SAR Target Open Set and Incremental Identification Method and Apparatus Based on Extremum Theory

Abstract

This application relates to a method and apparatus for SAR target open-set and incremental identification based on extreme value theory. The method includes: calculating the Euclidean distance between each target and other target categories based on target features in a SAR image, and using extreme value theory to determine the boundaries of each target category to obtain an open-set identification model; further, for each target category, transforming the distance boundary problem into a set coverage problem, selecting key samples for incremental learning and lightweighting the open-set identification model; then using the open-set identification model to classify known target categories while simultaneously identifying unknown target categories; when the model finds an unknown target category, updating the open-set identification model based on the unknown target to increase the known target categories in the model. This method can correctly classify known target categories and identify unknown target categories.

Description

Technical Field

[0001] This application relates to the field of SAR image target recognition technology, and in particular to a method and apparatus for SAR target open set and incremental recognition based on extreme value theory. Background Technology

[0002] As an active sensor, SAR has unique advantages such as all-weather, high-resolution, and wide-area imaging capabilities, making it widely used in various military and civilian fields. SAR target recognition is a typical example in many military applications, its main purpose being to identify targets in SAR images, such as fortifications, temporary command posts, tanks, armored vehicles, self-propelled artillery, infantry fighting vehicles, military trucks, military aircraft, and military ships.

[0003] Traditional SAR target recognition methods have two significant limitations: 1. They adhere to the closed-set assumption, meaning the training set is assumed to contain all target categories from the test set. However, real-world environments often present targets whose categories are not included in the training set—i.e., unknown category targets—and these target recognition methods typically cannot handle this situation. 2. Target recognition models are costly to update. After acquiring new category data, traditional batch training methods are required, resulting in high data storage and computation costs, long processing times, and poor real-time performance. Summary of the Invention

[0004] Therefore, it is necessary to provide a SAR target open set and incremental identification method and apparatus based on extreme value theory that can correctly classify known target categories while also identifying unknown target categories, in order to address the above-mentioned technical problems.

[0005] A SAR target open set and incremental identification method based on extreme value theory, the method comprising:

[0006] Acquire a SAR image dataset, which includes SAR images of multiple known target categories;

[0007] An autoencoder improved by a multi-scale structural similarity loss function is used to extract target features from each SAR image in the SAR image dataset to obtain target features of each SAR image sample, and the obtained target features are used as samples to construct a training sample set.

[0008] Based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated, and then the open set recognition model is constructed using the extreme value theory.

[0009] In the training sample set, representative key samples from each target category are selected, and the open set recognition model is compressed based on the key samples of each target category to obtain a lightweight open set recognition model.

[0010] A SAR image to be identified is acquired, and the features of the target to be identified in the SAR image are extracted. The lightweight open-set recognition model is used to classify the target in the SAR image according to the features of the target to be identified, and the classification result is obtained. If the classification result is an unknown target category, the SAR image to be identified is manually annotated, and the corresponding features of the target to be identified and the key samples are used to construct a new training sample set. The lightweight open-set recognition model is then updated using the new training sample set.

[0011] In one embodiment, the step of calculating the Euclidean distance between each sample in the training sample set and all samples in other target categories based on the feature space, and then constructing an open set recognition model using extreme value theory, includes:

[0012] Calculate the Euclidean distance between each sample in each target category and all samples in other target categories;

[0013] The smallest τ values ​​are selected to fit the Weibull distribution function to obtain the Weibull parameters for each sample;

[0014] The boundary of the corresponding target category is determined based on the Weibull parameters of all samples in each target category, and the open set recognition model is constructed based on the boundaries of each target category.

[0015] In one embodiment, the Weibull parameters of each sample are obtained according to the following formula:

[0016]

[0017] In the above formula, x i Let x represent the sample. j x represents i Samples from target categories other than their own target category, λ i and k i For sample x i The Weibull parameters.

