Small target detection method for high-altitude aerial image based on improved YOLOv8

Abstract

The invention provides a small target detection method for a high-altitude aerial image based on improved YOLOv8, and relates to the technical field of deep learning and target detection.The method comprises the steps that firstly, a data set is obtained and preprocessed, then key improvement is conducted on a traditional YOLOv8 model, and the improvement method comprises the steps that an advanced FocNeXt module is introduced into a backbone network of the model, and the advanced FocNeXt module is introduced into the backbone network of the model; the feature extraction capability is enhanced; a RepBiFPN structure is adopted to optimize a feature fusion process; an efficient and lightweight detection head LAEH is designed to improve the operation efficiency of the model; and the Shape-IoU is introduced as a loss function to optimize the prediction accuracy of the model, through the improvement, the detection precision is remarkably improved while the model keeps rapid detection, and the effectiveness and superiority of the method in high-altitude aerial image small target detection are verified.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and target detection, and in particular to a high-altitude aerial image target detection method based on improved YOLOv8, which can be used for high-altitude aerial detection of small ground targets. Background Art

[0002] With the rapid development of aerial photography technology, high-altitude aerial images are becoming increasingly important in many fields such as agricultural monitoring, urban planning, and disaster assessment. However, due to the particularity of high-altitude aerial images, such as small target size, wide viewing angle, and sparse target distribution, traditional target detection methods often fail to achieve satisfactory results. Therefore, it is of great practical significance to develop a method that can accurately detect small targets in high-altitude aerial images.

[0003] In the research of small target detection in high-altitude aerial images, traditional target detection methods, such as feature-based methods and early deep learning methods, often face the following difficulties when processing high-altitude aerial images: first, complex background environment; second, a high proportion of small targets; third, inconsistency in target size and distribution; fourth, the model is too large to be suitable for embedded devices. This brings great challenges to the task of small target detection in high-altitude aerial images, so it is of great practical significance to study high-altitude aerial target detection algorithms.

[0004] Object detection algorithms based on deep learning can be roughly divided into two categories: one is a two-stage detection method based on candidate regions. The two-stage detection method first generates candidate regions that may contain target objects, and then accurately locates and identifies targets in these regions through classification and target box regression. This type of method can maintain high accuracy, but the model is large and the performance required for detection reasoning is high. Representative algorithms include R-CNN, Faster R-CNN, Mask R-CNN, etc. This type of model is usually not suitable for application scenarios that require real-time processing of detection tasks.

[0005] The other type is a single-stage detection network based on regression. These single-stage detection networks use a single network structure to directly predict the location and category of the target from the input image, eliminating the step of generating candidate regions, and thus can achieve faster detection speeds. The main representatives of this type of method are SSD and the YOLO series, among which YOLO is widely used in practice due to its high efficiency, especially YOLOv8, which is famous for its good balance in the YOLO series.

[0006] A method for detecting small targets in high-altitude aerial images of the present invention is based on the characteristics of high-altitude aerial images, and specific improvements are made to YOLOv8 to improve the detection performance of small targets. These improvements may include adjustments to the backbone network and feature network, optimization of the detection head, and improvement of the loss function. Summary of the invention

[0007] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to propose a small target detection method for high-altitude aerial photography, aiming to enhance the detection accuracy of small targets and improve the target detection performance of high-altitude aerial images.

[0008] The technical solution of the present invention is achieved by the following methods:

[0009] The present invention discloses a small target detection method for high-altitude aerial images based on improved YOLOv8, comprising the following steps:

[0010] Step 1. Obtain the aerial image dataset publicly available from VisDrone, preprocess it, and divide the dataset into three parts: training set, validation set, and test set;

[0011] Step 2. Process the YOLOv8 model, including integrating the FocNeXt module in the backbone network to optimize the efficiency and accuracy of feature extraction, and applying the RepBiFPN structure to enhance the information flow between features of different scales and improve the target recognition ability. The lightweight detection head LAEH is used to reduce the computational burden and maintain high detection accuracy, so that the model can run efficiently in a resource-constrained environment. The Loss function is improved to make the parameter update more accurate during the model training process and improve the generalization ability of the model.

[0012] Step 3. Determine the required hyperparameter configuration before improving the network model training;

[0013] Step 4. Conduct model training and regularly evaluate the model using recognized detection performance indicators to ensure the traceability and quantification of training results. Fine-tune model parameters based on evaluation results to achieve gradual improvement of model performance and obtain a trained model.

