Pulse sequence identification method and device based on self-attention mechanism and medium

Abstract

The invention discloses a pulse sequence recognition method and device based on a self-attention mechanism and a medium, and relates to the field of digital signal processing and deep learning, on one hand, through recurrent neural network modeling based on the self-attention mechanism, the feature weight of the pulse time sequence dimension can be improved, and the recognition accuracy is effectively improved; and on the other hand, based on normalization of global parameters, the distinguishability of a pulse sequence in parameter dimensions such as frequency is weakened, and the distinguishability of a sample in the parameter dimensions such as frequency is improved by adopting a grouping multi-model thought in consideration of the data characteristic of non-uniform distribution of signals.

Description

Technical Field

[0001] The present invention relates to the fields of digital signal processing and deep learning. It mainly models pulse sequence samples through a recurrent neural network based on a self-attention mechanism to improve the feature weight of the time dimension. At the same time, combined with the signal distribution law, a pulse sequence recognition method, device and medium based on the self-attention mechanism are proposed. By constructing multiple lightweight recognition network models, the distinguishability of each group of sample data is improved, thereby improving the recognition accuracy of complex system targets. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] With the continuous development of electronic technology and digital signal processing technology, the current electromagnetic environment presents the characteristics of dense targets and complex and changeable objects, which brings great challenges to target recognition.

[0004] The traditional template matching recognition method builds a recognition knowledge base through expert knowledge, and realizes the matching recognition of radar radiation sources with known radiation sources in the knowledge base based on conventional parameters. This method is simple and fast, and is widely used in scenes with simple electromagnetic environment and low signal density.

[0005] At present, the changes in radiation source parameters are complex, and template matching recognition technology is no longer applicable. In recent years, with the rise of machine learning and artificial intelligence recognition technology, support vector machines (SVM), CNN, LSTM and other technologies have shown strong recognition capabilities and have gradually become a hot research direction in the field of radiation source recognition. Intelligent recognition technology has been continuously applied to the field of signal processing.

[0006] At present, the intelligent target recognition technology mainly adopts the following two technical ideas: (1) Model training is performed using a single pulse parameter as a sample. The advantage of this method is that the model is simple, but the disadvantage is that the sample dimension loses the timing information between pulses, which reduces the recognition accuracy. (2) Model training is performed using a pulse sequence as a sample, and a global model is constructed based on a serialized neural network. The advantage of this method is that the sample information is more complete, but the problem is that for unevenly distributed sample data, the indifferent normalization preprocessing reduces the distinguishability of targets in concentrated distribution areas. Summary of the invention

[0007] The purpose of the present invention is to solve the problem of identifying targets of replication systems. In engineering practice, radiation source identification uses pulses as input, and simple pulse-to-pulse parameters cannot fully represent target features. The pulse sequence contains implicit features before and after the timing sequence, so the sequence-based deep learning model is more suitable for solving the problem of radiation source target identification. As a neural network with short-term memory ability, the recurrent neural network has a good fitting and modeling effect on sequences that are sensitive to the order of precedence. Among them, the self-attention mechanism in the Transformer can extract the correlation between different positions, capture richer contextual information, and can achieve better results in combination with the recurrent neural network. Therefore, the present invention can improve the feature weight of the pulse time series dimension and effectively improve the recognition accuracy by modeling the recurrent neural network based on the self-attention mechanism. On the other hand, according to the conventional data preprocessing method in the field of deep learning, the normalization based on the global parameters will cause the pulse sequence to weaken in the parameter dimensions such as frequency. Therefore, the present invention provides a pulse sequence recognition method, device and medium based on the self-attention mechanism. According to the distribution characteristics of the target signal, multiple groups of network recognition models are constructed to ensure the distinguishability of the internal features of each group of models. On the one hand, the resolution loss caused by data normalization is reduced, and on the other hand, the convergence of the model is guaranteed, thereby ultimately improving the recognition accuracy of the replication system target.

[0008] The technical solution of the present invention is as follows:

[0009] A pulse train recognition method based on a self-attention mechanism, comprising:

[0010] Step S1: construct a network model based on self-attention mechanism;

[0011] Step S2: using the sample sequence to train the network model;

[0012] Step S3: Perform real-time model recognition on the current pulse sequence through the trained network model.

