SAR (Synthetic Aperture Radar) marine ship target identification device under incomplete data condition

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) marine ship target identification device under an incomplete data condition, which comprises a database, an upper computer and an SAR, and the database, the upper computer and the SAR are connected in sequence. The database comprises different types of set marine ship target data. The upper computer comprises a data enhancement module, a supervision training module, a ship identification module and a result display module, the data enhancement module enhances the data in the database, and the supervision training module uses the enhanced data to train a ship identification model and transmits the ship identification model to the ship identification module. The SAR radar obtains real-time monitored marine ship target data, transmits the data to the ship identification module of the upper computer for identification, and displays the data on the result display module of the upper computer. According to the method, a customized model is creatively used for generating a specific label target to make full use of features and rules of setting data, and SAR radar marine ship target recognition under the incomplete data condition with high accuracy and online recognition is achieved.

Description

Technical Field

[0001] The present invention relates to the field of pattern recognition, and particularly to a SAR radar maritime ship target recognition device under the condition of incomplete data. Background Art

[0002] Synthetic Aperture Radar (SAR) is an important remote sensing technology, which is widely used in the fields of earth observation, environmental monitoring, etc. Among them, SAR radar maritime ship target recognition is a key technology, which is of great significance for maritime traffic management, marine environmental protection, maritime security defense, etc. However, due to the complexity and uncertainty of the marine environment, as well as the multi-source and heterogeneity of SAR radar data, SAR radar maritime ship target recognition faces huge challenges. On the one hand, the SAR radar data may have the diversity of ship variants caused by external disturbances, affecting the credibility of the data. This problem may be caused by various factors such as the variants caused by the ship itself, the environment and the sensor. On the other hand, the SAR radar data may have incomplete massive monitoring data, affecting the effectiveness of the data. These problems may be caused by various factors such as a large amount of data without strict manual tuning or errors in label tuning. Therefore, how to perform SAR radar maritime ship target recognition under the condition of incomplete data and improve the accuracy and speed of recognition is the key technical problem to improve the efficiency and value of SAR radar maritime ship target recognition.

[0003] At present, there have been some research works on SAR radar maritime ship target recognition, mainly including rule-based, statistic-based, and machine learning-based methods. These methods have their own advantages and disadvantages, but they all have some common limitations: on the one hand, they lack full utilization of the effective data of SAR radar data, resulting in the recognition model being difficult to extract the core features of the target and the recognition effect being unsatisfactory; on the other hand, they cannot solve the limitations of small data volume and poor quality of incomplete SAR radar data, resulting in the recognition model being prone to overfitting. Therefore, there is an urgent need to develop a new type of SAR radar maritime ship target recognition device to overcome the deficiencies of the existing methods, improve the quality and usability of SAR radar maritime ship target data, and thus improve the accuracy and real-time performance of recognition. Summary of the Invention

[0004] The object of the present invention is to provide a SAR radar maritime ship target recognition device under the condition of incomplete data, aiming at the deficiencies of the existing SAR radar maritime ship target recognition device with unsatisfactory recognition effect and easy overfitting of the recognition model, which can make full use of the characteristics and laws of the already tuned effective data, realize automatic and intelligent target recognition, and the recognition results are credible, accurate and timely.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a SAR radar maritime ship target recognition device under incomplete data conditions, including a database and a host computer;

[0006] The database contains calibrated SAR radar maritime ship target data of different types;

[0007] The host computer includes a data enhancement module, which enhances the calibrated SAR radar maritime ship target data collected in the database;

[0008] The host computer further includes a supervised training module; the supervised training module uses the data in the enhanced database to train a SAR radar maritime ship target recognition model;

[0009] The host computer further includes a ship recognition module; the ship recognition module uses the trained SAR radar maritime ship target recognition model to identify the types of SAR radar maritime ship targets under real-time monitoring.