[0014] Step 5. Obtain the trained improved YOLOv8 network model, input the test set images in the dataset into the trained improved YOLOv8 network model, perform detection and reasoning, and detect the reasoning results.

[0015] The beneficial effects of the present invention are as follows: the improved YOLOv8 network in the method of the present invention optimizes the efficiency and accuracy of feature extraction by integrating the FocNeXt module in the backbone network, thereby providing more abundant and accurate feature information for subsequent detection tasks; further, the RepBiFPN structure is applied in the feature fusion network to enhance the information flow between features of different scales, and effectively improve the model's ability to recognize targets; at the same time, by adopting the lightweight detection head LAEH, the computational burden of the model is reduced while maintaining high detection accuracy, so that the model can run efficiently even in a resource-constrained environment; in addition, the improvement of the loss function makes the parameter update more accurate during the model training process, improves the generalization ability of the model, and thus can maintain stable detection performance in various complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] Figure 1 is an overall flow chart of an embodiment of the present invention;

[0018] Figure 2 This is a diagram showing the overall structure of an improved YOLOv8 network according to an embodiment of the present invention;

[0019] Figure 3 A structural diagram of a FocNeXt module according to an embodiment of the present invention;

[0020] Figure 4 It is a structural diagram of FocNeXtBlock in the FocNeXt module of an embodiment of the present invention;

[0021] Figure 5 This is a structural diagram of the LAEH module of the improved high-efficiency and lightweight detection head according to an embodiment of the present invention;

[0022] Figure 6 An aerial image data set analysis for an embodiment of the present invention;

[0023] Figure 7 This is a high-altitude aerial target detection effect diagram of the original model of the embodiment of the present invention;

[0024] Figure 8 This is a rendering of the high-altitude aerial target detection effect of the improved model in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] See also Figures 1 to 8 This example provides a high-altitude aerial image target detection method based on improved YOLOv8, which can be used for high-altitude aerial photography to detect small ground targets, including:

[0027] Step 1: First, obtain the aerial image dataset publicly available from VisDrone and perform image preprocessing on it; then divide the dataset into three parts: training set, validation set, and test set.

[0028] Step 2: Process the YOLOv8 model, including integrating the FocNeXt module in the backbone network to optimize the efficiency and accuracy of feature extraction, and applying the RepBiFPN structure to enhance the information flow between features of different scales and improve the target recognition ability. The lightweight detection head LAEH is used to reduce the computational burden and maintain high detection accuracy, so that the model can run efficiently in a resource-constrained environment. The Loss function is improved to make the parameter update more accurate during the model training process and improve the generalization ability of the model.

[0029] Step 3: Determine all required hyperparameter configurations before improving network model training.

[0030] Step 4: Conduct model training. During the training process, the model is regularly evaluated using recognized detection performance indicators to ensure the traceability and quantifiability of the training effect. The model parameters are fine-tuned based on the evaluation results to achieve a gradual improvement in model performance and ultimately obtain the best training model.

[0031] Step 5: Finally, the trained improved YOLOv8 network model is obtained, and the test set images in the data set are input into the trained improved YOLOv8 network model for detection and reasoning, and the reasoning results are detected.

[0032] In this embodiment, the step 1 is specifically as follows:

[0033] Get the dataset,VisDrone dataset is a large-scale benchmark dataset created by the AISKYEYE team of the Machine Learning and Data Mining Laboratory of Tianjin University. It contains carefully annotated ground truth data for various computer vision tasks related to aerial-based image and video analysis. VisDrone consists of 288 video clips containing 261,908 frames and 10,209 static images taken by various aerial cameras. Figure 6 As shown: This dataset covers a wide range of aspects, including location (14 different cities in China), environment (urban and rural), objects (pedestrians, vehicles, bicycles, etc.) and density (sparse and crowded scenes). The dataset was collected using various aerial photography platforms in different scenes, weather and lighting conditions. These frames are manually annotated with more than 2.6 million target bounding boxes, such as pedestrians, cars, bicycles, and tricycles, and also provide attributes such as scene visibility, object category, and occlusion to better utilize the data. This embodiment uses the training set, validation set, and test set in the dataset, and modifies the annotation format required for YOLOv8 network model training to obtain a high-altitude aerial image dataset and its annotations that can be trained on the YOLOv8 network model.

[0034] In this embodiment, the step 2 is specifically as follows:

[0035] In the detection task of high-altitude aerial images, considering the need for lightweight and deployment on embedded devices, this embodiment uses the original YOLOv8 model as the basic network model for improvement.

