An off-library target identification method and device
By using Bayesian neural networks and variational inference methods, the problem of identifying targets outside the database is solved, achieving accurate identification of targets outside the database and discovery of new target categories, and supporting the rapid expansion of the template library.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAVAL AVIATION UNIV
- Filing Date
- 2022-08-31
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies lack effective identification methods for external target recognition, resulting in identification errors and the inability to detect new target categories in a timely manner, thus hindering the rapid expansion of the template library.
A target recognition method is constructed by combining Bayesian neural networks with variational inference and Monte Carlo sampling. The uncertainty of the network output is calculated to determine whether the target is outside or inside the database, and the target category is determined by using multiple prediction results.
It enables effective identification of targets outside the database, ensures correct identification of targets within the database, and promptly discovers new target categories, rapidly expanding the template database.
Smart Images
Figure CN115718897B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to target recognition technology, and more specifically, to a method and apparatus for identifying targets outside a database, applicable to target recognition problems of radar, electronic reconnaissance, satellite remote sensing, photoelectric and other detection equipment. Background Technology
[0002] Target identification is the process of determining the identity of a target using detection equipment such as radar, electronic reconnaissance, satellite remote sensing, and optoelectronics, based on historical experience and knowledge. It has wide applications in national defense and civilian sectors. Taking the national defense and military field as an example, target identification specifically refers to determining the target's identity information, such as whether it is military or civilian, friend or foe, country or region, type, model, and name. For instance, electronic reconnaissance equipment uses the electromagnetic signals emitted by the target to identify its type and even model; optical, infrared, and high-resolution radar equipment uses acquired optical images, infrared images, SAR images, or one-dimensional range images to identify the target's type, model, and even name; long-range early warning radar uses acquired target echo intensity information to identify the target's size and even whether it is military or civilian.
[0003] Traditional target recognition technology mainly consists of three parts: feature engineering, template library construction, and template matching. Each part is specifically as follows: First, feature engineering is performed, which involves constructing target features related to the target's identity based on the principles of radar, electronic reconnaissance, satellite remote sensing, photoelectric, and other detection equipment and the characteristics of the information acquired. Next, a template library is built, utilizing historical data to establish a target recognition template library. Each sample data in the library represents target feature data, and the sample label represents the target's true identity corresponding to the target feature; these labels are typically manually annotated. Finally, for newly acquired information from detection equipment, target feature data is calculated and matched against the template library to achieve target identification. With the rapid development and unprecedented success of deep learning artificial intelligence technologies, represented by convolutional neural networks, recurrent neural networks, and attention mechanisms, deep learning technology is widely used in various industries. Currently, deep learning-based target recognition technologies are developing rapidly, such as target recognition technology for remote sensing images, SAR images, or optical images based on convolutional neural networks, and radar one-dimensional range image target recognition technology based on recurrent neural networks. In terms of specific technical implementation, deep learning-based target recognition technology differs significantly from traditional target recognition technology. Deep learning-based target recognition technology trains neural networks, transforming the patterns and knowledge contained in the training dataset into the weights of the neural network. This directly achieves end-to-end learning of feature engineering and direct determination of target identity, blurring the boundary between feature engineering and template matching in traditional target recognition technology. However, in terms of the underlying principle of target recognition, they are consistent: both rely on historical information for target recognition. Traditional target recognition technology requires building a template library, while deep learning-based target recognition technology requires building a training dataset. Furthermore, the template library and training dataset are essentially the same thing, differing only in name.
