SAR target recognition method and related device based on multi-level capsule fusion network

Through the multi-level capsule fusion network, the attribute information of amplitude, azimuth, pitch angle and shape is extracted in SAR target recognition, which solves the problem of missing details and poor generalization ability of feature extraction in the prior art, and achieves stronger recognition performance and characteristic distinction.

CN118135402BActive Publication Date: 2024-11-15LINYI UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410305046.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-18
Publication Date
2024-11-15
Estimated Expiration
2044-03-18

AI Technical Summary

Technical Problem

Existing SAR target recognition technology will lose detailed information in feature extraction in image transformation domain feature extraction, single transformation method has poor generalization ability, scattering center distribution feature extraction calculation is complex and deep features cannot be extracted, and image morphological feature extraction cannot effectively identify geometric models. Similar goals, and convolutional neural networks rely on a large amount of training data and ignore spatial position relationships, resulting in poor generalization performance.

Method used

Using a multi-level capsule fusion network, including feature extraction module, initial capsule, classification capsule and attribute attention reconstruction module, the weight allocation between the initial capsule and classification capsule is updated through a dynamic routing mechanism, and the amplitude, azimuth, pitch angle, and The four types of attribute information of the shape are enhanced by the attribute attention reconstruction module.

Benefits of technology

The feature extraction ability and recognition performance are improved, the ability to distinguish between similar targets is enhanced, the overall recognition ability of the network is significantly improved, the number of parameters is reduced, and the interpretability of the features and the credibility of the deep representation are enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118135402B_ABST
    Figure CN118135402B_ABST
Patent Text Reader

Abstract

The present invention discloses a SAR target recognition method and related device based on a multi-level capsule fusion network. The method comprises the following steps: acquiring a SAR image of a target area and inputting it into a multi-level capsule fusion network; a feature extraction module respectively extracts features through four extraction layers and constructs them in the form of capsules to form an initial capsule; the output information of the initial capsule is transmitted to a classification capsule through a dynamic routing mechanism, and the features include four types of attribute information: amplitude, azimuth, pitch angle, and shape; an attribute attention reconstruction module is composed of a self-attention module and two fully connected layers connected in sequence, and is used to reconstruct 10 4-dimensional vectors output by the classification capsule to form a reconstructed image. The present invention not only enhances the feature extraction capability and interpretability of the feature of the network, but also highlights the attribute parameters with high separability, enhances the distinction between similar targets, and significantly improves the overall recognition capability of the network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing technology, and in particular to a SAR target recognition method based on a multi-level capsule fusion network and a related device. Background Art

[0002] Synthetic Aperture Radar (SAR) is a system that uses radar echo data for high-resolution imaging. Compared with optical and infrared sensors, SAR has strong penetration capabilities and can achieve continuous observation throughout the day and all weather, and is widely used in civil and military monitoring. With the continuous accumulation of SAR data, especially in the case of large-scale data sets, the need to automatically process and analyze this data to identify targets has become increasingly important. Automatic target recognition technology can improve efficiency and reduce the burden of manual processing.

[0003] At present, the SAR automatic target recognition framework mainly includes three steps: data preprocessing, feature extraction and classification recognition. Among them, SAR image target feature extraction is the basis and key to determine the recognition effect. Typical SAR target feature extraction methods can be roughly divided into three categories, namely image transform domain feature extraction, scattering center distribution feature extraction and image morphology feature extraction. Disadvantages of existing technologies: 1. There are disadvantages in image transform domain feature extraction: 1) The transform domain method will smooth or compress the image, thereby losing some detail information and affecting the recognition performance; 2) The single transform method is only applicable to specific targets and scenes, and has poor generalization ability; 2. The scattering center distribution feature extraction method also has some limitations: 1) The high computational complexity restricts practical applications; 2) It is unable to extract deep features and cannot effectively identify targets with similar geometric models; 3. The image morphological feature extraction method still cannot extract the deep features of the target, and there is still the problem of poor feature differentiation. Although the convolutional neural network (CNN) has strong feature extraction capabilities and good recognition performance, there are still some problems: 1) Relying on a large amount of training data to brute force the classification rules; 2) The translation invariance is used in the feature extraction process, ignoring the spatial position relationship between features; 3) It is separated from the physical explanation and has poor generalization performance. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a SAR target recognition method and related devices based on a multi-level capsule fusion network, which not only enhances the feature extraction capability of the network and the interpretability of the features, but also highlights the attribute parameters with high separability, enhances the distinction between similar targets, and significantly improves the overall recognition capability of the network.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] The SAR target recognition method based on the multi-level capsule fusion network includes the following steps:

