An unmanned aerial vehicle aerial small target detection method based on an improved YOLOv11s algorithm

Abstract

The application belongs to the technical field of computer vision and target detection, and discloses a kind of unmanned aerial vehicle aerial small target detection method based on improved YOLOv11s algorithm, including constructing unmanned aerial vehicle aerial image dataset, constructing unmanned aerial vehicle aerial small target detection model, optimizing unmanned aerial vehicle aerial small target detection model, setting the training parameters of unmanned aerial vehicle aerial small target detection model, using training set to train the improved unmanned aerial vehicle aerial small target detection model, in the training process, gradually save training weight, after training, select the model weight of the minimum verification set loss function and load into unmanned aerial vehicle aerial small target detection model, and evaluate with test set.The application improves the detection and positioning ability of small target, and realizes the improvement of model on average precision mAP.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and object detection, and specifically relates to a UAV aerial small target detection method based on an improved YOLOv11s algorithm. BACKGROUND

[0002] In recent years, unmanned aerial vehicle technology has developed rapidly and has been widely applied in many fields, such as agricultural monitoring, disaster rescue, security patrol and environmental monitoring. Unmanned aerial vehicles have unique advantages in aerial detection tasks due to their high mobility, low cost and real-time data acquisition capabilities. Under the promotion of computer vision and deep learning, unmanned aerial vehicle aerial target detection has become a popular research direction. Unmanned aerial vehicles can carry high-resolution cameras for high-altitude photography in complex environments and achieve accurate identification of ground objects in combination with target detection algorithms. However, due to the complexity of the unmanned aerial vehicle aerial environment, such as changes in light, occlusion, and scale changes, target detection faces many challenges.

[0003] Currently, the mainstream target detection methods mainly fall into two categories: traditional computer vision-based methods and deep learning-based methods.

[0004] (1) Traditional methods: Early target detection mainly relies on handcrafted feature extraction and classifiers, such as SIFT, HOG, SURF features combined with SVM, Adaboost for target recognition. However, these methods have limited effectiveness for target detection in complex backgrounds and are difficult to adapt to the changing environment of unmanned aerial vehicle aerial photography.

[0005] (2) Deep learning methods: Deep learning methods mainly fall into two categories: two-stage algorithms: such as Faster R-CNN, which first generates candidate regions through a region proposal network (RPN), then performs fine classification and bounding box regression on these regions. Two-stage methods usually have high detection accuracy and are suitable for fine detection tasks, but due to the need for multiple forward propagations, the inference speed is slow, making it difficult to meet the real-time requirements of unmanned aerial vehicle aerial photography. One-stage algorithms: such as YOLO and SSD, these methods directly extract features from input images and perform target detection, compared with two-stage methods, have higher detection speed. YOLO series algorithms have been widely applied in unmanned aerial vehicle target detection tasks due to their end-to-end training, high-speed inference and good accuracy. For example, YOLOv3 improves small target detection capability through multi-scale prediction, YOLOv5 further optimizes detection accuracy and inference speed, and YOLOv7 improves detection performance through efficient model structure optimization.

[0006] Although existing research has made some progress in UAV aerial target detection, there are still the following challenges: (1) Limited small target detection capability: due to the high height of UAV shooting, the target occupies a small proportion in the image, resulting in weak feature information of small targets, and existing models are still difficult to accurately identify small targets; (2) Insufficient feature fusion: although existing research has introduced FPN, PAN and other structures for feature fusion, in the small target detection task, deep features are prone to lose local information, and shallow features are difficult to express high-level semantics, resulting in that the small target detection effect still needs to be improved; (3) The 5 times down-sampling of YOLO algorithm has certain advantages for large target detection. Down-sampling can reduce the amount of calculation, improve the speed, and at the same time help to capture the overall features of large targets and improve the detection accuracy; however, for small targets, excessive down-sampling brings significant disadvantages: multiple down-sampling causes the details of small targets to be gradually blurred and even disappear, making it difficult to accurately identify; after multiple down-sampling, the position of the small target becomes no longer accurate, and the positioning accuracy of the detection result decreases; the low-resolution image after down-sampling may cause the background to be confused with the small target, thereby affecting the detection accuracy. SUMMARY

[0007] In order to solve the above technical problems, the present application provides a UAV aerial small target detection method based on an improved YOLOv11s algorithm, which improves the detection and positioning ability of small targets by optimizing the YOLOv8s model framework, realizes the improvement of the model in the average precision mAP, solves the technical problems of large information loss risk, missed detection, false detection and low detection accuracy of the YOLOv11s algorithm when detecting images with complex background and small pixel points such as UAV aerial small targets, and improves the precision, robustness and real-time performance of UAV aerial target detection.

