Knowledge distillation method and apparatus for target detection in SAR images, target detection method and apparatus

Abstract

This disclosure provides a knowledge distillation method for target detection in SAR images, comprising: extracting features from image samples using feature extraction modules of a teacher model and a student model to obtain teacher feature maps and student feature maps; performing frequency domain transformation on the teacher feature maps and student feature maps to obtain teacher frequency domain features and student frequency domain features; extracting attention from the teacher feature maps to obtain teacher attention maps; performing detection on the teacher feature maps and student feature maps using detection modules of the teacher model and student model to obtain teacher detection results, student detection results, and intermediate features; inputting the student intermediate detection values ​​into the detection layer of the teacher model to obtain cross-detection results; determining a joint loss value based on teacher frequency domain features, student frequency domain features, teacher attention maps, teacher feature maps, student feature maps, teacher detection results, cross-detection results, student detection results, and sample labels; and adjusting the parameters of the student model based on the joint loss value.

Description

Technical Field

[0001] This disclosure relates to the field of deep learning technology, and more specifically to a knowledge distillation method and apparatus for SAR image target detection, and a target detection method and apparatus. Background Technology

[0002] SAR (Synthetic Aperture Radar) systems generate an enormous amount of data during the imaging process. In related technologies, deep neural network models are used to process SAR images, which incurs significant computational and storage overhead, making it difficult to deploy on spaceborne equipment and edge computing platforms where power consumption, computing power, and storage are all limited. Summary of the Invention

[0003] In view of the above problems, this disclosure provides a knowledge distillation method and apparatus for target detection in SAR images, and a target detection method and apparatus.

[0004] According to the first aspect of this disclosure, a knowledge distillation method for SAR image target detection is provided. A student model is trained using a pre-trained teacher model. Both the teacher and student models include a feature extraction module, a feature fusion module, and a detection module. The feature extraction module of the teacher model has more layers than the feature extraction module of the student model. The detection modules of both the teacher and student models each include N detection layers, where N is a positive integer greater than 1. The method includes: extracting features from synthetic aperture radar (SAR) image samples using the feature extraction module of the pre-trained teacher model to obtain a teacher feature map; and extracting features from SAR image samples using the feature extraction module of the student model to obtain a student feature map; performing a frequency domain transformation on the teacher feature map to obtain teacher frequency domain features; and performing a frequency domain transformation on the student feature map to obtain student frequency domain features. The teacher frequency domain features and student frequency domain features respectively characterize the edge details extracted by the teacher model and the student model for the SAR image samples. The teacher feature map is used to extract attention, resulting in a teacher attention map. This teacher attention map represents the degree of attention the teacher model pays to each sub-region of the synthetic aperture radar image sample. The teacher model's detection module is used to detect the teacher feature map, resulting in teacher detection results. The student model's detection module is used to detect the student feature map, resulting in student detection results output from the Nth detection layer and intermediate student detection values ​​output from the nth detection layer, where n is a positive integer greater than 1 and less than N. The intermediate student detection values ​​are input into the (n+1)th detection layer of the teacher model to obtain cross-detection results. Based on the teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection results, cross-detection results, student detection results, and sample labels, a joint loss value is determined. The sample labels indicate the actual detection results of the synthetic aperture radar image sample. The parameters of the student model are adjusted based on the joint loss value to obtain a trained student model.

[0005] According to embodiments of this disclosure, the above-mentioned frequency domain transformation of the teacher feature map to obtain teacher frequency domain features includes: performing a fast Fourier transform on the teacher feature map to obtain teacher transformed features; extracting high-frequency features from the teacher frequency features using a filter to obtain teacher high-frequency features; and performing an inverse fast Fourier transform on the teacher high-frequency features to obtain the teacher frequency domain features.

[0006] According to embodiments of this disclosure, the above-mentioned attention extraction of the teacher feature map to obtain a teacher attention map includes: performing a global average pooling operation on the teacher feature map to obtain a teacher average attention vector; performing a global max pooling operation on the teacher feature map to obtain a teacher maximum attention vector; determining a target attention vector based on the teacher average attention vector and the teacher maximum attention vector; and performing channel dimension expansion on the target attention vector to generate the teacher attention map.

[0007] According to embodiments of this disclosure, determining the joint loss value based on the teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection result, cross-detection result, student detection result, and sample labels includes: determining a frequency domain loss value based on the teacher frequency domain features and student frequency domain features; determining a spatial domain loss value based on the teacher attention map, teacher feature map, and student feature map; determining a cross-detection loss value based on the teacher detection result and cross-detection result; determining a prediction loss value based on the student detection result and sample labels; and determining the joint loss value based on the frequency domain loss value, spatial domain loss value, cross-detection loss value, and prediction loss value.

[0008] According to embodiments of this disclosure, the feature fusion modules of both the teacher model and the student model include M feature layers. The teacher frequency domain features include M sub-teacher frequency domain features corresponding to the M feature layers, and the student frequency domain features include M sub-student frequency domain features corresponding to the M feature layers, where M is a positive integer greater than 1. The determination of the frequency domain loss value based on the teacher frequency domain features and the student frequency domain features includes: for the m-th feature layer, calculating the square Euclidean norm of the m-th sub-teacher frequency domain feature and the m-th sub-student frequency domain feature to obtain the m-th sub-frequency domain loss value; and obtaining the frequency domain loss value based on the M sub-frequency domain loss values.

