A sar small sample target detection method and system

Abstract

The application discloses a SAR small sample target detection method and system, and the method comprises the following steps: training a DINO-ViT module for semantic feature extraction on a ViT network architecture; SAR , performing attribute transformation on the source image I SAR to obtain an image I target_crop and an image I background_crop , and inputting the images I target_crop and I background_crop into three branches of an Unet3+ network respectively; a loss function is constructed, the Unet3+ network is trained, and a detection network connected to the output end of the Unet3+ network is also constrained and trained through a regression classification loss to obtain a target detection model; real-time collected images are input into the trained target detection model to obtain a final detection result; the application has the advantages that problem fitting and precision improvement can be performed without requiring huge training data, and the application can be used in a data collection difficult situation and a specific small sample scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of synthetic aperture radar (SAR) automatic target recognition, and particularly relates to a SAR small sample target detection method and system. BACKGROUND

[0002] Synthetic aperture radar (SAR) has the unique ability to provide high-resolution ground images under any weather and day-night conditions. This makes SAR widely used in military and civilian fields. Among them, the automatic detection and recognition of ship targets is an important task, aiming to protect the safety of maritime traffic and crack down on illegal maritime activities. However, due to the characteristics of SAR images and the complexity of the marine environment, this field faces many challenges.

[0003] The early SAR target detection and recognition algorithm research focuses on traditional methods, mainly based on manual features and limited shallow learning representation, which can be divided into three steps: preprocessing, candidate region extraction and discrimination. The constant false alarm rate (CFAR) based method is a common candidate region extraction method, which divides the pixel region into two categories of ships and non-ships according to the threshold. It is difficult to model for small ships and complex sea scenes. The rise of deep learning solves the problem of lack of robustness and universality of traditional methods, and the end-to-end architecture simplifies the complex process of manual design. The SAR target detection and recognition method based on deep learning can be divided into two categories: one is two-stage detection, which uses CNN network to perform two regression and classification on the pre-selected frame. The classic two-stage detection algorithm includes Faster R-CNN, R-FCN, FPN and Mask R-CNN. The other is one-stage detection, which uses a full convolutional network to perform one classification and regression on the pre-selected frame. The representative algorithms include YOLO, SSD, RetinaNet and CornerNet. The two-stage detection algorithm has higher accuracy, and the one-stage algorithm has faster prediction speed. Zhou et al. proposed a multi-scale ship detection network based on YOLOv5, which uses a cross-stage partial network to improve the feature representation capability, and a feature pyramid network with a fusion coefficient module to adaptively fuse feature mapping, achieving a good balance between model size and inference time. The literature "Sun W, Huang X. Semantic Attention-Based Network for Inshore SAR Ship Detection [C] / / International Conference on Digital Image Processing. 2021." proposes SANet (semantic attention-based network), which combines semantic attention, focal loss, labels and anchor assignment to improve detection performance without increasing computational complexity. Most of the above existing algorithms are based on large training data to realize problem fitting and precision improvement, which is difficult to cope with data acquisition difficulties and specific small sample scene description. SUMMARY

[0004] The technical problem to be solved by the present application is that most of the existing SAR target detection algorithms are based on large training data to realize problem fitting and precision improvement, which is difficult to cope with data acquisition difficulties and specific small sample scene description.

[0005] The present application solves the above technical problems by the following technical means: a SAR small sample target detection method, comprising the following steps:

[0006] Step one: train a DINO-ViT module for semantic feature extraction on a ViT network architecture;

[0007] Step two: the source image I SAR is subjected to attribute transformation to obtain image I SAR , image I target_crop and image I background_crop , which are respectively input into three branches of the Unet3+ network, and feature enhancement and fusion are realized through parameter sharing;

[0008] Step three: semantic information of enhanced features is extracted by using the pre-trained DINO-ViT module, a loss function is constructed, the Unet3+ network is trained, and a detection network connected to the output end of the Unet3+ network is also trained through a regression classification loss, and the training is stopped when all loss function values are minimized, and a trained model is obtained, which is defined as a target detection model;

[0009] Step four: real-time collected images are input into the Unet3+ network in the trained target detection model, and three branches thereof generate images with different emphases, which are sent into a detection network as effective feature layers to obtain a target detection result.