[0008] In order to achieve the above purpose, the present application is realized by the following technical scheme:

[0009] The present application is a UAV aerial small target detection method based on an improved YOLOv11s algorithm, which specifically comprises the following steps:

[0010] Step 1, constructing a UAV aerial image dataset, and randomly dividing the UAV aerial image dataset into a training set, a validation set and a test set according to a predetermined proportion;

[0011] Step 2, constructing a UAV aerial small target detection model based on the YOLOv11s algorithm, wherein the UAV aerial small target detection model, i.e. the improved YOLOv11s algorithm, comprises a backbone feature extraction network, a neck fusion network and a head detection network;

[0012] Step 3, the unmanned aerial vehicle aerial small target detection model in step 2 is optimized, including optimization of the backbone feature extraction network, optimization of the neck fusion network and optimization of the loss function of the head detection network;

[0013] Step 4, the training parameters of the unmanned aerial vehicle aerial small target detection model are set, the improved unmanned aerial vehicle aerial small target detection model is trained using the training set in step 1, and the training weight is saved gradually during the training process;

[0014] Step 5, after the training is completed, the model weight with the minimum validation set loss function is loaded into the unmanned aerial vehicle aerial small target detection model, and the test set is used for evaluation.

[0015] Further improvement of the application is that in step 1, the format of the unmanned aerial vehicle aerial image data set is YOLO format.

[0016] Further improvement of the application is that in step 1, the unmanned aerial vehicle aerial image data set is divided into a training set, a validation set and a test set according to a ratio of 8:1:1.

[0017] Further improvement of the application is that in step 1, the backbone feature extraction network adopts four times of downsampling operation, the first downsampling layer has a step of 1, four feature maps of 320*320, 160*160, 80*80 and 40*40 are obtained, the information loss risk is reduced, the neck fusion network receives three feature maps of 160*160, 80*80 and 40*40, and through a larger resolution, the network can better detect and locate small targets on larger feature maps.

[0018] Further improvement of the application is that in the backbone feature extraction network, the output channel number of the first Conv module is 16, the output channel number of the second Conv module is 32, the output channel number of the first PPA_C3k2 module is 64, the output channel number of the third Conv module is 64, the output channel number of the second PPA_C3k2 module is 128, the output channel number of the fourth Conv module is 128, the output channel number of the third PPA_C3k2 module is 128, the output channel number of the fifth Conv module is 256, the output channel number of the fourth PPA_C3k2 module is 256, the output channel number of SPPF is 256, and the output channel number of C2PSA is 256.

[0019] Further improvement of the present application is that in step 3, the backbone feature extraction network in the UAV aerial small target detection model is optimized, specifically: introducing a PPA_C3k2 module in the backbone feature extraction network of the YOLOv11s algorithm, the PPA_C3k2 module includes a local branch, a global branch and a C3k2 branch, the local branch is responsible for capturing detailed information, the global branch is responsible for helping to obtain the context information of the global branch, and the C3k2 branch further enhances the expression ability of the features by introducing a convolutional skip connection; specifically, the PPA_C3k2 module adopts a parallel multi-branch method to realize the process of multi-branch extraction, and the process specifically includes the following steps:

[0020] Step 3.1.1, given an input feature tensor F' ∈ R H′×W′×C , adjust the input feature tensor by pointwise convolution to obtain a pointwise convolution output tensor F ∈ R H′×W′×C′ ;

[0021] Step 3.1.2, calculate the local branch output tensor F local ∈ R H′×W′×C′ , the global branch output tensor F global ∈ R H′×W′×C′ and the C3k2 branch output tensor F C3K2 ∈ R H′×W′×C′ respectively through the local branch, the global branch and the C3k2 branch;

