Remote sensing image target detection method and system based on improved YOLOv11 network

Abstract

The invention discloses a remote sensing image target detection method and system based on an improved YOLOv11 network, and belongs to the technical field of remote sensing image target detection. The method comprises the steps that S1, a remote sensing image data set is collected, each image corresponds to an annotation file in a YOLO format, and the annotation file comprises category information and position coordinates of each target in the images; s2, dividing the data set into a training set, a verification set and a test set according to a preset proportion; s3, inputting the training set image into an improved YOLOv11 network, and generating category information and a bounding box of a target; s4, calculating a network loss value according to the prediction frame and the real bounding box, and updating YOLOv11 network parameters through back propagation; s5, predicting the verification set image by using the updated YOLOv11 network, and calculating the AP value of each category and the overall mAP value; s6, the steps S3-S5 are repeated until the mAP value of the verification set tends to be stable, and YOLOv11 network training is completed; and S7, detecting the test set image by using the trained YOLOv11 network, and outputting a final detection result.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image target detection technology, specifically relating to a remote sensing image target detection method and system based on an improved YOLOv11 model, applicable to the detection of small targets in images acquired by remote sensing platforms such as satellites and UAVs. Background Technology

[0002] With the rapid development of high-resolution remote sensing technology, remote sensing images are increasingly widely used in fields such as military reconnaissance, environmental monitoring, and urban planning. Conducting research on target detection in remote sensing images can not only improve the accuracy and efficiency of target recognition, meeting the application needs of key areas such as military reconnaissance and disaster monitoring, but also promote the in-depth development of computer vision technology in the field of remote sensing, possessing significant theoretical value and application prospects.

[0003] The core challenges facing remote sensing target detection currently lie in key issues such as the difficulty in extracting features from small targets, interference from complex backgrounds, and the need for multi-scale adaptation. Long-distance imaging reduces targets to only a few pixels, making it difficult for traditional algorithms to capture subtle features, while complex terrain backgrounds increase the risk of false detections. Simultaneously, the coexistence of targets at different scales in the scene, the problem of missed detections in dense areas, and the real-time processing requirements of edge devices such as drones all place higher demands on detection technology.

[0004] In recent years, deep convolutional neural networks (CNNs) have demonstrated strong potential in the field of target detection in remote sensing images. One-stage algorithms, represented by the YOLO series, have attracted considerable attention due to their efficiency. Their multi-scale feature fusion mechanism effectively addresses the problem of diverse target scales in remote sensing images, while their anchor box design adapts to the dense distribution of targets. These characteristics make the YOLO series a reliable solution for target detection in remote sensing images. Based on this, this invention provides a remote sensing image target detection method and system based on an improved YOLOv11 network. Summary of the Invention

[0005] To address the aforementioned limitations of existing technologies, this invention proposes a remote sensing image target detection method and system based on an improved YOLOv11 network. This invention enhances multi-scale feature extraction capabilities by introducing a Large Selective Kernel (LSK) module and a Convolutional Block Attention Module (CBAM) at specific locations within the network. This allows the invention to significantly improve the detection performance of small targets in complex remote sensing scenarios while maintaining the high efficiency of the YOLO series algorithms. It is suitable for the rapid and accurate detection of targets such as vehicles, ships, and buildings in high-resolution remote sensing images.

[0006] This invention includes the following steps:

[0007] A remote sensing image target detection method based on an improved YOLOv11 network includes the following steps:

[0008] S1. Collect a remote sensing image dataset. Each image corresponds to a YOLO format annotation file, which contains the category and location of each target in the image.

[0009] S2. Divide the dataset according to a preset ratio to obtain a training set, a validation set, and a test set;

[0010] S3. Input the images from the training set in step S2 into the improved YOLOv11 network to generate category information and bounding boxes;

[0011] S4. Calculate the network loss value based on the predicted bounding box obtained in step S3 and the corresponding ground truth bounding box of the image, and update the YOLOv11 network parameters through backpropagation.

[0012] S5. Based on the updated YOLOv11 network, predict each image in the validation set in step S2, and output the AP value and mAP value of each category in the validation set after statistics.