[0009] According to embodiments of this disclosure, the teacher attention feature map includes M attention sub-feature maps corresponding to M feature layers, the teacher feature map includes M sub-teacher feature maps corresponding to M feature layers, and the student feature map includes M sub-student feature maps corresponding to M feature layers. Determining the spatial domain loss value based on the teacher attention map, the teacher feature map, and the student feature map includes: for the m-th feature layer, calculating the Hadamard product of the m-th attention sub-feature map and the m-th sub-student feature map to obtain the m-th sub-spatial domain student prediction feature map; obtaining the m-th sub-spatial domain loss value based on the m-th sub-teacher feature map and the m-th sub-spatial domain student prediction feature map; and obtaining the spatial domain loss value based on the M sub-spatial domain loss values.

[0010] According to an embodiment of this disclosure, determining the cross-loss value based on the teacher detection results and the cross-detection results includes: calculating the distribution difference between the teacher detection results and the cross-detection results to obtain the cross-loss value.

[0011] The second aspect of this disclosure provides a target detection method, comprising: acquiring a synthetic aperture radar image of a target; inputting the synthetic aperture radar image of the target into a pre-trained student model to obtain a target detection result, wherein the pre-trained student model is trained by the method of the first aspect.

[0012] The third aspect of this disclosure provides a knowledge distillation apparatus for SAR image target detection, which trains a student model using a pre-trained teacher model. Both the teacher and student models include a feature extraction module, a feature fusion module, and a detection module. The feature extraction module of the teacher model has more layers than the feature extraction module of the student model. The detection modules of both the teacher and student models each include N detection layers, where N is a positive integer greater than 1. The apparatus includes: an extraction module for extracting features from synthetic aperture radar (SAR) image samples using the feature extraction module of the pre-trained teacher model to obtain a teacher feature map, and extracting features from SAR image samples using the feature extraction module of the student model to obtain a student feature map; a frequency transformation module for performing a frequency domain transformation on the teacher feature map to obtain teacher frequency domain features, and performing a frequency domain transformation on the student feature map to obtain student frequency domain features, wherein the teacher frequency domain features characterize the edge details of the SAR image samples extracted by the teacher model; and an attention extraction module for performing attention extraction on the teacher feature map. Attention extraction yields a teacher attention map, representing the teacher model's focus on each sub-region of the synthetic aperture radar image sample. A cross-detection module uses the teacher model's detection module to detect the teacher feature map, obtaining teacher detection results, and uses the student model's detection module to detect the student feature map, obtaining the student detection result output from the Nth detection layer and the intermediate student detection value output from the nth detection layer, where n is a positive integer greater than 1 and less than N. An input module inputs the intermediate student detection value into the (n+1)th detection layer of the teacher model to obtain cross-detection results. A loss calculation module determines a joint loss value based on the teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection results, cross-detection results, student detection results, and sample labels, where the sample labels indicate the actual detection results of the synthetic aperture radar image sample. An adjustment module adjusts the parameters of the student model based on the joint loss value to obtain a trained student model.

[0013] A fourth aspect of this disclosure provides a target detection apparatus, comprising: an acquisition module for acquiring a synthetic aperture radar image of a target; and a detection module for inputting the synthetic aperture radar image of the target into a pre-trained student model to obtain a target detection result, wherein the pre-trained student model is trained using the method described in the first aspect.

[0014] A fifth aspect of this disclosure provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A sixth aspect of this disclosure also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The seventh aspect of this disclosure also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0017] According to embodiments of this disclosure, frequency domain transformation is performed on student and teacher feature maps to obtain teacher and student frequency domain features. Then, based on the loss value between the teacher and student frequency domain features, the student model learns the teacher model's processing of image edge detail features. Furthermore, based on the loss value between the teacher attention map, teacher feature map, and student feature map, the teacher model's emphasis on different feature channels is introduced into the student model. This guides the student model to focus on key regions that contribute significantly to target detection during training, suppressing interference from complex background clutter and enhancing feature discriminativeness. Further, the intermediate detection values ​​obtained from the student model's detection module are input into the teacher model's subsequent detection layer to generate cross-detection results. By constraining these cross-detection results to maintain consistency with the teacher model's detection results, the student model's prediction logic can be more robustly calibrated, enhancing the stability and reliability of its prediction results and effectively avoiding target conflict problems that may arise from traditional distillation supervision methods. Attached Figure Description

[0018] The foregoing contents, as well as other objects, features, and advantages of this disclosure, will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 This illustration schematically depicts an application scenario of a knowledge distillation method and apparatus for SAR image target detection, and a target detection method and apparatus according to embodiments of the present disclosure.

[0020] Figure 2 A flowchart illustrating a knowledge distillation method for SAR image target detection according to an embodiment of the present disclosure is shown schematically.

[0021] Figure 3 A flowchart illustrating student model training according to an embodiment of the present disclosure is shown schematically;

[0022] Figure 4A flowchart illustrating the calculation of frequency domain loss values ​​according to embodiments of the present disclosure is shown schematically.

[0023] Figure 5 A flowchart illustrating the calculation of spatial domain loss values ​​according to embodiments of the present disclosure is shown schematically.

[0024] Figure 6 A flowchart illustrating a target detection method according to an embodiment of the present disclosure is shown schematically.

[0025] Figure 7 This schematically illustrates a structural block diagram of a knowledge distillation apparatus for SAR image target detection according to an embodiment of the present disclosure;

[0026] Figure 8 A schematic block diagram of a target detection apparatus according to an embodiment of the present disclosure is shown.

[0027] Figure 9 A block diagram of an electronic device suitable for implementing a knowledge distillation method for SAR image target detection, according to an embodiment of the present disclosure, is shown schematically. Detailed Implementation

[0028] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0031] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0032] Synthetic aperture radar, with its unique microwave imaging method, has the advantage of all-weather and all-time imaging, and can overcome the limitations of optical imaging that are restricted by weather and lighting. Therefore, it is widely used in fields such as marine monitoring, resource exploration, disaster assessment and military reconnaissance.