[0013] Furthermore, the step S1 comprises:

[0014] First, the self-attention mechanism is used to re-characterize the correlation between sequences, and the extracted sequence features establish the dependency on the entire sequence;

[0015] After extracting sequence features through the self-attention mechanism, the network model is subsequently connected to a recurrent gating unit GRU. The recurrent gating unit GRU combines the forget gate and the input gate together on the basis of the long short-term memory neural network LSTM, reducing the number of gates and thus improving training efficiency;

[0016] After being processed by the recurrent gating unit GRU, a 32-dimensional recognition result is output. In order to map the conclusion to the classification dimension, a fully connected layer is connected after the recurrent gating unit GRU network to complete the modeling of the entire network model.

[0017] Furthermore, the calculation formula of the self-attention mechanism is as follows:

[0018]

[0019] The input of the self-attention mechanism is a pulse sample sequence Y=(y1, ..y i ...y 32 ), where y i =(Δpri i ,Δrf i ,Δpw i ), is the pulse sequence after data preprocessing; after extracting the correlation features according to the self-attention mechanism, an output Z of the same dimension is obtained = (z1, ..z i ...z 32 ).

[0020] Furthermore, the calculation formula of the recurrent gating unit GRU is as follows:

[0021] r t =σ(W ir x t +b ir +W hr h t-1 +b hr )

[0022] z t =σ(W iz x t +b iz +W hz h t-1 +b hz )

[0023] n t =tanh(W in x t +b in +r t ×(W hn h t-1 +b hn ))

[0024] h t =(1-z t )×n t +z t ×h t-1

[0025] The input of the GRU network is Z = (z1, ..z i ...z 32 ), so the number of neurons in the network is consistent with the length of the input sequence, the number of neurons is also 32, and the number of layers in the network is set to 1.

[0026] Furthermore, the step S2 comprises:

[0027] Step S21: perform data preprocessing;

[0028] Step S22: training the network model based on the preprocessed data.

[0029] Furthermore, the step S21 includes:

[0030] First input the sample sequence X=(x1,x2..x i ...x n ) is normalized and preprocessed, where x i =(pri i ,rf i ,pw i ), represents a set of pulse parameters; for each x i =(pri i ,rf i ,pw i ) to obtain the preprocessed y i =(Δpri i ,Δrf i ,Δpw i ).

[0031] Furthermore, the preprocessing includes the following formula:

[0032]

[0033]

[0034]

[0035] Among them, rf min ,rf max ,pri max ,pri min ,pw max ,pw min The value of is derived from the data distribution of the sample pile; taking rf as an example, first obtain the minimum value min(rf i ) and max(rf i), according to the difference between the maximum and minimum values, a certain proportion of expansion is made, where DeltaRf is 1 / 4 of the current sample value range and the sample frequency band range. The specific formula is as follows, where H rf , L rf Indicates the maximum and minimum values ​​of the frequency band of the sample set;

[0036] rf min =max(((1+DeltaRf)*min(rf i )-DeltaRf*max(rf i )),L rf )

[0037] rf max =min(((1+DeltaRf)*max(rf i )-DeltaRf*min(rf i )),H rf )

[0038]

[0039] The rf obtained through the sample training phase min ,rf max ,pri max ,pri min ,pw max ,pw min The data preprocessing that will be synchronized to the real-time recognition processing will be normalized with the same maximum and minimum values ​​to ensure the consistency of features after parameter normalization in the training and recognition stages.

[0040] Furthermore, the step S3 comprises:

[0041] First, determine whether the current pulse sequence length meets the input conditions. When the pulse sequence length is greater than or equal to the sequence length of the network model input, divide the pulses using the model input length as a window to construct multiple groups of inputs. For each group of input pulse sequence inter-pulse statistical parameters, select the corresponding lightweight model. Obtain the maximum and minimum values ​​generated in the network model training phase for normalization preprocessing of the real-time recognition sequence. Next, perform model recognition based on the self-attention mechanism to obtain the model recognition conclusion. Finally, based on the model recognition conclusion, fuse the results to form the target conclusion.

[0042] A pulse sequence recognition device based on a self-attention mechanism comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a pulse sequence recognition method based on a self-attention mechanism as described above are implemented.

[0043] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a pulse sequence recognition method based on a self-attention mechanism as described above are implemented.