[0013] S6. Repeat steps S3-S5 until the mAP counted in step S5 stabilizes at a certain value, and obtain the finally trained YOLOv11 network.

[0014] S7. Use the trained YOLOv11 network to detect all images in the test set of step S2 and obtain the detection results.

[0015] As a preferred embodiment, in step S1, a remote sensing image dataset is created. The remote sensing image dataset consists of several image files containing various objects taken by a drone and corresponding label files. The label files mark the real location information and category of each object in the corresponding image.

[0016] As a preferred option, in step S1, the remote sensing image dataset is collected by a UAV or satellite platform. The images contain various target types, such as vehicles, ships, and buildings. The image size is uniformly 640×640 pixels, and the annotation file is in YOLO format.

[0017] As a preferred embodiment, in step S2, the ratio of the training set, validation set, and test set is 7:1:2, and they are stored in the train, val, and test folders respectively. A YAML file is created to store the remote sensing target detection dataset parameters, such as the dataset path, category name, training set path, validation set path, and test set path. The dataset parameters stored in the YAML file include the paths to the dataset, the train file, the val file, the test file, and the specific categories of each object detected in the remote sensing image.

[0018] As a preferred embodiment, in step S3, the improved YOLOv11 network includes Backbone, Neck, and Head. The training set images are input into the improved YOLOv11 network and processed sequentially through Backbone, Neck, and Head to generate target category information and bounding boxes.

[0019] As a preferred embodiment, step S3 specifically includes the following steps:

[0020] S3.1. Adjust the image size of the training set from step S2 to 640×640 and input it into the Backbone part of the improved YOLOv11 network for feature extraction. The improved Backbone part is an 11-layer network structure consisting of a Conv layer, a C3k2-LSK module, a C3k2 module, an SPPF module, and a C2PSA module connected in sequence. The C3k2-LSK module is formed by introducing an LSK (Large Selective Kernel) module into the C3k2 module. After passing through this part, feature maps of three different sizes are obtained: 80×80, 40×40, and 20×20.

[0021] S3.2. The three effective feature maps obtained in step S3.1 (referred to as Q2, Q4, and Q8 according to their size from smallest to largest) are fed into the Neck part of the YOLOv11 network for feature fusion. After top-down and bottom-up fusion, the features of deep and shallow layers are fully mixed. For each layer output feature map P2, P4, and P8 of the feature fusion module, their scale is consistent with the input feature maps Q2, Q4, and Q8.

[0022] As a preferred embodiment, step S3.2 specifically includes the following steps:

[0023] S3.2.1 Input the smallest size feature map Q2 into the SPPF module for processing to obtain feature map M2; then perform an upsampling operation on M2 and concat it with feature map Q4. Input the fusion result into the C3k2 module for feature extraction and integration to obtain feature map M4.

[0024] S3.2.2. Perform upsampling operation on feature map M4 and concat it with feature map Q8. Input the fusion result into C3k2 module for feature extraction and integration to obtain the shallowest output feature map P8.

[0025] S3.2.3. Perform downsampling operation on feature map P8 and fuse it with feature map M4. Input the fusion result into module C3k2 for feature extraction and integration to obtain intermediate layer output feature map P4.

[0026] S3.2.4. Perform downsampling on feature map P4 and fuse it with feature map M2. Input the fusion result into module C3k2 for feature extraction and integration to obtain the deepest output feature map P2.

[0027] S3.3. Input the three fused feature maps P2, P4, and P8 output from the Neck part into the improved YOLOv11 network Head Detector CBAM-Detect to generate the final predicted feature maps. The improved Head Detector CBAM-Detect adopts a decoupled structure, adding a CBAM attention module to the original YOLOv11 network head, outputting the target's category information and bounding box information through parallel branches. The CBAM attention module includes a channel attention module and a spatial attention module. The channel attention module performs global max pooling and average pooling on the input feature layer, processes it through a shared fully connected layer (Shared MLP), adds the two results Maxout and Avgout and normalizes them to generate a channel weight matrix, dynamically enhancing the feature response of key channels. The spatial attention module takes the maximum and average values ​​in the channel dimension, concatenates them, compresses them into a single channel through convolution, and normalizes them to obtain a spatial weight matrix, accurately locating the key regions of the target.