[0010] Furthermore, the data enhancement module includes the following steps:

[0011] (1.1) There are c types of SAR radar maritime ship target data in the database sample space X where x represents the database sample, i = 1, 2,..., c represents the i-th type of SAR radar maritime ship target, and l i represents that the i-th type of SAR radar maritime ship target has l i samples, and the i-th type of SAR radar maritime ship target x in the database sample space X is taken out i ;

[0012] (1.2) For the i-th type of SAR radar maritime ship target x in the database sample space X i train the corresponding low-rank adapter model L i ; among them, the low-rank adapter model is a fine-tuning method designed for large models. Denote the pre-trained parameters of the large model as where d×k represents the rank of Φ 0 is d×k, and use low-rank decomposition to represent the parameter update ΔΦ, that is:

[0013] Φ 0 +ΔΦ = Φ 0 +BA (1)

[0014] where and represent the decomposed parameter update matrices, r << min(d, k) represents the "intrinsic rank" in the parameter update process, and the parameter Φ is frozen during the training process0 , only the parameters in A and B are trained. Since the dimension of r is much smaller than d and k, the number of parameters to be trained is also much smaller than the original number of parameters;

[0015] (1.3) The low-rank adapter model L trained using the data in the database i The large model to be adapted is a diffusion generation model, and its parameters correspond to the parameters Φ frozen during the training process in step (1.2) 0 , and the noise reduction diffusion probability model is used to estimate the probability distribution p θ (x 0 ), to approximate the distribution q(x 0 ) of the actual data, where x 0 is any sample in the database sample space of the latent variable. The encoder in the variational autoencoder is used to transform the features from the high-dimensional image space to the low-dimensional space, compress the data, and simplify the feature representation; j represents the jth sample in the ith type of SAR radar sea ship target:

[0016]

[0017] Among them, x 1 ,..., x T are latent variables with the same dimension as x 0 , where T represents the latent variable at the Tth time step; the forward process and the reverse process form the diffusion probability model, and both are composed of Markov chains; among them, the forward process is expressed as:

[0018]

[0019] Among them, x 1:T is the abbreviated representation of x 1 ,..., x T , represents the Gaussian distribution, I represents the Gaussian noise of the standard normal distribution, and β t represents the magnitude of the noise. The formula for the latent variable x t at any time step t is obtained through recursive derivation:

[0020]

[0021] Among them, represents the proportion of the original information retained by the data at each time step t, and α t represents the cumulative product of all the original information proportion values from time step 1 to t, represents the intermediate variable in the calculation of the cumulative product, that is, the original information proportion value at the kth time step. The reverse Markov process estimated by the noise is:

[0022]

[0023] Among them, x 0:T is the abbreviated representation of x 0 ,..., x T , μ θ (x t , t) and σ θ (x t , t) respectively represent the mean and variance in the estimated reverse process; when conditioned on x 0 , the posterior probability q(x t-1 |x t , x 0 ) of the true reverse process is:

[0024]

[0025] Among them, and are the mean and variance in the true reverse process; meanwhile, replacing x 0 with To reduce the KL divergence between the estimated probability and the true probability, reparameterize the probability expression of the estimated reverse process:

[0026]

[0027] Among them, based on ∈ θ (x t , t) represents the noise in the estimated reverse process; the final optimization objective is obtained as:

[0028]

[0029] During the reverse sampling process, the estimated target latent variable T is generated from the Gaussian noise x through formulas (12) and (13), and then the variable is restored from the latent space to the image space through the decoder in the variational autoencoder:

[0030]

[0031] Among them, is the estimated target sample; the device uses a diffusion generative model combined with a customized trained low-rank adapter model L i to generate augmented data of the i-th type of SAR radar maritime ship target, combining the generative ability of the diffusion model and the control ability of the low-rank adapter model to expand the database capacity and enhance the database quality; the database after data enhancement is denoted as Among them, represents the enhanced database sample, It is indicated that there are samples of the SAR radar maritime ship target of the i-th type after enhancement,

[0032] Furthermore, the SAR radar maritime ship target recognition model trained by the supervised training module is a Vision Transformer (ViT) model.

[0033] Furthermore, the host computer further includes a result display module; the result display module displays the recognition result of the ship recognition module on the host computer.

[0034] Furthermore, an SAR radar is further included, and the SAR radar acquires real-time monitored maritime ship target data and transmits it to the ship recognition module of the host computer for recognition.