[0018] In one embodiment, constructing the open set recognition model based on the boundaries of each target category includes:

[0019] Establish the probability equation for each target category based on the boundary of each target category;

[0020] The open set recognition model is constructed from the probability equations of each target category.

[0021] In one embodiment, the probability equation is:

[0022]

[0023] In the above formula, A is the target category, and x t For the SAR image to be identified, x i Let λ be a sample from target category A. i and k i For sample x i The Weibull parameters.

[0024] In one embodiment, the step of selecting representative key samples from each target category in the training sample set, and compressing the open set recognition model based on the key samples of each target category to obtain a lightweight open set recognition model includes:

[0025] Find the containing samples of each sample in each target category, and then select the sample set with the smallest sample size from each target category based on a greedy algorithm. The samples in this set are the key samples.

[0026] The boundaries of the corresponding target categories are redefined based on the key samples in each target category, and the lightweight open set recognition model is constructed based on the redefined boundaries of each target category.

[0027] In one embodiment, identifying the containing samples of each sample within each target category specifically includes:

[0028] For samples x in the same target category i and sample x j Make the following judgment: if Ψ(||x) is satisfied... i -x j ||;λ i ,k i )>δ and Ψ(||x j -x i ||;λ j ,k j If ) < δ, then sample x i Includes sample x j Where δ is a preset threshold, and λ i and k i For sample x i The Weibull parameter, λ j and k j For sample x j Weibull parameters.

[0029] In one embodiment, before updating the lightweight open set recognition model using the new training sample set, the samples corresponding to the unknown target categories in the new training sample set are manually labeled.

[0030] A SAR target open set and incremental identification device based on extreme value theory, the device comprising:

[0031] A SAR image dataset acquisition module is used to acquire a SAR image dataset, which includes SAR images of multiple different target categories;

[0032] The training sample set construction module is used to extract target features from each SAR image in the SAR image dataset using an autoencoder improved by a multi-scale structural similarity loss function, to obtain the target features of each SAR image sample, and to use the obtained target features as samples to construct a training sample set.

[0033] The open set recognition model construction module is used to calculate the Euclidean distance between each sample in the training sample set and all samples in other target categories based on the feature space, and then use the extreme value theory to construct the open set recognition model.

[0034] An open set recognition model construction compression module is used to select representative key samples from each target category in the training sample set, and compress the open set recognition model according to the key samples of each target category to obtain a lightweight open set recognition model.

[0035] The target recognition and incremental module is used to acquire a SAR image to be identified, extract the target features of the SAR image to be identified, use the lightweight open set recognition model to classify the SAR image to be identified based on the target features to be identified, and obtain the classification result. If the classification result is an unknown target category, after manually annotating the SAR image to be identified, the corresponding target features to be identified and the key samples are used to construct a new training sample set, and the lightweight open set recognition model is updated using the new training sample set.

[0036] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0037] Acquire a SAR image dataset, which includes SAR images of multiple known target categories;

[0038] An autoencoder improved by a multi-scale structural similarity loss function is used to extract target features from each SAR image in the SAR image dataset to obtain target features of each SAR image sample, and the obtained target features are used as samples to construct a training sample set.

[0039] Based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated, and then the open set recognition model is constructed using the extreme value theory.

[0040] In the training sample set, representative key samples from each target category are selected, and the open set recognition model is compressed based on the key samples of each target category to obtain a lightweight open set recognition model.

[0041] A SAR image to be identified is acquired, and the features of the target to be identified in the SAR image are extracted. The lightweight open-set recognition model is used to classify the target in the SAR image according to the features of the target to be identified, and the classification result is obtained. If the classification result is an unknown target category, the SAR image to be identified is manually annotated, and the corresponding features of the target to be identified and the key samples are used to construct a new training sample set. The lightweight open-set recognition model is then updated using the new training sample set.

[0042] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0043] Acquire a SAR image dataset, which includes SAR images of multiple known target categories;

[0044] An autoencoder improved by a multi-scale structural similarity loss function is used to extract target features from each SAR image in the SAR image dataset to obtain target features of each SAR image sample, and the obtained target features are used as samples to construct a training sample set.