[0022] Step 3.1.3, add the local branch output tensor F local , the global branch output tensor F global and the C3k2 branch output tensor F C3K2 to obtain a multi-branch fusion output tensor F ～ ∈ R H′×W′×C′ ;

[0023] Step 3.1.4, using the attention module CBAM to adaptively enhance the features obtained by multi-branch feature extraction, the multi-branch fusion output tensor F ～ ∈ R H′×W′×C′ , to obtain the PPA_C3k2 module output F PPA_C3K2 ∈ R H′×W′×C′ ;

[0024] Further improvement of the present application is that in step 3.1.2, the specific calculation method of the local branch output tensor F local and the global branch output tensor F global is:

[0025] Step 3.1.2.1, divide the input feature tensor F' into a group of spatially continuous blocks

[0026] Step 3.1.2.2, channel average is performed to obtain a channel average output tensor

[0027] Step 3.1.2.3, linear calculation is performed using a feedforward neural network FFN to obtain linear calculation features, and then an activation function is applied to obtain a probability distribution of the linear calculation features in the spatial dimension, the obtained probability distribution is multiplied with the linear calculation features, the weights of the linear calculation features are adjusted to obtain weighted features, a feature selection module (FeatureSelection) is used to select task-related features from tokens and channels, specifically, the feature selection module is set as The weighted result is denoted as Where t i ∈R d represents the i-th output token, the feature selection module selects to operate on each token, and the output is Where s∈R c′ and P∈R c′×C′ is a task-specific parameter, and sim(·,·) is a cosine similarity function with a range of [0,1];

[0028] Step 3.1.2.4, channel selection is applied to each token by applying a linear transformation of p, and then reshaping and interpolation operations are performed to finally produce a local branch output tensor F local ∈R H′×W′×C′ and a global branch output tensor F global ∈R H′×W′×C′ .

[0029] A further improvement of the application is that in step 3, the neck fusion network of the unmanned aerial vehicle aerial small target detection model is optimized, specifically: introducing a TriFPN structure in the YOLOv11s algorithm, the TriFPN structure includes three enhanced paths, namely two top-down path fusion and one bottom-up path enhancement, the TriFPN structure introduces an adaptive fusion module AFConcat, which adds an additional weight to each input feature, allowing the network to learn the importance of each feature, and the formula of the adaptive fusion module fusion mechanism is as follows:

[0030]

[0031] Where O is the output value, representing the result after the adaptive weight fusion mechanism; w i represents the weight of the i-th input feature map, i represents the index number of the i-th input feature map; I i is the input value, representing the i-th input feature map; in the above formula, w iThe TriFPN structure can improve accuracy while minimizing calculation cost through the adaptive weight fusion method.

[0032] The further improvement of the present application is that the optimization of the loss function of the head detection network in step 3 refers to using the combination of CIoU and Dot Distance (DotD) as the positioning loss function, and enhancing the small target positioning capability.

[0033] The beneficial effects of the present application are:

[0034] The present application optimizes the network module, introduces the PPA_C3k2 module, maintains and enhances the representation of small targets;

[0035] The present application optimizes the network structure, uses larger resolution feature maps, reduces the risk of information loss, and improves the detection and positioning capability of small targets;

[0036] The present application optimizes the network structure, adds a top-down path fusion, and makes the high-resolution feature layer fuse more advanced semantic features.

[0037] The present application optimizes the network structure, adjusts the width of the backbone network to half of YOLOv11s, balances the parameter amount of the network while improving the accuracy.

[0038] The present application improves the loss function, improves the detection accuracy and stability of small targets. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 It is the overall flowchart of the present application.

[0040] Figure 2 It is the network structure diagram of the present application.

[0041] Figure 3 It is the PPA_C3k2 module diagram in the backbone network of the present application.

[0042] Figure 4 It is the Patch-Aware module diagram in the PPA_C3k2 module of the present application.

[0043] Figure 5 It is the CBAM module diagram in the PPA_C3k2 module of the present application.

[0044] Figure 6 It is the MAP curve diagram of YOLOv11s.

[0045] Figure 7 It is the MAP curve diagram of the method of the present application.

[0046] Figure 8 is a training result graph of the method of the present application.

[0047] Figure 9 is a detection effect graph of YOLOv11s in a dense scene.