[0010] Further, the step one comprises:

[0011] Based on the SSDD dataset on the ViT network architecture, a semantic feature extraction module suitable for SAR images is trained using a label-free self-distillation method, and the module is defined as a DINO-ViT module.

[0012] Further, the step two comprises:

[0013] The source image I SAR is a collected SAR image, and the SAR image is subjected to random cropping, horizontal flipping and color shaking to generate image I target_crop , the SAR image is subjected to random cropping, horizontal flipping and Gaussian blur processing to generate image I background_crop .

[0014] Further, the step three of constructing a loss function comprises:

[0015] The loss function comprises background space structure constraint, target appearance structure constraint and key-value detail constraint, and the calculation formula is

[0016] L con = aL stru + bL app + cL key

[0017] Wherein, L con represents the content loss of feature extraction, L con is obtained from the background space structure loss L stru, target appearance structure loss L app and key-value loss L key , a, b, c are hyperparameters used to balance the loss.

[0018] Further, the background space structure loss L stru The key-value of the multi-head attention mechanism is calculated using the cosine autocorrelation, so that the generated image can retain more target structure information of the SAR image, and is defined as follows:

[0019] L stru =||(S(I SAR )-S(I f ))+(S(I background_crop )-S(I out_2 ))|| F

[0020] In the formula, I f , I out_1 respectively represent the fusion image and the attribute transformation image I target_crop corresponding output, S(I) represents the cosine autocorrelation of the image, and |||| represents the Frobenius norm. F

[0021] Further, the target appearance structure loss L app The [CLS] token is extracted from the ViT feature space, and the token is used to constrain the network to retain the target appearance texture of the SAR image, and the target appearance structure loss L app is defined as follows:

[0022] L app =||(T cls (I target_crop )-T cls (I f ))+(T cls (I target_crop )-T cls (I out_1 ))||2

[0023] In the formula, T cls (·) represents extracting the [CLS] token about the image, and ||||2 represents the l2 norm.

[0024] Further, the key-value of the ViT feature space describes some detailed structure information, and through the key-value loss L key the detailed texture features of the source image can be retained, and the key-value loss L key is defined as follows:

[0025] L key =(1-λ)||K(I​target_crop )-K(I out_1 || F +lambda||K(I SAR )-K(I f || F

[0026] In the formula, K(·) represents extracting key value information about the image, and lambda is an equalization parameter.

[0027] Further, the regression classification loss in the third step includes a classification loss function and a regression loss function.

[0028] The classification loss function is constructed using the VFL Loss function.

[0029] The regression loss function is constructed using the CIOU Loss+DFL function.

[0030] The application also provides a SAR small sample target detection system, comprising:

[0031] A pre-training unit is configured to train a DINO-ViT module for semantic feature extraction on a ViT network architecture.

[0032] A data processing unit is configured to input a source image I SAR , an image I SAR obtained by performing attribute transformation on the source image I target_crop , and an image I background_crop as inputs into three branches of an Unet3+ network respectively, and realize feature enhancement and fusion through parameter sharing.

[0033] A model training unit is configured to extract semantic information of enhanced features by using the pre-trained DINO-ViT module, construct a loss function, train the Unet3+ network, and simultaneously constrain a detection network connected to an output end of the Unet3+ network through a regression classification loss, stop training when all loss function values are minimized, and obtain a trained model, which is defined as a target detection model.

[0034] A detection output unit is configured to input a real-time collected image into the Unet3+ network in the trained target detection model, generate images with different emphases by three branches of the Unet3+ network, input the images as effective feature layers into the detection network, and obtain a target detection result.

