A SAR target detection method and device

Abstract

This application discloses a SAR target detection method and device. SAR and HRRP sample data are normalized and mapped to the same feature space through their respective encoders. The unlabeled original SAR image is uniformly sliced ​​into blocks. Unsupervised learning is performed on the sliced ​​areas to obtain a SAR full-image variational autoencoder model. The quantized embedding dictionary is optimized to focus some entries on target entries. Multi-task joint training is performed using contrastive learning, reconstruction learning, and triple learning to obtain model training weights. Finally, the original SAR image to be detected is uniformly sliced ​​into blocks and input into a variational autoencoder. The image blocks mapped to the quantized word embedding dictionary of the target category are output to obtain the potential target location. The variational autoencoder quantizes the redundant SAR image pixel-level information into a discrete feature space to obtain the common low-dimensional target features, reducing the interference of the target scene on the detection algorithm and improving the generalization ability of target detection.

Description

Technical Field

[0001] The present application relates to the technical field of signal image fusion recognition processing, and in particular to a SAR target detection method and device. Background Art

[0002] Synthetic Aperture Radar (SAR) images offer all-day, all-weather, and high-resolution capabilities. By actively transmitting and receiving microwaves in specific frequency bands, SAR imaging technology possesses a certain degree of penetrating power, ensuring stable and continuous ground observation while also enabling the continuous detection of concealed targets. With the advancement of deep learning technology, neural network-based computer vision target detection technology has flourished. However, in the field of SAR imagery, the mainstream approach to supervised target detection based on deep learning still draws heavily on classic networks in the field of computer vision, relying on manual annotation of large numbers of SAR images in the same statistical distribution space to improve target detection results.

[0003] Among existing technologies, supervised SAR target detection solutions mainly fall into two categories: one is to expand the number of samples, using data augmentation methods from the field of computer vision, such as scaling, cropping, and horizontal flipping, to expand target data, but the transformation combinations are relatively limited. If a reinforcement learning strategy is used to obtain more complex combinations of enhanced data, the augmented objects targeted by reinforcement learning are still pixel-level target features. Therefore, the algorithm has a huge search space and high computational cost, and cannot extract the commonalities of targets in different scenes. Instead, a physical simulation model is used to generate target simulated image samples with diverse types for specific scenes and targets. However, there are still differences in style and data feature spatial distribution between simulated samples and real samples. Therefore, simply mixing simulated samples with real samples as the augmented sample set will only improve the performance of the algorithm model to a limited extent.

[0004] The other approach is to pre-train the model using a large amount of data from different background distributions, and then fine-tune the model using SAR image data of the target scene. However, when the distribution of the fine-tuning scene data differs significantly from that of the pre-training scene, such as when different imaging data sources are used: optical, SAR, infrared, or when different data scenes of the same imaging data source are switched: mountains, forests, deserts, plains, farmland, cities, etc. Limited by the distribution differences caused by redundant pixel information in the image data itself, even models pre-trained using a large amount of data cannot be universally applied to different types of scenes. This is because existing methods have technical issues such as high annotation costs for SAR full-image target detection and poor adaptability to scene changes. This will become one of the many problems that those skilled in the art need to solve. Summary of the Invention

[0005] The purpose of this application is to provide a SAR target detection method and device in order to overcome the existing technical defects, which can solve the technical problems that the method of SAR full-image target detection has high annotation cost and poor adaptability of detection effect to scene changes.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] In a first aspect, the present application proposes a SAR target detection method, the method comprising:

[0008] S1, normalize and map the a priori aligned SAR sample data and HRRP sample data to the same feature space through their respective SAR encoders and HRRP encoders;

[0009] S2. The unlabeled original SAR image is evenly cut into blocks. A SAR full-image variational autoencoder model based on unsupervised learning is constructed for the cut-block areas using a SAR encoder, a quantized embedding dictionary, and a SAR image decoder.

