Method, device and equipment for detecting aircraft in optical remote sensing images

Abstract

The application provides a method, device and equipment for detecting an airplane in an optical remote sensing image, and relates to the technical field of optical remote sensing image processing. The method comprises the following steps: when detecting an airplane, an optical remote sensing image to be detected is acquired, and the optical remote sensing image to be detected is input into an airplane detection model to obtain an airplane detection result output by the airplane detection model; wherein the airplane detection model is obtained by training an initial airplane detection model based on an improved lightweight attention mechanism image denoising network ADNet extracting image features of different levels from optical remote sensing images in a training set and based on the extracted multi-level image features and corresponding labeled labels. By using the technical solution provided in the application, the defect of low detection accuracy of the traditional airplane detection algorithm in the optical remote sensing image can be solved, and thus the accuracy of the airplane detection result in the target optical remote sensing image is improved.

Description

Technical Field

[0001] This application relates to the field of image technology, and in particular to a method, apparatus and equipment for detecting aircraft in optical remote sensing images. Background Technology

[0002] Detecting aircraft in optical remote sensing images is crucial in scenarios such as traffic safety, disaster assessment, economic development, and military reconnaissance. However, due to their unique morphological characteristics, complex background environments, and variable attitude changes, detecting aircraft in optical remote sensing images is an extremely challenging task.

[0003] Traditional aircraft detection algorithms in optical remote sensing images: use a fixed scale to extract features of aircraft in optical remote sensing images, and then use machine learning to classify the extracted features, thereby achieving aircraft detection based on the classification results.

[0004] However, using the aforementioned traditional aircraft detection algorithm in optical remote sensing images results in low accuracy. Summary of the Invention

[0005] This application provides a method, apparatus, and device for aircraft detection in optical remote sensing images, which addresses the shortcomings of traditional aircraft detection algorithms in optical remote sensing images, namely, low accuracy of detection results, thereby improving the accuracy of aircraft detection results in optical remote sensing images.

[0006] This application provides a method for detecting aircraft in optical remote sensing images, including:

[0007] Acquire optical remote sensing images of the target to be detected;

[0008] The target optical remote sensing image is input into the aircraft detection model to obtain the aircraft detection result output by the aircraft detection model;

[0009] The aircraft detection model is obtained by using ADNet, an image denoising network with an improved lightweight attention mechanism, to extract image features at different levels from optical remote sensing images in the training set, and then training the initial aircraft detection model based on the extracted multi-level image features and corresponding labels.

[0010] According to the aircraft detection method in optical remote sensing images provided in this application, the initial aircraft detection model includes the image denoising network ADNet based on an improved lightweight attention mechanism, the bidirectional feature pyramid network BiFPN, and the detection head network.

[0011] The aircraft detection model was trained using the following method:

[0012] The optical remote sensing image is input into the image denoising network ADNet to obtain the multi-level image features output by the image denoising network ADNet; wherein, the multi-level image features include first-level image features, second-level image features and third-level image features;

[0013] The first-level image features, the second-level image features, and the third-level image features are all input into the Bidirectional Feature Pyramid Network (BiFPN) to obtain the first-scale features corresponding to the first-level image features, the second-scale features corresponding to the second-level image features, and the third-scale features corresponding to the third-level image features output by the BiFPN.

[0014] The first scale feature, the second scale feature, and the third scale feature are all input into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network.

[0015] Based on the classification loss function corresponding to the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, the model parameters in the initial aircraft detection model are updated to obtain the aircraft detection model.

[0016] According to the aircraft detection method in optical remote sensing images provided in this application, the image denoising network ADNet includes a sparse module SB, a feature enhancement module FEB, and an attention module based on an improved lightweight attention mechanism, which are connected in series.

[0017] The step of inputting the optical remote sensing image into the image denoising network ADNet to obtain the multi-level image features output by the image denoising network ADNet includes:

[0018] The optical remote sensing image is input into the sparse module SB to obtain the first-level image features output by the sparse module SB;

[0019] Both the first-level image features and the optical remote sensing image are input into the feature enhancement module FEB, and the second-level image features and convolutional image features are output through the feature enhancement module FEB.

[0020] The second-level image features and the convolutional image features are both input into the attention module based on the improved lightweight attention mechanism, and the third-level image features are output through the attention module.

[0021] According to the aircraft detection method in optical remote sensing images provided in this application, the feature enhancement module FEB includes a first convolutional normalization unit, a second convolutional normalization unit, a third convolutional normalization unit, a convolutional layer, and an activation function layer connected in series.

[0022] The step of inputting both the first-level image features and the optical remote sensing image into the feature enhancement module FEB, and outputting the second-level image features and convolutional image features through the feature enhancement module, includes:

[0023] The image features of the first level are sequentially input into the first convolutional normalization unit, the second convolutional normalization unit, and the third convolutional normalization unit to obtain the convolutional normalized features output by the third convolutional normalization unit.