[0035] Further, the pre-training unit is further configured to:

[0036] The semantic feature extraction module adapted to SAR images is trained in a label-free self-distillation manner on the ViT network architecture based on an SSDD dataset, and the module is defined as a DINO-ViT module.

[0037] Further, the data processing unit is further configured to:

[0038] The source image I SAR That is, for the collected SAR image, randomly crop, horizontally flip, and color shake the SAR image to generate an image I target_crop , randomly crop, horizontally flip, and Gaussian blur the SAR image to generate an image I background_crop .

[0039] Further, the model training unit comprises a loss function comprising:

[0040] The loss function comprises a background space structure constraint, a target appearance structure constraint, and a key-value detail constraint, and the calculation formula is

[0041] L con = aL stru + bL app + cL key

[0042] Wherein, L con represents the content loss of feature extraction, L con is composed of a background space structure loss L stru , a target appearance structure loss L app , and a key-value loss L key , and a, b, and c are all hyperparameters for balancing the loss.

[0043] Further, the background space structure loss L stru uses cosine autocorrelation to calculate the key-value of the multi-head attention mechanism, so that the generated image can retain more target structure information of the SAR image, and is defined as follows:

[0044] L stru = ||(S(I SAR )-S(I f ))+(S(I background_crop )-S(I out_2 ))|| F

[0045] In the formula, I f , I out_1 respectively represent the fusion image and the attribute transformed image I target_crop corresponding to the output, S(I) represents the calculation of the cosine autocorrelation of the image, and ||| F represents the calculation of the Frobenius norm.

[0046] Further, the target appearance structure loss L appThe [CLS] mark is extracted from the ViT feature space, the mark is used to constrain the network, the target appearance texture of the SAR image is reserved, and the target appearance structure loss L app The definitions are as follows:

[0047] L app =||(T cls (I target_crop )-T cls (I f ))+(T cls (I target_crop )-T cls (I out_1 ))||2

[0048] In the formula, T cls (·) represents extracting the [CLS] mark about the image, and |||2 calculates the l2 norm.

[0049] Further, the key value of the ViT feature space describes some detailed structure information, and the key value loss L key can reserve the detailed texture features of the source image, and the key value loss L key is defined as follows:

[0050] L key =(1-λ)||K(I target_crop )-K(I out_1 )| F +λ||K(I SAR )-K(I f )| F

[0051] In the formula, K(·) represents extracting the key value information about the image, and lambda is a balance parameter.

[0052] Further, the regression classification loss in the model training unit includes a classification loss function and a regression loss function;

[0053] The VFL Loss function is used to construct the classification loss function;

[0054] The CIOU Loss+DFL function is used to construct the regression loss function.

[0055] The present application has the following advantages:

[0056] (1) The present application performs attribute transformation on the source image I SAR to obtain the image I target_crop and the image I background_cropAs input, the three branches of the Unet3+ network are sent in, the attribute transformation mode not only increases the randomness of spatial information and reduces redundant information, but also expands the number of samples, and a loss function is constructed to train the model, so that problem fitting and precision improvement can be performed without a large amount of training data, and the method can be used in difficult data acquisition conditions and specific small sample scenarios.

[0057] (2) The application generates a semantic prior module for a SAR image by using the knowledge distillation mode of DINO-ViT, uses the [CLS] label of the ViT feature space to represent the foreground target, uses the self-similarity of the key to represent the background space structure, and uses a semantic loss function to constrain network fitting direction, so that the foreground and background semantic information of the source image is distinguished, and effective feature layers are provided for detection.