[0048] Figure 10 is a detection effect graph of the method of the present application in a dense scene. DETAILED DESCRIPTION

[0049] Embodiments of the present application will be described below with reference to drawings. Many practical details will be described in the following description in order to provide a thorough understanding of the present application. However, it should be understood that these practical details are not intended to limit the present application. That is, these practical details are not necessary in some embodiments of the present application.

[0050] As shown in Figures 1-2 , the present application is a UAV aerial small target detection method based on an improved YOLOv11s algorithm, which specifically comprises the following steps:

[0051] Step 1, constructing a UAV aerial image dataset, which is divided into a training set, a validation set and a test set according to a ratio of 8:1:1.

[0052] Step 2, constructing a UAV aerial small target detection model based on a YOLOv11s algorithm, wherein the UAV aerial small target detection model, i.e., the improved YOLOv11s algorithm, comprises a backbone feature extraction network, a neck fusion network and a head detection network.

[0053] Step 3, optimizing the UAV aerial small target detection model in Step 2, including optimizing the backbone feature extraction network, optimizing the neck fusion network and optimizing the loss function of the head detection network.

[0054] A PPA_C3k2 module is introduced into the original backbone feature extraction network of the YOLOv11s algorithm. The advantage of PPA_C3k2 lies in a multi-branch feature extraction strategy. A parallel multi-branch method is adopted, and each branch is responsible for extracting features of different scales and levels, which helps to capture multi-scale features of objects, thereby improving the accuracy of small object detection.

[0055] As shown in Figure 3 , the PPA_C3k2 module comprises a local branch, a global branch and a C3k2 branch. The local branch is responsible for capturing detailed information, the global branch is responsible for helping to obtain context information of the global branch, and the C3k2 branch further enhances the expression ability of features by introducing a convolutional skip connection.

[0056] The PPA_C3k2 module adopts a parallel multi-branch method to realize the process of multi-branch extraction, which specifically includes the following steps:

[0057] Step 3.1.1, given the input feature tensor F' ∈ R H′×W′×C , adjust the input feature tensor by point-wise convolution to obtain the point-wise convolution output tensor F ∈ R H′×W′×C′ ;

[0058] Step 3.1.2, calculate the local branch output tensor F local ∈ R H′×W′×C′ , the global branch output tensor F global ∈ R H′×W′×C′ and the C3k2 branch output tensor F C3K2 ∈ R H′×W′×C′ respectively through local branch, global branch and C3k2 branch. Specifically, the local and global branches are distinguished by controlling the patch size parameter p, and the determination of the patch size is realized by the aggregation and displacement of non-overlapping patches in the spatial dimension; then the attention matrix between non-overlapping patches is calculated to realize the extraction and interaction of local and global features. The specific calculation method of the local branch output tensor F local and the global branch output tensor F global is as follows:

[0059] First, divide the input feature tensor F' into a set of spatially continuous blocks

[0060] Then, perform channel averaging to obtain the channel average output tensor

[0061] Use the feedforward neural network FFN to perform linear calculation to obtain linear calculation features, and then apply the activation function to obtain the probability distribution of the linear calculation features in the spatial dimension. Multiply the obtained probability distribution points with the linear calculation features to adjust the weights of the linear calculation features, obtain the weighted features, and use the feature selection module to select the features related to the task from the tokens and channels. Specifically, set The weighted result is expressed as Where t i ∈ R d represents the i-th output token, and the feature selection module selects the operation on each token, and the output is Where s ∈ R c′ and P ∈ R c′×C′is a task-specific parameter, sim(·, ·) is a cosine similarity function with a range of [0, 1]; here, s acts as a task embedding, specifying which tokens are relevant to the task. Each token t i is re-weighted according to its relevance to the task embedding (measured by cosine similarity), effectively modeling token selection.

[0062] A linear transformation of p is then applied to each token for channel selection, followed by reshaping and interpolation operations, resulting in a local branch output tensor F local ∈R H′×W′×C′ and a global branch output tensor F global ∈R H′×W′×C′ , as shown in Figure 4 .

[0063] Step 3.1.3, add the local branch output tensor F local , the global branch output tensor F global , and the C3k2 branch output tensor F C3K2 to obtain a multi-branch fusion output tensor F ～ ∈R H′×W′×C′ .