[0033] Traditional SAR image target detection methods mainly include constant false alarm rate (CFAR) detection algorithms and modeling and discrimination based on manually extracted image features such as grayscale, texture, and shape. While these methods offered some interpretability and ease of implementation in early engineering projects, they generally rely on threshold settings and prior assumptions, are sensitive to clutter, speckle noise, and differences in scattering mechanisms, and have limited generalization capabilities. In large-scale, high-resolution image processing of complex land-sea boundaries or suburban areas, ensuring detection quality often requires significant manual intervention for parameter tuning and result verification, leading to low overall efficiency, insufficient detection consistency, and substantial consumption of human and material resources at the operational level, making it difficult to support continuous operation and maintenance for high-frequency, near-real-time services.

[0034] Deep learning methods offer significant advantages in target detection for SAR images, enabling the automatic extraction of multi-layered and complex image features through training. However, SAR systems generate an enormous amount of data during imaging. Transmitting all raw echoes or large volumes of image data back to ground for processing would result in massive bandwidth consumption and storage pressure, significantly increasing the burden on the satellite-to-ground link and introducing processing latency. This makes it difficult to meet the minute-level or even second-level response requirements of scenarios such as disaster monitoring and emergency response. Furthermore, the computational and storage costs of models are substantial, posing challenges for deployment on spaceborne equipment or on-orbit edge platforms where power consumption, computing power, and storage are limited.

[0035] In view of this, embodiments of this disclosure provide a knowledge distillation method for SAR image target detection. A pre-trained teacher model is used to train a student model. Both the teacher and student models include a feature extraction module, a feature fusion module, and a detection module. The feature extraction module of the teacher model has more layers than the feature extraction module of the student model. The detection modules of both the teacher and student models each include N detection layers, where N is a positive integer greater than 1. The method includes: extracting features from synthetic aperture radar (SAR) image samples using the feature extraction module of the pre-trained teacher model to obtain a teacher feature map; extracting features from SAR image samples using the feature extraction module of the student model to obtain a student feature map; performing a frequency domain transformation on the teacher feature map to obtain teacher frequency domain features; performing a frequency domain transformation on the student feature map to obtain student frequency domain features, where the teacher frequency domain features and student frequency domain features respectively characterize the edge detail features extracted by the teacher and student models for SAR image samples; extracting attention from the teacher feature map to obtain a teacher attention map, which represents the degree of attention the teacher model pays to each sub-region of the SAR image sample; and utilizing the detection module of the teacher model... The teacher feature map is detected to obtain the teacher detection result. The student model's detection module is then used to detect the student feature map, resulting in the student detection result output from the Nth detection layer and the intermediate student detection value output from the nth detection layer, where n is a positive integer greater than 1 and less than N. The intermediate student detection value is then input into the (n+1)th detection layer of the teacher model to obtain the cross-detection result. A joint loss value is determined based on the teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection result, cross-detection result, student detection result, and sample labels. The sample labels indicate the actual detection result of the synthetic aperture radar image sample. The parameters of the student model are adjusted based on the joint loss value to obtain the trained student model.

[0036] Figure 1 The illustration schematically depicts an application scenario of a knowledge distillation method and apparatus for SAR image target detection, and a target detection method and apparatus according to embodiments of the present disclosure.

[0037] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0038] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0039] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0040] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0041] It should be noted that the knowledge distillation method and target detection method for SAR image target detection provided in this disclosure embodiment can generally be executed by server 105. Correspondingly, the knowledge distillation device and target detection device for SAR image target detection provided in this disclosure embodiment can generally be located in server 105. The knowledge distillation method and target detection method for SAR image target detection provided in this disclosure embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the knowledge distillation device and target detection device for SAR image target detection provided in this disclosure embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0042] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0043] The following will be based on Figure 1 The described scene, through Figures 2-5 A knowledge distillation method for SAR image target detection according to the disclosed embodiments is described in detail.

[0044] Figure 2 A flowchart illustrating a knowledge distillation method for SAR image target detection according to an embodiment of the present disclosure is shown.

[0045] like Figure 2 As shown, the knowledge distillation method for SAR image target detection in this embodiment includes operations S210 to S270.

[0046] According to embodiments of this disclosure, a pre-trained teacher model can be used to train a student model. Both the teacher model and the student model include a feature extraction module (also called a backbone network), a feature fusion module (also called a neck), and a detection module (also called a detection head). The feature extraction module of the teacher model has more layers than the feature extraction module of the student model. The detection modules of both the teacher model and the student model include N detection layers, where N is a positive integer greater than 1.

[0047] According to embodiments of this disclosure, the student model is used on a resource-constrained platform.

[0048] In operation S210, the feature extraction module of the pre-trained teacher model is used to extract features from the synthetic aperture radar image samples to obtain the teacher feature map, and the feature extraction module of the student model is used to extract features from the synthetic aperture radar image samples to obtain the student feature map.

[0049] According to embodiments of this disclosure, multi-layer feature maps can be extracted using a feature pyramid network or equivalent structure of the feature fusion module. Furthermore, the teacher and student feature maps can be normalized to reduce amplitude differences between them.

[0050] In operation S220, the teacher feature map is transformed in the frequency domain to obtain the teacher's frequency domain features, and the student feature map is transformed in the frequency domain to obtain the student's frequency domain features.

[0051] Among them, the teacher frequency domain features and the student frequency domain features represent the edge detail features extracted by the teacher model and the student model for synthetic aperture radar image samples, respectively.

[0052] According to embodiments of this disclosure, a two-dimensional fast Fourier transform algorithm can be used to map the feature map to the frequency domain, further align the frequency features therein, and restore it to a frequency domain feature map (i.e., student frequency domain features and teacher frequency domain features) through inverse fast Fourier transform.

[0053] In operation S230, attention is extracted from the teacher feature map to obtain the teacher attention map.