[0024] The convolutional normalized features are input into the convolutional layer to obtain the convolutional image features output by the convolutional layer;

[0025] The sum of the optical remote sensing image and the convolutional image features is input into the activation function layer to obtain the second-level image features output by the activation function layer.

[0026] According to the method for detecting aircraft in optical remote sensing images provided in this application, the detection head network includes multiple detection heads of different scales connected in parallel, and the multiple detection heads include a detection head corresponding to a first scale, a detection head corresponding to a second scale, and a detection head corresponding to a third scale.

[0027] Specifically, the first scale feature, the second scale feature, and the third scale feature are all input into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network, including:

[0028] The first scale feature is input into the detection head corresponding to the first scale to obtain the corresponding first aircraft category prediction probability and the first aircraft detection box;

[0029] The second scale feature is input into the detection head corresponding to the second scale to obtain the corresponding second aircraft category prediction probability and the second aircraft detection box;

[0030] The third-scale features are input into the detection head corresponding to the third scale to obtain the corresponding third aircraft category prediction probability and the third aircraft detection box;

[0031] The predicted probability of the aircraft category is determined based on the predicted probability of the first aircraft category, the predicted probability of the second aircraft category, and the predicted probability of the third aircraft category;

[0032] The aircraft detection frame is determined based on the first aircraft detection frame, the second aircraft detection frame, and the third aircraft detection frame.

[0033] According to the aircraft detection method in optical remote sensing images provided in this application, the classification loss function corresponding to the predicted probability of the aircraft category can be seen from the following formula:

[0034]

[0035] in, This represents the classification loss function corresponding to the predicted probability of the aircraft category. This indicates the number of optical remote sensing images. Indicates the first An optical remote sensing image, Indicates the first Target aircraft category in optical remote sensing images The predicted probability, This represents the preset probability threshold.

[0036] This application also provides an aircraft detection device in optical remote sensing images, comprising:

[0037] The acquisition unit is used to acquire optical remote sensing images of the target to be detected;

[0038] The detection unit is used to input the target optical remote sensing image into the aircraft detection model and obtain the aircraft detection result output by the aircraft detection model.

[0039] The aircraft detection model is obtained by using ADNet, an image denoising network with an improved lightweight attention mechanism, to extract image features at different levels from optical remote sensing images in the training set, and then training the initial aircraft detection model based on the extracted multi-level image features and corresponding labels.

[0040] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aircraft detection method in optical remote sensing images as described above.

[0041] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aircraft detection method in optical remote sensing images as described above.

[0042] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the aircraft detection method in optical remote sensing images as described above.

[0043] The aircraft detection method, apparatus, and device provided in this application involve inputting the target optical remote sensing image to an aircraft detection model during aircraft detection, and obtaining the aircraft detection result output by the model. The aircraft detection model is obtained by using an improved lightweight attention-based image denoising network (ADNet) to extract image features at different levels from the optical remote sensing images in the training set, and training an initial aircraft detection model based on the extracted multi-level image features and corresponding labels. Given that the aircraft detection model has good feature enhancement capabilities and can effectively capture and utilize image features at different levels, when inputting the target optical remote sensing image to the aircraft detection model and determining the aircraft detection result, it can effectively capture and utilize image feature information at different levels of the target optical remote sensing image. This combination of image feature information at different levels to determine the aircraft detection result allows for rapid and accurate aircraft detection of the target optical remote sensing image, overcoming the low detection accuracy of traditional aircraft detection algorithms in optical remote sensing images, and thus improving the accuracy of aircraft detection results in target optical remote sensing images. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating a method for detecting aircraft in optical remote sensing images, provided as an embodiment of this application.

[0046] Figure 2 This is a schematic diagram of the structure of an initial aircraft detection model provided in an embodiment of this application.

[0047] Figure 3 This is a flowchart illustrating a training method for an aircraft detection model provided in an embodiment of this application.

[0048] Figure 4 This is a schematic diagram of the structure of an aircraft detection device in an optical remote sensing image, provided as an embodiment of this application.

[0049] Figure 5 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone, where A and B can be singular or plural. In the textual description of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0052] The technical solutions provided in this application can be applied to scenarios such as traffic safety, disaster assessment, economic development, and military reconnaissance. Traditional aircraft detection algorithms in optical remote sensing images extract features from the images using a fixed scale, and then classify the extracted features using machine learning to detect the aircraft based on the classification results.

[0053] However, the traditional aircraft detection algorithm used in optical remote sensing images has the problem of low detection accuracy.