[0045] Based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated, and then the open set recognition model is constructed using the extreme value theory.

[0046] In the training sample set, representative key samples from each target category are selected, and the open set recognition model is compressed based on the key samples of each target category to obtain a lightweight open set recognition model.

[0047] A SAR image to be identified is acquired, and the features of the target to be identified in the SAR image are extracted. The lightweight open-set recognition model is used to classify the target in the SAR image according to the features of the target to be identified, and the classification result is obtained. If the classification result is an unknown target category, the SAR image to be identified is manually annotated, and the corresponding features of the target to be identified and the key samples are used to construct a new training sample set. The lightweight open-set recognition model is then updated using the new training sample set.

[0048] The aforementioned SAR target open-set and incremental identification method and apparatus based on extreme value theory extracts target features from SAR images using an autoencoder improved with MS-SSIMloss (multi-scale structural similarity loss function). It then calculates the Euclidean distance between each target and other target categories, and uses extreme value theory to determine the boundaries of each target category, resulting in an open-set identification model. For each target category, the distance boundary problem is transformed into a set coverage problem, and key samples for each category are selected for incremental learning and lightweighting of the open-set identification model. The open-set identification model is then used to classify known target categories while simultaneously identifying unknown target categories. When the model finds unknown target categories, these categories are manually labeled, and a training set is constructed using these labels and key samples. This is then used to update the open-set identification model to increase the number of known target categories. This method can correctly classify known target categories and identify unknown target categories. Furthermore, after obtaining new target categories, the target identification model can be updated incrementally, avoiding the high storage and computational costs of traditional batch training methods. This method is highly versatile and can be flexibly applied to SAR data from various airborne or spaceborne platforms with different imaging modes. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a SAR target open set and incremental identification method based on extreme value theory in one embodiment.

[0050] Figure 2 This is a schematic diagram of the autoencoder structure improved by MS-SSIM loss in one embodiment;

[0051] Figure 3 This is a block diagram of a SAR target open set and incremental identification device based on extreme value theory in one embodiment;

[0052] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] like Figure 2 As shown, a SAR target open set and incremental identification method based on extreme value theory is provided, including the following steps:

[0055] Step S100: Obtain a SAR image dataset, which includes multiple SAR images of different known target categories;

[0056] Step S110: Use the autoencoder improved by the multi-scale structural similarity loss function to extract target features from each SAR image in the SAR image dataset, obtain the target features of each SAR image sample, and use the obtained target features as samples to construct a training sample set.

[0057] Step S120: Based on the feature space, calculate the Euclidean distance between each sample in the training sample set and all samples in other target categories, and then use the extreme value theory to construct an open set recognition model.

[0058] Step S130: In the training sample set, select representative key samples from each target category, and compress the open set recognition model according to the key samples of each target category to obtain a lightweight open set recognition model.

[0059] Step S140: Obtain the SAR image to be identified, extract the target features of the SAR image to be identified, use a lightweight open set recognition model to classify the SAR image to be identified based on the target features, and obtain the classification result. If the classification result is an unknown target category, construct a new training sample set with the corresponding target features and key samples, and update the lightweight open set recognition model using the new training sample set.

[0060] In step S100, the SAR images in the SAR image dataset can be obtained through simulation or by using SAR radar to detect different types of targets. The SAR image dataset includes SAR target images of multiple known types, and there are multiple SAR images for each target type.

[0061] In step S110, the SAR images in the SAR image dataset are processed. In this embodiment, an autoencoder improved with the multi-scale structural similarity loss function (MS-SSIM loss) is used to extract target features from each SAR image to obtain target feature vectors. This data serves as intermediate data for the modified autoencoder. In fact, the input to the improved autoencoder is the SAR image, and the output is the image reconstructed from the SAR image.

