A zero-shot target detection method for SAR images based on domain generalization

Abstract

The invention discloses a SAR image zero-sample target detection method based on domain generalization, which belongs to the field of image processing technology. The method comprises: expanding the SAR image data set by data enhancement; performing multimodal feature encoding according to the enhanced SAR image data set, including: inputting the SAR image and its text category label into the CLIP pre-training model to obtain image features and text category features; obtaining SAR image target features by learnable image encoding based on multi-head self-attention; performing multi-scale domain invariant feature learning on the SAR image features, using gradient matching technology to make the gradient direction of the classification loss of each image consistent; and realizing zero-sample target detection by the similarity between the SAR image target feature and the text category feature. The invention enables the target detection model to have the ability to process basic categories and new category targets of multiple scales under multiple environments.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a zero-sample target detection method for SAR images based on domain generalization. Background Art

[0002] Synthetic Aperture Radar (SAR) images refer to images of surface information obtained by using radar beams through synthetic aperture radar systems. Compared with traditional optical images, SAR images can be effectively collected at different time periods and under various weather conditions; in addition, SAR images can achieve high-resolution and high-quality surface detection. In the military field, images are an important source of information acquisition. The military field requires that high-resolution and high-quality images can be obtained at any time under various weather conditions to ensure the accuracy of information. Therefore, compared with optical images, SAR images are more suitable for application in the military field.

[0003] In the military field, various military targets are usually identified and located based on images to obtain accurate battlefield information. However, traditional target detection methods have the following problems: 1) Traditional target detection methods can only detect targets of known categories, but not unknown targets; 2) Known targets and unknown targets in unknown environments cannot be accurately identified and located; 3) They are insensitive to targets of different scales and are prone to reduce detection accuracy due to scale changes. Therefore, a target detection method that can adapt to various environments and has the ability to identify and locate known and unknown categories of different scales is needed. Summary of the invention

[0004] In order to solve the above-mentioned target detection problems, and at the same time, the target detection model has a reliable image data source, the present invention provides a SAR image zero-sample target detection method based on domain generalization, aiming to achieve zero-sample target detection of SAR images based on domain generalization technology and multimodal technology. The method is based on gradient matching to learn domain invariant features, and at the same time uses the multimodal matching method of the visual-language model to achieve the recognition and positioning of known and unknown targets in a variety of environments, thereby achieving accurate information acquisition of the entire target area.

[0005] The present invention provides a SAR image zero-sample target detection method based on domain generalization, comprising:

[0006] Expanding SAR image datasets through data augmentation;

[0007] Multimodal feature encoding is performed based on the enhanced SAR image dataset, including: inputting the SAR image and its text category label into the CLIP pre-trained model to obtain image features and text category features; obtaining SAR image target features through learnable image encoding based on multi-head self-attention;

[0008] Perform multi-scale domain invariant feature learning on SAR image features, and use gradient matching technology to make the gradient direction of classification loss of each image consistent;

[0009] Zero-shot target detection is achieved through the similarity between SAR image target features and text category features.

[0010] Furthermore, it also includes: expanding the SAR image data set by data enhancement; including:

[0011] Get the original SAR image dataset , where D represents the original data set, D k represents the kth data source, K represents the number of data sources contained in the dataset D, and each data source has a different data distribution; the data in the SAR image dataset exists in pairs (I, T), where I is the image and T is the text category label;

[0012] Data augmentation includes two operations: flipping and rotation;

[0013] Flip any image I in the original data set D to obtain image I f Perform the rotation operation to obtain the graph I r , and set I f and I r The corresponding text category labels are the same as I; the enhanced images and the original images together constitute the enhanced dataset ,in Denotes the dataset D k The corresponding enhanced dataset.