[0064] Step 3.1.4, the multi-branch fusion output tensor F ～ ∈R H′×W′×C′ obtained by multi-branch feature extraction is used to perform self-adaptive feature enhancement using the attention module CBAM, obtaining the PPA_C3k2 module output F PPA_C3K2 ∈R H′×W′×C′ , as shown in Figure 5 .

[0065] The step of the first down-sampling layer in YOLOv11s is changed to 1, and a total of 4 down-sampling is performed to obtain four feature maps of 320*320, 160*160, 80*80, and 40*40, reducing the risk of information loss. The neck network receives three feature maps of 160*160, 80*80, and 40*40, allowing the network to better detect and locate small targets on larger feature maps through greater resolution.

[0066] The width of the backbone network is adjusted to be half of YOLOv11s, specifically: the output channel number of the first Conv module in the backbone network is 16, the output channel number of the second Conv module is 32, the output channel number of the first PPA_C3k2 module is 64, the output channel number of the third Conv module is 64, the output channel number of the second PPA_C3k2 module is 128, the output channel number of the fourth Conv module is 128, the output channel number of the third PPA_C3k2 module is 128, the output channel number of the fifth Conv module is 256, the output channel number of the fourth PPA_C3k2 module is 256, the output channel number of SPPF is 256, and the output channel number of C2PSA is 256.

[0067] In step 3, the neck fusion network of the UAV aerial small target detection model is optimized, specifically: introducing a TriFPN structure in the YOLOv11s algorithm, the TriFPN structure includes three enhanced paths, i.e., two top-down path fusion and one bottom-up path enhancement, a total of three enhanced paths, which are used to build high-level semantic feature maps of all scales, and at the same time, the precise positioning signal of the lower layer is used to enhance the entire feature hierarchy. The TriFPN structure introduces an adaptive fusion module, AFConcat, which adds an additional weight to each input feature, allowing the network to learn the importance of each feature. The formula of the adaptive fusion module fusion mechanism is as follows:

[0068]

[0069] wherein O is an output value, representing the result after fusion by the adaptive weight fusion mechanism; w i represents the weight of the i-th input feature map, i represents the index number of the i-th input feature map; I i is an input value, representing the i-th input feature map; in the above formula, w i is a learnable weight that determines the contribution of each input feature to the final output feature. Through the above adaptive weight fusion method, the TriFPN structure can improve accuracy while minimizing computational cost.

[0070] Optimizing the loss function of the head detection network means using the combination of CIoU and Dot Distance (DotD) as the positioning loss function to enhance the small target positioning capability.

[0071] Dot Distance (DotD) measurement: Since a slight change in position relationship can make the CIOU value of the extremely small target become very large, using the CIOU index will reduce the detection of extremely small target detection, therefore, for small targets, a new measurement point distance Dot Distance (DotD) is introduced.

[0072] DotD is defined as the normalized Euclidean distance between the centers of two bounding boxes.

[0073] Small targets can be regarded as a point, and the importance of width and height is far lower than the position of the center point. Therefore, DotD is defined as:

[0074]

[0075] where D represents the Euclidean distance between the two center points, and S represents the average size of all targets in a specific dataset. In order to limit the range to [0, 1], an exponential form is used for normalization. D and S are represented as:

[0076]

[0077] where (x A -x B ) and (y A -y B ) represent the coordinates of the two center points, M represents the number of images, and N i represents the number of bounding boxes.

[0078] By using the Dot Distance metric, the sensitivity brought by the original CIoU metric is reduced, which helps the network effectively locate the area where the small target is located. In most UAV datasets, not only small targets exist, but also medium and large targets. A combination of CIoU and Dot Distance is proposed as a positioning loss function:

[0079] Loss = 1 - 0.5CIoU - 0.5DotD

[0080] The above loss function considers both medium and large targets and small targets. On the one hand, it can make the model accurately predict the position and size of the target, and on the other hand, it can speed up the convergence of the network.