[0058] (3) The application realizes foreground and background feature separation by the three-branch parameter sharing mode, realizes the balance and enhancement of effective information, and generates foreground images, background images and fusion images as effective feature layers for classification and regression, thereby improving the detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 An algorithm framework diagram of the SAR small sample target detection method is disclosed. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below in conjunction with the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0061] Embodiment 1

[0062] As shown in Figure 1 , the application provides a SAR small sample target detection method, including the following steps:

[0063] S1: Based on the SSDD dataset, a DINO-ViT module for semantic feature extraction suitable for SAR images is trained on the ViT network architecture using a label-free self-distillation method; the specific process is as follows:

[0064] A ViT semantic feature extraction module for the SSDD dataset is trained using the self-supervised method DINO, and a pre-trained module DINO-ViT extracts foreground information and background information of the SAR image as knowledge prior of the small sample target detection network. The foreground and background information of the source image I SARor the source image I target_crop , background_crop respectively as input, generate three different data streams, I target_crop representing random cropping, horizontal flipping, color jittering on SAR images, I background_crop representing random cropping, horizontal flipping and Gaussian blur on SAR images, and the above operations are performed synchronously.

[0065] S2: the source image I SAR , the image I SAR obtained by performing attribute transformation on the source image I target_crop , and the image I background_crop are respectively input into three branches of the Unet3+ network, and feature enhancement and fusion are performed, the three branches input different information but share parameters, so as to quickly learn to distinguish foreground and background information;

[0066] Based on the extracted spatial structure and target appearance representation, a loss function is constructed to make the generated features semantically emphasize the difference between foreground and background. Among them, the first data stream inputs I target_crop , and the loss function is used to force the network to enhance the impression of the foreground target features in the iteration, and to learn the target appearance features I SAR ; the second data stream inputs the original I SAR , so that the output features can retain the foreground and background information of the SAR image; the third data stream inputs I background_crop , to enhance the impression of the generated results on I background_crop spatial structure information; overall, a separate key-value is also used as a constraint to enhance the retention of spatial detail information of the attribute transformation image. The loss function constraint part will be described in detail in the following step S3, and will not be repeated here.

[0067] S3: the pre-trained DINO-ViT module is used to extract semantic information of enhanced features, a loss function is constructed to constrain feature enhancement effect, and effective information is provided for the detection network. At the same time, the detection network connected with the output end of the Unet3+ network is also constrained through regression classification loss, the whole model is trained, and the training is stopped when all loss function values are minimized, and the trained model is obtained. The model is defined as a target detection model; the specific process is as follows:

[0068] The three data streams update the model parameters in the gradient descent algorithm, which enhances the differentiation of foreground and background features, and adaptively fuses the two in the same image. The features generated by the three data streams are used as effective feature layers in the detection stage, and the accuracy of regression classification is improved according to the enhanced foreground and background information. The loss function construction process of each part is described in detail below.

[0069] The loss function is divided into feature extraction loss, classification loss and regression loss. Considering that the application is based on semantic feature extraction and fusion of foreground and background differences, the key value similarity of the ViT feature space is used to represent the background space structure, the [CLS] label represents the appearance of the foreground target, and a single key value is introduced to supplement the global details, the loss function includes background space structure constraint, target appearance structure constraint and key value detail constraint, and the calculation formula is

[0070] L con =aL stru +bL app +cL key

[0071] Wherein, L con represents the content loss of feature extraction, L con is composed of background space structure loss L stru , target appearance structure loss L app and key value loss L key , a, b and c are all super parameters for balancing loss.

[0072] The background space structure loss L stru uses cosine autocorrelation to calculate the key value of the multi-head attention mechanism, so that the generated image can retain more target structure information of the SAR image, and is defined as follows:

[0073] L stru ||(S(I SAR )-S(I f ))+(S(I background_crop )-S(I out_2 ))|| F

[0074] In the formula, I f , I out_1 respectively represent the fused image, the attribute transformed image I target_crop highlighting the background information, and the corresponding output, S(I) represents the cosine autocorrelation of the image, and || || F represents the calculation of Frobenius norm.

[0075] The target appearance structure loss L app extracts the [CLS] label from the ViT feature space, and uses the label to constrain the network to retain the target appearance texture of the SAR image, and the target appearance structure loss L app is defined as follows:

[0076] L app ||(T cls (I target_crop )-T cls (I f ))+(Tcls (I target_crop )-T cls (I out_1 ))||2

[0077] In the formula, T cls (·) represents extracting the [CLS] mark of the image, and || ||2 calculates the l2 norm.

