FFN of the C2PSA module in the backbone network with an EDFFN module that fuses frequency domain information to construct the C2PSA_EDFFN module; the feature enhancement unit introduces a FEMD feature enhancement module based on multi-branch convolution and dilated convolution into the p1 and p2 shallow feature layers of the backbone network; and the loss function optimization unit replaces the GIoU and DIoU loss functions used in the bounding box regression training of the YOLOv11n model with a Shape-IoU loss function that incorporates scale and shape factors.

[0048] The model training module is used to iteratively optimize and train the improved YOLOv11n model using training image data to obtain a UAV image target detection model. The target detection module is used to detect targets in the test image data using the UAV image target detection model and output the target detection results. Each module in the above system can be implemented through software programs, hardware circuits, or a combination of both, and its execution process corresponds to the method embodiments described above.

[0049] Combination Figure 2 The system module diagram shown illustrates that before the training image data output by the image acquisition module enters the model training module, the model improvement unit first determines the structure of the improved YOLOv11n model. The model training module performs iterative optimization based on the training image data and labeled bounding boxes to obtain the UAV image target detection model. During the inference phase, the target detection module calls the trained model and outputs the target detection results for the image data to be tested. Thus, each functional module in the system embodiment has a one-to-one correspondence with the S1-S4 processes in the method embodiment.

[0050] Example 8 To illustrate the function of the above modules, ablation experiments were conducted on the C3k2_StripCGLU module, C2PSA_EDFFN module, FEMD module, and Shape-IoU loss function while maintaining consistent training data and configuration. Evaluation metrics included Recall, mAP50 (mean precision @ 50), mAP50-95 (mean precision @ 50-95), Params (number of parameters), and FLOPs (floating-point operations). The experimental results are as follows.

[0051] Table 1 Ablation Experiment Results of C3k2_StripCGLU Module As shown in Table 1, compared with the C3k2 structure, the C3k2_StripCGLU structure has improved Recall, mAP50 and mAP50-95, while Params and FLOPs are reduced, indicating that strip convolution and gating mechanism can enhance feature extraction capability while reducing computational cost.

[0052] Specifically, the Recall of the C3k2_StripCGLU structure increased from 63.4% to 65.2%, an improvement of 1.8 percentage points, indicating a reduction in the model's false negative rate; mAP50 increased from 33.2% to 34.6%, and mAP50-95 increased from 19.1% to 20.3%; meanwhile, Params decreased from 2.6M to 2.34M, a reduction of approximately 9.9%, and FLOPs decreased from 6.4G to 5.9G, a reduction of approximately 7.8%. These results demonstrate that strip convolution and gating mechanisms can simultaneously compress the number of parameters and computational cost while improving feature representation capabilities.

[0053] Table 2 Ablation Experiment Results of C2PSA_EDFFN Module As shown in Table 2, compared with the C2PSA structure, the C2PSA_EDFFN structure has improvements in Recall, mAP50, and mAP50-95, indicating that the feature enhancement processing of frequency domain and spatial domain fusion can improve the ability to express occluded, blurred, and small targets.

[0054] Specifically, the Recall of the C2PSA_EDFFN structure increased from 63.4% to 66.1%, a 2.7 percentage point improvement; mAP50 increased from 33.2% to 34.8%, and mAP50-95 increased from 19.1% to 20.7%. Although Params increased slightly from 2.6M to 2.64M and FLOPs increased from 6.4G to 6.7G, the increases were small, indicating that the enhanced representation of occluded targets, blurred targets, and targets with weak textures brought about by frequency domain filtering has a high performance benefit.

[0055] Table 3 Ablation Experiment Results of FEMD Module As shown in Table 3, the introduction of the FEMD module significantly improves Recall and mAP50-95, while Params and FLOPs only increase slightly. This indicates that multi-branch convolution and dilated convolution enhancement in the shallow feature layers p1 and p2 are beneficial for preserving the texture details and local context information of small targets.

[0056] Specifically, after introducing the FEMD module, Recall increased from 63.4% to 67.4%, an improvement of 4.0 percentage points; mAP50 increased from 33.2% to 35.2%; and mAP50-95 increased from 19.1% to 21.3%, an improvement of 2.2 percentage points. Meanwhile, Params only increased from 2.6M to 2.62M, an increase of approximately 0.8%, and FLOPs increased from 6.4G to 6.5G, an increase of approximately 1.6%. These results indicate that the p1 and p2 shallow enhancements significantly contribute to small target recall and localization accuracy, while having a relatively small impact on lightweight deployment.

[0057] Table 4 Results of Loss Function Ablation Experiment As shown in Table 4, the Shape-IoU loss function outperforms other listed IoU loss functions on both mAP50 and mAP50-95, indicating that constraining the bounding box regression by using scale and shape factors can improve the localization accuracy of small targets in UAV images.