[0054] To address the low accuracy of traditional aircraft detection algorithms in optical remote sensing images and improve the accuracy of aircraft detection results in optical remote sensing images, this application provides an aircraft detection method in optical remote sensing images. The method involves pre-extracting image features at different levels from the optical remote sensing images in the training set using an improved lightweight attention-guided denoising convolutional neural network (ADNet). Based on the extracted multi-level image features and corresponding labels, an initial aircraft detection model is trained to obtain the aircraft detection model. Given that aircraft detection models possess strong feature enhancement capabilities and can effectively capture and utilize image features at different levels, when inputting the target optical remote sensing image into the aircraft detection model and determining the aircraft detection result through the model, image feature information at different levels of the target optical remote sensing image can be effectively captured and utilized. This combination of image feature information at different levels to determine the aircraft detection result allows for rapid and accurate acquisition of aircraft detection results from the target optical remote sensing image. This addresses the shortcomings of traditional aircraft detection algorithms in optical remote sensing images, which suffer from low detection accuracy, thereby improving the accuracy of aircraft detection results in target optical remote sensing images.

[0055] It is understood that the subject of this method can be an electronic device such as an aircraft detection device, a computer or server, or an aircraft detection device in an optical remote sensing image set in the electronic device. The aircraft detection device in the optical remote sensing image can be implemented by software, hardware or a combination of both, and can be set according to actual needs.

[0056] The aircraft detection method in optical remote sensing images provided in this application will be described in detail below through several specific embodiments. It is understood that these specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0057] Figure 1 This application provides a flowchart illustrating a method for detecting aircraft in optical remote sensing images. For example, please refer to [link to relevant documentation]. Figure 1 As shown, the aircraft detection method in this optical remote sensing image may include:

[0058] S101. Acquire the optical remote sensing image of the target to be detected.

[0059] For example, the optical remote sensing image of the target to be detected can be obtained by taking pictures by remote sensing satellites, or by obtaining the optical remote sensing image of the target to be detected from an open-source remote sensing image database, or by other means, such as obtaining the optical remote sensing image of the target to be detected from a remote sensing data sharing library, etc. The specific settings can be made according to actual needs.

[0060] Normally, before inputting the target optical remote sensing image into the aircraft detection model, the target optical remote sensing image can be preprocessed, such as denoising, geometric correction, atmospheric correction, resolution adjustment, etc. The preprocessed target optical remote sensing image is then used as input to the aircraft detection model so that the aircraft detection model can output the aircraft detection result, i.e., the following S102 is executed.

[0061] S102. Input the target optical remote sensing image into the aircraft detection model to obtain the aircraft detection results output by the aircraft detection model.

[0062] The aircraft detection model is obtained by using ADNet, an image denoising network with an improved lightweight attention mechanism, to extract image features at different levels from optical remote sensing images in the training set. The initial aircraft detection model is then trained based on the extracted multi-level image features and corresponding labels.

[0063] For example, aircraft detection results can include the aircraft's location in the optical remote sensing image, the probability of the aircraft's class, etc., which can be set according to actual needs.

[0064] As can be seen, in this embodiment, when detecting an aircraft, the optical remote sensing image of the target to be detected is input into the aircraft detection model to obtain the aircraft detection result output by the aircraft detection model. The aircraft detection model is obtained by using an improved lightweight attention-based image denoising network (ADNet) to extract image features at different levels from the optical remote sensing images in the training set, and by training the initial aircraft detection model based on the extracted multi-level image features and corresponding labels. Given that the aircraft detection model has good feature enhancement capabilities and can effectively capture and utilize image features at different levels, when the target optical remote sensing image to be detected is input into the aircraft detection model, and the aircraft detection result is determined by the aircraft detection model, image feature information at different levels of the target optical remote sensing image can be effectively captured and utilized. By combining image feature information at different levels to determine the aircraft detection result, the aircraft detection result of the target optical remote sensing image can be obtained quickly and accurately, solving the problem of low detection accuracy in traditional aircraft detection algorithms for optical remote sensing images, thereby improving the accuracy of aircraft detection results in target optical remote sensing images.

[0065] based on Figure 1As illustrated in the embodiments, in this application embodiment, the initial aircraft detection model can be an improved YOLOv10 network model. The backbone network in the YOLOv10 network model is replaced with an image denoising network ADNet based on an improved lightweight attention mechanism. At the same time, the AB module in the YOLOv10 network model is replaced with a lightweight attention mechanism. The neck structure in the YOLOv10 network model is modified by replacing the original path aggregation network (PAN) structure with a bidirectional feature pyramid network (BiFPN) structure. The improved YOLOv10 network model is then used as the initial aircraft detection model.

[0066] For example, see Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of an initial aircraft detection model provided in an embodiment of this application. The initial aircraft detection model may include an image denoising network ADNet based on an improved lightweight attention mechanism, a bidirectional feature pyramid network BiFPN, and a detection head network, combined with... Figure 2 The initial aircraft detection model shown below will be analyzed through the following... Figure 3 The illustrated embodiment describes a method for training an aircraft detection model.