[0004] It is evident that the construction of template libraries or training datasets is crucial for both traditional and deep learning-based target recognition technologies. For target recognition problems where the template library contains relevant category information (i.e., in-library target recognition), existing methods can achieve accurate identification. However, for target recognition problems where the template library does not contain relevant category information (i.e., out-of-library target recognition), existing methods lack effective technical measures for identification and often identify targets as if they were already in the library. This not only leads to target recognition errors but also prevents the timely discovery of new target categories, hindering the rapid expansion of the template library. Therefore, further research into out-of-library target recognition technologies is necessary. Summary of the Invention
[0005] The purpose of this invention is to provide a method and apparatus for identifying targets outside a database, offering an effective solution to the problem of identifying targets outside a database. The specific technical solution is as follows:
[0006] In a first aspect, embodiments of the present invention provide a method for identifying targets outside a database, the method comprising:
[0007] Based on the target recognition database, a target recognition Bayesian neural network f({W}) is constructed. The input dimension of the network input layer is determined by the dimension of the sample data in the target recognition database. The output dimension of the network output layer is consistent with the number of target categories in the target recognition database. The activation function of the output layer adopts the Softmax function. The network weights W follow a probability distribution, and the initial distribution is set to p0(W), where {W} represents the set of all weights in the target recognition Bayesian neural network.
[0008] The Bayesian network is approximated using variational inference, and the variational distribution of the network weights W is set as q. λ (W), the network loss function is Further employing the reparameterization technique, the variational distribution is represented as W = g(λ, ε), where λ is the variational distribution parameter to be optimized, KL(·) is the Kullback-Leibler divergence function, D is the training dataset constructed based on the target recognition database, consisting of sample data x and sample labels y, i.e., (x, y) ∈ D, y is a one-hot encoded vector with a size consistent with the number of target categories in the target recognition database, where the corresponding category bit of the sample data is 1 and the other bits are 0, q {λ} ({W}) represents the joint variational distribution of all network weights in the Bayesian neural network, p0({W}) represents the joint initial distribution of all network weights, p(D|f({W})) is the likelihood function corresponding to dataset D, g(·) is a differentiable function with respect to λ, ε is a random variable, and is a deterministic probability distribution that does not contain any parameters to be optimized.
[0009] In the training process of a Bayesian neural network, a single training sample is first obtained by sampling from the training dataset, and then the network weights W are determined by performing Monte Carlo sampling on ε. M =g(λ,ε), and then through the forward propagation of the neural network, calculate the network output corresponding to the network input sample data x. The Monte Carlo estimate of the loss function was then calculated. Where N is the number of training samples in a single run, and further based on the loss function estimation, a neural network optimization method is adopted. By using g(·) as an intermediate variable, the loss function is differentiated with respect to λ to achieve the backpropagation of the loss function gradient and the optimization of the variational distribution parameter λ. The above process is repeated to train the Bayesian neural network multiple times until the network's recognition ability meets the requirements, and finally the trained and optimized Bayesian neural network is obtained.
[0010] When using a Bayesian neural network to predict the target category of the data to be identified, the data to be identified is input into the trained Bayesian neural network M times, and the network output is calculated. For each input, ε is resampled using Monte Carlo methods, and then network predictions are made based on the obtained M iterations. Calculate the uncertainty of the target category prediction. If the uncertainty is greater than the set threshold, the target is determined to be an external target. If the uncertainty is less than the set threshold, the target is determined to be an internal target. Select the category with the highest average probability from M predictions as the target prediction category.
[0011] Secondly, embodiments of the present invention also provide an external target identification device, the device comprising:
[0012] The Bayesian neural network module is a specific target recognition neural network designed with the input dimension of the network input layer determined by the dimension of the sample data in the target recognition database, the output dimension of the network output layer consistent with the number of target categories in the target recognition database, and the network weights following a probability distribution.
[0013] The Bayesian neural network variational inference approximation module sets a variational distribution and uses reparameterization techniques to approximate the Bayesian neural network.
[0014] The Bayesian neural network loss function calculation module calculates the loss function during training according to the loss function calculation formula.
[0015] The Bayesian neural network training module optimizes the variational distribution parameters using a neural network optimization method based on the Bayesian neural network variational inference approximation results and a training dataset constructed using a target recognition database.