[0062] Specifically, the autoencoder improved by MS-SSIM loss has the following structure: Figure 2 As shown, unlike the commonly used MSE loss, MS-SSIM loss focuses more on the structural similarity between the input and output images, rather than calculating the differences between the two images pixel by pixel. MS-SSIM loss is an improvement on SSIM loss, and the calculation method of SSIM loss is as follows:

[0063] SSIM(x,y)=[l(x,y)] α ·[c(x,y)]β ·[s(x,y)] γ (1)

[0064] In formula (1), x and y are the input and output images of the autoencoder, respectively, and l(x,y), c(x,y) and s(x,y) represent the brightness similarity, contrast similarity and structural similarity between the input and output images, respectively.

[0065] in, And s(x,y)=(σ xy +C3) / (σ x σ y +C3), in addition μ x (μ y ) and σ x (σ y σ represents the mean and variance of the image x(y). xy This represents the standard deviation of the input and output images. In one embodiment, C1 = (0.01·L) 2 C2 = (0.03 L) 2 C3 = C2 / 2, L is the range of image pixel values, with a value of 225; α, β, and γ are weighting factors for the three-part loss. Analysis shows that the SSIM loss takes into account the image's brightness, contrast, and structural similarity, thus extracting more SAR image structural information, which is crucial for SAR target recognition.

[0066] In practical applications, the entire image is usually divided into M image blocks, and then the average SSIM (MeanSSIM, MSSIM) is calculated. Where x i and y i This represents the i-th image patch in images X and Y.

[0067] In the referenced example, when training the autoencoder using MS-SSIM loss, the input and output images of size 128×128 are downsampled to obtain images of sizes 64×64, 32×32, and 16×16. Contrast similarity and structural similarity are then calculated on the four image sizes, while brightness similarity is calculated only on the 16×16 image. The MS-SSIM loss is then calculated as follows:

[0068]

[0069] In formula (2), N = 4, α N =1, β1=γ1=0.4, β i =γ i =0.2 (i∈{2,3,4}).

[0070] After extracting features from each SAR image using an autoencoder improved by MS-SSIM loss, the obtained target features are used as samples to construct a training sample set.

[0071] In step S120, based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated. Then, an open-set recognition model is constructed using extreme value theory, including: calculating the Euclidean distance between each sample in each target category and all samples in other target categories; selecting the smallest τ values ​​to fit the Weibull distribution function to obtain the Weibull parameters for each sample; determining the boundary of the corresponding target category based on the Weibull parameters of all samples in each target category; and constructing the open-set recognition model based on the boundaries of each target category.

[0072] Specifically, for each sample x in category A i Calculate its value with all samples x in other target categories. j Euclidean distance between ||x i -x j Then, the smallest τ values ​​are selected to fit the Weibull distribution function, obtaining the Weibull parameters λ. i and k i In other words, the corresponding Weibull parameter λ can be obtained for each sample. i and k i The Weibull parameters for each sample are obtained using the following formula:

[0073]

[0074] In formula (3), x i Let x represent the sample. j x represents i Samples from target categories other than their own target category, λ i and k i For sample x i The Weibull parameters.

[0075] After obtaining sample x i After obtaining the Weibull parameters, the sample x to be detected can be calculated based on the Weibull parameters. t Is it related to sample x? i The probability of belonging to the same category is calculated using the following formula:

[0076]

[0077] By comparing the calculation result with the preset threshold δ, the sample x to be detected can be determined. t Is it related to sample x? iThey belong to the same category. When the calculation result is greater than the preset threshold δ, that is, Ψ(||x i -x t ||;λ i ,k i ) > δ, therefore x is considered t and x i Same category.

[0078] After obtaining the Weibull parameters of all samples, the boundary of the target class can be determined based on the Weibull parameters of all samples within the same target class. This is how the sample to be detected, x, is determined. t The probability formula for belonging to a target category is as follows:

[0079]

[0080] In formula (5), A is the target category, x t For the SAR image to be identified, x i Let λ be a sample from target category A. i and k i For sample x i The Weibull parameter, when P(A|x t When ) > δ, then the sample to be tested x t It belongs to category A.