[0036] YOLOv8 includes five versions of different sizes: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l and YOLOv8x. These models are composed of three key components: Backbone, Neck and Head. Backbone is responsible for extracting features from the input image, Neck integrates features at different levels to generate multi-scale feature maps with sizes of 80*80, 40*40 and 20*20, and Head uses these features for target detection. The loss function (Loss) is also an important part of the improved model. Loss provides guidance for model learning by accurately quantifying the difference between the predicted and true labels, thereby significantly improving the accuracy of target detection. The improvements in this embodiment focus on the core links of Backbone, Neck, Head and Loss.

[0037] This embodiment uses the FocNext module to improve the backbone network of YOLOv8, such as Figure 2As shown in the figure, the feature information of the input image is first extracted through the Conv convolution module to form a feature map, and then the feature map of the base layer is divided into two parts through the FocNeXt module, and then they are merged through a cross-stage hierarchy with 128 channels.

[0038] The specific steps of FocNeXtBlock construction are:

[0039] First, the first part uses two dilated depthwise convolutional modules and two residual connections.

[0040] The depth-wise separable convolution consists of two layers: depth-wise convolution and point-wise convolution. Depth-wise convolution is used to apply a filter to each input channel, and then 1*1 point-wise convolution is used to create a linear combination of the depth-wise layer output. BN and ReLU nonlinearity are used for both layers.

[0041] The depthwise convolution with one filter per input channel can be written as:

[0042]

[0043] Where K is the size of D k ×D k ×M depth convolution kernel, the m-th filter in K is applied to the m-th channel in F to generate the m-th channel of the filtered output feature map G.

[0044] The computational cost of depthwise convolution is:

[0045] D k *D k *M*D F *D F

[0046] The depth convolution generates new features through an additional layer of 1*1 convolution to calculate the linear combination of the depth convolution output. The combination of depth convolution and 1*1 point-by-point convolution is called depth-wise separable convolution dwconv.

[0047] The computational cost of depthwise separable convolution is:

[0048] D k *D k *M*D F *D F +M*N*D F *D F

[0049] Then add two Conv convolution modules of size 1*1, and finally use a residual connection.

[0050] Finally, the FocNeXt module F″(x) is constructed, and its expression is as follows:

[0051] F(x)=dwconv(x)+x

[0052] F′(x)=dwconv(F(x))+x

[0053] F″(x)=conv(conv(F′(x)))

[0054] Next, downsampling is performed through the Conv convolution module, and the feature map of the base layer is divided into two parts through the FocNeXt_P2 module, and then they are merged through the cross-stage hierarchy with a channel of 256.

[0055] Next, downsampling is performed through the Conv convolution module, and the feature map of the base layer is divided into two parts through the FocNeXt_P3 module, and then they are merged through the cross-stage hierarchy with a channel of 512.

[0056] Next, downsampling is performed through the Conv convolution module, and the feature map of the base layer is divided into two parts through the FocNeXt_P4 module, and then they are merged through the cross-stage hierarchy with a channel of 1024.

[0057] They are 128, 256, 512, and 1024 channels respectively. Finally, the features extracted by the backbone network (BackBone) are fused and passed into the feature fusion network (Neck).

[0058] Use the RepBiFPN structure to optimize the feature fusion network. Figure 2 As shown: In order to enhance the positioning signal without increasing too much computational burden, a bidirectional concatenation (BiFusion) module is used to integrate the feature maps of three adjacent layers, fusing additional low-level features from the backbone, which can retain more accurate positioning signals, which is of great significance for the positioning of small objects.

[0059] The specific process of improving the feature fusion network (Neck) is as follows:

[0060] The BiFusion fusion module is introduced in the feature fusion network (Neck) to integrate the feature maps of three adjacent layers and fuse the additional low-level features from the backbone to retain more accurate positioning signals.

[0061] In the first part, the Conv convolution module is first used to adjust the channel, and then the BiFusion fusion module is used. The features of the second feature extraction module, the third feature extraction module and the previous layer are used as input. By fusing the features of different feature layers, the feature maps of the three adjacent layers are integrated to retain a more accurate positioning signal.

[0062] Then, the features output by the BiFusion module are adjusted through the Conv convolution module and input into the feature fusion module. By screening and fusing the feature information of small targets, the detection accuracy of small targets is improved.