[0016] The target recognition and prediction module uses a trained Bayesian neural network to perform multiple network predictions on the data to be identified. Based on the results of these multiple predictions, it calculates the uncertainty of the target category prediction and determines whether the target is outside or inside the database based on the magnitude of the uncertainty. For targets inside the database, it further determines the target category based on the prediction results.
[0017] The present invention provides a method and apparatus for identifying targets outside the database, which can be applied to detection equipment such as radar, electronic reconnaissance, satellite remote sensing, and photoelectric devices. While ensuring the correct identification of targets within the database, it can effectively and automatically identify targets outside the database, thereby enabling the timely discovery of new target categories and rapid expansion of the target template database. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart of an off-database target identification method provided in an embodiment of the present invention;
[0020] Figure 2 This is a schematic diagram of the structure of an external target identification device provided in an embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] To effectively identify targets outside the database, embodiments of the present invention provide a method for identifying targets outside the database, such as... Figure 1 As shown, the method includes the following steps:
[0023] Step 1: Based on the target recognition database, construct a Bayesian neural network f({W}). The input dimension of the network input layer is determined by the dimension of the sample data in the target recognition database. The output dimension of the network output layer is consistent with the number of target categories in the target recognition database. The activation function of the output layer adopts the Softmax function. The network weights W follow a probability distribution, and the initial distribution is set to p0(W), where {W} represents the set of all weights in the target recognition Bayesian neural network.
[0024] Step 1.1: Based on the target recognition database, construct a target recognition Bayesian neural network f({W}). The input dimension of the network input layer is determined by the dimension of the sample data in the target recognition database. The output dimension of the network output layer is consistent with the number of target categories in the target recognition database. The activation function of the output layer adopts the Softmax function. The network weights W follow a probability distribution, and the initial distribution is set to p0(W), where {W} represents the set of all weights in the target recognition Bayesian neural network.
[0025] Step 1.2: The sample data types in the target recognition database can be video images, remote sensing, optical, infrared, SAR and other images, and vector or time-series vector data such as speech, one-dimensional range images, electromagnetic signals, and tracks;
[0026] Step 1.3: A feasible method for constructing a Bayesian neural network for target recognition is to modify existing typical target recognition networks and reset the network weights to follow a probability distribution, thereby quickly creating a Bayesian neural network for target recognition. Examples include VGG, Inception, ResNet, and DenseNet networks that can be used for image data classification, and LSTM, GRU, Transformer, and fully connected networks that can be used for vector or time-series vector data.
[0027] Step 1.4: Another feasible method for constructing a Bayesian neural network for target recognition is to design and construct a brand-new Bayesian neural network based on a target recognition database and neural network modules such as fully connected layers, convolutional layers, recurrent network layers, attention mechanisms, and self-attention mechanisms, setting the network weights to follow a probability distribution.
[0028] Step 1.5: The initial probability of the network weights in the Bayesian neural network for target recognition represents the initial information that the network weights can obtain. It can be set to any probability distribution. For the convenience of subsequent loss function calculation, it is generally set to a normal distribution with a mean of 0 and a variance of 1, i.e., p0(W) = N(0,1).
[0029] Step 1.6: In the constructed Bayesian neural network for target recognition, some network weights can be set to fixed weights. It should be noted that fixed weights can be considered as a special probability distribution, where the probability of a weight at a fixed value is 1, and the probability of other values is 0.
[0030] Step 2: Approximate the Bayesian network using variational inference, and set the variational distribution of the network weights W as q. λ (W), the network loss function is Further employing the reparameterization technique, the variational distribution is represented as W = g(λ, ε), where λ is the variational distribution parameter to be optimized, KL(·) is the Kullback-Leibler divergence function, D is the training dataset constructed based on the target recognition database, consisting of sample data x and sample labels y, i.e., (x, y) ∈ D, y is a one-hot encoded vector with a size consistent with the number of target categories in the target recognition database, where the corresponding category bit of the sample data is 1 and the other bits are 0, q {λ} ({W}) represents the joint variational distribution of all network weights in the Bayesian neural network, p0({W}) represents the joint initial distribution of all network weights, p(D|f({W})) is the likelihood function corresponding to dataset D, g(·) is a differentiable function with respect to λ, ε is a random variable, and is a deterministic probability distribution that does not contain any parameters to be optimized.