[0028] As a preferred option, step S4 specifically involves: generating prediction boxes and their class confidence scores from the prediction feature maps obtained in S3; matching the prediction boxes with the ground truth bounding boxes using a task-aligned dynamic sample allocation strategy (Task-Aligned SimOTA); filtering positive and negative samples; calculating the total loss value using classification loss and bounding box loss; and finally performing backpropagation through gradient descent (SGD) to update the YOLOv11 network parameters and achieve network optimization.

[0029] As a preferred embodiment, step S5 specifically involves: inputting each image in the validation set into the YOLOv11 network after parameter updates to generate predicted bounding boxes and their class confidence scores; using the non-maximum suppression (NMS) algorithm to remove redundant boxes and retaining the predicted boxes with the highest confidence scores; matching the NMS-processed predicted bounding boxes with the ground truth bounding boxes; and calculating the AP value and mAP value for each class.

[0030] As a preferred approach, in step S6, a maximum number of training epochs is set, and the mAP value of the validation set is recorded after each epoch. An early stopping mechanism is used to automatically determine whether the network has converged. When the mAP value does not significantly improve or shows a downward trend within several consecutive epochs, training is stopped in time and the current network parameters are saved, thereby effectively preventing overfitting and improving training efficiency and model generalization ability.

[0031] As a preferred approach, in step S7, the test set images are input into the trained YOLOv11 network to generate a predicted feature map. Then, redundant boxes are removed using non-maximum suppression (NMS) to obtain the final detection result. Subsequently, the predicted boxes at the feature map scale are mapped back to the original image scale, and the coordinates, category, and confidence of the detection boxes are output to form the final detection result file.

[0032] This invention also discloses a remote sensing image target detection system based on an improved YOLOv11 network for performing the above method, comprising the following modules:

[0033] Data acquisition module: Acquires remote sensing image datasets. Each image corresponds to a YOLO format annotation file, which contains the category information and location coordinates of each target in the image.

[0034] Dataset partitioning module: Divides the dataset into training set, validation set, and test set according to a preset ratio;

[0035] The target category information and bounding box generation module: Input the training set images into the improved YOLOv11 network to generate target category information and bounding boxes;

[0036] Network loss calculation module: Calculates the network loss value based on the predicted bounding box and the ground truth bounding box, and updates the YOLOv11 network parameters through backpropagation;

[0037] AP and mAP calculation module: Predicts validation set images using the updated YOLOv11 network, and calculates AP values ​​for each category and the overall mAP value;

[0038] Iterative module: The target category information and bounding box generation module, the network loss calculation module, and the AP value and mAP value calculation module are executed sequentially until the validation set mAP value tends to stabilize, thus completing the YOLOv11 network training;

[0039] Detection module: Uses the trained YOLOv11 network to detect images in the test set and outputs the final detection results.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] (1) This invention introduces the C3k2-LSK module into the Backbone of the YOLOv11 network and the CBAM attention mechanism into the Head part, which effectively enhances the network’s ability to extract and express features of small targets and significantly improves the detection accuracy of small targets in remote sensing images.

[0042] (2) In this invention, the Head detection head adopts a decoupled structure and introduces the CBAM module to optimize the classification and regression tasks respectively, thereby improving the bounding box positioning accuracy and category recognition accuracy, while maintaining the high efficiency of the YOLO series algorithms.

[0043] (3) The present invention introduces an early stopping mechanism and a dynamic loss weight adjustment strategy in the training process to improve training efficiency, prevent overfitting, and enhance the generalization ability of the model.