[0014] Furthermore, the multimodal feature encoding includes a text encoding network and an image encoding network based on the CLIP multimodal model; the parameters of the text encoding network and the image encoding network do not participate in the training, and the acquired text category features and image features are used for supervision signals;

[0015] The text encoding includes:

[0016] Text category label T uses CLIP pre-trained text encoding network E t For feature extraction, the formula is as follows:

[0017]

[0018] Where D E represents the total enhanced dataset, represents the kth augmented dataset, express The i-th text category label T in E t represents the CLIP pre-trained text encoding network, Represents the extracted text category features.

[0019] The image encoding comprises:

[0020] Image I uses the image encoding network E pre-trained by CLIP v For feature extraction, the formula is as follows:

[0021]

[0022] Where E v represents the CLIP pre-trained image encoding network, Represents the extracted image features.

[0023] Furthermore, the text encoding network and the image encoding network are both implemented using Transformer, and the Transformer is composed of N repeated Transformer Blocks (B), so the text encoding network and the image encoding network can be expressed as B N , that is, N consecutive B operations; B consists of layer normalization, self-attention, multi-layer perceptron, and layer normalization; the input image x is divided into P image blocks, x represents the image or text category label of the enhanced data set, and B processes it as follows:

[0024]

[0025] Among them, E represents the fully connected layer, E pos represents position encoding; s represents the feature after encoding in the fully connected layer E, s' represents the feature after processing by the multi-head self-attention module, and s'' represents the feature after processing by the perceptron; x cls It is a category code, which is used to integrate the information of the entire input image or text; out It represents the intermediate features of the input x after passing through a B; MLP represents a multi-layer perceptron; LN represents layer normalization; MHSA represents a multi-head self-attention module, which is calculated as follows:

[0026]

[0027] Among them, the normalized entire image or text category features are split into H parts according to the channel dimension, represented as head1~head h , Q h , K h , V h Respectively represent the Query, Key and Value of the h-th head, represents the normalization term; W ois the output transformation matrix, Concat represents the concatenation operation, Attention represents the attention operation, MHSA represents the multi-head self-attention operation, and softmax represents the normalization operation.

[0028] Furthermore, the learnable image encoding includes:

[0029] The learnable image encoding network includes a convolutional network ResNet and a multi-head self-attention module; the convolutional network ResNet includes F convolutional blocks with residual connections connected in series, and the convolutional block operation formula is:

[0030]

[0031] in Represents the enhanced dataset Any image in, Conv represents the convolution operation, f represents the fth convolution layer, c f represents the features extracted by the f-th convolutional layer;

[0032] The learnable image encoding network outputs SAR image target features: C3, C4, C5 and , the calculation formula is as follows:

[0033]

[0034] in Represents the feature maps of different layers of the convolutional network ResNet output, express The result after the global average pooling operation is Represents the output of the multi-head self-attention module MHSA.

[0035] Furthermore, the multi-scale domain invariant feature learning is performed through feature processing, classification processing, regression processing, and gradient matching;

[0036] The SAR image target features output by the learnable image coding network are processed; the feature processing is implemented by the convolutional network ResNet, and the formula is:

[0037]

[0038] in, , is the SAR image target feature output by the learnable image encoding network, f d is the dimension reduction function, Represents output features of 5 different scales;

[0039] The output features of feature processing are classified to obtain target features of different scales; the classification processing is performed through a fully connected layer network Ec To achieve this, the formula is:

[0040]

[0041] in represents the output features of feature processing, Represent target features at different scales;

[0042] The feature maps of different scales are regressed to obtain the position features of the target in the image; the regression process is performed through the fully connected layer network E r To achieve this, the formula is:

[0043]

[0044] in represents the output features of feature processing, Represents the location characteristics of the target in the image;

[0045] The gradient matching includes:

[0046] The gradient of the loss function of aligning data from different data sets is:

[0047]

[0048] Where, L match represents the gradient matching loss, d represents the gradient calculation, d()·d() represents the dot product operation, || represents the modulus operation, K represents the number of data sets, Representation dataset The classification loss corresponding to the data in Representation dataset The regression loss corresponding to the data in .