[0031] Step 2.1: Approximate the Bayesian network using variational inference, and set the variational distribution of the network weights W as q. λ (W), the network loss function is KL(·) is the Kullback-Leibler divergence function, D is the training dataset constructed based on the target recognition database, consisting of sample data x and sample labels y, i.e., (x,y)∈D, y is a one-hot encoded vector, the size of which is consistent with the number of target categories in the target recognition database, the corresponding category bit of the sample data is 1, and the other bits are 0, q {λ}( {W}) represents the joint variational distribution of all network weights in the Bayesian neural network, p0({W}) represents the joint initial distribution of all network weights, and p(D|f({W})) is the likelihood function corresponding to dataset D;
[0032] Step 2.2: Further employ the reparameterization technique to represent the variational distribution as W = g(λ, ε), so that the loss gradient can be backpropagated during network training. g(·) is a differentiable function with respect to λ, and ε is a random variable, which is a deterministic probability distribution and does not contain any parameters to be optimized.
[0033] Step 2.3: Generally, it can be assumed that the weights of the Bayesian neural network are independent of each other. In this case, the network loss function can be expressed as: The network loss function includes KL[q] λ The exact or approximate expression for the part (W)||p0(W)] can generally be obtained;
[0034] Step 2.4: One method for setting the variational distribution of network weights is to set q. λ (W) is a Gaussian distribution, i.e. Using the reparameterization technique, the network weights can be expressed as W = g(λ, ε) = λ μ +λ σ ·ε, where ε follows a normal distribution with mean 0 and variance 1. During training, g(λ,ε) varies with respect to λ. μ The derivative is 1 with respect to λ. σ The derivative is ε M , ε M For Monte Carlo sampling values, further, when the initial distribution p0(W) is a normal distribution with a mean of 0 and a variance of 1, i.e., p0(W) = N(0,1), the following can be obtained.
[0035] Step 2.5: One method for setting the variational distribution of network weights is to set q. λ (W) is a mixture Gaussian distribution, i.e. Where ρ is the hybrid weight, and using the reparameterization technique, the network weight can be expressed as W = g(λ, ε) = ε. B(λ μ +λ σ ·ε N )+(1-ε B )(λ σ ·ε N ), where ε B It follows a 0-1 Bernoulli distribution with parameter ρ, i.e., ε B The probability of ε being 1 is ρ, and the probability of ε being 0 is 1-ρ. During the training process, ε B When the Monte Carlo sampling is 1, g(λ,ε) with respect to λ μ The derivative is 1 with respect to λ. σ The derivative is ε B When the Monte Carlo sampling is 0, g(λ,ε) with respect to λ μ The derivative is 0 with respect to λ. σ The derivative is ε is a normal distribution N The Monte Carlo sampling values can be further obtained when the initial distribution p0(W) is a normal distribution with a mean of 0 and a variance of 1, i.e., p0(W) = N(0,1).
[0036] Step 2.6: The Dropout method, which randomly sets the network weights to 0 or keeps them at their original values with a certain probability, can be approximated as the variational inference approximation Bayesian neural network corresponding to Step 2.5. Thus, the Dropout method can be used to approximate the Bayesian neural network.
[0037] Step 3: During the training of the Bayesian neural network, firstly, single training samples are sampled from the training dataset, and then the network weights W are determined by performing Monte Carlo sampling on ε. M =g(λ,ε), and then through the forward propagation of the neural network, calculate the network output corresponding to the network input sample data x. The Monte Carlo estimate of the loss function was then calculated. Where N is the number of training samples in a single run, and further based on the loss function estimation, a neural network optimization method is adopted. By using g(·) as an intermediate variable, the loss function is differentiated with respect to λ to achieve the backpropagation of the loss function gradient and the optimization of the variational distribution parameter λ. The above process is repeated to train the Bayesian neural network multiple times until the network's recognition ability meets the requirements, and finally the trained and optimized Bayesian neural network is obtained.