[0044] (4) This invention is particularly applicable to high-resolution images acquired by remote sensing platforms such as satellites and drones. It can efficiently and accurately detect various targets such as vehicles, ships, and buildings, meeting the practical application needs of military reconnaissance, disaster monitoring, and urban planning. It has broad practical value and promotion prospects. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a remote sensing image target detection method based on an improved YOLOv11 network according to a preferred embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the improved YOLOv11 network structure proposed in a preferred embodiment of the present invention;

[0048] Figure 3This is a schematic diagram of the C3k2-LSK module structure proposed in a preferred embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of the CBAM-Detect module structure in the improved Head detection head proposed in a preferred embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the LSK module structure used in a preferred embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the channel attention submodule and spatial attention submodule of the CBAM module used in a preferred embodiment of the present invention;

[0052] Figure 7 This is a remote sensing image target detection result diagram of a preferred embodiment of the present invention;

[0053] Figure 8 This is a block diagram of a remote sensing image target detection system based on an improved YOLOv11 network, according to a preferred embodiment of the present invention. Detailed Implementation

[0054] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0055] This embodiment provides a remote sensing image target detection method based on an improved YOLOv11 network, referring to... Figure 1 The diagram shown is a flowchart of the method. (Refer to...) Figures 2 to 4 The figures show the improved YOLOv11 network structure, the C3k2-LSK module structure, and the improved Head Detector CBAM-Detect structure, respectively, as described in this invention. The method of this invention will be described in detail below with reference to the accompanying drawings.

[0056] This embodiment presents a remote sensing image target detection method based on an improved YOLOv11 network, comprising the following steps:

[0057] S1. Collect remote sensing image datasets. Each image corresponds to a YOLO format annotation file, which contains the category information and location coordinates of each target in the image.

[0058] Specifically, the remote sensing image dataset described in this embodiment is collected by a drone or satellite platform. The images contain various target types, such as vehicles, ships, and buildings. The image size is uniformly 640×640 pixels, and the annotation file is in YOLO format, recording the category number, center point coordinates, width and height information of each target.

[0059] S2. Divide the dataset into training set, validation set and test set according to a preset ratio. In this embodiment, the ratio is 7:1:2. Generate train, val, test folders and yaml configuration files to store dataset path and category information.

[0060] S3. Input the training set images into the improved YOLOv11 network, and process them sequentially through the Backbone, Neck and Head parts to generate the target's category information and bounding box;

[0061] Specifically, step S3 includes the following steps:

[0062] S3.1. Adjust the image size to 640×640 and input it into the Backbone part of the improved YOLOv11 network for feature extraction. The Backbone part consists of a Conv layer, a C3k2-LSK module, a C3k2 module, an SPPF module, and a C2PSA module, outputting feature maps Q2, Q4, and Q8 at three different scales: 80×80, 40×40, and 20×20. The C3k2-LSK module is constructed by introducing an LSK (Large Selective Kernel) module into the C3k2 module to enhance the network's ability to perceive small targets.

[0063] S3.2, Input feature maps Q2, Q4, and Q8 into the Neck part for feature fusion, specifically including:

[0064] S3.2.1 Input Q2 into the SPPF module to obtain feature map M2. Upsample feature map M2 and concatenate it with feature map Q4. Input it into the C3k2 module to obtain feature map M4.

[0065] S3.2.2. Upsample feature map M4 and concatenate it with feature map Q8, then input it into the C3k2 module to obtain feature map P8;

[0066] S3.2.3. After downsampling feature map P8, fuse it with feature map M4 and input it into C3k2 module to obtain feature map P4;

[0067] S3.2.4. After downsampling feature map P4, fuse it with feature map M2 and input it into C3k2 module to obtain feature map P2;

[0068] S3.3. Input feature maps P2, P4, and P8 into the improved Head Detection Head to generate the final predicted feature map. The Head Detection Head adopts a decoupled structure, introducing a CBAM attention module on top of the original YOLOv11 detection head, and outputting the target's category information and bounding box information through parallel branches. For example... Figure 6 As shown, the CBAM attention module includes a channel attention module and a spatial attention module. The channel attention module performs global max pooling and average pooling on the input feature layer, processes it through a shared fully connected layer (Shared MLP), adds the two results Maxout and Avgout, and normalizes them to generate a channel weight matrix, dynamically enhancing the feature responses of key channels. The spatial attention module takes the maximum and average values ​​in the channel dimension, concatenates them, and then convolves them to compress them into a single channel. Normalization yields the spatial weight matrix, accurately locating key target regions.