[0078] The key value of the ViT feature space describes some detailed structure information, and the key value loss L key can retain the detailed texture features of the source image, and the key value loss L key is defined as follows:

[0079] L key =(1-λ)||K(I target_crop )-K(I out_1 )|| F +λ||K(I SAR )-K(I f )|| F

[0080] In the formula, K(·) represents extracting the key value information of the image, and λ is a balance parameter.

[0081] The regression classification loss includes a classification loss function and a regression loss function;

[0082] The classification loss function is constructed using the VFL Loss function, the classification loss is calculated using the VFL Loss, and the non-symmetric weighting operation is performed for the positive and negative sample imbalance and the sample unequal weight condition.

[0083] The regression loss function is constructed using the CIOU Loss+DFL function, the regression loss is calculated using the CIOU Loss+DFL, and the position of the pre-selected frame is modeled as the overall data distribution, so that the network quickly focuses on the distribution close to the target position.

[0084] When all the loss function values reach the minimum, stop training, and obtain the trained target detection model.

[0085] S4: input the real-time collected image into the Unet3+ network in the trained target detection model, and the three branches generate images with different emphases, which are sent into the detection network as effective feature layers to obtain the target detection result. The detection network can be a classifier, such as a softmax classifier.

[0086] It needs to be explained that the present application mainly improves the input image part of the model and the loss function involved in the training process, and the network architecture itself is not improved. The above-mentioned ViT network architecture and Unet3+ network are prior art, and the specific architecture composition is not described in detail.

[0087] Through the above technical scheme, the present application adopts a small sample training method to weaken the dependence of the model on the data volume. By limiting the threshold to randomly crop the image and simultaneously performing attribute transformation, the spatial information randomness is increased, the redundant information is reduced, and the foreground and background features are extracted specifically. The DINO-ViT knowledge distillation method is used to generate a semantic prior module for SAR images, the [CLS] label of the ViT feature space is used to represent the foreground target, and the key self-similarity represents the background space structure. The network fitting direction is constrained by the semantic loss function, so that the foreground and background semantic information of the source image is distinguished to provide effective feature layers for detection. Overall, the network realizes foreground and background feature separation through the three branches sharing parameters, realizes the balance and enhancement of effective information, and generates foreground images, background images and fusion images as effective feature layers for classification and regression to improve the detection accuracy. It can well cope with the situation of difficult data collection, and can better learn scene features when describing specific scenes to improve detection accuracy.

[0088] Embodiment 2

[0089] Based on embodiment 1, the present application embodiment 2 further provides a SAR small sample target detection system, comprising:

[0090] A pre-training unit is configured to train a DINO-ViT module for semantic feature extraction on a ViT network architecture.

[0091] A data processing unit is configured to input the source image I SAR , the image I SAR obtained by performing attribute transformation on the source image I target_crop , and the image I background_crop as inputs into three branches of the Unet3+ network respectively, and realize feature enhancement and fusion through parameter sharing.

[0092] A model training unit is configured to extract semantic information of enhanced features by using the pre-trained DINO-ViT module, construct a loss function, train the Unet3+ network, and also constrain a detection network connected to the output end of the Unet3+ network through regression classification loss, stop training when all loss function values are minimized, and obtain a trained model, which is defined as a target detection model.

[0093] The detection output unit is configured to input the real-time collected image into an Unet3+ network in the trained target detection model, and three branches of the Unet3+ network generate images with different emphases as effective feature layers to be input into a detection network to obtain a target detection result.

[0094] Specifically, the pre-training unit is further configured to:

[0095] Based on the SSDD dataset and the ViT network architecture, a semantic feature extraction module suitable for SAR images is trained by using a label-free self-distillation manner, and the module is defined as a DINO-ViT module.

