A domain adaptation rotating target detection method based on pixel-instance information migration

Abstract

The application discloses a domain adaptation rotating target detection method based on pixel-instance information migration, belongs to the technical field of target detection, and solves the problem that the SAR target detection precision is limited due to the lack of labeled training data; the coherent speckle noise in a SAR image is first mixed into a visible light image; the visual content difference between the visible light domain and the SAR domain is relieved by adaptively synthesizing an intermediate domain image; then, multi-scale and multi-direction features are extracted based on the intermediate domain image and the SAR image; on the basis, rotating instance features are acquired; and the decoupled instance features are used to enhance the distinguishability of the domain invariant features; the technical scheme of the application combines the effective migration of visual content and rotating instances, and makes up for the information distribution difference between the visible light domain and the SAR domain; the application can effectively realize the information distribution fitting of the visible light domain-SAR domain, enhances the generalization of cross-domain target detection, and effectively makes up for the difference in information distribution between the visible light domain and the SAR domain.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology and relates to a domain-adaptive rotation target detection method based on pixel-instance information transfer. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging sensor with strong penetration capabilities, enabling long-range observation in all weather conditions and around the clock, making it an important tool for Earth observation. However, due to its unique observation method and imaging mechanism, SAR target detection faces two major challenges. First, due to the special overhead viewing angle and long-range shooting mode, SAR images suffer from large differences in target scale, dense arrangement, and strong directionality. When multiple targets are densely arranged at an inclined angle, traditional horizontal detection boxes will have overlapping boundaries, leading to over-coverage of background areas or unexpected targets. Second, due to its imaging mechanism, SAR images are difficult to identify and interpret, resulting in most SAR images being unlabeled and unusable. Target detection performance often relies on a large amount of labeled data, making it difficult to improve the detection accuracy of unlabeled SAR targets.

[0003] Rotating target detection and unsupervised domain adaptation are effective solutions to the two challenges mentioned above. Specifically, rotating target detection uses oriented bounding boxes to represent rotating targets, providing more complete geometric location information compared to horizontal bounding boxes. Furthermore, the annotation cost of visible light images is far lower than that of SAR images, and unsupervised domain adaptation can transfer domain information from the visible light domain to the SAR domain, thereby mitigating the domain gap between the label-rich domain and the new unlabeled domain. Ideally, a model trained in the visible light domain can effectively detect unlabeled SAR targets. To significantly improve the exploration depth and interpretation reliability of rotating SAR target information, rotating target detection and unsupervised domain adaptation can be combined, resulting in a domain-adaptive rotating target detection method.

[0004] Since the detection accuracy of SAR rotating targets depends not only on the alignment of cross-domain information distribution, but also on the accurate acquisition of rotation angle, exploring specific domain information and extracting rotating target angle information is the key to achieving accurate and stable domain information transfer. Summary of the Invention

[0005] The technical solution of this invention is used to solve the problem of limited SAR target detection accuracy due to the lack of labeled training data.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions: A domain-adaptive target detection method based on pixel-instance information migration includes the following steps: Step 1: Extract the SAR image domain separately Inherent speckle noise and visible light image domain The basic structural information in the two is fused to obtain the intermediate domain. This helps to mitigate the difference in visual content between the visible light domain and the SAR domain. Step 2: Extract the intermediate domain and SAR image domain Multi-scale and multi-directional features ; Step 3: Obtain rotation instance features Furthermore, instance features are decoupled into domain-invariant features and domain-specific features, thereby enhancing the cross-domain recognition capability of rotated instances and improving the generalization of cross-domain target detection.

[0007] Furthermore, the extraction of the SAR image domain in step 1 is as follows: Inherent speckle noise and visible light image domain The basic structural information in the two is fused to obtain the intermediate domain. The specific method is as follows: Step 1.1, Input has Zhang's Visible Light Image Dataset and have Zhang's SAR image dataset For visible light images By applying contrast-limited adaptive histogram equalization, an enhanced visible light image is obtained through contrast enhancement. Improve visual effects and highlight structural details of the target; Step 1.2: Design an adaptive speckle noise background extraction method to extract speckle noise backgrounds of varying sizes introduced during SAR imaging. Step 1.3: Combine enhanced visible light images and speckle noise background image Synthetic intermediate domain .