[0069] S4. Calculate the network loss value based on the predicted box and the ground truth bounding box (GT box), and update the YOLOv11 network parameters through backpropagation;

[0070] Specifically, a task-aligned dynamic allocation strategy (Task-Aligned SimOTA) is adopted. A cost matrix is ​​constructed by combining classification scores and IoU, and predicted bounding boxes are adaptively matched for each ground truth (GT) box. Positive and negative samples are filtered, and the loss function includes classification loss and bounding box loss. These are weighted summed and then updated using SGD. The specific calculation process is as follows:

[0071] (1) Classification loss: solves the problem of class imbalance;

[0072] L cls =-α t (1-p t ) γ log(p t )

[0073] Parameter meaning:

[0074] p t The target class probability predicted by the model after activation by the sigmoid function:

[0075] For positive samples: p t = p (probability of predicting the true class)

[0076] For negative samples: p t =1-p

[0077] α tClass balancing weights are used to mitigate class imbalance.

[0078] For positive samples: α t =α

[0079] For negative samples: α t =1-α

[0080] γ: Focusing parameter, suppresses gradient contributions from easily separable samples.

[0081] (2) Bounding box loss: CIoU Loss combined with DFL (Discrete Distribution Loss) is adopted. CIoU Loss is a comprehensive measure that integrates center point distance, aspect ratio, and overlap rate. DFL supervises the discrete distribution of bounding box coordinates and obtains continuous coordinates through Softmax weighted summation.

[0082] A.CIoU Loss:

[0083]

[0084] Parameter meaning:

[0085] IoU: Intersection over Union (IoU) between predicted bounding boxes and ground truth bounding boxes;

[0086] ρ: Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box;

[0087] c: The length of the diagonal of the smallest bounding rectangle;

[0088] v: Aspect Ratio Consistency Metric

[0089]

[0090] α: Weighting coefficient (automatically calculated):

[0091]

[0092] B.DFL

[0093] L dfl =-((y) i+1 -y)log(P i )+(yy i )log(P i+1 ))

[0094] Parameter meaning:

[0095] y: Continuous values ​​of the normalized true bounding box coordinates;

[0096] y i ,y i+1 : The endpoints of the discrete interval where y is located;

[0097] Pi ,P i+1 : The discrete probability distribution predicted by the model;

[0098] C. Total bounding box loss:

[0099]

[0100] (3) Weighted summation of the total loss function:

[0101] L total =γ cls L cls +γ box L box

[0102] S5. Use the updated network to predict the validation set images, use non-maximum suppression (NMS) to remove redundant boxes, calculate the AP value of each category and the overall mAP value. AP is calculated based on the standard mathematical formula PR curve integral and is calculated directly using the PyTorch toolkit. mAP is the average of AP for each category and is used to evaluate the model performance.

[0103] S6. Repeat steps S3-S5 until the validation set mAP value tends to stabilize, and the network training is complete.

[0104] Specifically, the training process employs an early stopping mechanism. When the mAP value does not significantly improve or shows a downward trend within several consecutive epochs, training is stopped and the optimal model parameters are saved.

[0105] S7. Use the trained YOLOv11 network to detect images on the test set and output the final detection results;

[0106] Specifically, the test image is input into the network, a prediction box is generated and processed by NMS, and finally the prediction box at the feature map scale is mapped back to the original image scale. The coordinates, category and confidence of the detection box are output to form a detection result file.

[0107] In this embodiment, the network training environment is configured as follows:

[0108] A: Configure the network training environment using Python 3.8, the deep learning framework PyTorch 1.12, and CUDA 12.2 for GPU acceleration;

[0109] B: Set the number of network training epochs to 300, set the initial learning rate to 0.01, and use cosine annealing to dynamically decay the learning rate.

[0110] C: Set the number of images input to the network in each batch to 8;

[0111] D: Pre-trained weights are used during network training. After each training epoch, the overall network loss is calculated, and the mAP metric is evaluated on the validation set. This process is iterated until the mAP metric on the validation set stabilizes at a certain value, at which point training of the YOLOv11 network is stopped.