[0010] S3. Encode the SAR sample and / or HRRP sample through their respective SAR encoders and / or HRRP encoders and input them into the SAR full-image variational autoencoder model, and optimize the quantized embedding dictionary so that some entries are focused on target entries, wherein the SAR sample includes SAR sample data and the HRRP sample includes HRRP sample data;

[0011] S4, using the contrast learning of S1, the reconstruction learning of S2, and the triplet learning of S3 to perform multi-task joint training to obtain model training weights, and training the SAR full-image variational autoencoder model to obtain a multi-task joint trained variational autoencoder;

[0012] S5. The original SAR image to be detected is uniformly cut into blocks and input into the variational autoencoder trained with the multi-task joint method, and the image blocks mapped to the quantized word embedding dictionary of the target category are output to obtain the potential target position.

[0013] In an optional embodiment, the step of normalizing and mapping the a priori aligned SAR sample data and HRRP sample data to the same feature space through respective SAR encoders and HRRP encoders includes:

[0014] Normalizing the SAR sample data to an N*N size by interpolation scaling;

[0015] The HRRP sample data is normalized to 1*M size by truncation and zero padding;

[0016] Inputting the normalized SAR sample data and HRRP sample data into respective SAR encoders and HRRP encoders to obtain first initial coding features;

[0017] The first initial coding features are normalized and the cosine similarity loss is calculated to obtain the weights of the SAR encoder and the HRRP encoder.

[0018] In an optional embodiment, the step of uniformly slicing the unlabeled original SAR image into blocks, and performing unsupervised learning on the slicing areas using a SAR encoder, a quantized embedding dictionary, and a SAR image decoder to obtain a SAR full-image variational autoencoder model includes:

[0019] The unlabeled original SAR image is uniformly cut into blocks to obtain the first sub-image sequence;

[0020] Inputting the first sub-image sequence into the SAR encoder to extract initial coding features to obtain second initial coding features;

[0021] discretize the second initial coding feature using a quantum word embedding table in a quantized embedding dictionary, query the mapping to obtain a mapped second initial coding feature;

[0022] Constructing a SAR image decoder to perform image block pixel restoration on the mapped second initial coding feature to obtain a restored second initial coding feature;

[0023] The first sub-image sequence is compared with the restored second initial coding features, and the second initial coding features are compared with the mapped second initial coding features to obtain a SAR full-image variational autoencoder model.

[0024] In an optional embodiment, the step of discretizing the second initial coding feature using a quantum word embedding table in a quantized embedding dictionary and querying the mapping to obtain the mapped second initial coding feature includes:

[0025] Keeping the word vector dimension in the quantized word embedding table consistent with the dimension of the second initial encoding feature;

[0026] The second initial coding feature is replaced with the discrete quantized vector in the quantized word embedding table by looking up the table, and query mapping is performed to obtain the mapped second initial coding feature.

[0027] In an optional embodiment, the step of triple learning in S3 is:

[0028] Sample a sample x in the database and obtain its quantized encoding feature e(x);

[0029] Select a sample x from the database that belongs to the same category as x p , select samples x that do not belong to the same category as x n , and x p and x n Quantize and encode the three elements e(x), e(x p )、e(x n ) forms a triple <e(x),e(x p ), e(x n )>.

[0030] In an optional embodiment, the steps of uniformly slicing the original SAR image to be detected and inputting the image into the multi-task jointly trained variational autoencoder, outputting the image blocks mapped to the quantized word embedding dictionary of the target category, and obtaining the potential target location include:

[0031] The original SAR image to be detected is evenly cut into blocks to obtain a second sub-image sequence;

[0032] Inputting the second sub-image sequence into a SAR encoder to extract initial coding features to obtain third initial coding features;

[0033] Use the target word in the quantum word embedding table to query and map the third initial coding feature to obtain the mapped third initial coding feature;

[0034] The image block output that matches the quantized term of the target category term in the mapped third initial coding feature is filtered to obtain the potential target position.