[0049] Furthermore, the zero-sample target detection includes:

[0050] By calculating the similarity, we can find the target category of each image and get:

[0051] The similarity between the target feature of the a-th scale of the i-th sample in the k-th augmented dataset and the category l for:

[0052]

[0053] Where L represents the number of categories, represents the target feature of the ath scale, t l Represents the text category feature of the lth category.

[0054] Furthermore, it also includes: calculating the loss function to achieve model optimization. The loss function includes 4 parts: image loss Limage , gradient matching loss L match , classification loss L c , regression loss L r , so the overall loss Loss is:

[0055]

[0056] in, is the kth enhanced data set, K is the number of enhanced data sets; SAR image target features extracted by the learnable feature encoding network, is the image feature used for supervision signal, where represents the similarity between the target feature of the a-th scale of the ith sample in the k-th augmented dataset and the category l, is the true label of the corresponding category; Represents the position feature of the ath target feature of the ith sample in the kth augmented dataset, is the true label of the corresponding position.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention proposes a zero-shot target detection method for SAR images based on domain generalization. This method is based on gradient matching to learn domain-invariant features, and uses a multimodal matching method of a visual-language model to achieve recognition and positioning of known and unknown targets in a variety of environments, thereby achieving accurate information acquisition of the entire target area. Its advantages are as follows:

[0059] (1) Compared with traditional optical images, SAR images can be obtained at any time and under any weather conditions. Therefore, target detection based on SAR images makes the input of the target detection method stable, thereby ensuring higher accuracy of the target detection method.

[0060] (2) The present invention uses a visual-language multimodal combination strategy to migrate the CLIP model to the target detection task. By matching the similarity between visual features and text category features, it can realize the recognition and positioning of new targets, which is more in line with the application requirements of military scenarios.

[0061] (3) Data diversity is increased through data augmentation. A pre-trained feature encoding network is used for basic encoding. The gradient matching technology is used to enable the model to learn image features that are applicable to a variety of distributions. This makes the target detection model robust to data distribution and can ensure the stability and accuracy of target detection under actual application conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present disclosure and do not constitute improper limitations on the present invention.

[0063] Figure 1 It is a flow chart of the zero-sample target detection method of SAR images based on domain generalization provided by the present invention.

[0064] Figure 2 It is a modular schematic diagram of the domain generalization-based zero-sample target detection method for SAR images provided by the present invention.

[0065] Figure 3 It is a schematic diagram of the SAR image data enhancement process provided by the present invention.

[0066] Figure 4 It is a schematic diagram of extracting image features and text category features based on a vision-language pre-trained network coding network provided by the present invention.

[0067] Figure 5 It is a schematic diagram of the domain invariant feature learning process based on gradient matching provided by the present invention.

[0068] Figure 6 Schematic diagram of zero-sample target detection provided by the present invention. DETAILED DESCRIPTION

[0069] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0070] A zero-shot target detection method for SAR images based on domain generalization according to an embodiment of the present invention is described below with reference to the accompanying drawings.

[0071] like Figure 1 , 2 As shown, an embodiment of the present invention provides a SAR image zero-sample target detection method based on domain generalization. The method comprises the following steps:

[0072] Step 101, expanding the SAR image data set by data enhancement;

[0073] Step 102, performing multi-modal feature encoding according to the enhanced SAR image data set;

[0074] Specifically, the SAR image and its text category label are input into the CLIP pre-trained network to obtain image features and text category features; the SAR image target features are obtained through learnable image encoding based on multi-head self-attention;

[0075] Step 103, performing multi-scale domain invariant feature learning on the SAR image features, and using gradient matching technology to make the gradient direction of the classification loss of each image consistent;

[0076] Step 104: Implement zero-sample target detection through the similarity between the SAR image target features and the text category features.