[0067] Typically, training an initial aircraft detection model involves acquiring optical remote sensing images, preprocessing them, and labeling them into a dataset. The dataset is then randomly divided into training, validation, and test sets, with different aircraft categories identified by rotating bounding boxes according to a predefined ratio. The training and validation sets are used to train the initial aircraft detection model, while the test data is used to test and validate the trained model, resulting in the final aircraft detection model. For example, the ratio of training, validation, and test sets can be 8:1:1, but this can be adjusted based on specific needs.

[0068] Combination Figure 2 The initial aircraft detection model shown can be used as the basis for training the aircraft detection model. For example, see [link to example]. Figure 3 As shown, Figure 3 This is a flowchart illustrating a training method for an aircraft detection model provided in an embodiment of this application. The training method for the aircraft detection model may include:

[0069] S301. Input the optical remote sensing image into the image denoising network ADNet to obtain the multi-level image features output by the image denoising network ADNet; wherein, the multi-level image features include the first-level image features, the second-level image features, and the third-level image features.

[0070] The image denoising network can be the ADNet network. In this embodiment, the backbone network in the YOLOv10 network model can be changed to an image denoising network based on an improved lightweight attention mechanism (Coordinate Attention), and the AB module in ADNet can be replaced with a lightweight attention mechanism.

[0071] For example, in the embodiments of this application, it can be combined with Figure 2 As shown, the image denoising network includes a sparse block (SB), a feature enhancement block (FEB), and an attention (Coordinate Attention) module based on an improved lightweight attention mechanism, which are connected in series.

[0072] When inputting an optical remote sensing image into the image denoising network ADNet to obtain multi-level image features output by ADNet, the optical remote sensing image can first be input into the sparse module SB to obtain the first-level image features output by the sparse module; then, both the first-level image features and the optical remote sensing image are input into the feature enhancement module FEB, through which the feature enhancement module FEB outputs the second-level image features and convolutional image features; finally, both the second-level image features and the convolutional image features are input into the attention module based on an improved lightweight attention mechanism, through which the attention module outputs the third-level image features.

[0073] In this process, optical remote sensing images are input into the sparse module. The sparse module SB can use dilated convolution and ordinary convolution to achieve a sparsity mechanism to obtain sparse convolution features. These sparse convolution features are the first-level image features output by the sparse module SB, thus achieving a balance between efficiency and performance.

[0074] Both the first-level image features and the optical remote sensing image are input into the feature enhancement module FEB. The feature enhancement module FEB integrates shallow and deep feature information through a long-range path to obtain enhanced convolutional features. These enhanced convolutional features are the second-level image features output by the feature enhancement module FEB to enhance the feature representation capability of the aircraft detection model.

[0075] The second-level image features and convolutional image features are both input into the attention module based on the improved lightweight attention mechanism to capture long-range dependencies between space and channels, obtain the local and global relationships of pixels in space, and obtain the third-level image features output by the attention module.

[0076] It is understood that in the embodiments of this application, a deep residual structure and a deeply supervised 1×1 convolutional layer are introduced for the above-mentioned sparse module SB, feature enhancement module FEB and attention module based on the improved lightweight attention mechanism.

[0077] The deep residual structure adds residual branch outputs to the sparse module SB, the feature enhancement module FEB, and the attention module based on the improved lightweight attention mechanism for multi-scale feature extraction; the deep supervised 1×1 convolutional layer is the connection layer after the deep residual structure, used to adjust the number of input feature channels.

[0078] For example, the second-level image features and convolutional image features are both input into an attention module based on an improved lightweight attention mechanism. When the third-level image features are output through the attention module, each channel is encoded in both the horizontal and vertical coordinate directions. The encoding expressions for these two spatial directions can be found in Equations 1 and 2 below.

[0079] Formula 1

[0080] Formula 2

[0081] in, Indicates the width of the input feature. H Indicates the height of the input feature. Represents pixels coordinates This indicates that the second-level image features and convolutional image features are fused together, and the results are applied horizontally to the channels. Features This indicates that the second-level image features and convolutional image features are fused together, and the results are applied to the channels in the vertical plane. Features Indicates the channel The horizontal feature map obtained by encoding This indicates the channel. The vertical feature map obtained from encoding.

[0082] The encoded horizontal and vertical feature maps are transformed using a shared 1×1 convolution to obtain the aggregated feature map, as shown in Formula 3 below.

[0083] Formula 3

[0084] in, Represents the aggregated feature map. Represents the horizontal feature map. This represents a vertical feature map, without distinguishing between channels.

[0085] The above aggregated feature map Divided along the spatial dimension into two separate tensors and The feature map is processed using two 1×1 convolutions. and feature map The feature map with the same number of channels as the input features of the transformation and attention module can be seen in Equation 4 below.