[0035] In a second aspect, the present application further proposes a SAR target detection device, comprising:

[0036] The contrastive learning module is used to normalize and map the a priori aligned SAR sample data and HRRP sample data into the same feature space through their respective SAR encoders and HRRP encoders;

[0037] The reconstruction learning module is used to uniformly slice the unlabeled original SAR image into blocks. The unsupervised learning of the block areas is performed using a SAR encoder, a quantized embedding dictionary, and a SAR image decoder to obtain a SAR full-image variational autoencoder model.

[0038] a triplet learning module for encoding SAR samples and / or HRRP samples through respective SAR encoders and / or HRRP encoders and inputting the encoded data into the SAR full-image variational autoencoder model, and optimizing the quantized embedding dictionary so that some entries are focused on target entries, wherein the SAR samples include SAR sample data and the HRRP samples include HRRP sample data;

[0039] A training module is used to perform multi-task joint training using contrast learning of S1, reconstruction learning of S2, and triple learning of S3 to obtain model training weights, and train the SAR full-image variational autoencoder model to obtain a multi-task joint trained variational autoencoder;

[0040] The output module is used to uniformly slice the original SAR image to be detected into blocks and input the blocks into the variational autoencoder trained with the multi-task joint method, output the image blocks mapped to the quantized word embedding dictionary of the target category, and obtain the potential target position.

[0041] In a third aspect, the present application further proposes a computer device, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the SAR target detection method as described in any one of the first aspects.

[0042] In a fourth aspect, the present application further proposes a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is loaded and executed by a processor to implement the SAR target detection method as described in any one of the first aspects.

[0043] The above-mentioned main scheme of this application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and protected by this application; and in this application, (non-conflicting options) can also be freely combined with each other and with other options. After understanding the scheme of this application, those skilled in the art will understand that there are many combinations based on existing technology and common knowledge, all of which are technical solutions to be protected by this application, and they are not exhaustive here.

[0044] This application discloses a SAR target detection method and device. SAR and HRRP sample data are normalized and mapped to the same feature space through their respective encoders. The unlabeled original SAR image is uniformly sliced ​​into blocks. Unsupervised learning is performed on the sliced ​​areas to obtain a SAR full-image variational autoencoder model. The quantized embedding dictionary is optimized to focus some entries on target entries. Multi-task joint training is performed using contrastive learning, reconstruction learning, and triple learning to obtain model training weights. Finally, the original SAR image to be detected is uniformly sliced ​​into blocks and input into a variational autoencoder. The image blocks mapped to the quantized word embedding dictionary of the target category are output to obtain the potential target location. The variational autoencoder quantizes the redundant SAR image pixel-level information into a discrete feature space to obtain the common low-dimensional target features, reducing the interference of the target scene on the detection algorithm and improving the generalization ability of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of the SAR target detection method proposed in an embodiment of the present application is shown.

[0046] Figure 2 The figure shows the algorithm training flow chart of the SAR target detection method proposed in the embodiment of the present application.

[0047] Figure 3 The flowchart of the algorithm test of SAR target image proposed in the embodiment of the present application is shown. DETAILED DESCRIPTION

[0048] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0049] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0050] In existing technologies, due to differences in style and spatial distribution of data features between simulated and real samples, simply mixing simulated and real samples as the augmented sample set only improves algorithm model performance to a limited extent. Furthermore, due to the distribution differences caused by redundant pixel information in image data itself, even models pre-trained with large amounts of data still cannot be universally applied to different scenarios.

[0051] To this end, this application proposes a SAR target detection method, which is based on the knowledge-assisted principle. First, the idea of ​​unsupervised quantized variational autoencoder is adopted to quantize the redundant SAR image pixel-level information processing into a discrete feature space to obtain the commonality of low-dimensional target features, reduce the interference of the target scene on the detection algorithm, and improve the generalization ability of target detection. At the same time, based on the ideas of reference contrast learning and triple learning, the existing SAR target samples of other scenes and the radar high-resolution range image (HRRP) knowledge information are uniformly distributed in the space mapping, and the quantized word embedding table is used to assist in target clustering. In the inference stage, the target detection of the entire image is completed by cutting the entire SAR image to be detected through a sliding window and matching and mapping it with the quantized dictionary of known categories. It is described in detail below.