[0054] The teacher attention map represents the degree of attention the teacher model pays to each sub-region of the synthetic aperture radar image sample.

[0055] According to embodiments of this disclosure, an attention heatmap can be generated based on a teacher feature map, wherein high-value regions in the attention heatmap represent sub-regions that the teacher model considers important.

[0056] In operation S240, the detection module of the teacher model is used to detect the teacher feature map to obtain the teacher detection result, and the detection module of the student model is used to detect the student feature map to obtain the student detection result output by the Nth detection layer and the intermediate detection value of the student output by the nth detection layer.

[0057] Where n is a positive integer greater than 1 and less than N.

[0058] In operation S250, the intermediate detection values ​​of students are input into the (n+1)th detection layer of the teacher model to obtain the cross-detection results.

[0059] According to embodiments of this disclosure, the detection modules of both the teacher model and the student model can include 5 detection layers, and the intermediate detection values ​​of students output from the 4th detection layer can be input into the 5th detection layer of the teacher model.

[0060] The cross-detection result can be represented by the following formula (1).

[0061] (1)

[0062] in, This represents the intermediate detection value of the student output by the nth detection layer. () indicates that the processing is performed by the (n+1)th detection layer of the teacher model. This indicates the cross-detection result.

[0063] In the S260 operation, the joint loss value is determined based on teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection results, cross-detection results, student detection results, and sample labels.

[0064] The sample label indicates the actual detection result of the synthetic aperture radar image sample.

[0065] According to embodiments of this disclosure, the sample labels may include data such as the bounding box and category of manually labeled synthetic aperture radar image samples.

[0066] In operation S270, the parameters of the student model are adjusted based on the joint loss value to obtain the trained student model.

[0067] According to embodiments of this disclosure, frequency domain transformation is performed on student and teacher feature maps to obtain teacher and student frequency domain features. Then, based on the loss value between the teacher and student frequency domain features, the student model learns the teacher model's processing of image edge detail features. Furthermore, based on the loss value between the teacher attention map, teacher feature map, and student feature map, the teacher model's emphasis on different feature channels is introduced into the student model. This guides the student model to focus on key regions that contribute significantly to target detection during training, suppressing interference from complex background clutter and enhancing feature discriminativeness. Further, the intermediate detection values ​​obtained from the student model's detection module are input into the teacher model's subsequent detection layer to generate cross-detection results. By constraining these cross-detection results to maintain consistency with the teacher model's detection results, the student model's prediction logic can be calibrated, enhancing the stability and reliability of its prediction results and effectively avoiding target conflict problems that may arise from traditional distillation supervision methods.

[0068] According to embodiments of this disclosure, the above-mentioned frequency domain transformation of the teacher feature map to obtain teacher frequency domain features includes: performing a fast Fourier transform on the teacher feature map to obtain teacher transformed features; extracting high-frequency features from the teacher frequency features using a filter to obtain teacher high-frequency features; and performing an inverse fast Fourier transform on the teacher high-frequency features to obtain teacher frequency domain features.

[0069] According to the embodiments of this disclosure, high-frequency features can be extracted using a high-pass filter, which can be defined by the following formula (2).

[0070] (2)

[0071] in, Let H be a high-pass filter, (x, y) be the coordinates in the high-frequency feature map, H be the height of the high-frequency feature map, and W be the width of the high-frequency feature map. , , This is the proportionality coefficient.

[0072] According to embodiments of this disclosure, high-frequency features in the teacher's frequency features are extracted by a filter, so that during the training process of the student model, the student model can learn the detailed texture and edge structure information contained in the high-frequency features of the model, thereby enhancing the student model's ability to perceive subtle features.

[0073] According to an embodiment of this disclosure, the above-mentioned attention extraction from the teacher feature map to obtain a teacher attention map includes: performing a global average pooling operation on the teacher feature map to obtain a teacher average attention vector; performing a global max pooling operation on the teacher feature map to obtain a teacher maximum attention vector; determining a target attention vector based on the teacher average attention vector and the teacher maximum attention vector; and expanding the target attention vector by channel dimension to generate a teacher attention map.

[0074] According to an embodiment of this disclosure, the teacher feature map obtained after being extracted by the l-layer feature extraction module... The average teacher attention vector is obtained by performing global average pooling and global max pooling respectively. And the teacher's maximum attention vector Where C, H, and W represent the number of channels, height, and width of the feature map, respectively. Furthermore, different weights can be assigned to the teacher's average attention vector and the teacher's maximum attention vector to obtain the target attention vector.

[0075] The target attention vector can be obtained using the following formula (3).

[0076] (3)

[0077] in, For the target attention vector, It is the balance coefficient. The average attention vector of teachers. This represents the teacher's maximum attention vector.

[0078] According to embodiments of this disclosure, the target attention vector can be extended in the channel dimension (also known as broadcasting) to generate a teacher attention map.

[0079] According to embodiments of this disclosure, by performing global average pooling and global max pooling on the teacher feature map, attention weights for each channel are extracted, and a target attention vector that can characterize the global distribution characteristics and local significant activation features of the teacher feature map in the channel dimension is constructed. Then, the target attention vector is expanded in the channel dimension to obtain the teacher attention map. This allows the student model and the teacher model to align their attention in key feature regions, enabling the student model to accurately distinguish between target and non-target regions in complex backgrounds, reducing interference caused by background clutter, and enhancing the discriminativeness and robustness of the overall features.

[0080] According to embodiments of this disclosure, the determination of the joint loss value based on teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection result, cross-detection result, student detection result, and sample label includes: determining a frequency domain loss value based on teacher frequency domain features and student frequency domain features; determining a spatial domain loss value based on teacher attention map, teacher feature map, and student feature map; determining a cross-detection loss value based on teacher detection result and cross-detection result; determining a predicted loss value based on student detection result and sample label; and determining a joint loss value based on the frequency domain loss value, spatial domain loss value, cross-detection loss value, and predicted loss value.