[0035] The beneficial effects of the present invention are as follows: Aiming at the problem of SAR radar maritime ship target recognition under incomplete data conditions, the present invention utilizes the powerful generation ability of the diffusion generation model to expand the data volume, and uses the customized trained low-rank adapter model to control the diffusion generation model to generate ship targets of specific types, making full use of the inherent data characteristics and laws of the calibrated data to improve the quantity and quality of the data, and using the expanded data for the training of the recognition model, finally realizing the real-time intelligent monitoring and recognition of SAR radar maritime ship targets; it has the following advantages: 1. High recognition accuracy and high credibility; 2. Making full use of data characteristics and laws, and the recognition model is robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0037] Figure 1 is a schematic diagram of the hardware structure of the device proposed by the present invention.

[0038] Figure 2 is a schematic diagram of the functional modules of the host computer proposed by the present invention.

[0039] Figure 3 is a schematic diagram of the functional modules of the data enhancement module proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following will describe the present invention in detail with reference to the drawings. Without conflict, the features in the following embodiments and implementation manners can be combined with each other.

[0041] The present invention aims at the problem of SAR radar maritime ship target recognition under the condition of incomplete data. It utilizes the powerful generation ability of the diffusion generative model to expand the data volume, and uses a customized trained low-rank adapter model to control the diffusion generative model to generate specific types of ship targets. It makes full use of the inherent data characteristics and laws of the calibrated data to improve the quantity and quality of the data, and uses the expanded data for the training of the recognition model, ultimately realizing the real-time intelligent monitoring and recognition of SAR radar maritime ship targets.

[0042] The following specifically describes the present invention with reference to the accompanying drawings. Figure 1 、 Figure 2 、 Figure 3 A SAR radar maritime ship target recognition device under the condition of incomplete data includes a database 1, a host computer 2, and a SAR radar 3, which are connected in sequence; the database 1 contains calibrated maritime ship target data of different types; the host computer 2 includes:

[0043] (1) A data enhancement module 4. Since there is a large amount of uncalibrated or wrongly labeled data for SAR radar maritime ship targets, the amount of effectively calibrated data is limited. The data enhancement module enhances the calibrated SAR radar maritime ship target data collected in the database. The data enhancement module includes the following steps:

[0044] (1.1) There are c types of SAR radar maritime ship target data in the database sample space X where x represents the database sample, i = 1, 2,..., c represents the i-th type of SAR radar maritime ship target, and l i represents that the i-th type of SAR radar maritime ship target has l i samples. Take out the i-th type of SAR radar maritime ship target x i from the database sample space X;

[0045] (1.2) Train the corresponding low-rank adapter model L i for the i-th type of SAR radar maritime ship target x i in the database sample space X; among them, the low-rank adapter model is a low-resource and efficient fine-tuning method designed for large models. Since the enhanced data needs to be the same as the original type or the type can be controlled, the low-rank adapter model trained with specific types of SAR radar maritime ship target data can control the large model to generate customized results. Denote the pre-trained parameters of the large model as where d×k represents the rank of Φ 0 is d×k. The result of using full-parameter fine-tuning by the traditional method is Φ 0 +ΔΦ, where ΔΦ is the update amount of the large model parameters. At this time That is, the rank of ΔΦ is also d×k, and the fine-tuning update of the parameters consumes a large amount of computing resources. The pre-trained model has a very small intrinsic dimension, that is, there is a parameter with a very low dimension, and fine-tuning it has the same effect as fine-tuning in the full parameter space. There is also an "intrinsic rank" in the parameter update process. For the pre-trained parameter matrix Use low-rank decomposition to represent the parameter update ΔΦ, that is:

[0046] Φ 0 +ΔΦ = Φ 0 +BA (1)

[0047] Where and represent the decomposed parameter update matrices, r << min(d,k) represents the "intrinsic rank" in the parameter update process, and the parameters Φ are frozen during the training process 0 , and only the parameters in A and B are trained. Since the dimension of r is much smaller than d and k, the number of parameters to be trained is also much smaller than the original number of parameters;