[0044] Pulse sequence recognition is an important output in the field of digital signal processing, and its results directly affect situation generation. Different from the traditional target recognition idea, this invention effectively combines the sequence recognition model in the field of deep learning, introduces the self-attention mechanism and the recurrent neural network, constructs a network structure that is sensitive to the correlation between the previous and next sequences, and improves the recognition weight of the repetitive cycle parameter dimension. In terms of data preprocessing, considering the data characteristics of the uneven distribution of the signal itself, the idea of ​​grouping multiple models is adopted to improve the distinguishability of samples in parameter dimensions such as frequency. This method has better scalability and robustness, ensures that each dimensional parameter can maintain a good resolution, and effectively improves the accuracy of complex target recognition.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. Compared with the radiation source identification algorithm based on template matching, the model is used to extract parameter features, which can better characterize the parameter characteristics of complex system targets, adapt to signals with more complex changes in inter-pulse parameter combinations, and have a higher recognition accuracy.

[0047] 2. Compared with the algorithm using neural network for radiation source identification, the self-attention mechanism is introduced to increase the receptive field of time dimension features and improve the recognition weight of repeated period dimension features. In addition, the preprocessing method of grouping multiple models avoids the information loss caused by normalization and maintains the distinguishability of data features;

[0048] 3. The present invention, by combining signal recognition technology with deep learning technology, can adapt to target recognition of complex signals, extract feature information of the before-and-after correlation of sequence time dimensions that is difficult to describe by methods such as template matching through the self-attention mechanism, and use the preprocessing method of grouping lightweight small models to improve the distinguishability of the recognition space, which can have a better recognition effect on complex changing parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Compute the architecture for the self-attention mechanism;

[0050] Figure 2 It is the GRU unit network structure;

[0051] Figure 3 To identify flowcharts in real time based on self-attention mechanism;

[0052] Figure 4 This is the S-band model training accuracy curve in Example 2;

[0053] Figure 5 This is the S-band model training loss curve in Example 2;

[0054] Figure 6 The confusion matrix for the S-band test set identification in Example 2;

[0055] Figure 7 This is the X-band model training accuracy curve in Example 2;

[0056] Figure 8 This is the X-band model training loss curve in Example 2;

[0057] Fig. 9 This is the confusion matrix for identifying the X-band test set in Example 2. DETAILED DESCRIPTION

[0058] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0059] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0060] Embodiment 1

[0061] A pulse train recognition method based on a self-attention mechanism specifically comprises the following steps:

[0062] Step S1: construct a network model based on self-attention mechanism;

[0063] Step S2: using the sample sequence to train the network model;

[0064] Step S3: Perform real-time model recognition on the current pulse sequence through the trained network model.

[0065] In this embodiment, specifically, step S1 includes:

[0066] First, the self-attention mechanism is used to re-characterize the correlation between sequences, and the extracted features establish the dependency on the entire sequence. The computational architecture of the self-attention mechanism is as follows: Figure 1 The calculation formula is as follows:

[0067]

[0068] The input of the self-attention mechanism is a pulse sample sequence Y=(y1, ..y i ...y 32 ), where y i =(Δpri i ,Δrf i ,Δpw i ), is the pulse sequence after data preprocessing. After extracting the correlation features according to the self-attention mechanism, an output Z of the same dimension is obtained = (z1, ..z i ...z 32 After extracting sequence features through the self-attention mechanism, the model is subsequently connected to a recurrent gate unit (GRU). GRU combines the forget gate and the input gate together on the basis of LSTM, reducing the number of gates and thus improving training efficiency. Figure 2 As shown, the calculation formula for each GRU unit is as follows:

[0069] r t =σ(W ir x t +b ir +W hr h t-1 +b hr )

[0070] z t =σ(W iz x t +b iz +W hz h t-1 +b hz )

[0071] n t =tanh(W in x t +b in +r t ×(W hn h t-1 +b hn ))

[0072] h t =(1-z t )×n t +zt ×h t-1

[0073] The input of the GRU network is Z = (z1, ..z i ...z 32 ), so the number of neurons in the network is consistent with the length of the input sequence, the number of neurons is also 32, and the number of layers of the network is set to 1. After GRU processing, a 32-dimensional recognition result is output. In order to map the conclusion to the classification dimension, a fully connected layer is connected after the GRU network to complete the modeling of the entire model.