[0008] Furthermore, the aforementioned visible light image By applying contrast-limited adaptive histogram equalization, an enhanced visible light image is obtained through contrast enhancement. The specific method is as follows: Will Divided into sizes The sub-blocks are processed using the following transformations to calculate the equalization results:

[0009] in, , The value isu The conversion results of the pixels, The value is u The number of pixels, N The total number of pixels in each sub-block; To prevent excessive amplification of contrast in smooth regions, which would lead to noise amplification, the resulting histogram is cropped at the contrast limit, and the cropped portion is evenly redistributed back into the histogram. Each sub-block is determined according to its input. The corresponding positions in the image are combined to form an enhanced visible light image. .

[0010] Furthermore, the aforementioned adaptive speckle noise background extraction method extracts speckle noise backgrounds of varying sizes introduced during SAR imaging, as detailed below: For SAR images An adaptive threshold is used for binary processing to dynamically generate a masking image, i.e., images larger than the threshold are processed. The pixel value is set to 255, and the remaining pixels are set to 0; for small speckle noise, Defined as the threshold of interference perceptible to the human eye To capture minute imaging differences; for large speckle noise, sorted pixel values... A threshold used to balance foreground information removal and background information preservation; a threshold for adaptive speckle noise background extraction. The options are as follows:

[0011] in, For input The top 60% of the pixel values, sorted by pixel value. Based on experience, it is set to 70.

[0012] After obtaining the masked image, its contour information representing the foreground target is used. By defining these regions as distorted regions and using proximity information for inpainting, the foreground target in the SAR image is removed, generating a speckle noise background image. .

[0013] Furthermore, the aforementioned synthetic intermediate domain The method is as follows: Weighted combination of the first j Zhang Zengqiang's visible light images With randomly selected speckle noise background image , obtained the j Zhang's intermediate domain image The formula is as follows:

[0014] in, Gradually synthesize intermediate domains .

[0015] Furthermore, the extraction of intermediate domains and SAR image domain Multi-scale and multi-directional features The method is as follows: Input intermediate field In the SAR image domain, multi-directional Gabor features are generated through a Gabor convolutional network. The formula is as follows:

[0016] in, F For the convolutional features of the input image, O For the total number of directions, for o A Gabor directional filter, consisting of a convolutional filter C and a Gabor filter bank. The composition, as shown in the formula:

[0017]

[0018] in, , This represents a pixel-by-pixel product operation. Let be the standard deviation of the Gaussian function. This indicates the direction information of the filter. Directional index; Multi-scale spatial features of input images are extracted based on feature pyramid network. Integrating multi-scale spatial features and multidirectional Gabor features Generate multi-scale, multi-directional features The fusion formula is as follows:

[0019] in, Conv for The convolution kernel.

[0020] Furthermore, the step 3 involves obtaining the rotation instance features. The method of decoupling instance features into domain-invariant features and domain-specific features to improve the cross-domain recognition capability of rotated instances and enhance the generalization of cross-domain target detection is as follows: Step 3.1: Targeting the extracted multi-scale, multi-directional features Rotation instance features of images are obtained based on the RoIAlign module. Then, the common-specific decomposition module is applied to decompose the instance features into domain-invariant features and domain-specific features, i.e., designing an invariant classification compensation loss. and specific category compensation loss ; Furthermore, for each image, an orthogonal regularizer is used to make the domain-invariant classifier... Domain-specific classifiers Orthogonalize the softmax parameters to improve numerical stability, while reducing orthogonality loss. The definition is as follows:

[0021] in, , and They represent domain-invariant classifiers respectively. Domain-specific classifiers The softmax parameter; I The identity matrix is ​​L=2, which represents the number of participating domains, namely the visible light domain and the SAR domain. Classification of compensation for losses Defined as:

[0022] in, , and It is a hyperparameter that balances these compensating loss components.

[0023] Step 3.2: Calculate Gabor information adversarial loss based on the gradient reversal layer. Combating loss against global instances This improves the consistency of feature distribution between the visible light domain and the SAR domain, as shown in the following formula:

[0024] in, and This represents multi-directional Gabor features and rotational instance features obtained from the visible light domain. and This represents multi-directional Gabor features and rotational instance features obtained from the SAR domain. and This indicates the corresponding domain discriminator.