[0081] Step 4, set the training parameters, and use the training set to train the improved YOLOv11s model. The specific parameter settings are: the input image size is 640x640, the optimizer is Adam, the initial learning rate is 0.01, the momentum is 0.92, the iteration number is set to 300 times, and the batch normalization size is 8. During the training process, the training weights are saved step by step;

[0082] Step 5, after the training is completed, select the model weight with the minimum validation set loss function to load into the UAV aerial small target detection model, and use the test set for evaluation. The detection accuracy and visual results are used to evaluate the performance of the model.

[0083] In order to comprehensively evaluate the performance of the improved algorithm, a series of comparative experiments are carried out, and the optimized algorithm is compared and analyzed with the current mainstream target detection algorithm. In the experiment, the mAP50, mAP50-90 and Parmas indicators are mainly evaluated. The results are shown in Table 1.

[0084] Table 1 Experimental results of the improved model in the VisDrone2019 dataset

[0085]

[0086] From Figure 6 the PR curve diagram of the YOLOv11s and Figure 7 the PR curve diagram of the method of the present application shown in the comparison diagram and Figure 8 It can be seen that the parameter amount of the method of the present application is only 2.75M, the mAP50 reaches 46.97%, the mAP50-95 reaches 29.17%, especially for small target pedestrians (pedes) and people (people) and other categories, the model performs excellently, the mAP50 is as high as 53.97% and 44%, showing that the model has strong small target detection capability.

[0087] From the detection results of Figure 9 and Figure 10 It can be seen that since the small target occupies fewer pixels in the image, it is difficult to accurately identify and locate it. Therefore, the original YOLOv11s algorithm has a missed detection situation when facing small targets. However, through the method of the present application, the features of small targets can be better extracted, and small targets can be more accurately detected and located. Therefore, the method of the present application effectively improves the ability to detect and locate small targets.

[0088] In summary, the present application can reduce the computational complexity of the model and improve the small target detection accuracy of unmanned aerial vehicle aerial images.

[0089] The above only describes the embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. An unmanned aerial vehicle aerial small target detection method based on an improved YOLOv11s algorithm, characterized in that: The unmanned aerial vehicle aerial small target detection method specifically comprises the following steps: Step 1, constructing an unmanned aerial vehicle aerial image dataset, and randomly dividing the unmanned aerial vehicle aerial image dataset into a training set, a validation set and a test set according to a preset ratio; Step 2, constructing an unmanned aerial vehicle aerial small target detection model based on a YOLOv11s algorithm, wherein the unmanned aerial vehicle aerial small target detection model comprises a backbone feature extraction network, a neck fusion network and a head detection network; Step 3, optimizing the unmanned aerial vehicle aerial small target detection model in step 2, including optimizing the backbone feature extraction network, optimizing the neck fusion network and optimizing the loss function of the head detection network; Step 4, setting the training parameters of the unmanned aerial vehicle aerial small target detection model, training the improved unmanned aerial vehicle aerial small target detection model using the training set in step 1, and saving the training weights step by step during the training process; Step 5, after the training is completed, loading the model weight with the minimum validation set loss function into the unmanned aerial vehicle aerial small target detection model, and evaluating it with the test set, wherein: The backbone feature extraction network in the unmanned aerial vehicle aerial small target detection model is optimized, specifically: introducing a PPA_C3k2 module in the backbone feature extraction network of the YOLOv11s algorithm, the PPA_C3k2 module includes a local branch, a global branch and a C3k2 branch, the local branch is responsible for capturing detailed information, the global branch is responsible for helping to obtain the context information of the global branch, and the C3k2 branch further enhances the expression ability of the features by introducing a convolutional jump connection; In step 3, the optimization of the neck fusion network of the unmanned aerial vehicle aerial small target detection model is specifically: introducing a TriFPN structure in the YOLOv11s algorithm, the TriFPN structure includes three enhanced paths, namely two top-down path fusion and one bottom-up path enhancement, the TriFPN structure introduces an adaptive fusion module AFConcat, and the adaptive fusion module fusion mechanism formula is as follows: wherein O is an output value, representing a result after an adaptive weight fusion mechanism; w i represents a weight of the i-th input feature map, i represents an index number of the i-th input feature map of the input; i is an input value, representing the i-th input feature map; In step 3, the optimization of the loss function of the head detection network means using the combination of CIoU and point distance DotDistance (DotD) as the positioning loss function to enhance the small target positioning ability, and DotD is defined as the normalized Euclidean distance between the centers of two bounding boxes.