[0063] The second part first uses the Conv convolution module to adjust the channel, and then uses the BiFusion fusion module. It uses the features of the first feature extraction module, the second feature extraction module, and the previous layer as inputs. By fusing the features of different feature layers, it integrates the feature maps of three adjacent layers to retain a more accurate positioning signal.

[0064] Then, the features output by the BiFusion module are adjusted through the Conv convolution module and input into the feature fusion module. By screening and fusing the feature information of small targets, the detection accuracy of small targets is improved.

[0065] Then, the output features are adjusted through the Conv convolution module and input into the feature fusion module. The small target feature information is screened and fused, and the execution is repeated twice. The obtained feature map is directly input into the detection head.

[0066] The detection head (Head) performs regression prediction and classification prediction on the feature map extracted after fusion by the feature fusion network (Neck). In view of the characteristics of small targets in high-altitude aerial images, this embodiment uses a detection head based on efficient and lightweight LAEH. Compared with the classification branch, its regression branch uses fewer convolution modules, thereby reducing the amount of calculation. The classification branch focuses on determining the category of the extracted features that are most similar to the target category, while the localization branch focuses on refining the center point coordinates, width, and height information of the box to correct the bounding box parameters, which makes the classification focus more on the center content and the regression focus more on the edge information.

[0067] The specific process of improving the detection head is as follows:

[0068] In the IoU localization branch, a Conv convolution is first used, followed by a SaELayerSqueezeAggregated Excitation (SaE) module, which enhances the network’s ability to capture channel patterns and global knowledge, leading to better feature representation.

[0069] Then, a multi-branch fully connected layer of the same size is introduced. The FC operation starts after the global average pooling layer, which extracts channel statistics. These channel information is then directed to an FC operation, where the dimension of the input is reduced. Next is the Reshape layer. In the architecture, a residual module is constructed by repeating the convolution layer after a specific interval to form a structured module. These modal units are repeated in cycles to ensure that the learned gradients are effectively propagated without being affected by the vanishing gradient problem that may occur in deep networks. The predicted positioning results are output through a convolution module.

[0070] In the Cls classification branch, a Conv convolution module is first used, and then two depth-wise separable convolution (DWConv) modules are used continuously to adjust the channel so that the classification focuses more on the central content. The predicted classification result is output through a convolution module, and finally the positioning and recognition results predicted by the detection head are output.

[0071] This embodiment adopts the Shape-IoU loss function as the loss function of the improved YOLOv8 network model.

[0072] The Shape-IoU loss function is as follows:

[0073] L shape-IoU =1-IoU+distance shape +0.5*ρ shape

[0074] Where IoU represents the following intersection-over-union function:

[0075]

[0076] B gt represents the area of ​​the true bounding box, and B represents the area of ​​the predicted box.

[0077]

[0078]

[0079] Among them, the value of θ defines the Shape Cost of the dataset and its value is unique, indicating the degree of attention paid to the Shape Cost.

[0080]

[0081]

[0082]

[0083] scale is a scaling factor, which is related to the size of the target in the dataset. ww and hh represent the weight coefficients in the horizontal and vertical directions, and their values ​​are related to the shape of the GT box.

[0084] In this embodiment, step 3 is specifically as follows: before starting the training of the improved network model, it is necessary to determine all relevant hyperparameter configurations, set the learning rate to 0.001, the batch size to 16, the training epoch to 200, the input image to 640×640, the initial learning rate lr0 to 0.01, the final learning rate lr to 0.01, the momentum to 0.937, the optimizer weight decay weight_decay to 0.0005, the warm-up epoch to 3.0, the warm-up initial momentum warmup_momentum to 0.8, and the warm-up initial deviation warmup_bias_lr to 0.1.

[0085] In this embodiment, step 4 is specifically as follows: performing model training, continuously optimizing the network during the training process, performing epoch training, and using common detection performance indicators to evaluate the training effect during the training process, such as precision P (Precision), recall R (Recall), accuracy index mAP.50 and average accuracy index mAP.50-95 as key indicators for evaluating model performance.

[0086] In this embodiment, step 5 is specifically as follows: obtaining a trained improved YOLOv8 network model, inputting the test set images in the high-altitude aerial image data set into the trained improved YOLOv8 network model, performing detection and reasoning, and obtaining the detection and reasoning results of the high-altitude aerial image targets.