[0077] This embodiment uses a data enhancement module to supplement the quantity and diversity of training data, thereby increasing the robustness of the target detection model; a feature encoding module based on vision-language realizes the encoding of basic category and new category target images and corresponding texts; a domain-invariant feature learning module based on gradient matching learns common features across data distributions, so that the target detection model has the ability to handle new data distributions; and enables target detection to identify and locate basic category and new category targets in images. The present invention utilizes the powerful representation capabilities of the vision-language pre-training model and learns domain-invariant features based on gradient matching, so that the target detection model has the ability to handle basic category and new category targets of multiple scales in multiple environments.

[0078] Optionally, as an embodiment of the present invention, the data enhancement method is specifically as follows.

[0079] Compared with optical images, SAR images have the advantage that they are not easily affected by weather conditions and time points when they are acquired. Therefore, the present invention provides a target detection method based on SAR images. In order to enhance the magnitude and data diversity of SAR image datasets, thereby further increasing the robustness of the target detection method, the present invention adopts the following method: Figure 3 The data augmentation process shown is for the original dataset Perform data enhancement, where D represents the original data set, D k represents the kth data source, K represents the number of data sources contained in the data set D, and each data source has a different data distribution.

[0080] Data augmentation includes two operations: flipping and rotation. For any image I in the dataset D, the operations are as follows:

[0081] I f = Flip(I)

[0082] I r = Rotate(I)

[0083] In the formula is any image in the original dataset D, I f is the image after the flip operation of image I, I r For images The image after the rotation operation, If and I r The corresponding text category label is the same as I. The enhanced images and the original images together constitute the enhanced dataset ,in Denotes the dataset D k The corresponding enhanced dataset.

[0084] In this embodiment, for each image, a new related image is generated by image flipping and image rotation, which is used to expand the original SAR image data set, thereby increasing the diversity of the data and facilitating the generalization of the model.

[0085] Optionally, as an embodiment of the present invention, the multimodal feature encoding method is specifically as follows.

[0086] Utilizing multimodal information is the key to achieving zero-shot target detection. The present invention provides an improved feature encoding method based on the visual-language pre-trained network CLIP, which migrates the CLIP multimodal network to the target detection task and can achieve zero-shot target detection with a high accuracy. Figure 4 As shown, the encoding process includes the following steps:

[0087] (1) Text encoding

[0088] The data in the dataset exists in pairs of images and labels (I, T), where I is the image and T is the text category label, such as ship, airplane, etc. For the text T, the CLIP pre-trained text encoding network is used for feature extraction, and the network parameters are not involved in the object detection model training. The feature extraction process is as follows:

[0089]

[0090] Where D E represents the total enhanced dataset, represents the kth augmented dataset, express The i-th text category label T in E t represents the CLIP pre-trained text encoding network, Represents the extracted text category features.

[0091] (2) Image Coding

[0092] For image I, the CLIP pre-trained image encoding network is used for feature extraction. The network parameters do not participate in the object detection model training. The feature extraction process is as follows:

[0093]

[0094] In the formula, D Erepresents the total enhanced dataset, represents the kth augmented dataset.

[0095] Text Encoding Network t and image encoding network E v Both are implemented using Transformer. Transformer consists of N repeated Transformer Blocks (B), so the text encoding network and image encoding network can be represented as B N , that is, N consecutive B operations; B consists of layer normalization, self-attention, multi-layer perceptron, and layer normalization; the input image x is divided into P image blocks, x represents the image or text category label of the enhanced data set, and B processes it as follows:

[0096]

[0097] Among them, E represents the fully connected layer, E pos represents position encoding; s represents the feature after encoding in the fully connected layer E, s' represents the feature after processing by the multi-head self-attention module, and s'' represents the feature after processing by the perceptron; x cls It is a category code, which is used to integrate the information of the entire input image or text; out It represents the intermediate features of the input x after passing through a B; MLP represents a multi-layer perceptron; LN represents layer normalization; MHSA represents a multi-head self-attention module, which is calculated as follows:

[0098]

[0099] The entire image or text category feature LN(s) is split into H parts according to the channel dimension, represented as head1~head h , Q h , K h , V h Respectively represent the Query, Key and Value of the h-th head, represents the normalization term; W o is the output transformation matrix, Concat represents concatenation operation, Attention represents attention operation, MHSA represents multi-head self-attention operation, and softmax represents normalization operation (3) Learnable image encoding

[0100] The encoding network parameters of the first two steps do not participate in the training, and the acquired text category features and image features are used as supervisory signals, i.e., labels. In order to achieve target detection, the present invention provides a learnable image encoding network, and the learned features can be used in the subsequent target detection process.