[0058] Specifically, the mAP50 for Shape-IoU is 34.7%, and the mAP50-95 is 20.6%. Its mAP50-95 is 1.0 percentage point higher than PIoUv2's 19.6%, and 1.5 percentage points higher than the original YOLOv11n's 19.1%. This result indicates that by incorporating the scale factor and width / height shape factor into the bounding box regression, the model can more stably handle variations in the size and shape of small target bounding boxes.

[0059] Table 5 Ablation Experiment Results of Each Module Combination As shown in Table 5, when the C3k2_StripCGLU module, C2PSA_EDFFN module, FEMD module and Shape-IoU loss function work together, Recall, mAP50 and mAP50-95 reach the highest values ​​in the table. At the same time, Params are lower than the original baseline model, and FLOPs remain within the range suitable for lightweight detection. This indicates that the above modules have a synergistic effect in feature enhancement, frequency domain filtering, shallow small target preservation and bounding box localization optimization.

[0060] The combined results show that the recall of the complete combination reaches 69.8%, an improvement of 6.4 percentage points compared to the benchmark of 63.4%; mAP50 reaches 35.8%, and mAP50-95 reaches 21.9%. Meanwhile, the params of the complete combination are 2.36M, lower than the benchmark of 2.6M, a reduction of approximately 9.2%; FLOPs are 6.2G, still within the acceptable range for lightweight detection. These results indicate that C3k2_StripCGLU mainly improves spatial feature extraction and reduces redundant computation, C2PSA_EDFFN mainly enhances frequency domain and global dependency representation, FEMD mainly improves the preservation of details of shallow small targets, and Shape-IoU enhances the localization of small targets from the regression loss perspective. These modules form a synergistic optimization effect that is not simply an additive combination.

[0061] This paper further explains the differences in performance between the present invention and conventional YOLOv11n modification schemes. Traditional FFN modules in C2PSA typically focus on spatial domain channel mixing, lacking explicit frequency domain modeling, and are prone to losing high-frequency information such as edges and textures after multiple downsampling. The present invention combines spatial domain local convolution, gated fusion, 8×8 local block fast Fourier transform, learnable frequency domain filtering, and inverse transform reconstruction through C2PSA_EDFFN, enabling features to be repeatedly enhanced between the frequency and spatial domains, thereby improving the ability to recognize small targets, low-contrast targets, and blurred boundaries in complex backgrounds.

[0062] The Bottleneck structure in the traditional C3k2 module has a limited local convolutional receptive field, making it difficult to fully capture slender targets and long-distance spatial dependencies in UAV overhead scenes. This invention combines ordinary convolutional branches and strip convolutional branches in parallel using C3k2_StripCGLU, and dynamically weights important channels through CGLU gating. This expands the effective receptive field while controlling the number of parameters, thus solving the balance problem between model computational complexity and feature representation capability.

[0063] For small target enhancement methods that simply add convolutional layers in shallow layers, the enhancement effect is limited or the number of additional parameters is too large. The FEMD module of this invention uses residual mapping, 1×3 convolution, 3×1 convolution, 3×3 convolution and dilated convolution to work together to expand the local context receptive field without changing the size of p1 and p2 feature maps. This preserves the key details of small targets and reduces the impact on inference speed.

[0064] Conventional bounding box regression losses such as GIoU and DIoU are insufficient in adapting to the scale, width, height, shape differences, and blurred boundaries of small targets in UAV images. This invention adopts Shape-IoU, which dynamically adjusts the loss weight of small targets through scale factors and shape factors, amplifies the localization gradient contribution of small targets, and enables the model to obtain a more stable bounding box regression direction even under conditions of occlusion, low contrast, or perturbation of labeled boundaries.

[0065] Therefore, the technical effect of this invention is not an isolated improvement resulting from the replacement of a single module, but rather the result of the combined effects of long-distance spatial dependence enhancement, high-frequency information preservation in the frequency domain, shallow small target detail enhancement, and shape perception regression optimization. It can balance the requirements of detection accuracy, positioning stability, and lightweight deployment in scenarios such as UAV aerial photography and mapping, power line inspection, traffic monitoring, and emergency search and rescue.