[0052] Please refer to Figure 1 , Figure 1 The flowchart of the SAR target detection method proposed in the embodiment of the present application is shown. The SAR target detection method is a method for performing SAR satellite full image target detection using known SAR image samples or HRRP signal fragment knowledge. The method is applied to synthetic aperture radar. The specific steps are as follows:

[0053] S1. The a priori aligned SAR sample data and HRRP sample data are normalized and mapped to the same feature space through their respective SAR encoders and HRRP encoders.

[0054] First, comparative learning can determine the encoders of the two sources, which remain fixed. The comparative learning of step S1 is: SAR sample data obtains a feature through the SAR encoder, and HRRP sample data obtains another feature through the HRRP encoder. Since the input SAR sample data and HRRP sample data are paired, the two mapping features can be constrained to be infinitely close. By inputting the paired data, the feature of the SAR encoder and the other feature of the HRRP encoder can be forced to be infinitely close. SAR sample data (two-dimensional) and HRRP sample data (one-dimensional) are expressions of the same potential target in different feature spaces, and have the premise of mapping to the same feature space for unified expression. The mapping method is completed using the text-image unified mapping algorithm CLIP idea. The specific steps of using comparative learning technology to uniformly map the prior-aligned SAR sample data and HRRP sample data into a unified distribution space are as follows:

[0055] First, the data pairs (x SAR ,x HRRP ), the data pair (x SAR ,x HRRP) are normalized to the same size, interpolation scaling is used to normalize the SAR sample data to N*N size, and truncation and zero padding are used to normalize the HRRP sample data to 1*M size, and two independent encoders are constructed. SAR (), Encoder HRRP (), input the normalized SAR sample data and HRRP sample data into their respective encoders SAR (), Encoder HRRP () Get the first initial encoding feature Encoder SAR (x SAR ), Encoder HRRP (x HRRP ), at this time, the feature dimension of the first initial coding feature is 1×n (n is much smaller than N and M), and this step has the effect of plane space dimensionality reduction.

[0056] Then the first initial encoding feature Encoder SAR (x SAR ), Encoder HRRP (x HRRP ) performs L2 normalization processing and calculates the SAR and HRRP features Gl2 (Encoder SAR (x SAR ))、Gl2(Encoder HRRP (x HRRP ))'s cosine similarity loss can be used to obtain the SAR encoder Encoder through network learning SAR () and HRRP signal encoder Encoder HRRP ().

[0057] S2. The unlabeled original SAR image is evenly divided into blocks. A SAR full-image variational autoencoder model based on unsupervised learning is constructed for the block areas through a SAR encoder, a quantized embedding dictionary, and a SAR image decoder.

[0058] First, a quantized variational autoencoder (VAE) is constructed. This VE is used to uniformly slice the existing, large, and unlabeled raw SAR images. Unsupervised learning is then performed on the sliced ​​regions to obtain a SAR full-image VEE model. The SAR full-image VEE model includes an encoder and decoder for the SAR sliced ​​images, as well as a quantized word embedding dictionary. The encoder for the SAR sliced ​​images remains consistent with the SAR sample data encoder in S1.

[0059] S2 specifically includes the following steps:

[0060] The sliding window technique is used to uniformly cut the unlabeled original SAR image into blocks to obtain the first sub-image sequence (I SAR1 ,I SAR2 ...I SARn );

[0061] Input the first sub-image sequence to the SAR encoder Encoder SAR () Extract the initial encoding feature to obtain the second initial encoding feature (Encoder SAR (I SAR1 ),Encoder SAR (I SAR2 )...Encoder SAR (I SARn ));

[0062] In order to enable some vectors in the quantized word embedding table to have the ability to classify, constrained learning of the discrete quantized vector word embedding table is added to the variational autoencoder model. In addition, due to the high redundancy of image information, in order to obtain the common potential information of the target besides background noise, the quantized variational autoencoder DQ-VAE mechanism is introduced to discretize the initial encoding features. After the discretization, query mapping is required. The steps are as follows:

[0063] Keep the word vector dimension in the quantized word embedding table consistent with the dimension of the second initial encoding feature;

[0064] The second initial coding feature is replaced by the discrete quantized vector in the quantized word embedding table by looking up the table to obtain the mapped second initial coding feature.