2. The unmanned aerial vehicle aerial small target detection method based on the improved YOLOv11s algorithm according to claim 1, characterized in that: In step 1, the unmanned aerial vehicle aerial image dataset is divided into a training set, a validation set and a test set according to a ratio of 8:1:

1.

3. The unmanned aerial vehicle aerial small target detection method based on the improved YOLOv11s algorithm according to claim 1, characterized in that: In step 3, the PPA_C3k2 module adopts a parallel multi-branch method to realize the process of multi-branch extraction, specifically comprising the following steps: Step 3.1.1, given an input feature tensor F ′ ∈R H′×W′×C , adjust the input feature tensor by point-wise convolution to get a point-wise convolution output tensor F∈R H′×W′×C′ ; Step 3.1.

2. Compute the local branch output tensor Floc, the global branch output tensor Fglo, and the C3k2 branch output tensor Fc3k2, respectively, through the local branch, the global branch, and the C3k2 branch local ∈ R H′×W′×C′ , the global branch output tensor Fglo global ∈ R H′×W′×C′ , and the C3k2 branch output tensor Fc3k2 C3K2 ∈ R H′×W′×C′ ; Step 3.1.3, adding the local branch output tensors F local , the global branch output tensors F global , and the C3k2 branch output tensors F C3K2 , to obtain the multi-branch fusion output tensor F ～ ∈R H′×W′×C′ ; Step 3.1.

4. Obtain a multi-branch fusion output tensor F through multi-branch feature extraction ～ ∈R H′×W′×C′ Adaptively enhance the features using the attention module CBAM to obtain the PPA_C3k2 module output F PPA_C3K2 ∈R H′×W′×C′。 4. The unmanned aerial vehicle aerial small target detection method based on the improved YOLOv11s algorithm according to claim 3, characterized in that: In step 3.1.2, the local branch output tensor F local and the global branch output tensor F global The specific calculation method is: Step 3.1.2.1, partitioning the input feature tensor F ′ into a set of spatially contiguous chunks Step 3.1.2.

2. Perform channel averaging to derive a channel-averaged output tensor Step 3.1.2.

3. Linear computation using a feed-forward neural network (FFN) to obtain linear computation features, followed by an activation function to obtain a probability distribution of the linear computation features over the spatial dimensions, point-wise multiplication of the resulting probability distribution with the linear computation features to adjust the weights of the linear computation features to obtain weighted features, use of a feature selection module (FeatureSelection) to select task-relevant features from tokens and channels, specifically set Let the weighted result be denoted as where t i ∈R d denotes the i-th output token, the feature selection module selects to operate on each token, and the output is where s∈R c′ and P∈R c′×C′ is a task-specific parameter, and sim(·,·) is a cosine similarity function with a range of [0,1]. Step 3.1.2.

4. Channel selection followed by reshaping and interpolation operations are applied to each token applying the linear transformation of p, resulting in a local branch output tensor F local ∈ R H′×W′×C′ and a global branch output tensor F global ∈ R H′×W′×C′ .

5. The unmanned aerial vehicle aerial small target detection method based on the improved YOLOv11s algorithm according to claim 1, characterized in that: In step 1, the backbone feature extraction network adopts four downsampling operations, the first downsampling layer has a step of 1, and four feature maps of 320*320, 160*160, 80*80 and 40*40 are obtained, which reduces the risk of information loss, and the neck fusion network receives three feature maps of 160*160, 80*80 and 40*40, so that the network can better detect and locate small targets on larger feature maps.

6. The unmanned aerial vehicle aerial small target detection method based on the improved YOLOv11s algorithm according to claim 1, characterized in that: The output channel number of the first Conv module in the backbone feature extraction network is 16, the output channel number of the second Conv module is 32, the output channel number of the first PPA_C3k2 module is 64, the output channel number of the third Conv module is 64, the output channel number of the second PPA_C3k2 module is 128, the output channel number of the fourth Conv module is 128, the output channel number of the third PPA_C3k2 module is 128, the output channel number of the fifth Conv module is 256, the output channel number of the fourth PPA_C3k2 module is 256, the output channel number of the SPPF is 256, and the output channel number of the C2PSA is 256.

Images