[0048] (1.3) Due to the powerful generation ability of the diffusion generative model, the large model adapted by the low-rank adapter model L trained using the data in the database i is a diffusion generative model, and its parameters correspond to the parameters Φ frozen during the training process in step (1.2) 0 . The principle of the diffusion generative model is the diffusion process, and the diffusion process is to use a denoising diffusion probability model to estimate the probability distribution p θ (x 0 ), to approach the distribution q(x 0 ) of the actual data, where x 0 is any sample in the database sample space of the latent variable. The encoder in the variational autoencoder is used to transform the features from the high-dimensional image space to the low-dimensional space, compress the data, and simplify the feature representation; j represents the jth sample in the ith type of SAR radar maritime ship target:

[0049]

[0050] Among them, x 1 ,..., x T are latent variables with the same dimension as x 0 , where T represents the latent variable at the Tth time step; the forward process and the reverse process form the diffusion probability model, and both are composed of Markov chains; among them, the forward process is expressed as:

[0051]

[0052] Among them, represents the Gaussian distribution, I represents the Gaussian noise of the standard normal distribution, βt Indicates the magnitude of the noise, and the latent variable x at any time step t is obtained through recursive derivation t The formula for

[0053]

[0054] where noise The estimated reverse Markov process is

[0055]

[0056] where μ θ (x t , t) and σ θ (x t , t) represent the mean and variance in the estimated reverse process respectively; when conditioned on x 0 , the posterior probability q(x t-1 |x t , x 0 ) of the true reverse process is

[0057]

[0058] where and are the mean and variance in the true reverse process; meanwhile, replacing x 0 with To reduce the KL divergence between the estimated probability and the true probability, reparameterize the probability expression of the estimated reverse process

[0059]

[0060] where, based on ∈ θ (x t , t) represents the noise in the estimated reverse process; the final optimization objective is

[0061]

[0062] During the reverse sampling process, the estimated target latent variable is generated from the Gaussian noise x T through formulas (12) and (13) Then, the variable is restored from the latent space to the image space through the decoder in the variational autoencoder

[0063]

[0064] where is the estimated target sample; the device uses a diffusion generative model combined with a customized low-rank adapter model L iGenerate augmented data for the i-th type of SAR radar maritime ship target, combine the generation ability of the diffusion model and the control ability of the low-rank adapter model to expand the database capacity and enhance the database quality; Denote the database after data augmentation as Among them, represents the samples in the augmented database, represents that there are samples of the i-th type of SAR radar maritime ship target after augmentation,

[0065] Specifically, the host computer further includes a supervision training module 5; The supervision training module uses the database after data augmentation to train the SAR radar maritime ship target recognition model; The SAR radar maritime ship target recognition model trained by the supervision training module 5 is a Vision Transformer (ViT) model.

[0066] Specifically, the host computer further includes a ship recognition module 6; The ship recognition module uses the trained SAR radar maritime ship target recognition model to identify the types of SAR radar maritime ship targets monitored in real time.

[0067] Specifically, the host computer further includes a result display module 7; The result display module displays the recognition result of the ship recognition module 6 on the host computer 2.

[0068] Specifically, the SAR radar 3 acquires the data of the maritime ship target monitored in real time and transmits it to the ship recognition module 6 of the host computer 2 for recognition, and displays it on the result display module 7 of the host computer.

[0069] The above embodiments are used to explain the present invention, rather than limit the present invention. Any modification and change made within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A SAR radar marine ship target recognition device under incomplete data conditions, characterized in that: Including database and host computer; The database contains different types of SAR radar marine ship target data that have been adjusted; The host computer includes a data enhancement module, which performs data enhancement on the SAR radar marine ship target data collected in the database; The host computer also includes a supervised training module; the supervised training module uses the data of the data-enhanced database to train the SAR radar marine ship target recognition model; The host computer also includes a ship identification module; the ship identification module uses a trained SAR radar maritime ship target identification model to identify the type of SAR radar maritime ship target monitored in real time.