[0066] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A UAV image target detection method based on YOLOv11n improvement, characterized in that, include: Acquire image data captured by drones, and obtain training image data from the image data for training the improved YOLOv11n model; The YOLOv11n model is improved to obtain an improved YOLOv11n model. The improvement process includes: replacing the Bottleneck structure in the C3k2 module of the model backbone network with a StripCGLU module based on strip convolution and gating mechanism to construct the C3k2_StripCGLU module; replacing the feedforward network FFN in the C2PSA module of the backbone network with an EDFFN module that integrates frequency domain information to construct the C2PSA_EDFFN module; introducing a FEMD feature enhancement module based on multi-branch convolution and dilated convolution into the p1 shallow feature layer and p2 shallow feature layer of the backbone network; and replacing the GIoU and DIoU loss functions used by the YOLOv11n model in bounding box regression training with a Shape-IoU loss function that introduces scale factor and shape factor. The improved YOLOv11n model is iteratively optimized and trained using the training image data to obtain a UAV image target detection model. The UAV image target detection model is then used to detect the image data to be tested, and the target detection results are output.

2. The UAV image target detection method based on YOLOv11n as described in claim 1, characterized in that, The improved YOLOv11n model includes a backbone network, a neck network, and a detection head along the image data processing direction; The backbone network receives the training image data and, through a stepwise downsampling feature extraction path including the C3k2_StripCGLU module, forms a p1 shallow feature layer, a p2 shallow feature layer, and a backbone network terminal feature layer. The shallow features output by the p1 and p2 shallow feature layers are processed by the FEMD feature enhancement module and then input as shallow enhancement features into the neck network. The high-level features output by the backbone network terminal feature layer are processed by the C2PSA_EDFFN module and then input as high-level enhancement features into the neck network. The neck network performs multi-scale feature fusion on the shallow enhancement features and the high-level enhancement features, and inputs the fused feature map into the detection head; the detection head generates a predicted bounding box based on the fused feature map. During training, the annotation information of the training image data provides the true bounding boxes, and the Shape-IoU loss function calculates the bounding box regression loss based on the predicted bounding boxes and the true bounding boxes.

3. The UAV image target detection method based on YOLOv11n as described in claim 1, characterized in that, The processing steps of the C3k2_StripCGLU module include: using the feature map input to the C3k2_StripCGLU module as the C3k2 input feature map, and dividing the C3k2 input feature map into a first input sub-feature map and a second input sub-feature map; performing convolution processing on the first input sub-feature map to obtain a first convolutional feature map; performing depth feature extraction on the second input sub-feature map through multiple cascaded StripCGLU structures to obtain a second depth feature map; and fusing the first convolutional feature map and the second depth feature map to obtain the output feature map of the C3k2_StripCGLU module. The processing procedure for any of the StripCGLU structures includes: using the feature map input to the StripCGLU structure as the StripCGLU input feature map, and assigning the StripCGLU input feature map to a first strip convolution branch and a second strip convolution branch to form a first branch input feature map and a second branch input feature map; processing the first branch input feature map with a 1×1 strip convolution to obtain a first strip convolution feature map; processing the second branch input feature map with a 19×1 strip convolution to obtain a second strip convolution feature map; concatenating the first strip convolution feature map and the second strip convolution feature map in the channel dimension to obtain a concatenated feature map; performing batch normalization on the concatenated feature map to obtain a normalized feature map; and inputting the normalized feature map into the CGLU activation function to obtain the StripCGLU output feature map. The formula for calculating the CGLU activation function is: CGLU(x) = σ(Conv(x)) Conv(x), where x is the feature map input to the CGLU activation function, σ is the Sigmoid function, and Conv is the convolution operation. This indicates element-wise multiplication.

4. The UAV image target detection method based on YOLOv11n improvement according to claim 1, characterized in that, The processing steps of the EDFFN module in the C2PSA_EDFFN module include: using the feature map input to the EDFFN module as the EDFFN input feature map; inputting the EDFFN input feature map into a 1×1 pointwise convolution to expand the number of channels, obtaining a channel-expanded feature map; inputting the channel-expanded feature map into a 3×3 depthwise separable convolution to obtain a local neighborhood feature map; segmenting the local neighborhood feature map into a first segmentation feature x1 and a second segmentation feature x2; performing GELU activation processing on the first segmentation feature x1 to obtain a gated control feature GELU(x1); and combining the gated control feature GELU(x1) with the second segmentation feature x2.

2. Perform element-wise multiplication to obtain a gated fusion feature map; divide the gated fusion feature map into multiple 8×8 local image blocks; perform a two-dimensional fast Fourier transform on each of the 8×8 local image blocks to map the features from the spatial domain to the frequency domain, obtaining a frequency domain feature block; multiply the frequency domain feature block element-wise with the learnable complex tensor parameter W to obtain a filtered frequency domain feature block; perform an inverse fast Fourier transform on the filtered frequency domain feature block to restore the filtered frequency domain feature block back to the spatial domain, obtaining a spatial domain reconstructed feature block; reassemble the spatial domain reconstructed feature blocks into the original resolution feature map, and restore the number of channels to the target dimension through 1×1 convolution to obtain the EDFFN output feature map.