[0025] Combating loss with Gabor information Combating loss against global instances The fusion yields an adversarial loss, defined as follows:

[0026] in, and It is a hyperparameter that balances these adversarial loss components; The formula for calculating the overall loss function is as follows:

[0027] in, and These represent classification loss and regression loss, respectively.

[0028] Furthermore, the design-invariant classification compensation loss is described. and specific category compensation loss The details are as follows:

[0029] in, Represents the cross-entropy loss function. and For domain-invariant classifiers and domain-specific classifiers; and These represent instance characteristics in the visible light domain and the SAR domain, respectively. and This includes pseudo-label information for the visible light domain and the SAR domain.

[0030] An electronic device includes a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the above-described domain-adaptive target detection method based on pixel-instance information migration, and the processor is configured to execute the program stored in the memory.

[0031] A storage medium storing a computer program, characterized in that the computer program, when run by a processor, executes the steps of the above-described domain adaptation target detection method based on pixel-instance information migration.

[0032] The advantages of this invention are: The technical solution of this invention combines effective transfer of visual content and rotation instances to bridge the information distribution gap between the visible light domain and the SAR domain. First, speckle noise from the SAR image is mixed into the visible light image, and an intermediate domain image is adaptively synthesized to alleviate the visual content difference between the visible light domain and the SAR domain. Then, multi-scale and multi-directional features are extracted based on the intermediate domain image and the SAR image. Based on this, rotation instance features are obtained, and the discriminability of domain-invariant features is enhanced by decoupling instance features. This invention can effectively achieve information distribution fitting between the visible light domain and the SAR domain, enhance the generalization of cross-domain target detection, and effectively bridge the information distribution gap between the visible light domain and the SAR domain. Attached Figure Description

[0033] Figure 1 This is a flowchart of the domain-adaptive target detection method based on pixel-instance information migration according to the present invention; Figure 2 This is a schematic diagram of the framework of the domain-adaptive rotating target detection method based on pixel-instance information migration of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 See Figure 1 and Figure 2 As shown, a domain-adaptive target detection method based on pixel-instance information migration has the following specific steps: Step 1: Extract the SAR image domain separately Inherent speckle noise and visible light image domain The basic structural information in the two is fused to obtain the intermediate domain. To mitigate the difference in visual content between the visible light domain and the SAR domain, the following measures are taken: Step 1.1, Input has Zhang's Visible Light Image Dataset and have Zhang's SAR image dataset ,in x , y , b These represent the image, label, and bounding box annotation, respectively. For visible light images... By applying Contrast Limiting Adaptive Histogram Equalization (CLAHE), an enhanced visible light image is obtained through contrast enhancement. This improves visual appeal and highlights the structural details of the target.

[0036] Specifically, will Divided into sizes The sub-blocks are processed using the following transformations to calculate the equalization results:

[0037] in, , The value is u The conversion results of the pixels, The value is u The number of pixels, N The total number of pixels for each sub-block. To prevent excessive contrast amplification in smooth regions, which could lead to noise amplification, the resulting histogram is cropped at contrast limits, and the cropped portions are evenly redistributed back into the histogram. Each sub-block is then assigned a number of pixels based on its position in the input... The corresponding positions in the image are combined to form an enhanced visible light image. .

[0038] Step 1.2: Design an adaptive speckle noise background extraction method to extract speckle noise backgrounds of varying sizes introduced during SAR imaging. Details are as follows: For SAR images An adaptive threshold is used for binary processing to dynamically generate a masking image, i.e., images larger than the threshold are processed. The pixel value (representing foreground information) is set to 255, and the remaining pixels (representing background information) are set to 0. For small speckle noise, Defined as the threshold of interference perceptible to the human eye To capture minute imaging differences; for large speckle noise, sorted pixel values... A threshold is used to balance foreground information removal and background information preservation.

[0039] Threshold for adaptive speckle noise background extraction The options are as follows:

[0040] in, For input The top 60% of the pixel values, sorted by pixel value. Based on experience, it is set to 70.