[0038] Step 3.1: In the Bayesian neural network training process, firstly, single training samples are sampled from the training dataset, and then the network weights W are determined by performing Monte Carlo sampling on ε. M=g(λ,ε), and then through the forward propagation of the neural network, calculate the network output corresponding to the network input sample data x. The Monte Carlo estimate of the loss function was then calculated. Where N is the number of training samples in a single run;
[0039] Step 3.2: Further based on the loss function estimation, a neural network optimization method is adopted. By taking g(·) as an intermediate variable, the derivative of the loss function with respect to λ is calculated to realize the backpropagation of the loss function gradient and the optimization of the variational distribution parameter λ.
[0040] Step 3.3: Repeat steps 3.1 and 3.2 to train the Bayesian neural network multiple times until the network's recognition ability meets the requirements, and finally obtain the trained and optimized Bayesian neural network.
[0041] Step 3.4: A reasonable method for calculating the log-likelihood component of the loss function is to use the cross-entropy function, i.e. The corresponding Monte Carlo estimation of the loss function is: Where y k This represents the k-th element of the truth vector y for the target category. Represents the target category prediction vector The value of the kth element, where K represents the number of target categories;
[0042] Step 3.5: During training, a single training sample can be a single sample, and the corresponding loss function is... It can also be a batch of multiple samples. In this case, the Flipout technique can be used to accelerate the learning process, reduce the correlation between the calculation results of the batch samples, and improve the training and learning speed of the network.
[0043] Step 3.6: The neural network optimization methods that can be used include batch gradient descent, stochastic gradient descent, momentum method, Nesterov accelerated gradient descent, Adagrad algorithm, Adadelta algorithm, RMSprop algorithm, and Adam algorithm;
[0044] Step 4: When using a Bayesian neural network to predict the target category of the data to be identified, the data to be identified is input into the trained Bayesian neural network M times, and the network output is calculated. For each input, ε is resampled using Monte Carlo methods, and then network predictions are made based on the obtained M iterations. Calculate the uncertainty of target category prediction. If the uncertainty is greater than the set threshold, the target is determined to be an external target. If the uncertainty is less than the set threshold, the target is determined to be an internal target. Select the category with the highest average probability of M predictions as the target prediction category.
[0045] Step 4.1: When using a Bayesian neural network to predict the target category of the data to be identified, the data to be identified is input into the trained Bayesian neural network M times, and the network output is calculated. For each input, ε is resampled using Monte Carlo methods. This means that for the same input, the Bayesian neural network weights used in the output calculation are different each time, resulting in different network outputs. Ultimately, this yields M network predictions. That is, there will be M prediction results for each target category;
[0046] Step 4.2: Then, based on the obtained M network predictions... Calculate the uncertainty of the target category prediction. If the uncertainty is greater than a set threshold, the target is determined to be an external target; if the uncertainty is less than the set threshold, the target is determined to be an internal target. The category with the highest average probability from M predictions is selected as the target prediction category. in k F Predict the category for the target;
[0047] Step 4.3: A reasonable method for calculating the uncertainty of target category prediction is as follows: in This represents the value of the k-th element in the prediction result vector in the m-th Bayesian network prediction, where M represents the number of predictions made by the same input Bayesian network.
[0048] Step 4.4: Another reasonable method for calculating the uncertainty of target category prediction is as follows: in This represents the value of the k-th element in the prediction result vector in the m-th Bayesian network prediction, where M represents the number of predictions made by the same input Bayesian network.