[0101] The learnable image encoding network includes a convolutional network ResNet and a multi-head self-attention module; the convolutional network ResNet includes F convolutional blocks with residual connections connected in series, and the convolutional block operation formula is:

[0102]

[0103] in Represents the enhanced dataset Any image in, Conv represents the convolution operation, f represents the fth convolution layer, c f Represents the features extracted by the fth convolutional layer

[0104] like Figure 4 As shown, the output of the learnable image encoding network consists of 5 parts: C3, C4, C5 and , the calculation formula is as follows:

[0105]

[0106] in Represents the feature maps of different layers of the convolutional network ResNet output, express The result after the global average pooling operation is Represents the output of the multi-head self-attention module MHSA, and the multi-head self-attention module refers to the above.

[0107] The multimodal feature encoding provided in this embodiment. The image encoding network and the text encoding network of the CLIP pre-trained network (comparative language-vision pre-trained network) are respectively used for SAR images and their text category labels to obtain image features and text category features. Since the CLIP network is pre-trained on large-scale data, the encoded image features and text category features have rich information and can be migrated to SAR image target detection.

[0108] Optionally, as an embodiment of the present invention, this embodiment provides a domain-invariant feature learning module based on gradient matching, which is intended to learn SAR image data with different distributions, that is, SAR images acquired in different environments. Figure 5 As shown in Figure 2, the process of learning multi-scale domain invariant features includes the following steps:

[0109] (1) Feature processing The feature processing module FPN is implemented by the convolutional network ResNet, and the formula is:

[0110]

[0111] in, , is the SAR image target feature output by the learnable image encoding network, f d is the dimension reduction function, Represents output features of 5 different scales;

[0112] (2) Classification processing Classification processing is performed by the fully connected layer network E c To achieve this, the formula is:

[0113]

[0114] in represents the output features of feature processing, Represent target features at different scales;

[0115] (3) Regression processing The regression processing is performed through a fully connected layer network E r To achieve this, the formula is:

[0116]

[0117] in represents the output features of feature processing, Represents the location characteristics of the target in the image;

[0118] (4) Gradient matching

[0119] like Figure 5 As shown, from different data sets The features obtained from the data After feature processing, classification and regression modules, the loss is calculated with the true label and .in and Represents the data set The classification loss and regression loss functions corresponding to the data in the example will be described in detail in the following embodiments. By aligning the gradients of the loss functions of different data sets through the following formula, the network parameters of the feature processing, classification and regression modules can be regularized, thereby realizing the learning of domain invariant features.

[0120]

[0121] Where, L match represents the gradient matching loss, d represents the gradient calculation, d()·d() represents the dot product operation, || represents the modulus operation, K represents the number of data sets, Representation dataset The classification loss corresponding to the data in Representation dataset The regression loss corresponding to the data in .

[0122] The multi-scale domain invariant feature learning method provided in this embodiment needs to perform further feature processing to realize multi-scale target recognition when the features obtained by multi-modal feature coding are transferred to SAR image target detection, and the target is identified and located by learning the classification processing and regression processing parts. For the classification and regression processing parts, images containing the same target from different data sets are used as input, and the gradient matching technology is used to make the gradient direction of the classification loss of each image consistent, so that the target detection model learns domain invariant features across data distribution and enhances the robustness of the classification model.