[0086] Formula 4

[0087] in, For feature map The feature map obtained after transformation For feature map The feature map obtained after transformation.

[0088] The transformed feature map is then calculated, as shown in Formula 5 below, to obtain the third-level image features output by the attention module.

[0089] Formula 5

[0090] in, Indicates in the channel The third level of image features, This represents the feature obtained by fusing the second-level image features and the convolutional image features. Feature map of channel c The feature map obtained after transformation Feature map of channel c The feature map obtained after transformation.

[0091] As can be seen from the embodiments of this application, when training the aircraft detection model, by improving the YOLOv10 model and replacing the backbone network in the YOLOv10 network model with an image denoising network ADNet based on an improved lightweight attention mechanism, not only can noise interference be effectively dealt with, enhancing the stability and reliability of the aircraft detection system, but also the introduced ADNet, the attention module based on the improved lightweight attention mechanism, and the bidirectional feature pyramid network BiFPN have lower computational complexity, significantly improving the aircraft detection efficiency, meeting the application scenarios with high real-time requirements, greatly improving detection accuracy, and reducing the probability of false detection and missed detection, providing an innovative and efficient solution for aircraft target detection in optical remote sensing images.

[0092] For example, in the embodiments of this application, it can be combined with Figure 2 As shown, the Feature Enhancement Module (FEB) may include a first convolutional normalization unit, a second convolutional normalization unit, a third convolutional normalization unit, a convolutional layer, and an activation function layer, which are connected in series.

[0093] Both the first-level image features and the optical remote sensing image are input into the Feature Enhancement Module (FEB). When the Feature Enhancement Module (FEB) outputs the second-level image features and convolutional image features, the first-level image features can be sequentially input into the first convolutional normalization unit, the second convolutional normalization unit, and the third convolutional normalization unit to obtain the convolutional normalized features output by the third convolutional normalization unit. The convolutional normalized features are then input into the convolutional layer to obtain the convolutional image features output by the convolutional layer. Finally, the sum of the optical remote sensing image and the convolutional image features is input into the activation function layer to obtain the second-level image features output by the activation function layer. These second-level image features are the enhanced convolutional features output by the Feature Enhancement Module.

[0094] After obtaining the first-level image features through the sparse module SB, the second-level image features through the feature enhancement module FEB, and the third-level image features through the attention module, the following S302 can be executed:

[0095] S302. Input the first-level image features, the second-level image features, and the third-level image features into the Bidirectional Feature Pyramid Network (BiFPN) to obtain the first-scale features corresponding to the first-level image features, the second-scale features corresponding to the second-level image features, and the third-scale features corresponding to the third-level image features output by the BiFPN.

[0096] In this embodiment, the neck structure in the YOLOv10 network model can be modified by replacing the original path aggregation network (PAN) structure with a Bidirectional Feature Pyramid Network (BiFPN) structure. BiFPN is an improved Feature Pyramid Network (FPN) that achieves bidirectional feature fusion by introducing three top-down and bottom-up paths. This results in first-level image features corresponding to the first scale, second-level image features corresponding to the second scale, and third-level image features corresponding to the third scale. Learnable weights are added to better utilize feature information from different levels, while simplifying the feature fusion path and nodes, thus reducing computational complexity.

[0097] S303. Input the first-scale feature, the second-scale feature, and the third-scale feature into the detection head network to obtain the aircraft category prediction probability and the aircraft detection box output by the detection head network.

[0098] For example, in the embodiments of this application, it can be combined with Figure 2 As shown, the detection head network includes multiple detection heads of different scales connected in parallel. The multiple detection heads include a detection head corresponding to the first scale, a detection head corresponding to the second scale, and a detection head corresponding to the third scale.

[0099] When inputting the first-scale features, second-scale features, and third-scale features into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network, the first-scale features can be input into the detection head corresponding to the first scale to obtain the corresponding first aircraft category prediction probability and first aircraft detection box; the second-scale features can be input into the detection head corresponding to the second scale to obtain the corresponding second aircraft category prediction probability and second aircraft detection box; the third-scale features can be input into the detection head corresponding to the third scale to obtain the corresponding third aircraft category prediction probability and third aircraft detection box; then, based on the first aircraft category prediction probability, the second aircraft category prediction probability, and the third aircraft category prediction probability, the aircraft category prediction probability is determined; and based on the first aircraft detection box, the second aircraft detection box, and the third aircraft detection box, the aircraft detection box is determined.

[0100] Taking the determination of the aircraft category prediction probability based on the prediction probability of the first aircraft category, the prediction probability of the second aircraft category, and the prediction probability of the third aircraft category as an example, the prediction probabilities of these three aircraft categories can be weighted and the weighted result can be used as the final aircraft category prediction probability. The specific settings can be set according to actual needs.