[0074] In this embodiment, specifically, step S2 includes:

[0075] First input the sample sequence X=(x1,x2..x i ...x n ) is normalized and preprocessed, where x i =(pri i ,rf i ,pw i ), represents a set of pulse parameters. For each x i =(pri i ,rf i ,pw i ) According to the following formula, we can get the preprocessed y i =(Δpri i ,Δrf i ,Δpw i )

[0076]

[0077]

[0078]

[0079] Among them, rf min ,rf max ,pri max ,pri min ,pw max ,pw min The value of comes from the data distribution of the sample pile. Taking rf as an example, first get the minimum value min(rf i ) and max(rf i ), according to the difference between the maximum and minimum values, a certain proportion of expansion is made, where DeltaRf is 1 / 4 of the current sample value range and the sample frequency band range. The specific formula is as follows, where H rf , L rf Indicates the maximum and minimum values ​​of the frequency band of the sample set.

[0080] rf min =max(((1+DeltaRf)*min(rf i )-DeltaRf*max(rf i )),L rf )

[0081] rf max =min(((1+DeltaRf)*max(rf i )-DeltaRf*min(rf i )),H rf )

[0082]

[0083] The rf obtained through the sample training phase min ,rf max ,pri max ,pri min ,pw max ,pw min The data preprocessing that will be synchronized to the real-time recognition processing will be normalized with the same maximum and minimum values ​​to ensure the consistency of features after parameter normalization in the training and recognition stages.

[0084] In this embodiment, specifically, the overall process of real-time model recognition based on the self-attention mechanism is as follows: Figure 3 As shown, the step S3 includes:

[0085] Since the input requirement of the model is a full pulse sequence, we first determine whether the current pulse sequence length meets the input conditions. When the pulse sequence length is greater than or equal to the sequence length of the network model input, the pulses are divided into windows using the model input length to construct multiple groups of inputs. For each group of input pulse sequence inter-pulse statistical parameters, the corresponding lightweight model is selected. The maximum and minimum values ​​generated during the network model training phase are obtained for normalization preprocessing of the real-time recognition sequence. Next, model recognition based on the self-attention mechanism is performed to obtain the model recognition conclusion. Finally, based on the model recognition conclusion, the results are fused to form the target conclusion.

[0086] This embodiment also proposes a pulse sequence recognition device based on the self-attention mechanism, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of a pulse sequence recognition method based on the self-attention mechanism as described above when executing the computer program; preferably, the computer program can be executed on a terminal device, such as a personal computer.

[0087] This embodiment also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of a pulse sequence recognition method based on a self-attention mechanism as described above are implemented.

[0088] However, the apparatus of the present invention is not limited thereto, and in this document, a readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0089] The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more conductors, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0090] The computer readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by an instruction execution system, an apparatus, or a device or used in combination with it. The program code contained on the readable storage medium may be transmitted with any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0091] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0092] Embodiment 2

[0093] The present invention provides a radiation source identification method based on the self-attention mechanism, which can be applied to radiation source identification to complete the identification of complex signals. In order to prove the effectiveness of the method, the model was tested on the training set and the test set for two sample subsets with frequency distribution in the S and X bands. For the sample subsets, the samples were first divided into a test set and a training set in a ratio of 3:7, and the two were independent of each other. Figure 4 and Figure 5 The accuracy and loss curves of S-band model training are given. Figure 7 and Figure 8 The accuracy and loss rate curves of the X-band model training are given. As can be seen from the figure, after 100 rounds of iterative training, the model gradually converged, the loss on the training set and the validation set continued to decrease, and the recognition accuracy continued to increase. Finally, the accuracy of the model after training on the training set and the validation set can reach 97%. Then, the S and X-band models were tested on the test set. Figure 6 and Fig. 9 The confusion matrix of the model after recognition on the test set is given respectively. As can be seen from the figure, the S-band model recognized a total of 5796 sequences, of which 5789 were correctly recognized, 7 were incorrectly recognized, and the recognition accuracy was 99.87%. The X-band model recognized a total of 2478 sequences, of which 2458 were correctly recognized, 20 were incorrectly recognized, and the recognition accuracy was 99.19%.

[0094] In summary, the method proposed in the present invention can improve the accuracy of radiation source identification through continuous iterative training of the network model, and has certain scalability and practicality.

[0095] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

[0096] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A pulse train recognition method based on self-attention mechanism, characterized in that: include: Step S1: construct a network model based on self-attention mechanism; Step S2: using the sample sequence to train the network model; Step S3: Perform real-time model recognition on the current pulse sequence through the trained network model.