[0112] E: such as Figure 7 As shown, this embodiment improves mAP by 1.3 percentage points on the standard test set compared to the existing YOLOv11. Particularly noteworthy is the even more significant mAP improvement of 2.1 percentage points for the highly challenging task of small target detection (4-16 pixel targets) in remote sensing images. This improvement is primarily due to the following technological innovations:

[0113] (1) The C3k2-LSK module introduced in Backbone effectively enhances the network’s feature extraction capability for small-scale targets through a dynamic kernel selection mechanism, increasing the feature response value of 4-16 pixel targets by 28%.

[0114] (2) The improved CBAM-Detect head reduces the false detection rate of small targets by 19% and the false negative rate by 23% through channel-spatial attention co-optimization.

[0115] Experimental data show that this embodiment significantly improves the detection performance of small targets in remote sensing scenarios while maintaining the model's computational efficiency. It has important practical value for applications requiring high-precision small target identification, such as military reconnaissance and disaster monitoring.

[0116] like Figure 8 As shown, this embodiment discloses a remote sensing image target detection system based on an improved YOLOv11 network, used to perform the above method, including the following modules:

[0117] Data acquisition module: Acquires remote sensing image datasets. Each image corresponds to a YOLO format annotation file, which contains the category information and location coordinates of each target in the image.

[0118] Dataset partitioning module: Divides the dataset into training set, validation set, and test set according to a preset ratio;

[0119] The target category information and bounding box generation module: Input the training set images into the improved YOLOv11 network to generate target category information and bounding boxes;

[0120] Network loss calculation module: Calculates the network loss value based on the predicted bounding box and the ground truth bounding box, and updates the YOLOv11 network parameters through backpropagation;

[0121] AP and mAP calculation module: Predicts validation set images using the updated YOLOv11 network, and calculates AP values ​​for each category and the overall mAP value;

[0122] Iterative module: The target category information and bounding box generation module, the network loss calculation module, and the AP value and mAP value calculation module are executed sequentially until the validation set mAP value tends to stabilize, thus completing the YOLOv11 network training;

[0123] Detection module: Uses the trained YOLOv11 network to detect images in the test set and outputs the final detection results.

[0124] Other aspects of this embodiment can be found in the above method embodiments.

[0125] The beneficial effects of this invention are as follows:

[0126] (1) The C3k2-LSK module is introduced in the Backbone, and the CBAM attention mechanism is introduced in the Neck and Head to effectively enhance the network's feature extraction capability for small targets and improve detection accuracy;

[0127] (2) The head detection head adopts a decoupled structure and introduces the CBAM module to optimize the classification and regression tasks respectively, improve the bounding box localization accuracy and category recognition accuracy, while maintaining the efficiency of the algorithm;

[0128] (3) The training process introduces an early stopping mechanism and a dynamic loss weight adjustment strategy to improve training efficiency, prevent overfitting, and enhance the model's generalization ability.

[0129] (4) This invention is applicable to high-resolution images acquired by remote sensing platforms such as satellites and drones. It can efficiently and accurately detect various targets such as vehicles, ships, and buildings, meeting the practical application needs of military reconnaissance, disaster monitoring, and urban planning. It has broad practical value and promotion prospects.

[0130] In summary, this invention enhances the network's ability to perceive small targets by introducing the LSK module and CBAM attention mechanism, thereby improving detection accuracy and efficiency. It is applicable to high-resolution images acquired by remote sensing platforms such as satellites and UAVs, and can efficiently detect various targets such as vehicles, ships, and buildings, thus having broad practical value.

[0131] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope of the present invention.

Claims

1. A remote sensing image target detection method based on an improved YOLOv11 network, characterized in that, Includes the following steps: S1. Collect remote sensing image datasets. Each image corresponds to a YOLO format annotation file, which contains the category information and location coordinates of each target in the image. S2. Divide the dataset into training set, validation set and test set according to the preset ratio; S3. Input the training set images into the improved YOLOv11 network to generate target category information and bounding boxes; S4. Calculate the network loss value based on the bounding box obtained in step S3 and the corresponding real bounding box of the image, and update the YOLOv11 network parameters through backpropagation. S5. Use the updated YOLOv11 network to predict the validation set images, and calculate the AP value of each category and the overall mAP value. S6. Repeat steps S3-S5 until the validation set mAP value counted in step S5 tends to stabilize, and complete the YOLOv11 network training. S7. Use the trained YOLOv11 network to detect images on the test set and output the final detection results.

2. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 1, characterized in that: In step S1, the remote sensing image dataset is collected by UAV or satellite platform. The images contain multiple target types, the image size is uniformly 640×640 pixels, and the annotation file is in YOLO format.

3. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 1, characterized in that: In step S2, the dataset is divided into training set, validation set and test set in a ratio of 7:1:2, and stored in the train, val and test folders respectively. A YAML configuration file is created to store the dataset path, class name, training set path, validation set path and test set path.

4. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 1, characterized in that: In step S3, the improved YOLOv11 network includes Backbone, Neck, and Head, where: The backbone consists of a Conv layer, a C3k2-LSK module, a C3k2 module, an SPPF module, and a C2PSA module, forming an 11-layer feature extraction structure that outputs feature maps of three different scales: Q2, Q4, and Q8. Neck employs a bidirectional feature fusion structure, which achieves multi-scale feature fusion through upsampling, downsampling and concatenation operations, and outputs feature maps P2, P4 and P8. Head employs a decoupled structure to handle classification and regression tasks separately, and introduces the CBAM attention module.

5. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 4, characterized in that: The C3k2-LSK module is constructed by introducing an LSK module into the C3k2 module.

6. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 4, characterized in that: The minimum scale feature map Q2 is processed by Backbone's SPPF and C2PSA modules to obtain feature map M2. Neck upsamples feature map M2 and concatenates it with feature map Q4. The concatenation is then input into Neck's C3k2 module for feature extraction to obtain feature map M4. After upsampling feature map M4, it is concatenated with feature map Q8 and input into C3k2 module to obtain the shallowest output feature map P8; After downsampling feature map P8, it is fused with feature map M4 and input into C3k2 module to obtain intermediate layer output feature map P4; After downsampling feature map P4, it is fused with feature map M2 and input into module C3k2 to obtain the deepest output feature map P2.

7. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 4, characterized in that: The Head adopts a decoupled structure, outputting the target's category information and bounding box information through two independent branches, and introducing a CBAM attention module in each branch. The CBAM attention module includes a channel attention module and a spatial attention module.

8. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 1, characterized in that: In step S4, the loss function includes classification loss and bounding box loss. The IoU matching strategy is used to filter positive and negative samples, and the parameters are updated through the SGD optimizer.

9. The remote sensing image target detection method based on the improved YOLOv11 network as described in claim 1, characterized in that: In step S6, the training process adopts an early stopping mechanism. When the mAP value of the validation set does not increase for several consecutive epochs or shows a downward trend, the training is stopped and the current optimal model parameters are saved.

10. A remote sensing image target detection system based on an improved YOLOv11 network, used to perform the method as described in any one of claims 1-9, characterized in that, Includes the following modules: Data acquisition module: Acquires remote sensing image datasets. Each image corresponds to a YOLO format annotation file, which contains the category information and location coordinates of each target in the image. Dataset partitioning module: Divides the dataset into training set, validation set, and test set according to a preset ratio; The target category information and bounding box generation module: Input the training set images into the improved YOLOv11 network to generate target category information and bounding boxes; Network loss calculation module: Calculates the network loss value based on the predicted bounding box and the ground truth bounding box, and updates the YOLOv11 network parameters through backpropagation; AP and mAP calculation module: Predicts validation set images using the updated YOLOv11 network, and calculates AP values ​​for each category and the overall mAP value; Iterative module: The target category information and bounding box generation module, the network loss calculation module, and the AP value and mAP value calculation module are executed sequentially until the validation set mAP value tends to stabilize, thus completing the YOLOv11 network training; Detection module: Uses the trained YOLOv11 network to detect images in the test set and outputs the final detection results.

Images