[0065] Construct the quantized word embedding table E to ensure that the word vector dimension is consistent with the second initial encoding feature Encoder SAR (I SARi ) dimension is consistent, the number of words is a manually set hyperparameter, and the second initial encoding feature Encoder is converted into SAR (I SARi ) is replaced by the closest discrete quantized vector e(I SARi ), query mapping to obtain the mapped second initial coding feature (e(I SAR1 ),e(I SAR2 )...e(I SARn )) and construct a SAR image decoder Decoder() to restore the image block pixels of the quantized image sequence coding features to obtain the restored second initial coding features (Decoder(e(I SAR1 )),Decoder(e(I SAR2 ))...Decoder(e(ISARn ))), compare the first sub-image sequence (I SAR1 ,I SAR2 ...I SARn ) and the restored second initial coding feature (Decoder(e(I SAR1 )),Decoder(e(I SAR2 ))...Decoder(e(I SARn ))), the second initial encoding feature (Encoder SAR (I SAR1 ),Encoder SAR (I SAR2 )...Encoder SAR (I SARn )) and the mapped second initial coding feature (e(I SAR1 ),e(I SAR2 )...e(I SARn ))A SAR full-image variational autoencoder model is obtained to realize unsupervised autoencoding.

[0066] It is worth noting that the SAR encoder in step S2 and the encoder in step S1 SAR (), Encoder HRRP ()Weight sharing.

[0067] S3. Encode the SAR samples and / or HRRP samples through their respective SAR encoders and / or HRRP encoders and then input them into the SAR full-image variational autoencoder model. Use triplet learning to optimize the quantized embedding dictionary so that some entries are focused on target entries.

[0068] SAR samples include SAR sample data, and HRRP samples include HRRP sample data. S1 constructs the encoder Encoder SAR (), Encoder HRRP () Encode the first initial encoding feature after encoding the SAR sample and HRRP segment SAR (x SAR ), Encoder HRRP (x HRRP ) are mapped to the same feature space, S3 no longer distinguishes the original data modality, but considers the encoded features in a unified feature space.

[0069] The triplet is mainly composed of inputting SAR samples or HRRP samples of the same category, as well as SAR or HRRP samples of different categories, to construct a sample triplet. <e(x),e(x p ), e(x n )>,x∈XSAR ∪X HRRP Data pairs, perform spatial clustering on the quantum word embedding table. The steps of triple learning are: randomly sample a sample x in the database and obtain its quantized encoding feature e(x); select samples x from the database that belong to the same category as x p , select samples x that do not belong to the same category as x n , the three elements e(x), e(x p )、e(x n ) forms a triple <e(x),e(x p ), e(x n )>.

[0070] S4, using the contrast learning of S1, reconstruction learning of S2 and triple learning of S3 to perform multi-task joint training to obtain the model training weights, and train the SAR full-image variational autoencoder model to obtain a multi-task joint training variational autoencoder.

[0071] Contrastive learning, reconstruction learning, and triplet learning mainly perform multi-task joint training on the algorithm structures in S1, S2, and S3 to obtain the final training weight of the model.

[0072] Contrastive learning is the first initial encoding feature Encoder that aligns the SAR and HRRP priori in S1 SAR (x SAR ), Encoder HRRP (x HRRP ) CLIP algorithm process mapped to the same feature space;

[0073] Reconstruction learning refers to the process of comparing the first sub-image sequence (I SAR1 ,I SAR2 ...I SARn ) and the restored second initial coding feature (Decoder(e(I SAR1 )),Decoder(e(I SAR2 ))...Decoder(e(I SARn ))), the second initial encoding feature (Encoder SAR (I SAR1 ),Encoder SAR (I SAR2 )...Encoder SAR (I SARn )) and the mapped second initial coding feature (e(I SAR1 ),e(I SAR2 )...e(I SARn)) Obtain the SAR full-image variational autoencoder model to implement the VQ-VAE algorithm process of unsupervised autoencoding;

[0074] Triplet learning refers to the triplet data pairs after quantization encoding in S3 <e(x),e(x p ), e(x n )> Triplet Loss algorithm learning, so that e(x) and e(x p ) as small as possible, e(x) and e(x n ) as large as possible.