[0096] Specifically, the data processing unit is further configured to:

[0097] The source image I SAR That is, the SAR image is collected, and the SAR image is randomly cropped, horizontally flipped, and color-shaken to generate an image I target_crop The SAR image is randomly cropped, horizontally flipped, and Gaussian blurred to generate an image I background_crop .

[0098] Specifically, the loss function in the model training unit comprises:

[0099] The loss function comprises background space structure constraint, target appearance structure constraint, and key-value detail constraint, and the calculation formula is

[0100] L con =aL stru +bL app +cL key

[0101] Wherein, L con represents the content loss of feature extraction, L con is composed of background space structure loss L stru , target appearance structure loss L app and key-value loss L key , a, b and c are all hyperparameters for balancing the loss.

[0102] More specifically, the background space structure loss L stru uses cosine autocorrelation to calculate the key-value of the multi-head attention mechanism, so that the generated image can retain more target structure information of the SAR image, and is defined as follows:

[0103] L stru =||(S(I SAR )-S(I f ))+(S(I background_crop )-S(I out_2 ))|| F

[0104] where I f , I out_1 respectively represent the fusion image, the attribute transformation image highlighting the background information target_crop The corresponding output, S(I) represents the cosine self-correlation of the calculated image, ||| F represents the calculation of the Frobenius norm.

[0105] More specifically, the target appearance structure loss L app The [CLS] token is extracted from the ViT feature space, and the token is used to constrain the network to retain the target appearance texture of the SAR image, and the target appearance structure loss L app is defined as follows:

[0106] L app =||(T cls (I target_crop )-T cls (I f ))+(T cls (I target_crop )-T cls (I out_1 ))||2

[0107] where T cls (·) represents the extraction of the [CLS] token about the image, and |||2 represents the calculation of the l2 norm.

[0108] More specifically, the key value of the ViT feature space describes some detailed structure information, and the key value loss L key can retain the detailed texture features of the source image, and the key value loss L key is defined as follows:

[0109] L key =(1-λ)||K(I target_crop )-K(I out_1 )| F +λ||K(I SAR )-K(I f )| F

[0110] where K(·) represents the extraction of the key value information about the image, and λ is the balance parameter.

[0111] Specifically, the regression classification loss in the model training unit includes a classification loss function and a regression loss function.

[0112] The VFL Loss function is used to construct the classification loss function.

[0113] The CIOU Loss+DFL function is used to construct the regression loss function.

[0114] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those ordinarily skilled in the art should understand: the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A SAR small sample target detection method, characterized in that, Includes the following steps: Step 1: Train the DINO-ViT module for semantic feature extraction on the ViT network architecture; Step 2: Transfer the source image I SAR For the source image I SAR Image I obtained by performing attribute transformation target_crop And image I background_crop The three components are fed into the Unet3+ network as inputs, forming three branches and yielding the corresponding outputs. , , Feature enhancement and fusion are achieved through parameter sharing; Step 3: Extract semantic information of the enhanced features using the pre-trained DINO-ViT module, construct a loss function, train the Unet3+ network, and simultaneously train the detection network connected to the output of the Unet3+ network using regression classification loss constraints. Training stops when all loss function values ​​are minimized, resulting in the trained model, which is defined as the object detection model. The construction of the loss function includes: The loss function includes background spatial structure constraints, target appearance structure constraints, and key-value detail constraints, and the calculation formula is as follows: in, This represents the content loss from feature extraction. Loss due to background spatial structure Target appearance and structural loss and key-value loss The components a, b, and c are all hyperparameters used to balance the loss; The background space structure loss The key value of the multi-head attention mechanism is calculated using cosine autocorrelation, enabling the generated image to retain more target structure information from the SAR image, as defined below: In the formula , These represent the fused image and the attribute-transformed image that highlights background information, respectively. The corresponding output, This represents the calculation of the cosine autocorrelation of an image. This indicates the calculation of the Frobenius norm; The target appearance structure loss [CLS] labels are extracted from the ViT feature space, and a label-constrained network is used to preserve the target appearance texture of the SAR image, while the target appearance structure loss is applied. The definition is as follows: In the formula, This indicates the extraction of the [CLS] tag related to the image. calculate Norm, Representing an image The corresponding output; The key values ​​in the ViT feature space describe some detailed structural information, which is obtained through key-value loss. The key-value loss can preserve the detailed texture features of the source image. The definition is as follows: In the formula, This indicates the extraction of key-value information about the image. It is the equilibrium parameter; Step 4: Input the real-time acquired images into the Unet3+ network of the trained object detection model. Its three branches generate images with different focuses, which are then sent into the detection network as effective feature layers to obtain the object detection results.