[0081] According to embodiments of this disclosure, the feature fusion modules of the teacher model and the student model both include M feature layers. The teacher frequency domain features include M sub-teacher frequency domain features corresponding to the M feature layers, and the student frequency domain features include M sub-student frequency domain features corresponding to the M feature layers, where M is a positive integer greater than 1. Determining the frequency domain loss value based on the teacher frequency domain features and the student frequency domain features includes: for the m-th feature layer, calculating the square Euclidean norm of the m-th sub-teacher frequency domain feature and the m-th sub-student frequency domain feature to obtain the m-th sub-frequency domain loss value; and obtaining the frequency domain loss value based on the M sub-frequency domain loss values.

[0082] The frequency domain loss value can be determined by the following formula (4).

[0083] (4)

[0084] in, This is the frequency domain loss value. For student frequency domain characteristics, For teacher frequency domain characteristics, It is the square Euclidean norm.

[0085] According to embodiments of this disclosure, by aligning the high-frequency feature representation of the teacher model layer by layer with the student model through mean square error constraints, the student model's ability to represent the edge details and contour direction of SAR images can be improved, thereby enabling it to accurately capture target information and reduce missed detections even when faced with speckle noise and blurred boundaries.

[0086] According to an embodiment of this disclosure, the aforementioned teacher attention feature map includes M attention sub-feature maps corresponding to the M-layer feature layer, the teacher feature map includes M sub-teacher feature maps corresponding to the M-layer feature layer, and the student feature map includes M sub-student feature maps corresponding to the M-layer feature layer. Determining the spatial domain loss value based on the teacher attention map, teacher feature map, and student feature map includes: for the m-th feature layer, calculating the Hadamard product of the m-th attention sub-feature map and the m-th sub-student feature map to obtain the m-th sub-spatial domain student prediction feature map; obtaining the m-th sub-spatial domain loss value based on the m-th sub-teacher feature map and the m-th sub-spatial domain student prediction feature map; and obtaining the spatial domain loss value based on the M sub-spatial domain loss values.

[0087] The spatial domain loss value can be determined by the following formula (5).

[0088] (5)

[0089] in, This represents the spatial domain loss value. Predict the feature map for the student in the m-th subspace. Let m be the feature map of the m-th sub-teacher.

[0090] The student prediction feature map of the m-th subspace domain is obtained by the following formula (6).

[0091] (6)

[0092] in, Predict the feature map for the student in the m-th subspace. For the feature map of the m-th sub-student, For the m-th attention sub-feature map, It represents the Hadamardi (or Hadama) stack.

[0093] According to an embodiment of this disclosure, the determination of the cross-loss value based on the teacher detection results and the cross-detection results includes: calculating the distribution difference between the teacher detection results and the cross-detection results to obtain the cross-loss value.

[0094] According to embodiments of this disclosure, distribution differences can be determined by the cross-entropy loss function or the KL divergence.

[0095] The cross loss value can be calculated using the following formula (7).

[0096] (7)

[0097] in, This is the cross loss value. Due to distribution differences, For the cross-detection results, For teachers' test results.

[0098] Furthermore, the joint loss value can be obtained by the following formula (8).

[0099] (8)

[0100] in, For the joint loss value, To predict the loss value, This is the frequency domain loss value. This represents the spatial domain loss value. This is the cross loss value. This is the balance coefficient.

[0101] According to embodiments of this disclosure, the student model is optimized by obtaining cross-loss values ​​through corresponding layers of the detection modules across the teacher and student models. This allows the student model to learn the prediction output of the teacher model, thereby enhancing the stability and reliability of the student model's prediction results and avoiding performance degradation caused by conflicts during the distillation process. This results in more robust target recognition in actual detection tasks.

[0102] Figure 3 A flowchart illustrating student model training according to an embodiment of the present disclosure is shown schematically.

[0103] like Figure 3 As shown, the synthetic aperture radar image is processed by feature extraction modules of the teacher model and student model to obtain teacher feature maps and student feature maps, respectively. Then, the teacher feature maps and student feature maps are aligned in the frequency domain and spatial domain using a frequency domain alignment distillation module and a spatial domain feature distillation module. Further, the student feature map is processed by the detection module of the student model to obtain intermediate student detection values ​​and student detection results. The intermediate student detection results are then processed by the (n+1)th detection layer of the teacher model to obtain cross-detection results. The teacher feature map is processed by the detection module of the teacher model to obtain teacher detection results. The predicted loss value can be determined from the student detection results and sample labels. The cross-loss value can be determined from the teacher's test results and the cross-test results. .

[0104] Figure 4 A flowchart illustrating the calculation of frequency domain loss values ​​according to an embodiment of the present disclosure is shown.

[0105] like Figure 4As shown, taking the teacher feature map as an example to obtain the teacher's frequency domain features, the teacher feature map undergoes a Fast Fourier Transform (FFT) to obtain the teacher's transformed features. These transformed features are then filtered by a high-pass filter to extract the high-frequency features, and then subjected to an Inverse Fast Fourier Transform (IFT) to obtain the teacher's frequency domain features. Similarly, the student's frequency domain features can be obtained. Finally, based on the teacher's and student's frequency domain features, the frequency domain loss value can be calculated. .

[0106] Figure 5 A flowchart illustrating the calculation of spatial domain loss values ​​according to an embodiment of the present disclosure is shown.