[0075] S5. The original SAR image to be detected is evenly divided into blocks and input into a variational autoencoder trained with multiple tasks. The image blocks mapped to the quantized word embedding dictionary of the target category are output to obtain the potential target position.

[0076] The purpose of S1 to S4 is to learn the model weights. S5 is the inference stage, which aims to realize the function of target detection and recognition on unlabeled SAR original large images by using SAR image samples or HRRP signal samples of known modes through the idea of ​​comparative retrieval.

[0077] Step S5 specifically includes the following steps:

[0078] Slicing the original SAR image to be detected to obtain a second sub-image sequence;

[0079] Inputting the second sub-image sequence into the SAR encoder to extract initial coding features to obtain third initial coding features;

[0080] Using the quantum word embedding table to query and map the third initial coding feature to obtain the mapped third initial coding feature;

[0081] The image block output that matches the quantized term of the target category term in the mapped third initial coding feature is filtered to obtain the potential target position.

[0082] Only a portion of the words in the quantum word embedding table are related to the target. The target-related words are obtained through triple learning of HRRP samples or SAR samples of known categories in S3. The remaining entries are just mappings of irrelevant pixels in the image.

[0083] Please refer to Figure 2 , Figure 2The algorithm training flow chart of the SAR target detection method proposed in the embodiment of the present application is shown. The encoder used for the original SAR image slices and the SAR sample knowledge in the database is the same encoder, and only one public quantized word embedding table is maintained globally. Therefore, the three task learning in the present invention can be learned through a network model. The prior paired HRRP target signal samples are passed through the signal encoder to obtain the original HRRP knowledge sample encoding features, and the SAR target image samples are passed through the image encoder to obtain the original SAR knowledge sample encoding features. The original HRRP knowledge sample encoding features and the original SAR knowledge sample encoding features are compared and learned to constrain the encoders of the two data sources to be mapped to the same feature space. The original SAR image to be learned is uniformly cut into blocks to obtain a first sub-image sequence, and the first sub-image sequence is input into the SAR encoder to obtain a second initial encoding feature. A quantum word embedding table is used for query mapping to obtain the second initial SAR image block encoding feature after table lookup quantization. After reconstruction by the decoder, the second sub-image sequence is reconstructed and learned. After the SAR sample data and / or HRRP sample data of known categories in the database are passed through their respective modal encoders, feature triplets are constructed, and the quantum word embedding table is subjected to triple constraint learning, so that some entries are focused on potential target category entries for query use in the detection stage.

[0084] based on Figure 2 , Figure 3 The flowchart of the algorithm test of SAR target image proposed in the embodiment of the present application is shown. Figure 2 On this basis, the judgment of whether it matches a specific category term is introduced. When it matches, the potential target position and category are obtained. When it does not match, the original SAR image block (the second sub-image sequence) is returned.

[0085] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0086] First, we adopt the idea of ​​unsupervised quantized variational autoencoder to quantize the redundant pixel-level information processing of SAR images into a discrete feature space to obtain the common characteristics of low-dimensional target features, reduce the interference of the target scene on the detection algorithm, and improve the generalization ability of target detection.

[0087] Second, by referring to the ideas of contrastive learning and triplet learning, the existing SAR target samples of other scenes and the knowledge information of radar high-resolution range profile (HRRP) are uniformly mapped into the distribution space and assisted by the quantized word embedding table to complete target clustering.

[0088] Third, in the inference stage, the entire image is sliced ​​by sliding a window and compared with images of quantized word embedding strips of known categories to complete target detection in the entire image.