[0081] Similarly, the boundaries of all known categories can be calculated in the above manner. That is, the probability equations of each target category are established according to the boundaries of each target category. Finally, the open set recognition model is constructed from the probability equations of each target category.

[0082] Next, in step S130, the obtained open set recognition model is compressed to prepare for subsequent incremental learning. When compressing the open set recognition model, representative key samples from each target category are first selected from the training sample set. Specifically, this includes finding the containing samples of each target category, which transforms the model compression problem into a set coverage problem.

[0083] Specifically, identifying the containing samples within each target category includes: for samples x in the same target category... i and sample x j Make the following judgment: if two samples satisfy Ψ(||x) i -x j ||;λ i ,k i )>δ and Ψ(||x j -x i ||;λ j ,k j If ) < δ, then sample x i Includes sample xj Where δ is a preset threshold, and λ i and k i For sample x i The Weibull parameter, λ j and k j For sample x j Weibull parameters.

[0084] After obtaining the included samples for each sample, a greedy algorithm is used to select the sample set with the smallest number of samples from each target category. The samples in this set are then considered key samples. Similarly, representative key samples can be selected from each known category in the same way. The boundaries of the corresponding target categories are then redefined using the key samples in each target category, and a lightweight open-set recognition model is constructed based on the redefined boundaries of each target category.

[0085] In step S140, the obtained lightweight open-set recognition model is applied to classify targets in the SAR image to be identified. First, target features are extracted from the SAR image using an autoencoder improved with MS-SSIM loss. Then, the target features are input into the lightweight open-set recognition model. If the target in the SAR image is a known category target, the model will classify it correctly. If the target in the SAR image is an unknown category target, the model will identify it. The model is then updated again with the SAR image containing the unknown target category to increase the number of known target categories in the model.

[0086] When updating the open-set recognition model using SAR images with unknown target categories, a new training sample set is constructed using the target features of the SAR image and the key samples obtained in step S130. Steps S120 and S130 are then repeated to update the open-set recognition model. It is important to note that when updating the model using the new training sample set, the "key sample selection" step still needs to be performed, but this step is only performed on the newly added categories. In step S130, sample selection is only performed on the newly added categories, meaning that model compression is performed once for each category's samples.

[0087] In other embodiments, after obtaining the lightweight open-set recognition model, unknown category targets can be manually labeled, and then a new training sample set can be constructed using these labels and the key samples obtained in step S130. Steps S120 and S130 are then repeated to update the open-set recognition model, incrementally expanding the target types that the open-set recognition model can recognize. Similarly, when performing step S130, samples are selected only for newly added categories, meaning that model compression is performed once for each category of samples.

[0088] The aforementioned SAR target open-set and incremental identification method based on extreme value theory extracts target features from SAR images using an autoencoder improved with MS-SSIM loss (multi-scale structural similarity loss function). It then calculates the Euclidean distance between each target and other target categories, and uses extreme value theory to determine the boundaries of each target category, resulting in an open-set identification model. For each target category, the distance boundary problem is transformed into a set coverage problem, and key samples for each category are selected for incremental learning and lightweighting of the open-set identification model. The open-set identification model is then used to classify known target categories while simultaneously identifying unknown target categories. When the model finds unknown target categories, these categories are manually labeled, and a training set is constructed using these labels and key samples. This is then used to update the open-set identification model, increasing the number of known target categories. This method can correctly classify known target categories and identify unknown target categories. Furthermore, after obtaining new target categories, the target identification model can be updated incrementally, avoiding the high storage and computational costs of traditional batch training methods. This method is highly versatile and can be flexibly applied to SAR data from various airborne or spaceborne platforms with different imaging modes.