[0087] In order to verify the improvement effect of the present invention on the YOLOv8 high-altitude aerial image detection model, complex scene images in the high-altitude aerial image test set were specially selected for evaluation. The test results show that Figure 7 and Figure 8 As shown in the figure, there is a significant difference in detection effect between the original YOLOv8 model and the improved model of the present invention. Figure 7 The detection results of the original YOLOv8 model are shown, while Figure 8 The detection results of the improved model are presented. In the comparison between the improved model and the baseline model, this embodiment achieves a 17.21% reduction in the number of parameters and improves the detection accuracy by 7.1%.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A small target detection method for high-altitude aerial images based on improved YOLOv8, characterized in that: The following steps are involved: Step 1. Obtain the aerial image dataset publicly available from VisDrone, preprocess it, and divide the dataset into three parts: training set, validation set, and test set; Step 2. Process the YOLOv8 model, including integrating the FocNeXt module in the backbone network to optimize the efficiency and accuracy of feature extraction, and applying the RepBiFPN structure to enhance the information flow between features of different scales and improve the target recognition ability. The lightweight detection head LAEH is used to reduce the computational burden and maintain the detection accuracy, so that the model can run efficiently in a resource-constrained environment. The Loss function is improved to make the parameters accurate during the model training process and improve the generalization ability of the model. Step 3. Determine the required hyperparameter configuration before improving the network model training; Step 4. Conduct model training and regularly evaluate the model using recognized detection performance indicators to ensure the traceability and quantification of training results. Fine-tune model parameters based on evaluation results to achieve gradual improvement of model performance and obtain a trained model. Step 5. Obtain the trained improved YOLOv8 network model, input the test set images in the dataset into the trained improved YOLOv8 network model, perform detection and reasoning, and detect the reasoning results.

2. The small target detection method for high-altitude aerial images based on improved YOLOv8 according to claim 1 is characterized in that: The specific steps of step 2 include: S21. Select the original YOLOv8 network model as the basis and make improvements on this basis; S22. Use the FocNeXt module to improve the backbone network of YOLOv8. The improved network is constructed by the following method: The Conv convolution module extracts the feature information of the input image to form a feature map, and then the FocNeXt module divides the feature map of the base layer into two parts, and then merges them through a cross-stage hierarchy with 128 channels. Among them, the specific steps of FocNeXtBlock construction are: The first part uses two dilated depthwise convolutional modules and two residual connections; The depth-wise separable convolution consists of two layers: depth-wise convolution and point-wise convolution. A filter is applied to each input channel using depth-wise convolution. A linear combination of the outputs of the depth layer is created using 1*1 point-wise convolution. BN and ReLU nonlinearity are used for both layers. The depthwise convolution with one filter per input channel can be written as: Where K is the size of D k ×D k ×M depth convolution kernel, the m-th filter in K is applied to the m-th channel in F to generate the m-th channel of the filtered output feature map G; The computational cost of depthwise convolution is: D k *D k *M*D F *D F The depth convolution generates new features by calculating the additional layer of linear combination of the depth convolution output through 1*1 convolution. The combination of depth convolution and 1*1 point-by-point convolution is called depth separable convolution dwconv. The computational cost of depthwise separable convolution is: D k *D k *M*D F *D F +M*N*D F *D F Then add two Conv convolution modules of size 1*1, and finally use a residual connection; Finally, the FocNeXt module F″(x) is constructed, and its expression is as follows: F(x)=dwconv(x)+x F′(x)=dwconv(F(x))+x F″(x)=conv(conv(F′(x))) Down-sampling is performed through the Conv convolution module, and the feature map of the base layer is divided into two parts through the FocNeXt_P2 module, and they are merged through the cross-stage hierarchy with 256 channels; Down-sampling is performed through the Conv convolution module, and the feature map of the base layer is divided into two parts through the FocNeXt_P3 module, which are merged through the cross-stage hierarchy with a channel of 512; Down-sampling is performed through the Conv convolution module, and the feature map of the base layer is divided into two parts through the FocNeXt_P4 module, and they are merged through the cross-stage hierarchy with a channel of 1024; The features extracted by the backbone network are fused through 128, 256, 512, and 1024 channels and then passed into the feature fusion network; S23. Use the RepBiFPN structure to optimize the feature fusion network. The specific improvement process of the feature fusion network is as follows: A bidirectional concatenation (BiFusion) module is introduced into the feature fusion network to integrate the feature maps of three adjacent layers, fusing additional low-level features from the backbone network to retain accurate positioning signals; Use the Conv convolution module to adjust the channel, and then use the BiFusion fusion module. Use the second feature extraction module features, the third feature extraction module features, and the previous layer features as inputs. By fusing the features of different feature layers, the feature maps of three adjacent layers are integrated to retain accurate positioning signals. The features output by the BiFusion module are adjusted through the Conv convolution module and then input into the feature fusion module. The detection accuracy of small targets is improved by screening and fusing the feature information of small targets. Use the Conv convolution module to adjust the channel, and then use the BiFusion fusion module. Use the first feature extraction module features, the second feature extraction module features, and the previous layer features as inputs. By fusing the features of different feature layers, the feature maps of three adjacent layers are integrated to retain the positioning signal. The features output by the BiFusion module are adjusted through the Conv convolution module and then input into the feature fusion module. The detection accuracy of small targets is improved by screening and fusing the feature information of small targets. After the output features are adjusted through the Conv convolution module, they are input into the feature fusion module. The small target feature information is filtered and fused, and the resultant feature map is directly input into the detection head. S24. Improve the detection head (Head), perform regression prediction and classification prediction on the feature map extracted after the feature fusion network (Neck) is fused, and use the efficient and lightweight detection head LAEH based on the characteristics of small targets in high-altitude aerial images. The specific process of improving the detection head (Head) is as follows: In the IoU localization branch, a Conv convolution is first used, followed by a SaELayerSqueeze AggregatedExcitation (SaE) module, which enhances the network’s ability to capture channel patterns and global knowledge, leading to better feature representation; The FC operation starts after the global average pooling layer, which extracts channel statistics. These channel information are then directed to an FC operation, where the dimension of the input is reduced, followed by a Reshape layer. In the architecture, a residual module is constructed by repeating the convolution layer after a specific interval to form a structured module, which outputs the predicted positioning result through a convolution module. In the Cls classification branch, a Conv convolution module is first used, and then two depth-wise separable convolution (DWConv) modules are used in succession to adjust the channel so that the classification focuses more on the central content. The predicted classification result is output through a convolution module; finally, the positioning and recognition results predicted by the detection head are output; S25. Improve the loss function (Loss), and use the Shape-IoU loss function as the loss function of the improved YOLOv8 network model. The Shape-IoU loss function is improved as follows: L Shape-IoU =1-IoU+distance shape +0.5*p shape Among them, IoU represents the following intersection-over-union function: B gt Represents the area size of the real bounding box, and B represents the area size of the predicted box; Among them, the value of θ defines the Shape cost of the data set and its value is unique, indicating the degree of attention paid to the Shape cost; Among them, scale is the scaling factor, which is related to the size of the target in the dataset, ww and hh represent the weight coefficients in the horizontal and vertical directions, and their values ​​are related to the shape of the GT box.