[0123] Optionally, as an embodiment of the present invention, zero-sample target detection includes:

[0124] like Figure 6 As shown in the figure, similarity is calculated based on the multi-scale domain invariant image features and text category features calculated in the previous steps. Through the similarity score, the target category contained in each image can be obtained. The position of the target predicted by the regression processing process can complete the detection of targets of known and unknown categories at multiple scales in any image.

[0125] By calculating the similarity, we can find the target category of each image and get:

[0126] The similarity between the target feature of the a-th scale of the i-th sample in the k-th augmented dataset and the category l for:

[0127]

[0128] Where L represents the number of categories, represents the target feature of the ath scale, t l Represents the text category feature of the lth category.

[0129] After steps 101-104, all objects in any image can be detected. The optimization of the object detection model requires the calculation of the loss function. The loss function of the method proposed in the present invention includes four parts: image loss L image , gradient matching loss L match , classification loss L c , regression loss L r , so the overall loss Loss is:

[0130]

[0131] in, is the kth enhanced data set, K is the number of enhanced data sets; SAR image target features extracted by the learnable feature encoding network, is the image feature used for supervision signal, where represents the similarity between the target feature of the a-th scale of the ith sample in the k-th augmented dataset and the category l, is the true label of the corresponding category; Represents the position feature of the ath target feature of the ith sample in the kth augmented dataset, is the true label of the corresponding position.

[0132] In this embodiment, based on the previously learned image features and text category features, the target category in the image is identified by calculating the similarity of the image features and text category features, and the error between the target position predicted by the regression module and the true label is calculated, and the error is optimized using the above overall loss Loss. The optimized target detection model can realize the recognition and positioning of basic category and new category targets.

Claims

1. A zero-shot target detection method for SAR images based on domain generalization, characterized in that: include: Expanding SAR image datasets through data augmentation; Multimodal feature encoding is performed based on the enhanced SAR image dataset, including: inputting the SAR image and its text category label into the CLIP pre-trained model to obtain image features and text category features; obtaining SAR image target features through learnable image encoding based on multi-head self-attention; Perform multi-scale domain invariant feature learning on SAR image features, and use gradient matching technology to make the gradient direction of classification loss of each image consistent; Zero-shot target detection is achieved through the similarity between SAR image target features and text category features; The multimodal feature encoding includes a text encoding network and an image encoding network based on the CLIP multimodal model; the parameters of the text encoding network and the image encoding network do not participate in the training, and the acquired text category features and image features are used for supervision signals; The text encoding includes: Text category label T uses CLIP pre-trained text encoding network E t For feature extraction, the formula is as follows: ； Where D E represents the total enhancement data, represents the kth augmented dataset, express The i-th text category label T in E t represents the CLIP pre-trained text encoding network, Represents the extracted text category features; The image encoding comprises: Image I uses the image encoding network E pre-trained by CLIP v For feature extraction, the formula is as follows: ； Where E v represents the CLIP pre-trained image encoding network, represents the extracted image features; The text encoding network and the image encoding network are both implemented using Transformer, which is composed of N repeated Transformer Blocks (B). Therefore, the text encoding network and the image encoding network can be expressed as B N , that is, N consecutive B operations; B consists of layer normalization, self-attention, multi-layer perceptron, and layer normalization; the input image x is divided into P image blocks, x represents the image or text category label of the enhanced data set, and B processes it as follows: ； Among them, E represents the fully connected layer, E pos represents position encoding; s represents the feature after encoding in the fully connected layer E, s' represents the feature after processing by the multi-head self-attention module, and s'' represents the feature after processing by the perceptron; x cls It is a category code, which is used to integrate the information of the entire input image or text; out It represents the intermediate features of the input x after passing through a B; MLP represents a multi-layer perceptron; LN represents layer normalization; MHSA represents a multi-head self-attention module, which is calculated as follows: ； Among them, the normalized entire image or text category features are split into H parts according to the channel dimension, represented as head1~head h , Q h , K h , V h Respectively represent the Query, Key and Value of the h-th head, represents the normalization term; W o is the output transformation matrix, Concat represents the concatenation operation, Attention represents the attention operation, MHSA represents the multi-head self-attention operation, and softmax represents the normalization operation.