2. The SAR radar marine ship target recognition device under incomplete data conditions according to claim 1 is characterized in that: The data enhancement module comprises the following steps: (1.1) There are c types of SAR radar marine ship target data in the database sample space X. Where x represents the database sample, i = 1, 2, ..., c represents the i-th type of SAR radar marine ship target, l i It means that the number of SAR radar marine ship targets of type i is l i samples, take out the SAR radar marine ship target x of the i-th type in the database sample space X i ; (1.2) For the i-th type of SAR radar marine ship target x in the database sample space X i Train the corresponding low-rank adapter model L i Among them, the low-rank adapter model is a fine-tuning method designed for large models, which records the pre-trained parameters of the large model as Where d×k indicates that the rank of Φ0 is d×k, and low-rank decomposition is used to represent the parameter update ΔΦ, that is: Φ0+ΔΦ=Φ0+BA (1) in and represents the decomposed parameter update matrix, r<<min(d,k) represents the "intrinsic rank" in the parameter update process, freezes the parameter Φ0 during the training process, and only trains the parameters in A and B. Since the dimension of r is much smaller than d and k, the number of trained parameters is also much smaller than the original number of parameters; (1.3) The low-rank adapter model L trained using the database data i The large model adapted is a diffusion generation model, whose parameters correspond to the parameters Φ0 frozen in the training process in step (1.2). The denoising diffusion probability model is used to estimate the probability distribution p θ (x0) to approximate the distribution of actual data q(x0), where x0 is any sample in the database sample space. The latent variable of the variational autoencoder is used to transform the features from the high-dimensional image space to the low-dimensional space, compress the data and simplify the feature representation; j represents the jth sample in the i-th type of SAR radar marine ship target: where x1,...,x T is a hidden variable with the same dimension as x0, where T represents the hidden variable at the Tth time step; the forward process and the reverse process constitute a diffusion probability model, and both are composed of Markov chains; the forward process is expressed as: Among them, x 1:T is x1,...,x T The abbreviation of represents Gaussian distribution, I represents Gaussian noise of standard normal distribution, β t Represents the magnitude of the noise, and the hidden variable x at any time step t is obtained by recursive derivation t The formula is: in, Indicates the proportion of original information retained by the data at each time step t, α t Represents the cumulative multiplication of all original information ratio values ​​from time step 1 to t, Represents the intermediate variable in the process of calculating the cumulative multiplication, that is, the original information ratio value at the kth time step; noise The estimated inverse Markov process is: Among them, x 0:T is x0,...,x T The abbreviation of μ θ (x t ,t) and σ θ (x t , t) represent the mean and variance of the estimated reverse process respectively; when x0 is used as the condition, the posterior probability q(x t-1 |x t ,x0) is: in, and is the mean and variance of the true inverse process; at the same time, replace x0 with In order to reduce the KL divergence between the estimated probability and the true probability, the probability expression of the inverse process of the reparameterized estimation is: Among them, based on ∈ θ (x t , t) represents the noise in the estimated inverse process; the final optimization goal is: In the reverse sampling process, the Gaussian noise x is obtained by using formulas (12) and (13). T Generate estimated target latent variables Then, the decoder in the variational autoencoder is used to restore the variables from the latent space to the image space: in, is the estimated target sample; the device uses a diffusion generation model with a customized trained low-rank adapter model L i Generate the augmented data of the i-th type SAR radar marine ship target, combine the generation ability of the diffusion model and the control ability of the low-rank adapter model, expand the database capacity and enhance the database quality; the enhanced database is recorded as in, represents the enhanced database sample, It means that after enhancement, the i-th type of SAR radar sea ship target has samples, 3. The SAR radar marine ship target recognition device under incomplete data conditions according to claim 1 is characterized in that: The SAR radar maritime ship target recognition model trained by the supervised training module is a visual converter ViT model.

4. The SAR radar marine ship target recognition device under incomplete data conditions according to claim 1 is characterized in that: The host computer also includes a result display module; the result display module displays the recognition result of the ship recognition module on the host computer.

5. The SAR radar marine ship target recognition device under incomplete data conditions according to claim 1 is characterized in that: It also includes a SAR radar, which acquires real-time monitored marine ship target data and transmits the data to the ship identification module of the host computer for identification.

Images