[0107] like Figure 5 As shown, the teacher feature map is subjected to global average pooling to obtain the teacher average attention vector, and the teacher feature map is subjected to global maximization to obtain the teacher maximum attention vector. The teacher average attention vector and the teacher maximum attention vector are added bit by bit to obtain the target attention vector. The target attention vector is further broadcast (channel dimension expansion) to obtain the teacher attention map. The teacher attention map is applied to the student feature map to obtain M sub-space domain student prediction feature maps. The spatial domain loss value is further obtained based on the M sub-teacher feature maps and the M sub-space domain student prediction feature maps.

[0108] Figure 6 A flowchart illustrating a target detection method according to an embodiment of the present disclosure is shown schematically.

[0109] The S610 is used to acquire synthetic aperture radar images of the target.

[0110] By operating the S620, the target synthetic aperture radar image is input into the pre-trained student model to obtain the target detection result.

[0111] The pre-trained student model is trained using the knowledge distillation method for SAR image target detection described above.

[0112] Based on the aforementioned knowledge distillation method for SAR image target detection, this disclosure also provides a knowledge distillation apparatus for SAR image target detection. The following will be combined with... Figure 7 The device is described in detail.

[0113] Figure 7 A schematic block diagram of a knowledge distillation apparatus for SAR image target detection according to an embodiment of the present disclosure is shown.

[0114] like Figure 7As shown, the knowledge distillation device 700 for SAR image target detection in this embodiment includes an extraction module 710, a frequency transformation module 720, an attention extraction module 730, a cross detection module 740, an input module 750, a loss value calculation module 760, and an adjustment module 770.

[0115] The extraction module 710 is used to extract features from synthetic aperture radar image samples using the feature extraction module of the pre-trained teacher model to obtain a teacher feature map, and to extract features from synthetic aperture radar image samples using the feature extraction module of the student model to obtain a student feature map. In one embodiment, the extraction module 710 can be used to perform the operation S210 described above, which will not be repeated here.

[0116] The frequency transformation module 720 is used to perform frequency domain transformation on the teacher feature map to obtain teacher frequency domain features, and to perform frequency domain transformation on the student feature map to obtain student frequency domain features. The teacher frequency domain features characterize the edge details of the synthetic aperture radar image samples extracted by the teacher model. In one embodiment, the frequency transformation module 720 can be used to perform the operation S220 described above, which will not be repeated here.

[0117] The attention extraction module 730 is used to extract attention from the teacher feature map to obtain a teacher attention map, which represents the degree of attention the teacher model pays to each sub-region of the synthetic aperture radar image sample. In one embodiment, the attention extraction module 730 can be used to perform the operation S230 described above, which will not be repeated here.

[0118] The cross-detection module 740 is used to detect the teacher feature map using the detection module of the teacher model to obtain the teacher detection result, and to detect the student feature map using the detection module of the student model to obtain the student detection result output by the Nth detection layer and the intermediate student detection value output by the nth detection layer, where n is a positive integer greater than 1 and less than N. In one embodiment, the cross-detection module 740 can be used to perform the operation S240 described above, which will not be repeated here.

[0119] The input module 750 is used to input the intermediate detection values ​​of students into the (n+1)th detection layer of the teacher model to obtain the cross-detection results. In one embodiment, the input module 750 can be used to perform the operation S250 described above, which will not be repeated here.

[0120] The loss value calculation module 760 is used to determine a joint loss value based on teacher frequency domain features, student frequency domain features, teacher attention map, teacher feature map, student feature map, teacher detection result, cross-detection result, student detection result, and sample label. The sample label indicates the actual detection result of the synthetic aperture radar image sample. In one embodiment, the loss value calculation module 760 can be used to perform the operation S260 described above, which will not be repeated here.

[0121] The adjustment module 770 is used to adjust the parameters of the student model based on the joint loss value to obtain a trained student model. In one embodiment, the adjustment module 770 can be used to perform the operation S270 described above, which will not be repeated here.

[0122] According to an embodiment of this disclosure, the frequency transformation module 720 includes: a transformation submodule for performing a fast Fourier transform on the teacher feature map to obtain teacher transformed features; an extraction submodule for extracting high-frequency features from the teacher frequency features using a filter to obtain teacher high-frequency features; and an inverse transformation submodule for performing an inverse fast Fourier transform on the teacher high-frequency features to obtain teacher frequency domain features.

[0123] According to an embodiment of this disclosure, the attention extraction module 730 includes: an average pooling submodule for performing a global average pooling operation on the teacher feature map to obtain a teacher average attention vector; a max pooling submodule for performing a global max pooling operation on the teacher feature map to obtain a teacher maximum attention vector; a determination submodule for determining a target attention vector based on the teacher average attention vector and the teacher maximum attention vector; and an expansion submodule for expanding the target attention vector in the channel dimension to generate a teacher attention map.

[0124] According to embodiments of this disclosure, the loss value calculation module 760 includes: a frequency domain loss value calculation submodule, used to determine a frequency domain loss value based on teacher frequency domain features and student frequency domain features; a spatial domain loss value calculation submodule, used to determine a spatial domain loss value based on teacher attention map, teacher feature map, and student feature map; a cross-loss value determination submodule, used to determine a cross-loss value based on teacher detection results and cross-detection results; a prediction loss value determination submodule, used to determine a prediction loss value based on student detection results and sample labels; and a joint loss value determination submodule, used to determine a joint loss value based on the frequency domain loss value, spatial domain loss value, cross-loss value, and prediction loss value.

[0125] According to embodiments of this disclosure, the feature fusion modules of both the teacher model and the student model include M feature layers. The teacher frequency domain features include M sub-teacher frequency domain features corresponding to the M feature layers, and the student frequency domain features include M sub-student frequency domain features corresponding to the M feature layers, where M is a positive integer greater than 1. The frequency domain loss value calculation submodule includes: a sub-loss value calculation unit, used to calculate the square Euclidean norm of the m-th sub-teacher frequency domain feature and the m-th sub-student frequency domain feature for the m-th feature layer to obtain the m-th sub-frequency domain loss value; and a frequency domain loss value determination unit, used to obtain the frequency domain loss value based on the M sub-frequency domain loss values.