[0049] Step 4.5: Another reasonable method for calculating the uncertainty of target category prediction is as follows: in This represents the value of the k-th element in the prediction result vector in the m-th Bayesian network prediction, where M represents the number of predictions made by the same input Bayesian network.
[0050] This invention also provides an external target identification device, such as... Figure 2 As shown, the device includes:
[0051] The Bayesian neural network module is a specific target recognition neural network designed with the input dimension of the network input layer determined by the dimension of the sample data in the target recognition database, the output dimension of the network output layer consistent with the number of target categories in the target recognition database, and the network weights following a probability distribution.
[0052] The Bayesian neural network variational inference approximation module sets a variational distribution and uses reparameterization techniques to approximate the Bayesian neural network.
[0053] The Bayesian neural network loss function calculation module calculates the loss function during training according to the loss function calculation formula.
[0054] The Bayesian neural network training module optimizes the variational distribution parameters using a neural network optimization method based on the Bayesian neural network variational inference approximation results and a training dataset constructed using a target recognition database.
[0055] The target recognition and prediction module uses a trained Bayesian neural network to perform multiple network predictions on the data to be identified. Based on the results of these multiple predictions, it calculates the uncertainty of the target category prediction and determines whether the target is outside or inside the database based on the magnitude of the uncertainty. For targets inside the database, it further determines the target category based on the prediction results.
[0056] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.
[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A method for identifying targets outside a database, characterized in that, Applicable to radar, satellite remote sensing, or photoelectric detection equipment, including the following steps: Step 1: Construct a Bayesian neural network for target recognition based on the target recognition database. The input dimension of the network input layer is determined by the dimension of the sample data in the target recognition database, and the output dimension of the network output layer is consistent with the number of target categories in the target recognition database. The activation function of the output layer adopts... Functions, network weights It follows a probability distribution, with the initial distribution set as follows: ,in This represents the set of all weights of the Bayesian neural network for target recognition; the sample data types in the target recognition database include video images, remote sensing, infrared, SAR images, speech, and one-dimensional distance images. Step 2: Approximate the Bayesian network using variational inference and set the network weights. The variational distribution is The network loss function is Furthermore, by employing reparameterization techniques, the variational distribution is represented as ,in The parameters to be optimized are for a variational distribution. The Kullback-Leibler divergence function, The training dataset is constructed based on a target recognition database, consisting of sample data. and sample labels Composition, that is , This is a one-hot encoded vector, with a size consistent with the number of target categories in the target recognition database. The corresponding category bit in the sample data is 1, and the other bits are 0. This represents the joint variational distribution of all network weights in a Bayesian neural network. This represents the joint initial distribution of all network weights. For dataset The corresponding likelihood function, For about Differentiable functions, For random variables, with a defined probability distribution, there are no parameters to be optimized; Step 3: In the training process of the Bayesian neural network, firstly, single training samples are sampled from the training dataset, and then... Monte Carlo sampling is performed to determine the network weights. Then, through the forward propagation of the neural network, the input sample data of the network is calculated. Corresponding network output The Monte Carlo estimate of the loss function was calculated. ,in To determine the number of training samples per run, further optimization is performed based on loss function estimation using a neural network approach. As an intermediate variable, the loss function is related to By performing differentiation, the backpropagation of the loss function gradient and the variational distribution parameters are realized. To optimize the Bayesian neural network, repeat the above process multiple times until the network's recognition ability meets the requirements, and finally obtain the optimized Bayesian neural network. Step 4: When using a Bayesian neural network to predict the target category of the data to be identified, The input data to be identified is fed into the trained Bayesian neural network, and the network output is calculated. For each input, re-process it. Monte Carlo sampling was performed, and then based on the obtained Sub-network prediction The system calculates the uncertainty of target category prediction. If the uncertainty is greater than a set threshold, the target is determined to be an external target; if the uncertainty is less than the set threshold, the target is determined to be an internal target and selected. The category with the highest average probability in the predictions is selected as the target prediction category.
2. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 1, a feasible method for constructing a Bayesian neural network for target recognition is to modify existing typical target recognition networks, reset the network weights to follow a probability distribution, and thus quickly create a Bayesian neural network for target recognition. This includes VGG networks, Inception networks, ResNet networks, and DenseNet networks for image data classification, and LSTM networks, GRU networks, Transformer networks, and fully connected networks for vector or temporal vector data.
3. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 2, it can be assumed that the weights of the Bayesian neural network are independent of each other. In this case, the network loss function can be expressed as: Among them, the network loss function The expression is partially obtained.
4. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 2, one method for setting the variational distribution of network weights is to set... It is a Gaussian distribution, that is Using the reparameterization technique, the network weights can be expressed as: ,in It follows a normal distribution with a mean of 0 and a variance of 1. During the training process, about The derivative is 1, about The derivative is , For Monte Carlo sampling values, further, when the initial distribution When the distribution is a normal distribution with a mean of 0 and a variance of 1, that is... When, it can be obtained .
5. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 2, one method for setting the variational distribution of network weights is to set... It is a mixture Gaussian distribution, i.e. ,in To achieve hybrid weights, a reparameterization technique is employed, and the network weights can be represented as follows: ,in Obtain the parameter as The 0-1 Bernoulli distribution, i.e. The probability of being 1 is The probability that it is 0 is During the training process, When the Monte Carlo sampling is 1, about The derivative is 1, about The derivative is , When the Monte Carlo sampling is 0, about The derivative is 0, with respect to The derivative is , Normal distribution Monte Carlo sampling values, further in the initial distribution When the distribution is a normal distribution with a mean of 0 and a variance of 1, that is... When, it can be obtained .
6. The method for identifying targets outside a database as described in claim 5, characterized in that, The Dropout method can be used to approximate a Bayesian neural network.
7. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 3, the log-likelihood component of the loss function is calculated using the cross-entropy function, i.e. The corresponding Monte Carlo estimation of the loss function is: ,in Represents the truth vector of the target category The Each element value Represents the target category prediction vector The Each element value Indicates the number of target categories.
8. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 4, a reasonable method for calculating the uncertainty of target category prediction is as follows: ,in , Indicates the first In the prediction of the second-order Bayesian network, the prediction result vector is the first... Each element value This indicates the number of predictions made by the Bayesian network for the same input; Indicates the number of target categories.
9. The method for identifying targets outside a database as described in claim 1, characterized in that, In step 4, a reasonable method for calculating the uncertainty of target category prediction is as follows: ,in , Indicates the first In the prediction of the second-order Bayesian network, the prediction result vector is the first... Each element value This indicates the number of predictions made by the Bayesian network for the same input; Indicates the number of target categories.
10. A target identification device outside a storage area, characterized in that, Suitable for radar, satellite remote sensing, or photoelectric detection equipment, the device includes: The Bayesian neural network module is a specific target recognition neural network. The input dimension of the network input layer is determined by the dimension of the sample data in the target recognition database, and the output dimension of the network output layer is consistent with the number of target categories in the target recognition database. The network weights follow a probability distribution. The sample data types in the target recognition database include video images, remote sensing, infrared, SAR images, speech, and one-dimensional distance images. The Bayesian neural network variational inference approximation module sets a variational distribution and uses reparameterization techniques to approximate the Bayesian neural network. The Bayesian neural network loss function calculation module calculates the loss function during training according to the loss function calculation formula. The Bayesian neural network training module optimizes the variational distribution parameters using a neural network optimization method based on the Bayesian neural network variational inference approximation results and a training dataset constructed using a target recognition database. The target recognition and prediction module uses a trained Bayesian neural network to perform multiple network predictions on the data to be identified. Based on the results of these multiple predictions, it calculates the uncertainty of the target category prediction and determines whether the target is outside or inside the database based on the magnitude of the uncertainty. For targets inside the database, it further determines the target category based on the prediction results.