5. The UAV image target detection method based on YOLOv11n improvement according to claim 1, characterized in that, The processing steps of the FEMD feature enhancement module include: taking the output features of the p1 shallow feature layer or p2 shallow feature layer of the FEMD feature enhancement module as the FEMD input feature map, and allocating the FEMD input feature map to four feature branches; using a residual structure in the first feature branch to form an equivalent mapping of the FEMD input feature map, obtaining a residual mapping feature map; and performing a cascaded standard convolution operation with a kernel size of 1×3 on the FEMD input feature map through the second feature branch, followed by a 3×3 dilated convolution layer. The process involves processing the input feature map to obtain a first dilated convolutional feature map; performing a cascaded standard convolution operation with a kernel size of 3×1 on the FEMD input feature map through a third feature branch, followed by processing through a 3×3 dilated convolutional layer to obtain a second dilated convolutional feature map; performing a cascaded standard convolution operation with a kernel size of 3×3 on the FEMD input feature map through a fourth feature branch to obtain a standard convolutional feature map; and fusing the residual mapping feature map, the first dilated convolutional feature map, the second dilated convolutional feature map, and the standard convolutional feature map to obtain the FEMD output feature map.

6. The UAV image target detection method based on YOLOv11n improvement according to claim 1, characterized in that, The calculation process of the Shape-IoU loss function includes: during the bounding box regression training of the improved YOLOv11n model, obtaining the predicted bounding boxes output by the improved YOLOv11n model and the ground truth bounding boxes corresponding to the training image data; determining a scale factor based on the size of the predicted bounding boxes and the ground truth bounding boxes relative to the image; determining a width shape factor based on the similarity between the predicted bounding boxes and the ground truth bounding boxes in the width direction; determining a height shape factor based on the similarity between the predicted bounding boxes and the ground truth bounding boxes in the height direction; calculating the bounding box regression loss based on the scale factor, width shape factor, height shape factor, and the overlap relationship between the predicted bounding boxes and the ground truth bounding boxes; and updating the parameters of the improved YOLOv11n model based on the bounding box regression loss.

7. The UAV image target detection method based on YOLOv11n improvement according to claim 2, characterized in that, The training process of the improved YOLOv11n model includes: During forward propagation, the input feature map in the downsampling feature extraction path of the backbone network is processed by the C3k2_StripCGLU module to obtain the C3k2 enhanced feature map. The C3k2 enhanced feature map is then passed along the backbone network to the end feature layer of the backbone network, and the high-level features output from the end feature layer of the backbone network are input into the C2PSA_EDFFN module to obtain the high-level enhanced feature map. The output features of the p1 shallow feature layer and the p2 shallow feature layer are input into the FEMD feature enhancement module to obtain the p1 shallow enhanced feature map and the p2 shallow enhanced feature map, respectively. The p1 shallow enhanced feature map, the p2 shallow enhanced feature map, and the high-level enhanced feature map are input into the neck network for multi-scale feature fusion to obtain the fused feature map. The fused feature map is then input into the detection head to generate the predicted bounding box. During backpropagation, the bounding box regression loss is calculated using the Shape-IoU loss function based on the predicted bounding box and the ground truth bounding box corresponding to the training image data, and the parameters of the improved YOLOv11n model are updated based on the bounding box regression loss.

8. A UAV image target detection system based on an improved YOLOv11n, characterized in that, include: The image acquisition module is used to acquire image data captured by the drone, and to obtain training image data for training the improved YOLOv11n model and test image data for detection from the image data. The model improvement unit is used to improve the YOLOv11n model to obtain an improved YOLOv11n model. The YOLOv11n model includes a backbone network, a neck network, and a detection head. The model improvement unit includes: a first module reconstruction unit, used to replace the Bottleneck structure within the C3k2 module of the model backbone network with a StripCGLU module based on strip convolution and gating mechanisms to construct the C3k2_StripCGLU module; and a second module replacement unit, used to replace the structure within the backbone network... The feedforward network FFN of the C2PSA module in the network is replaced with the EDFFN module that fuses frequency domain information to construct the C2PSA_EDFFN module; the feature enhancement unit is used to introduce the FEMD feature enhancement module based on multi-branch convolution and dilated convolution into the p1 shallow feature layer and p2 shallow feature layer in the backbone network; the loss function optimization unit is used to replace the GIoU and DIoU loss functions used by the YOLOv11n model in bounding box regression training with the Shape-IoU loss function that introduces scale factor and shape factor; The model training module is used to iteratively optimize and train the improved YOLOv11n model using the training image data to obtain the UAV image target detection model. The target detection module is used to detect the image data to be tested using the UAV image target detection model and output the target detection result.

Images