3. The small target detection method for high-altitude aerial images based on improved YOLOv8 according to claim 1 is characterized in that: The step 1 specifically includes: obtaining a data set, using a training set, a validation set and a test set in the data set, and modifying the annotation format required for YOLOv8 network model training to obtain a high-altitude aerial image data set and its annotations that can be trained on the YOLOv8 network model.

4. The small target detection method for high-altitude aerial images based on improved YOLOv8 according to claim 1 is characterized in that: The step 3 is specifically: before starting the training of the improved network model, it is necessary to determine the relevant hyperparameter configuration, set the learning rate to 0.001, the batch size to 16, the training epoch to 200, the input image to 640×640, the initial learning rate lr0 to 0.01, the final learning rate lr to 0.01, the momentum to 0.937, the optimizer weight decay weight_decay to 0.0005, the warm-up epoch to 3.0, the warm-up initial momentum warmup_momentum to 0.8, and the warm-up initial bias warmup_bias_lr to 0.

1.

5. The small target detection method for high-altitude aerial images based on improved YOLOv8 according to claim 1 is characterized in that: The step 4 is specifically: perform model training, continuously optimize the network during the training process, perform epoch training, and use common detection performance indicators to evaluate the training effect during the training process, such as precision P (Precision), recall R (Recall), accuracy index mAP.50 and average accuracy index mAP.50-95 as key indicators for evaluating model performance.

6. The small target detection method for high-altitude aerial images based on improved YOLOv8 according to claim 1 is characterized in that: The step 5 specifically includes: obtaining a trained improved YOLOv8 network model, inputting the test set images in the high-altitude aerial image data set into the trained improved YOLOv8 network model, performing detection and reasoning, and obtaining the detection and reasoning results of the high-altitude aerial image targets.

Images