[0126] According to embodiments of this disclosure, the teacher attention feature map includes M attention sub-feature maps corresponding to the M-layer feature layer, the teacher feature map includes M sub-teacher feature maps corresponding to the M-layer feature layer, the student feature map includes M sub-student feature maps corresponding to the M-layer feature layer, and the spatial domain loss value calculation submodule includes: a spatial domain student prediction feature map determination module, used to calculate the Hadamard product of the m-th attention sub-feature map and the m-th sub-student feature map for the m-th feature layer to obtain the m-th sub-spatial domain student prediction feature map; a sub-spatial domain loss value calculation unit, used to obtain the m-th sub-spatial domain loss value based on the m-th sub-teacher feature map and the m-th sub-spatial domain student prediction feature map; and a base spatial domain loss value calculation unit, used to obtain the spatial domain loss value based on the M sub-spatial domain loss values.

[0127] According to an embodiment of this disclosure, the cross-loss value determination submodule includes: a distribution difference calculation unit, used to calculate the distribution difference between the teacher detection results and the cross-detection results to obtain the cross-loss value.

[0128] According to embodiments of this disclosure, any multiple modules among the extraction module 710, frequency conversion module 720, attention extraction module 730, cross-detection module 740, input module 750, loss value calculation module 760, and adjustment module 770 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this disclosure, at least one of the extraction module 710, frequency conversion module 720, attention extraction module 730, cross-detection module 740, input module 750, loss value calculation module 760, and adjustment module 770 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the extraction module 710, frequency transformation module 720, attention extraction module 730, cross-detection module 740, input module 750, loss value calculation module 760, and adjustment module 770 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0129] Based on the above-described target detection method, this disclosure also provides a target detection device. The following will be combined with... Figure 8 The device is described in detail.

[0130] Figure 8 A schematic block diagram of a target detection apparatus according to an embodiment of the present disclosure is shown.

[0131] like Figure 8 As shown, the target device 800 in this embodiment includes an acquisition module 810 and a detection module 820.

[0132] The acquisition module 810 is used to acquire a synthetic aperture radar image of the target. In one embodiment, the acquisition module 810 can be used to perform the operation S610 described above, which will not be repeated here.

[0133] The detection module 820 is used to input the target synthetic aperture radar image into a pre-trained student model to obtain the target detection result. The pre-trained student model is trained using the knowledge distillation method for SAR image target detection described above. In one embodiment, the detection module 820 can be used to perform the operation S620 described above, which will not be repeated here.

[0134] Figure 9 A block diagram of an electronic device suitable for implementing a knowledge distillation method for SAR image target detection, according to an embodiment of the present disclosure, is shown schematically.

[0135] like Figure 9 As shown, an electronic device 900 according to an embodiment of the present disclosure includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage portion 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0136] RAM 903 stores various programs and data required for the operation of electronic device 900. Processor 901, ROM 902, and RAM 903 are interconnected via bus 904. Processor 901 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 902 and / or RAM 903. It should be noted that programs may also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in one or more memories.

[0137] According to embodiments of this disclosure, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to a bus 904. The electronic device 900 may also include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 910 as needed so that computer programs read from it can be installed into the storage section 908 as needed.

[0138] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0139] According to embodiments of this disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this disclosure, the computer-readable storage medium may include ROM 902 and / or RAM 903 and / or one or more memories other than ROM 902 and RAM 903 described above.

[0140] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the knowledge distillation method for SAR image target detection provided in embodiments of this disclosure.

[0141] When the computer program is executed by the processor 901, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0142] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 909, and / or installed from a removable medium 911. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0143] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 909, and / or installed from the removable medium 911. When the computer program is executed by the processor 901, it performs the functions defined in the system of this disclosure embodiment. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0144] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on a user's computing device, partially on a user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0145] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0146] Those skilled in the art will understand that the features described in the various embodiments of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments of this disclosure can be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0147] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A knowledge distillation method for target detection in SAR images, characterized in that, A student model is trained using a pre-trained teacher model. Both the teacher and student models include a feature extraction module, a feature fusion module, and a detection module. The feature extraction module of the teacher model has more layers than the feature extraction module of the student model. The detection modules of both the teacher and student models each include N detection layers, where N is a positive integer greater than 1. The method includes: The feature extraction module of the pre-trained teacher model is used to extract features from the synthetic aperture radar image samples to obtain the teacher feature map, and the feature extraction module of the student model is used to extract features from the synthetic aperture radar image samples to obtain the student feature map. The teacher feature map is transformed in the frequency domain to obtain the teacher frequency domain features, and the student feature map is transformed in the frequency domain to obtain the student frequency domain features. The teacher frequency domain features and the student frequency domain features respectively characterize the edge detail features extracted by the teacher model and the student model for the synthetic aperture radar image sample. Attention is extracted from the teacher feature map to obtain a teacher attention map, which represents the degree of attention the teacher model pays to each sub-region of the synthetic aperture radar image sample; The teacher feature map is detected using the detection module of the teacher model to obtain the teacher detection result, and the student feature map is detected using the detection module of the student model to obtain the student detection result output by the Nth detection layer and the intermediate student detection value output by the nth detection layer, where n is a positive integer greater than 1 and less than N; The intermediate detection values ​​of the students are input into the (n+1)th detection layer of the teacher model to obtain the cross-detection results; The joint loss value is determined based on the teacher frequency domain features, the student frequency domain features, the teacher attention map, the teacher feature map, the student feature map, the teacher detection result, the cross-detection result, the student detection result, and the sample label, wherein the sample label indicates the actual detection result of the synthetic aperture radar image sample; The parameters of the student model are adjusted based on the joint loss value to obtain the trained student model.