2. The method for zero-shot target detection in SAR images based on domain generalization as claimed in claim 1, characterized in that: Also includes: Expanding SAR image datasets through data augmentation; including: Get the original SAR image dataset , where D represents the original data set, D k represents the kth data source, K represents the number of data sources contained in the dataset D, and each data source has a different data distribution; the data in the SAR image dataset exists in pairs (I, T), where I is the image and T is the text category label; Data augmentation includes two operations: flipping and rotation; Flip any image I in the original data set D to obtain image I f Perform the rotation operation to obtain the graph I r , and set I f and I r The corresponding text category labels are the same as I; the enhanced images and the original images together constitute the enhanced dataset ,in Denotes the dataset D k The corresponding augmented dataset.

3. The SAR image zero-shot target detection method based on domain generalization as claimed in claim 1, characterized in that: The learnable image encoding includes: The learnable image encoding network includes a convolutional network ResNet and a multi-head self-attention module; the convolutional network ResNet includes F convolutional blocks with residual connections connected in series, and the convolutional block operation formula is: ； in Represents the enhanced dataset Any image in, Conv represents the convolution operation, f represents the fth convolution layer, c f represents the features extracted by the f-th convolutional layer; The learnable image encoding network outputs SAR image target features: C3, C4, C5 and , the calculation formula is as follows: ； in Represents the feature maps of different layers of the convolutional network ResNet output, express The result after the global average pooling operation is Represents the output of the multi-head self-attention module MHSA.

4. The SAR image zero-shot target detection method based on domain generalization as claimed in claim 1, characterized in that: The multi-scale domain invariant feature learning is performed through feature processing, classification processing, regression processing, and gradient matching; The SAR image target features output by the learnable image coding network are processed; the feature processing is implemented by the convolutional network ResNet, and the formula is: ； in, , is the SAR image target feature output by the learnable image encoding network, f d is the dimension reduction function, Represents output features of 5 different scales; The output features of feature processing are classified to obtain target features of different scales; the classification processing is performed through a fully connected layer network E c To achieve this, the formula is: ； in represents the output features of feature processing, Represent target features at different scales; The feature maps of different scales are regressed to obtain the position features of the target in the image; the regression process is performed through the fully connected layer network E r To achieve this, the formula is: ； in represents the output features of feature processing, Represents the location characteristics of the target in the image; The gradient matching includes: The gradient of the loss function of aligning data from different data sets is: ； Where, L match represents the gradient matching loss, d represents the gradient calculation, d()·d() represents the dot product operation, || represents the modulus operation, K represents the number of data sets, Representation dataset The classification loss corresponding to the data in Representation dataset The regression loss corresponding to the data in .

5. The method for zero-shot target detection in SAR images based on domain generalization as claimed in claim 1, characterized in that: The zero-sample target detection includes: By calculating the similarity, we can find the target category of each image and get: The similarity between the target feature of the a-th scale of the i-th sample in the k-th augmented dataset and the category l for: ； Where L represents the number of categories, represents the target feature of the ath scale, t l Represents the text category feature of the lth category.

6. The SAR image zero-shot target detection method based on domain generalization as claimed in claim 1, characterized in that: Also includes: Calculate the loss function to optimize the model. The loss function consists of four parts: image loss L image , gradient matching loss L match , classification loss L c , regression loss L r , so the overall loss Loss is: ； in, is the kth enhanced data set, K is the number of enhanced data sets; SAR image target features extracted by the learnable feature encoding network, is the image feature used for supervision signal, where represents the similarity between the target feature of the a-th scale of the ith sample in the k-th augmented dataset and the category l, is the true label of the corresponding category; Represents the position feature of the ath target feature of the ith sample in the kth augmented dataset, is the true label of the corresponding position.

Images