2. The SAR small sample target detection method according to claim 1, characterized in that, Step one includes: Based on the SSDD dataset, a semantic feature extraction module adapted to SAR images is trained on the ViT network architecture using a label-free self-distillation method. This module is defined as the DINO-ViT module.

3. The SAR small sample target detection method according to claim 1, characterized in that, Step two includes: The source image I SAR This refers to the acquired SAR image, which is then randomly cropped, horizontally flipped, and color-dithered to generate image I. target_crop Image I is generated by randomly cropping, horizontally flipping, and Gaussian blurring SAR images. background_crop .

4. The SAR small sample target detection method according to claim 1, characterized in that, In step three, the regression classification loss includes a classification loss function and a regression loss function; Use the VFL Loss function to construct a classification loss function; Use the CIOU Loss+DFL function to construct the regression loss function.

5. A SAR small sample target detection system, characterized in that, include: Pre-trained units are used to train DINO-ViT modules for semantic feature extraction on the ViT network architecture; The data processing unit is used to process the source image I SAR For the source image I SAR Image I obtained by performing attribute transformation target_crop And image I background_crop The three components are fed into the Unet3+ network as inputs, forming three branches and yielding the corresponding outputs. , , Feature enhancement and fusion are achieved through parameter sharing. The model training unit is used to extract semantic information of enhanced features using the pre-trained DINO-ViT module, construct loss functions, train the Unet3+ network, and train the detection network connected to the output of the Unet3+ network through regression classification loss constraints. Training stops when all loss function values ​​are minimized, and the trained model is obtained. This model is defined as the object detection model. The construction loss function includes: The loss function includes background spatial structure constraints, target appearance structure constraints, and key-value detail constraints, and the calculation formula is as follows: in, This represents the content loss from feature extraction. Loss due to background spatial structure Target appearance and structural loss and key-value loss The components a, b, and c are all hyperparameters used to balance the loss; The background space structure loss The key value of the multi-head attention mechanism is calculated using cosine autocorrelation, enabling the generated image to retain more target structure information from the SAR image, as defined below: In the formula , These represent the fused image and the attribute-transformed image that highlights background information, respectively. The corresponding output, This represents the calculation of the cosine autocorrelation of an image. This indicates the calculation of the Frobenius norm; The target appearance structure loss [CLS] labels are extracted from the ViT feature space, and a label-constrained network is used to preserve the target appearance texture of the SAR image, while the target appearance structure loss is applied. The definition is as follows: In the formula, This indicates the extraction of the [CLS] tag related to the image. calculate Norm, Representing an image The corresponding output; The key values ​​in the ViT feature space describe some detailed structural information, which is obtained through key-value loss. The key-value loss can preserve the detailed texture features of the source image. The definition is as follows: In the formula, This indicates the extraction of key-value information about the image. It is the equilibrium parameter; The detection output unit is used to input the real-time acquired image into the Unet3+ network in the trained target detection model. Its three branches generate images with different focuses, which are sent into the detection network as effective feature layers to obtain the target detection result.

6. The SAR small sample target detection system according to claim 5, characterized in that, The pre-training unit is also used for: Based on the SSDD dataset, a semantic feature extraction module adapted to SAR images is trained on the ViT network architecture using a label-free self-distillation method. This module is defined as the DINO-ViT module.

Images