2. The method according to claim 1, characterized in that, The step of performing a frequency domain transformation on the teacher feature map to obtain the teacher's frequency domain features includes: Perform a Fast Fourier Transform on the teacher feature map to obtain the teacher transformed features; The high-frequency features of the teacher are extracted using a filter to obtain the teacher's high-frequency features; The teacher's high-frequency features are subjected to inverse fast Fourier transform to obtain the teacher's frequency domain features.

3. The method according to claim 2, characterized in that, The step of extracting attention from the teacher feature map to obtain a teacher attention map includes: Perform global average pooling on the teacher feature map to obtain the teacher average attention vector; Perform global max pooling on the teacher feature map to obtain the teacher maximum attention vector; The target attention vector is determined based on the teacher's average attention vector and the teacher's maximum attention vector; The target attention vector is expanded in the channel dimension to generate the teacher attention map.

4. The method according to claim 3, characterized in that, The determination of the joint loss value based on the teacher frequency domain features, the student frequency domain features, the teacher attention map, the teacher feature map, the student feature map, the teacher detection result, the cross-detection result, the student detection result, and the sample labels includes: The frequency domain loss value is determined based on the teacher's frequency domain characteristics and the student's frequency domain characteristics; Based on the teacher attention map, the teacher feature map, and the student feature map, a spatial domain loss value is determined; The cross-loss value is determined based on the teacher detection results and the cross-detection results; The predicted loss value is determined based on the student test results and the sample labels; The joint loss value is determined based on the frequency domain loss value, the spatial domain loss value, the cross loss value, and the prediction loss value.

5. The method according to claim 4, characterized in that, Both the teacher model and the student model feature fusion module contain M feature layers. The teacher frequency domain features include M sub-teacher frequency domain features corresponding to the M feature layers, and the student frequency domain features include M sub-student frequency domain features corresponding to the M feature layers, where M is a positive integer greater than 1. Determining the frequency domain loss value based on the teacher frequency domain features and the student frequency domain features includes: For the m-th feature layer, calculate the squared Euclidean norm of the frequency domain features of the m-th sub-teacher and the m-th sub-student to obtain the loss value of the m-th sub-frequency domain. The frequency domain loss value is obtained based on M of the sub-frequency domain loss values.

6. The method according to claim 5, characterized in that, The teacher attention feature map includes M attention sub-feature maps corresponding to M feature layers, the teacher feature map includes M sub-teacher feature maps corresponding to M feature layers, and the student feature map includes M sub-student feature maps corresponding to M feature layers. The step of determining the spatial domain loss value based on the teacher attention map, the teacher feature map, and the student feature map includes: For the m-th feature layer, calculate the Hadamard product of the m-th attention sub-feature map and the m-th sub-student feature map to obtain the m-th sub-spatial domain student prediction feature map; The loss value of the m-th sub-space domain is obtained based on the m-th sub-teacher feature map and the m-th sub-space domain student prediction feature map; The spatial domain loss value is obtained based on the M sub-spatial domain loss values.

7. The method according to claim 4, characterized in that, The determination of the cross-loss value based on the teacher detection results and the cross-detection results includes: The distribution difference between the teacher detection results and the cross detection results is calculated to obtain the cross loss value.

8. A target detection method, characterized in that, The method includes: Acquire synthetic aperture radar image of the target; The target synthetic aperture radar image is input into a pre-trained student model to obtain the target detection result, wherein the pre-trained student model is trained by any one of the methods in claims 1 to 7.

9. A knowledge distillation device for target detection in SAR images, characterized in that, A student model is trained using a pre-trained teacher model. Both the teacher and student models include a feature extraction module, a feature fusion module, and a detection module. The feature extraction module of the teacher model has more layers than the feature extraction module of the student model. The detection modules of both the teacher and student models each include N detection layers, where N is a positive integer greater than 1. The device includes: The feature extraction module is used to extract features from synthetic aperture radar image samples using the feature extraction module of the pre-trained teacher model to obtain teacher feature maps, and to extract features from synthetic aperture radar image samples using the feature extraction module of the student model to obtain student feature maps. The frequency transformation module is used to perform frequency domain transformation on the teacher feature map to obtain teacher frequency domain features, and to perform frequency domain transformation on the student feature map to obtain student frequency domain features. The teacher frequency domain features characterize the edge details of the synthetic aperture radar image samples extracted by the teacher model. An attention extraction module is used to extract attention from the teacher feature map to obtain a teacher attention map, which represents the degree of attention the teacher model pays to each sub-region of the synthetic aperture radar image sample. The cross-detection module is used to detect the teacher feature map using the detection module of the teacher model to obtain the teacher detection result, and to detect the student feature map using the detection module of the student model to obtain the student detection result output by the Nth detection layer and the intermediate student detection value output by the nth detection layer, where n is a positive integer greater than 1 and less than N; The input module is used to input the intermediate detection values ​​of the students into the (n+1)th detection layer of the teacher model to obtain the cross-detection results; The loss value calculation module is used to determine a joint loss value based on the teacher frequency domain features, the student frequency domain features, the teacher attention map, the teacher feature map, the student feature map, the teacher detection result, the cross detection result, the student detection result, and the sample label, wherein the sample label indicates the actual detection result of the synthetic aperture radar image sample; The adjustment module is used to adjust the parameters of the student model based on the joint loss value to obtain a trained student model.

10. A target detection device, characterized in that, The device includes: The acquisition module is used to acquire synthetic aperture radar images of the target; The detection module is used to input the synthetic aperture radar image of the target into a pre-trained student model to obtain the target detection result, wherein the pre-trained student model is trained by any one of the methods in claims 1 to 7.

Images