[0089] In addition, the present application also proposes a SAR target detection device, which includes:

[0090] The contrastive learning module is used to normalize and map the a priori aligned SAR sample data and HRRP sample data into the same feature space through their respective SAR encoders and HRRP encoders;

[0091] The reconstruction learning module is used to uniformly slice the unlabeled original SAR image into blocks. The unsupervised learning of the block areas is performed using a SAR encoder, a quantized embedding dictionary, and a SAR image decoder to obtain a SAR full-image variational autoencoder model.

[0092] A triplet learning module is configured to encode the SAR samples and / or HRRP samples through their respective SAR encoders and / or HRRP encoders and then input them into a SAR full-image variational autoencoder model, thereby optimizing the quantized embedding dictionary so that some entries are focused on target entries. The SAR samples include SAR sample data, and the HRRP samples include HRRP sample data.

[0093] The training module is used to perform multi-task joint training using contrast learning of S1, reconstruction learning of S2, and triple learning of S3 to obtain model training weights, and train the SAR full-image variational autoencoder model to obtain a multi-task joint training variational autoencoder;

[0094] The output module is configured to uniformly slice the entire original SAR image to be detected and input the resulting image into a variational autoencoder trained for multi-task joint training. The module then outputs the image blocks mapped to the target category and embeds the quantized word into the dictionary to obtain potential target locations. This preferred embodiment provides a computer device capable of implementing the steps of any of the embodiments of the SAR target detection method provided in the embodiments of this application, thereby achieving the beneficial effects of the SAR target detection method provided in the embodiments of this application. See the previous embodiments for details and will not be repeated here.

[0095] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be accomplished through instructions, or by controlling related hardware through instructions. The instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. To this end, the present invention provides a storage medium storing a plurality of instructions that can be loaded by a processor to execute the steps of any of the SAR target detection methods provided in the present invention.

[0096] The storage medium may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0097] Since the instructions stored in the storage medium can execute the steps in any of the SAR target detection method embodiments provided in the embodiments of the present application, the beneficial effects that can be achieved by any of the SAR target detection methods provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0098] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A SAR target detection method, characterized in that: The method comprises: S1, normalize and map the a priori aligned SAR sample data and HRRP sample data to the same feature space through their respective SAR encoders and HRRP encoders; S2. The unlabeled original SAR image is evenly cut into blocks. A SAR full-image variational autoencoder model based on unsupervised learning is constructed for the cut-block areas using a SAR encoder, a quantized embedding dictionary, and a SAR image decoder. S3. Encode the SAR sample and / or HRRP sample through their respective SAR encoders and / or HRRP encoders and input them into the SAR full-image variational autoencoder model, and optimize the quantized embedding dictionary so that some entries are focused on target entries, wherein the SAR sample includes SAR sample data and the HRRP sample includes HRRP sample data; S4, using the contrast learning of S1, the reconstruction learning of S2, and the triplet learning of S3 to perform multi-task joint training to obtain model training weights, and training the SAR full-image variational autoencoder model to obtain a multi-task joint trained variational autoencoder; S5. The original SAR image to be detected is uniformly cut into blocks and input into the variational autoencoder trained with the multi-task joint method, and the image blocks mapped to the quantized word embedding dictionary of the target category are output to obtain the potential target position.

2. The SAR target detection method according to claim 1, wherein: The steps of normalizing and mapping the a priori aligned SAR sample data and HRRP sample data to the same feature space through their respective SAR encoders and HRRP encoders include: Normalizing the SAR sample data to an N*N size by interpolation scaling; The HRRP sample data is normalized to 1*M size by truncation and zero padding; Inputting the normalized SAR sample data and HRRP sample data into respective SAR encoders and HRRP encoders to obtain first initial coding features; The first initial coding features are normalized and the cosine similarity loss is calculated to obtain the weights of the SAR encoder and the HRRP encoder.