[0101] After determining the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, the model parameters in the initial aircraft detection model can be updated based on the classification loss function corresponding to the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, so as to train the aircraft detection model, i.e., execute the following S304.

[0102] S304. Based on the classification loss function corresponding to the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, update the model parameters in the initial aircraft detection model to obtain the aircraft detection model.

[0103] For example, in the embodiments of this application, the OHEM loss function can be used to update the model parameters in the initial aircraft detection model. The classification loss function corresponding to the aircraft category prediction probability can be seen in Formula 6 below.

[0104] Formula 6

[0105] in, This represents the classification loss function corresponding to the predicted probability of aircraft category. Indicates the number of optical remote sensing images. Indicates the first An optical remote sensing image, Indicates the first Target aircraft category in optical remote sensing images The predicted probability, This represents a preset probability threshold. For example, It can be set to 0.7, but the specific setting can be adjusted according to actual needs.

[0106] This process updates the model parameters in the initial aircraft detection model based on the classification loss function corresponding to the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, until the aircraft detection model is trained. Given that the aircraft detection model has good feature enhancement capabilities and can effectively capture and utilize image features at different levels, when the target optical remote sensing image to be detected is input into the aircraft detection model, and the aircraft detection result is determined by the aircraft detection model, it can effectively capture and utilize image feature information at different levels of the target optical remote sensing image. By combining image feature information at different levels to determine the aircraft detection result, the aircraft detection result of the target optical remote sensing image can be obtained quickly and accurately, solving the deficiency of low detection accuracy in traditional aircraft detection algorithms for optical remote sensing images, thereby improving the accuracy of aircraft detection results in target optical remote sensing images.

[0107] The aircraft detection device in optical remote sensing images provided in this application is described below. The aircraft detection device in optical remote sensing images described below can be referred to in correspondence with the aircraft detection method in optical remote sensing images described above.

[0108] Figure 4 This application provides a schematic diagram of the structure of an aircraft detection device in an optical remote sensing image, as exemplified in this application. See also: Figure 4 As shown, the aircraft detection device 40 in the optical remote sensing image may include:

[0109] Acquisition unit 401 is used to acquire optical remote sensing images of the target to be detected;

[0110] The detection unit 402 is used to input the target optical remote sensing image into the aircraft detection model to obtain the aircraft detection result output by the aircraft detection model;

[0111] The aircraft detection model is obtained by using ADNet, an image denoising network with an improved lightweight attention mechanism, to extract image features at different levels from optical remote sensing images in the training set, and then training the initial aircraft detection model based on the extracted multi-level image features and corresponding labels.

[0112] For example, in an embodiment of this application, the initial aircraft detection model includes the image denoising network ADNet, the bidirectional feature pyramid network BiFPN, and the detection head network based on an improved lightweight attention mechanism;

[0113] The aircraft detection model was trained using the following method:

[0114] The optical remote sensing image is input into the image denoising network ADNet to obtain the multi-level image features output by the image denoising network ADNet; wherein, the multi-level image features include first-level image features, second-level image features and third-level image features;

[0115] The first-level image features, the second-level image features, and the third-level image features are all input into the Bidirectional Feature Pyramid Network (BiFPN) to obtain the first-scale features corresponding to the first-level image features, the second-scale features corresponding to the second-level image features, and the third-scale features corresponding to the third-level image features output by the BiFPN.

[0116] The first scale feature, the second scale feature, and the third scale feature are all input into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network.

[0117] Based on the classification loss function corresponding to the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, the model parameters in the initial aircraft detection model are updated to obtain the aircraft detection model.

[0118] For example, in an embodiment of this application, the image denoising network ADNet includes a sparse module SB, a feature enhancement module FEB, and an attention module based on an improved lightweight attention mechanism, which are connected in series.

[0119] The step of inputting the optical remote sensing image into the image denoising network ADNet to obtain the multi-level image features output by the image denoising network ADNet includes:

[0120] The optical remote sensing image is input into the sparse module SB to obtain the first-level image features output by the sparse module SB;

[0121] Both the first-level image features and the optical remote sensing image are input into the feature enhancement module FEB, and the second-level image features and convolutional image features are output through the feature enhancement module FEB.

[0122] The second-level image features and the convolutional image features are both input into the attention module based on the improved lightweight attention mechanism, and the third-level image features are output through the attention module.

[0123] For example, in an embodiment of this application, the feature enhancement module FEB includes a first convolutional normalization unit, a second convolutional normalization unit, a third convolutional normalization unit, a convolutional layer, and an activation function layer connected in series.

[0124] The step of inputting both the first-level image features and the optical remote sensing image into the feature enhancement module FEB, and outputting the second-level image features and convolutional image features through the feature enhancement module, includes:

[0125] The image features of the first level are sequentially input into the first convolutional normalization unit, the second convolutional normalization unit, and the third convolutional normalization unit to obtain the convolutional normalized features output by the third convolutional normalization unit.