[0089] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0090] In one embodiment, such as Figure 3 As shown, a SAR target open set and incremental recognition device based on extreme value theory is provided, including: a SAR image dataset acquisition module 200, a training sample set construction module 210, an open set recognition model construction module 220, an open set recognition model construction compression module 230, and a target recognition and incremental module 240, wherein:

[0091] SAR image dataset acquisition module 200 is used to acquire SAR image dataset, which includes SAR images of multiple different target categories;

[0092] The training sample set construction module 210 is used to extract target features from each SAR image in the SAR image dataset using an autoencoder improved by a multi-scale structural similarity loss function, to obtain the target features of each SAR image sample, and to use the obtained target features as samples to construct a training sample set.

[0093] The open set recognition model construction module 220 is used to calculate the Euclidean distance between each sample in the training sample set and all samples in other target categories based on the feature space, and then use the extreme value theory to construct the open set recognition model.

[0094] The open set recognition model construction compression module 230 is used to select representative key samples from each target category in the training sample set, and compress the open set recognition model according to the key samples of each target category to obtain a lightweight open set recognition model.

[0095] The target recognition and incremental module 240 is used to acquire a SAR image to be identified, extract the target features of the SAR image to be identified, classify the target in the SAR image according to the target features using the lightweight open-set recognition model, and obtain the classification result. If the classification result is an unknown target category, after manually annotating the SAR image to be identified, a new training sample set is constructed by combining the corresponding target features with the key samples, and the lightweight open-set recognition model is updated using the new training sample set.

[0096] Specific limitations regarding the SAR target open set and incremental identification device based on extremum theory can be found in the limitations of the SAR target open set and incremental identification method based on extremum theory mentioned above, and will not be repeated here. Each module in the aforementioned SAR target open set and incremental identification device based on extremum theory can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0097] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a SAR target open set and incremental identification method based on extreme value theory. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0098] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0099] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0100] Acquire a SAR image dataset, which includes SAR images of multiple known target categories;

[0101] An autoencoder improved by a multi-scale structural similarity loss function is used to extract target features from each SAR image in the SAR image dataset to obtain target features of each SAR image sample, and the obtained target features are used as samples to construct a training sample set.

[0102] Based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated, and then the open set recognition model is constructed using the extreme value theory.

[0103] In the training sample set, representative key samples from each target category are selected, and the open set recognition model is compressed based on the key samples of each target category to obtain a lightweight open set recognition model.

[0104] A SAR image to be identified is acquired, and the features of the target to be identified in the SAR image are extracted. The lightweight open-set recognition model is used to classify the target in the SAR image according to the features of the target to be identified, and the classification result is obtained. If the classification result is an unknown target category, the SAR image to be identified is manually annotated, and the corresponding features of the target to be identified and the key samples are used to construct a new training sample set. The lightweight open-set recognition model is then updated using the new training sample set.

[0105] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0106] Acquire a SAR image dataset, which includes SAR images of multiple known target categories;

[0107] An autoencoder improved by a multi-scale structural similarity loss function is used to extract target features from each SAR image in the SAR image dataset to obtain target features of each SAR image sample, and the obtained target features are used as samples to construct a training sample set.

[0108] Based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated, and then the open set recognition model is constructed using the extreme value theory.

[0109] In the training sample set, representative key samples from each target category are selected, and the open set recognition model is compressed based on the key samples of each target category to obtain a lightweight open set recognition model.

[0110] A SAR image to be identified is acquired, and the features of the target to be identified in the SAR image are extracted. The lightweight open-set recognition model is used to classify the target in the SAR image according to the features of the target to be identified, and the classification result is obtained. If the classification result is an unknown target category, the SAR image to be identified is manually annotated, and the corresponding features of the target to be identified and the key samples are used to construct a new training sample set. The lightweight open-set recognition model is then updated using the new training sample set.