3. The SAR target detection method according to claim 1, wherein: The steps of uniformly slicing the unlabeled original SAR image into blocks, and performing unsupervised learning on the slicing areas through a SAR encoder, a quantized embedding dictionary, and a SAR image decoder to obtain a SAR full-image variational autoencoder model include: The unlabeled original SAR image is uniformly cut into blocks to obtain the first sub-image sequence; Inputting the first sub-image sequence into the SAR encoder to extract initial coding features to obtain second initial coding features; discretize the second initial coding feature using a quantum word embedding table in a quantized embedding dictionary, query the mapping to obtain a mapped second initial coding feature; Constructing a SAR image decoder to perform image block pixel restoration on the mapped second initial coding feature to obtain a restored second initial coding feature; The first sub-image sequence is compared with the restored second initial coding features, and the second initial coding features are compared with the mapped second initial coding features to obtain a SAR full-image variational autoencoder model.

4. The SAR target detection method according to claim 3, wherein: The steps of discretizing the second initial coding feature using the quantum word embedding table in the quantized embedding dictionary and querying the mapping to obtain the mapped second initial coding feature include: Keeping the word vector dimension in the quantized word embedding table consistent with the dimension of the second initial encoding feature; The second initial coding feature is replaced by the discrete quantized vector in the quantized word embedding table by looking up the table, and query mapping is performed to obtain the mapped second initial coding feature.

5. The SAR target detection method according to claim 1, wherein: The steps of triple learning of S3 are: Sample a sample x in the database and obtain its quantized encoding feature e(x); Select a sample x from the database that belongs to the same category as x p , select samples x that do not belong to the same category as x n , and for x p and x n Quantize and encode the three elements e(x), e(x p )、e(x n ) forms a triple <e(x),e(x p ), e(x n )>.

6. The SAR target detection method according to claim 1, wherein: The steps of uniformly slicing the original SAR image to be detected into blocks and inputting the blocks into the multi-task joint trained variational autoencoder, outputting the image blocks mapped to the quantized word embedding dictionary of the target category, and obtaining the potential target location include: The original SAR image to be detected is evenly cut into blocks to obtain a second sub-image sequence; Inputting the second sub-image sequence into a SAR encoder to extract initial coding features to obtain third initial coding features; Use the target word in the quantum word embedding table to query and map the third initial coding feature to obtain the mapped third initial coding feature; The image block output that matches the quantized term of the target category term in the mapped third initial coding feature is filtered to obtain the potential target position.

7. A SAR target detection device, characterized in that: The device comprises: The contrastive learning module is used to normalize and map the a priori aligned SAR sample data and HRRP sample data into the same feature space through their respective SAR encoders and HRRP encoders; The reconstruction learning module is used to uniformly slice the unlabeled original SAR image into blocks. The unsupervised learning of the block areas is performed using a SAR encoder, a quantized embedding dictionary, and a SAR image decoder to obtain a SAR full-image variational autoencoder model. a triplet learning module for encoding SAR samples and / or HRRP samples through respective SAR encoders and / or HRRP encoders and inputting the encoded data into the SAR full-image variational autoencoder model, and optimizing the quantized embedding dictionary so that some entries are focused on target entries, wherein the SAR samples include SAR sample data and the HRRP samples include HRRP sample data; A training module is used to perform multi-task joint training using contrast learning of S1, reconstruction learning of S2, and triple learning of S3 to obtain model training weights, and train the SAR full-image variational autoencoder model to obtain a multi-task joint trained variational autoencoder; The output module is used to uniformly slice the original SAR image to be detected into blocks and input the blocks into the variational autoencoder trained with the multi-task joint method, output the image blocks mapped to the quantized word embedding dictionary of the target category, and obtain the potential target position.

8. The SAR target detection device according to claim 7, wherein: The contrastive learning module is specifically used to: Normalizing the SAR sample data to an N*N size by interpolation scaling; The HRRP sample data is normalized to 1*M size by truncation and zero padding; Inputting the normalized SAR sample data and HRRP sample data into respective SAR encoders and HRRP encoders to obtain first initial coding features; The first initial coding features are normalized and the cosine similarity loss is calculated to obtain the weights of the SAR encoder and the HRRP encoder.

9. A computer device, characterized in that: The computer device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the SAR target detection method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which is loaded and executed by a processor to implement the SAR target detection method according to any one of claims 1 to 6.

Images