[0126] The convolutional normalized features are input into the convolutional layer to obtain the convolutional image features output by the convolutional layer;

[0127] The sum of the optical remote sensing image and the convolutional image features is input into the activation function layer to obtain the second-level image features output by the activation function layer.

[0128] For example, in an embodiment of this application, the detection head network includes multiple detection heads of different scales connected in parallel, including a detection head corresponding to a first scale, a detection head corresponding to a second scale, and a detection head corresponding to a third scale;

[0129] Specifically, the first scale feature, the second scale feature, and the third scale feature are all input into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network, including:

[0130] The first scale feature is input into the detection head corresponding to the first scale to obtain the corresponding first aircraft category prediction probability and the first aircraft detection box;

[0131] The second scale feature is input into the detection head corresponding to the second scale to obtain the corresponding second aircraft category prediction probability and the second aircraft detection box;

[0132] The third-scale features are input into the detection head corresponding to the third scale to obtain the corresponding third aircraft category prediction probability and the third aircraft detection box;

[0133] The predicted probability of the aircraft category is determined based on the predicted probability of the first aircraft category, the predicted probability of the second aircraft category, and the predicted probability of the third aircraft category;

[0134] The aircraft detection frame is determined based on the first aircraft detection frame, the second aircraft detection frame, and the third aircraft detection frame.

[0135] For example, in an embodiment of this application, the classification loss function corresponding to the predicted probability of the aircraft category can be seen in the following formula:

[0136]

[0137] in, This represents the classification loss function corresponding to the predicted probability of the aircraft category. This indicates the number of optical remote sensing images. Indicates the first An optical remote sensing image, Indicates the first Target aircraft category in optical remote sensing images The predicted probability, This represents the preset probability threshold.

[0138] The aircraft detection device 40 in the optical remote sensing image provided in this application embodiment can execute the technical solution of the aircraft detection method in the optical remote sensing image in any of the above embodiments. Its implementation principle and beneficial effects are similar to those of the aircraft detection method in the optical remote sensing image. Please refer to the implementation principle and beneficial effects of the aircraft detection method in the optical remote sensing image. It will not be repeated here.

[0139] Figure 5 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute an aircraft detection method in optical remote sensing images. This method includes: acquiring an optical remote sensing image of the target to be detected; inputting the target optical remote sensing image into an aircraft detection model to obtain an aircraft detection result output by the aircraft detection model; wherein the aircraft detection model is obtained by extracting image features at different levels from optical remote sensing images in the training set using an improved lightweight attention mechanism image denoising network ADNet, and training an initial aircraft detection model based on the extracted multi-level image features and corresponding labels.

[0140] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0141] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the aircraft detection method in optical remote sensing images provided by the above methods. The method includes: acquiring an optical remote sensing image of a target to be detected; inputting the target optical remote sensing image into an aircraft detection model to obtain an aircraft detection result output by the aircraft detection model; wherein the aircraft detection model is obtained by extracting image features at different levels from optical remote sensing images in the training set based on an improved lightweight attention mechanism image denoising network ADNet, and training an initial aircraft detection model based on the extracted multi-level image features and corresponding labels.

[0142] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is implemented to perform the aircraft detection method in optical remote sensing images provided by the above methods. The method includes: acquiring an optical remote sensing image of a target to be detected; inputting the target optical remote sensing image into an aircraft detection model to obtain an aircraft detection result output by the aircraft detection model; wherein the aircraft detection model is obtained by extracting image features at different levels from optical remote sensing images in a training set based on an improved lightweight attention mechanism image denoising network ADNet, and training an initial aircraft detection model based on the extracted multi-level image features and corresponding labels.

[0143] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting aircraft in optical remote sensing images, characterized in that, Acquire optical remote sensing images of the target to be detected; The target optical remote sensing image is input into the aircraft detection model to obtain the aircraft detection result output by the aircraft detection model; The aircraft detection model is obtained by using ADNet, an image denoising network with an improved lightweight attention mechanism, to extract image features at different levels from optical remote sensing images in the training set, and then training the initial aircraft detection model based on the extracted multi-level image features and corresponding labels. The image denoising network ADNet comprises a sparse module SB, a feature enhancement module FEB, and an attention module based on an improved lightweight attention mechanism, all connected in series. ADNet extracts the multi-level image features in the following manner: The optical remote sensing image is input into the sparse module SB to obtain the first-level image features output by the sparse module SB; Both the first-level image features and the optical remote sensing image are input into the feature enhancement module FEB, and the second-level image features and convolutional image features are output through the feature enhancement module FEB. The second-level image features and the convolutional image features are both input into the attention module based on the improved lightweight attention mechanism, and the third-level image features are output through the attention module.