[0041] After obtaining the masked image, its contour information representing the foreground target is used. These regions are defined as distorted areas and repaired using proximity information to remove SAR image distortion. Foreground target, generating a speckle noise background image. .

[0042] Step 1.3, weighted combination j Zhang Zengqiang's visible light images With randomly selected speckle noise background image , obtained the j Zhang's intermediate domain image The formula is as follows:

[0043] in, Gradually synthesize intermediate domains .

[0044] Step 2: Extract the intermediate domain and SAR image domain Multi-scale and multi-directional features The details are as follows: Input intermediate field In the SAR image domain, multi-directional Gabor features are generated through a Gabor convolutional network. ,as follows:

[0045] in, F For the convolutional features of the input image, This represents the total number of directions used. for o A Gabor directional filter (GoF) consists of a convolutional filter C and a Gabor filter bank. The composition, as shown in the formula:

[0046]

[0047] in, , This represents a pixel-by-pixel product operation. Let be the standard deviation of the Gaussian function. This indicates the direction information of the filter. This is the directional index.

[0048] Subsequently, multi-scale spatial features of the input image are extracted based on the Feature Pyramid Network (FPN). Integrating multi-scale spatial features and multidirectional Gabor features Generate multi-scale, multi-directional features The fusion formula is as follows:

[0049] in, Conv for The convolution kernel.

[0050] Step 3: Obtain rotation instance features Furthermore, instance features are decoupled into domain-invariant features and domain-specific features to enhance the cross-domain recognition capability of rotated instances and improve the generalization of cross-domain target detection, as detailed below: Step 3.1: Targeting the extracted multi-scale, multi-directional features Rotation instance features of images are obtained based on the RoIAlign module. Then, the Common Specific Decomposition (CSD) module is applied to decompose instance features into domain-invariant features and domain-specific features, i.e., designing an invariant classification compensation loss. and specific category compensation loss The definition is as follows:

[0051] in, Represents the cross-entropy loss function. and For domain-invariant classifiers and domain-specific classifiers. and These represent instance characteristics in the visible light domain and the SAR domain, respectively. and This includes visible light domain labels and SAR domain pseudo-label information. By enhancing the discriminability of domain-invariant features, the cross-domain detection generalization of rotated instances is greatly improved.

[0052] Furthermore, for each image, an orthogonal regularizer is used to make the domain-invariant classifier... Domain-specific classifiers Orthogonalize the softmax parameters to improve numerical stability, while reducing orthogonality loss. The definition is as follows:

[0053] in, , and They represent domain-invariant classifiers respectively. Domain-specific classifiers The softmax parameter. I The identity matrix is ​​L=2, which represents the number of participating domains, namely the visible light domain and the SAR domain.

[0054] Classification of compensation for losses It can be defined as:

[0055] in, , and It is a hyperparameter that balances these compensating loss components.

[0056] Step 3.2: Calculate Gabor information adversarial loss based on the gradient inversion layer (GRL). Combating loss against global instances This improves the consistency of feature distribution between the visible light domain and the SAR domain, as shown in the following formula:

[0057] in, and This represents multi-directional Gabor features and rotational instance features obtained from the visible light domain. and This represents multi-directional Gabor features and rotational instance features obtained from the SAR domain. and This indicates the corresponding domain discriminator.

[0058] Combating loss with Gabor information Combating loss against global instances The fusion yields an adversarial loss, defined as follows:

[0059] in, and It is a hyperparameter that balances these adversarial loss components.

[0060] The formula for calculating the overall loss function of the model is as follows:

[0061] in, and These represent classification loss and regression loss, respectively.

[0062] Experimental verification This invention employs a ResNet-101 architecture pre-trained on ImageNet as the basic feature extractor. Experiments were conducted using the PyTorch framework, trained with stochastic gradient descent (SGD) at a weight decay of 0.0001 and a momentum of 0.9. During training, a total of 24 epochs were performed, with an initial learning rate of 0.005, divided by 10 at epochs 16 and 22. Each training batch contained two images, one from the intermediate domain and the other from the SAR image domain.