[0111] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0112] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0113] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A SAR target open set and incremental identification method based on extreme value theory, characterized in that, The method includes: Acquire a SAR image dataset, which includes SAR images of multiple known target categories; An autoencoder improved by a multi-scale structural similarity loss function is used to extract target features from each SAR image in the SAR image dataset to obtain target features of each SAR image sample, and the obtained target features are used as samples to construct a training sample set. Based on the feature space, the Euclidean distance between each sample in the training sample set and all samples in other target categories is calculated. Then, an open set recognition model is constructed using extreme value theory, including: calculating the Euclidean distance between each sample in each target category and all samples in other target categories; selecting the smallest value to fit the Weibull distribution function to obtain the Weibull parameter of each sample; determining the boundary of the corresponding target category based on the Weibull parameter of all samples in each target category; and constructing the open set recognition model based on the boundaries of each target category. The Weibull parameter of each sample is obtained according to the following formula: In the above formula, Indicates a sample, express Samples from target categories other than their own target category. as well as For the sample The Weibull parameters; In the training sample set, representative key samples from each target category are selected, and the open set recognition model is compressed based on the key samples of each target category to obtain a lightweight open set recognition model. Specifically, the containing samples of each sample in each target category are found, and then, based on a greedy algorithm, the sample set with the smallest sample size is selected from each target category. The samples in this set are the key samples. The boundaries of the corresponding target categories are redefined based on the key samples in each target category, and the lightweight open set recognition model is constructed based on the redefined boundaries of each target category. Finding the containing samples of each sample in each target category specifically includes, for samples in the same target category... and samples Perform the following judgment: if the following conditions are met... and Then the sample Includes samples ,in, For the preset threshold, as well as For the sample Weibull parameters, as well as For the sample Weibull parameters; A SAR image to be identified is acquired, and the features of the target to be identified in the SAR image are extracted. The lightweight open-set recognition model is used to classify the target in the SAR image according to the features of the target to be identified, and the classification result is obtained. If the classification result is an unknown target category, the SAR image to be identified is manually annotated, and the corresponding features of the target to be identified and the key samples are used to construct a new training sample set. The lightweight open-set recognition model is then updated using the new training sample set.

2. The SAR target open set and incremental identification method according to claim 1, characterized in that, The process of constructing the open set recognition model based on the boundaries of each target category includes: Establish the probability equation for each target category based on the boundary of each target category; The open set recognition model is constructed from the probability equations of each target category.

3. The SAR target open set and incremental identification method according to claim 2, characterized in that, The probability equation is: In the above formula, For the target category, For the SAR image to be identified, for A sample from the target category, as well as For the sample The Weibull parameters.

4. The SAR target open set and incremental identification method according to claim 1, characterized in that, Before updating the lightweight open set recognition model using the new training sample set, the samples corresponding to the unknown target categories in the new training sample set are manually labeled.

5. A SAR target open set and incremental identification device based on extreme value theory, characterized in that, The device implements the SAR target open set and incremental identification method based on extreme value theory as described in any one of claims 1-4, including: A SAR image dataset acquisition module is used to acquire a SAR image dataset, which includes SAR images of multiple different target categories; The training sample set construction module is used to extract target features from each SAR image in the SAR image dataset using an autoencoder improved by a multi-scale structural similarity loss function, to obtain the target features of each SAR image sample, and to use the obtained target features as samples to construct a training sample set. The open set recognition model construction module is used to calculate the Euclidean distance between each sample in the training sample set and all samples in other target categories based on the feature space, and then use the extreme value theory to construct the open set recognition model. An open set recognition model construction compression module is used to select representative key samples from each target category in the training sample set, and compress the open set recognition model according to the key samples of each target category to obtain a lightweight open set recognition model. The target recognition and incremental module is used to acquire a SAR image to be identified, extract the target features of the SAR image to be identified, use the lightweight open set recognition model to classify the SAR image to be identified based on the target features to be identified, and obtain the classification result. If the classification result is an unknown target category, after manually annotating the SAR image to be identified, the corresponding target features to be identified and the key samples are used to construct a new training sample set, and the lightweight open set recognition model is updated using the new training sample set.

Images