2. The method for detecting aircraft in optical remote sensing images according to claim 1, characterized in that, The initial aircraft detection model includes the image denoising network ADNet, the bidirectional feature pyramid network BiFPN, and the detection head network based on an improved lightweight attention mechanism; The aircraft detection model was trained using the following method: The optical remote sensing image is input into the image denoising network ADNet to obtain the multi-level image features output by the image denoising network ADNet; wherein, the multi-level image features include first-level image features, second-level image features and third-level image features; The first-level image features, the second-level image features, and the third-level image features are all input into the Bidirectional Feature Pyramid Network (BiFPN) to obtain the first-scale features corresponding to the first-level image features, the second-scale features corresponding to the second-level image features, and the third-scale features corresponding to the third-level image features output by the BiFPN. The first scale feature, the second scale feature, and the third scale feature are all input into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network. Based on the classification loss function corresponding to the predicted probability of the aircraft category and the loss function corresponding to the aircraft detection box, the model parameters in the initial aircraft detection model are updated to obtain the aircraft detection model.

3. The method for detecting aircraft in optical remote sensing images according to claim 1, characterized in that, The Feature Enhancement Module (FEB) includes a first convolutional normalization unit, a second convolutional normalization unit, a third convolutional normalization unit, a convolutional layer, and an activation function layer, which are connected in series. The step of inputting both the first-level image features and the optical remote sensing image into the feature enhancement module FEB, and outputting the second-level image features and convolutional image features through the feature enhancement module, includes: The image features of the first level are sequentially input into the first convolutional normalization unit, the second convolutional normalization unit, and the third convolutional normalization unit to obtain the convolutional normalized features output by the third convolutional normalization unit. The convolutional normalized features are input into the convolutional layer to obtain the convolutional image features output by the convolutional layer; The sum of the optical remote sensing image and the convolutional image features is input into the activation function layer to obtain the second-level image features output by the activation function layer.

4. The method for detecting aircraft in optical remote sensing images according to claim 2, characterized in that, The detection head network includes multiple detection heads of different scales connected in parallel, including a detection head corresponding to the first scale, a detection head corresponding to the second scale, and a detection head corresponding to the third scale. Specifically, the first scale feature, the second scale feature, and the third scale feature are all input into the detection head network to obtain the aircraft category prediction probability and aircraft detection box output by the detection head network, including: The first scale feature is input into the detection head corresponding to the first scale to obtain the corresponding first aircraft category prediction probability and the first aircraft detection box; The second scale feature is input into the detection head corresponding to the second scale to obtain the corresponding second aircraft category prediction probability and the second aircraft detection box; The third-scale features are input into the detection head corresponding to the third scale to obtain the corresponding third aircraft category prediction probability and the third aircraft detection box; The predicted probability of the aircraft category is determined based on the predicted probability of the first aircraft category, the predicted probability of the second aircraft category, and the predicted probability of the third aircraft category; The aircraft detection frame is determined based on the first aircraft detection frame, the second aircraft detection frame, and the third aircraft detection frame.

5. The method for detecting aircraft in optical remote sensing images according to any one of claims 2-4, characterized in that, The classification loss function corresponding to the predicted probability of the aircraft category can be seen in the following formula: Where OhemCrossEntropyLoss represents the classification loss function corresponding to the predicted probability of the aircraft category, N represents the number of optical remote sensing images, i represents the i-th optical remote sensing image, and q(m i ) represents the predicted probability of target aircraft category m in the i-th optical remote sensing image, and threshold represents the preset probability threshold.

6. An aircraft detection device in optical remote sensing images, characterized in that, include: The acquisition unit is used to acquire optical remote sensing images of the target to be detected; The detection unit is used to input the target optical remote sensing image into the aircraft detection model and obtain the aircraft detection result output by the aircraft detection model. The aircraft detection model is obtained by using ADNet, an image denoising network with an improved lightweight attention mechanism, to extract image features at different levels from optical remote sensing images in the training set, and then training the initial aircraft detection model based on the extracted multi-level image features and corresponding labels. The image denoising network ADNet comprises a sparse module SB, a feature enhancement module FEB, and an attention module based on an improved lightweight attention mechanism, all connected in series. ADNet extracts the multi-level image features in the following manner: The optical remote sensing image is input into the sparse module SB to obtain the first-level image features output by the sparse module SB; Both the first-level image features and the optical remote sensing image are input into the feature enhancement module FEB, and the second-level image features and convolutional image features are output through the feature enhancement module FEB. The second-level image features and the convolutional image features are both input into the attention module based on the improved lightweight attention mechanism, and the third-level image features are output through the attention module.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the aircraft detection method in optical remote sensing images as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the aircraft detection method in optical remote sensing images as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the aircraft detection method in optical remote sensing images as described in any one of claims 1 to 5.

Images