[0063] To verify the effectiveness of this invention, the domain adaptation performance of the proposed method was evaluated on two benchmark datasets specifically designed for domain-adaptive rotating target detection. These two benchmark datasets consist of a visible light dataset (HRSC2016) and two SAR datasets (SSDD and RSDD-SAR). The experiments used the classic rotating target detection models OrientedRCNN and Gliding Vertex as the basic detection models, and compared the proposed method with models trained only using visible light domain data (Scource-only). Furthermore, four commonly used evaluation metrics were employed to assess the accuracy of rotating target detection, thereby measuring the domain adaptation performance between the visible light and SAR domains: precision, recall, average precision (AP), and F1 score.

[0064] The experimental results are shown in Tables 1 and 2. For the domain adaptation task of HRSC2016→SSDD, the AP value of the method of this invention is improved by 85.13% and 90.91% respectively compared with the source-only model. For the case of HRSC2016→RSDD-SAR, the detection accuracy of the model is relatively low, that is, the AP values ​​are only 9.09 and 12.95. This indicates that there is a huge domain difference between the two datasets, and the method of this invention greatly reduces this difference, with AP values ​​improved to 19.69 and 20.31, with gains of 116.61% and 56.83% respectively. This shows that for the case of unlabeled SAR images, the method of this invention can significantly improve the accuracy of rotating SAR target detection.

[0065] Table 1. Results (%) of the method of this invention in the HRSC2016→SSDD domain adaptation task.

[0066] Table 2 Results (%) of the method of this invention in the HRSC2016→RSDD-SAR domain adaptation task

[0067] An electronic device includes a memory and a processor, the memory being used to store a program that supports the processor in executing the domain-adaptive target detection method based on pixel-instance information migration as described in Embodiment 1, the processor being configured to execute the program stored in the memory.

[0068] A storage medium storing a computer program, which, when executed by a processor, performs the steps of the domain-adaptive target detection method based on pixel-instance information migration in Embodiment 1.

[0069] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. A domain-adaptive target detection method based on pixel-instance information transfer, characterized in that, Includes the following steps: Step 1: Extract the SAR image domain separately Inherent speckle noise and visible light image domain The basic structural information in the two is fused to obtain the intermediate domain. To mitigate the visual content differences between the visible light domain and the SAR domain; the specific methods are as follows: Step 1.1, Input has Zhang's Visible Light Image Dataset and have Zhang's SAR image dataset For visible light images By applying contrast-limited adaptive histogram equalization, an enhanced visible light image is obtained through contrast enhancement. This improves visual appeal and highlights the structural details of the target; among them, Indicates the target category label. Indicates bounding box annotation; Step 1.2: Design an adaptive speckle noise background extraction method to extract speckle noise backgrounds of varying sizes introduced during SAR imaging. Step 1.3: Combine enhanced visible light images and speckle noise background image Synthetic intermediate domain ; Step 2: Extract the intermediate domain and SAR image domain Multi-scale and multi-directional features ; Step 3: Obtain rotation instance features Furthermore, instance features are decoupled into domain-invariant features and domain-specific features, thereby enhancing the cross-domain recognition capability of rotated instances and improving the generalization of cross-domain target detection.

2. The domain-adaptive target detection method based on pixel-instance information migration according to claim 1, characterized in that, The visible light image By applying contrast-limited adaptive histogram equalization, an enhanced visible light image is obtained through contrast enhancement. The specific method is as follows: Will Divided into sizes The sub-blocks are processed using the following transformations to calculate the equalization results: in, , The value is u The conversion results of the pixels, The value is u The number of pixels, N The total number of pixels in each sub-block; To prevent excessive amplification of contrast in smooth regions, which would lead to noise amplification, the resulting histogram is cropped at the contrast limit, and the cropped portion is evenly redistributed back into the histogram. Each sub-block is determined according to its input. The corresponding positions in the image are combined to form an enhanced visible light image. .

3. The domain-adaptive target detection method based on pixel-instance information migration according to claim 2, characterized in that, The aforementioned adaptive speckle noise background extraction method extracts speckle noise backgrounds of varying sizes introduced during SAR imaging, as detailed below: For SAR images An adaptive threshold is used for binary processing to dynamically generate a masking image, i.e., images larger than the threshold are processed. The pixel value is set to 255, and the remaining pixels are set to 0; for small speckle noise, Defined as the threshold of interference perceptible to the human eye To capture minute imaging differences; for large speckle noise, sorted pixel values... Used as a threshold to balance foreground information removal and background information preservation; Threshold for adaptive speckle noise background extraction The options are as follows: in, For input The top 60% of the pixel values, sorted by pixel value. Based on experience, it is set to 70; After obtaining the masked image, its contour information representing the foreground target is used. By defining these regions as distorted regions and using proximity information for inpainting, the foreground target in the SAR image is removed, generating a speckle noise background image. .