2. A pulse sequence recognition method based on self-attention mechanism according to claim 1, characterized in that: The step S1 comprises: First, the self-attention mechanism is used to re-characterize the correlation between sequences, and the extracted sequence features establish the dependency on the entire sequence; After extracting sequence features through the self-attention mechanism, the network model is subsequently connected to a recurrent gating unit GRU. The recurrent gating unit GRU combines the forget gate and the input gate together on the basis of the long short-term memory neural network LSTM, reducing the number of gates and thus improving training efficiency; After being processed by the recurrent gating unit GRU, a 32-dimensional recognition result is output. In order to map the conclusion to the classification dimension, a fully connected layer is connected after the recurrent gating unit GRU network to complete the modeling of the entire network model.

3. A pulse sequence recognition method based on self-attention mechanism according to claim 2, characterized in that: The calculation formula of the self-attention mechanism is as follows: The input of the self-attention mechanism is a pulse sample sequence Y=(y1, ..y i ...y 32 ), where y i =(Δpri i ,Δrf i ,Δpw i ), is the pulse sequence after data preprocessing; after extracting the correlation features according to the self-attention mechanism, an output Z of the same dimension is obtained = (z1, ..z i ...z 32 ).

4. The pulse sequence recognition method based on the self-attention mechanism according to claim 3, characterized in that: The calculation formula of the recurrent gating unit GRU is as follows: r t =σ(W ir x t +b ir +W hr h t-1 +b hr ) z t =σ(W iz x t +b iz +W hz h t-1 +b hz ) n t =tanh(W in x t +b in +r t ×(W hn h t-1 +b hn )) h t =(1-z t )×n t +z t ×h t-1 The input of the GRU network is Z = (z1, ..z i ...z 32 ), so the number of neurons in the network is consistent with the length of the input sequence, the number of neurons is also 32, and the number of layers in the network is set to 1.

5. The pulse sequence recognition method based on the self-attention mechanism according to claim 1, characterized in that: The step S2 comprises: Step S21: perform data preprocessing; Step S22: training the network model based on the preprocessed data.

6. A pulse sequence recognition method based on self-attention mechanism according to claim 5, characterized in that: The step S21 comprises: First input the sample sequence X=(x1,x2..x i ...x n ) is normalized and preprocessed, where x i =(pri i ,rf i ,pw i ), represents a set of pulse parameters; for each x i =(pri i ,rf i ,pw i ) to obtain the preprocessed y i =(Δpri i ,Δrf i ,Δpw i ).

7. The pulse sequence recognition method based on the self-attention mechanism according to claim 6, characterized in that: The preprocessing includes the following formula: Among them, rf min ,rf max ,pri max ,pri min ,pw max ,pw min The value of is derived from the data distribution of the sample pile; taking rf as an example, first obtain the minimum value min(rf i ) and max(rf i ), according to the difference between the maximum and minimum values, a certain proportion of expansion is made, where DeltaRf is 1 / 4 of the current sample value range and the sample frequency band range. The specific formula is as follows, where H rf , L rf Indicates the maximum and minimum values ​​of the frequency band of the sample set; rf min =max(((1+DeltaRf)*min(rf i )-DeltaRf*max(rf i )),L rf ) rf max =min(((1+DeltaRf)*max(rf i )-DeltaRf*min(rf i )),H rf ) The rf obtained through the sample training phase min ,rf max ,pri max ,pri min ,pw max ,pw min The data preprocessing that will be synchronized to the real-time recognition processing will be normalized with the same maximum and minimum values ​​to ensure the consistency of features after parameter normalization in the training and recognition stages.

8. The pulse sequence recognition method based on the self-attention mechanism according to claim 1, characterized in that: The step S3 comprises: First, determine whether the current pulse sequence length meets the input conditions. When the pulse sequence length is greater than or equal to the sequence length of the network model input, divide the pulses using the model input length as a window to construct multiple groups of inputs. For each group of input pulse sequence inter-pulse statistical parameters, select the corresponding lightweight model. Obtain the maximum and minimum values ​​generated in the network model training phase for normalization preprocessing of the real-time recognition sequence. Next, perform model recognition based on the self-attention mechanism to obtain the model recognition conclusion. Finally, based on the model recognition conclusion, fuse the results to form the target conclusion.

9. A pulse sequence recognition device based on a self-attention mechanism, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of a pulse sequence recognition method based on a self-attention mechanism as described in any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a pulse sequence recognition method based on a self-attention mechanism as described in any one of claims 1 to 8 are implemented.

Images