4. The domain-adaptive target detection method based on pixel-instance information migration according to claim 3, characterized in that, The aforementioned synthetic intermediate domain The method is as follows: Weighted combination of the first j Zhang Zengqiang's visible light images With randomly selected speckle noise background image , obtained the j Zhang's intermediate domain image The formula is as follows: in, Gradually synthesize intermediate domains .

5. The domain-adaptive target detection method based on pixel-instance information migration according to claim 4, characterized in that, Step 2 describes the extraction of the intermediate domain. and SAR image domain Multi-scale and multi-directional features The method is as follows: Input intermediate field In the SAR image domain, multi-directional Gabor features are generated through a Gabor convolutional network. The formula is as follows: in, F For the convolutional features of the input image, O For the total number of directions, for o A Gabor directional filter, consisting of a convolutional filter C and a Gabor filter bank. The composition, as shown in the formula: in, , This represents a pixel-by-pixel product operation. Let be the standard deviation of the Gaussian function. This indicates the direction information of the filter. Directional index; Multi-scale spatial features of input images are extracted based on feature pyramid network. Integrating multi-scale spatial features and multidirectional Gabor features Generate multi-scale, multi-directional features The fusion formula is as follows: in, Conv for The convolution kernel.

6. The domain-adaptive target detection method based on pixel-instance information migration according to claim 5, characterized in that, Step 3 describes obtaining rotation instance features. The method of decoupling instance features into domain-invariant features and domain-specific features to improve the cross-domain recognition capability of rotated instances and enhance the generalization of cross-domain target detection is as follows: Step 3.1: Targeting the extracted multi-scale, multi-directional features Rotation instance features of images are obtained based on the RoIAlign module. Then, the common-specific decomposition module is applied to decompose the instance features into domain-invariant features and domain-specific features, i.e., designing an invariant classification compensation loss. and specific category compensation loss ; Furthermore, for each image, an orthogonal regularizer is used to make the domain-invariant classifier... Domain-specific classifiers Orthogonalize the softmax parameters to improve numerical stability, while reducing orthogonality loss. The definition is as follows: in, , and They represent domain-invariant classifiers respectively. Domain-specific classifiers The softmax parameter; I The identity matrix is ​​L=2, which represents the number of participating domains, namely the visible light domain and the SAR domain. Classification of compensation for losses Defined as: in, , and It is a hyperparameter that balances these compensating loss components; Step 3.2: Calculate Gabor information adversarial loss based on the gradient reversal layer. Combating loss against global instances This improves the consistency of feature distribution between the visible light domain and the SAR domain, as shown in the following formula: in, and This represents multi-directional Gabor features and rotational instance features obtained from the visible light domain. and This represents multi-directional Gabor features and rotational instance features obtained from the SAR domain. and Indicates the corresponding domain discriminator; Combating loss with Gabor information Combating loss against global instances The fusion yields an adversarial loss, defined as follows: in, and It is a hyperparameter that balances these adversarial loss components; The formula for calculating the overall loss function is as follows: in, and These represent classification loss and regression loss, respectively.

7. The domain-adaptive target detection method based on pixel-instance information migration according to claim 6, characterized in that, The design-invariant classification compensation loss and specific category compensation loss The details are as follows: in, Represents the cross-entropy loss function. and For domain-invariant classifiers and domain-specific classifiers; and These represent instance characteristics in the visible light domain and the SAR domain, respectively. and This includes pseudo-label information for the visible light domain and the SAR domain.

8. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the domain-adaptive target detection method based on pixel-instance information migration as described in any one of claims 1 to 7, and the processor is configured to execute the program stored in the memory.

9. A storage medium storing a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the domain adaptation target detection method based on pixel-instance information migration as described in any one of claims 1 to 7.

Images