[0007] Acquire SAR images of the target area;

[0008] The SAR image of the target area is input into a multi-level capsule fusion network, which includes a feature extraction module, an initial capsule, a classification capsule and an attribute attention reconstruction module. A dynamic routing mechanism is used to iteratively update the weight distribution between the initial capsule and the classification capsule. The feature extraction module extracts features through four Basic block layers and constructs them in the form of capsules to form an initial capsule; the output information of the initial capsule is transmitted to the classification capsule through the dynamic routing mechanism, and the features include four types of attribute information: amplitude, azimuth, pitch angle and shape; the attribute attention reconstruction module is composed of a self-attention module and two fully connected layers connected in sequence, and is used to reconstruct 10 4-dimensional vectors output by the classification capsule to form a reconstructed image.

[0009] Preferably, the four extraction layers include a first extraction layer, a second extraction layer, a third extraction layer, and a fourth extraction layer, and the first extraction layer, the second extraction layer, the third extraction layer, and the fourth extraction layer all use a Basic block structure, wherein the first extraction layer, the second extraction layer, and the third extraction layer all include 2 Basic block structures, and the fourth extraction layer includes 3 Basic block structures; the Basic Block structure includes two 3x3 convolutional layers and jump connections, wherein the stride of the two convolutional layers is 1, and a ReLU activation function is used.

[0010] Preferably, the self-attention module is used to use point convolution to perform three nonlinear transformations on the input classification capsule to generate a query Q vector, a key K vector and a value V vector respectively, perform a dot product operation on the query Q vector and the key K vector to obtain a similarity matrix and scale the similarity matrix, and obtain an attention weight matrix through a Softmax function; multiply the value V vector by the attention weight matrix to obtain a classification capsule with high discrimination; the first fully connected layer of the two fully connected layers adopts a ReLU activation function, and the second fully connected layer adopts a sigmoid activation function.

[0011] Preferably, during the training process of the multi-level capsule fusion network, L2 distance is used to measure the similarity of attribute parameters between capsules, and the formula is as follows:

[0012]

[0013] In the formula, The vector corresponds to four types of attribute information of capsule attribute parameter M1: amplitude, azimuth, pitch angle and shape; The vector corresponds to the four types of attribute information of the capsule attribute parameter M2, namely, amplitude, azimuth, pitch angle and shape, and t indicates that the sum is performed on the first four elements.

[0014] Based on the above content, the present invention also discloses a SAR target recognition device based on a multi-level capsule fusion network, including: an acquisition module and a multi-level capsule fusion network, wherein:

[0015] The acquisition module is used to acquire a SAR image of the target area;

[0016] The multi-level capsule fusion network is used to input the SAR image of the target area into the multi-level capsule fusion network. The multi-level capsule fusion network includes a feature extraction module, an initial capsule, a classification capsule and an attribute attention reconstruction module. The weight distribution between the initial capsule and the classification capsule is continuously updated iteratively by using a dynamic routing mechanism. The feature extraction module extracts features through four Basic block layers and constructs them in the form of capsules to form an initial capsule; the output information of the initial capsule is transmitted to the classification capsule through the dynamic routing mechanism, and the features include four types of attribute information: amplitude, azimuth, pitch angle and shape; the attribute attention reconstruction module is composed of a self-attention module and two fully connected layers connected in sequence, and is used to reconstruct 10 4-dimensional vectors output by the classification capsule to form a reconstructed image.

[0017] Preferably, the Basic Block layer includes two 3x3 convolutional layers and a skip connection, wherein the stride of the two convolutional layers is 1 and a ReLU activation function is used.

[0018] Preferably, the self-attention module is used to use dot convolution to perform three nonlinear transformations on the input classification capsule to respectively generate a query Q vector, a key K vector and a value V vector, perform dot product operation on the query Q vector and the key K vector to obtain a similarity matrix and scale the similarity matrix, and obtain an attention weight matrix through a Softmax function; multiply the value V vector by the attention weight matrix to obtain a classification capsule with high discrimination; the first fully connected layer of the two fully connected layers uses a ReLU activation function, and the second fully connected layer uses a sigmoid activation.

[0019] Preferably, during the training process of the multi-level capsule fusion network, L2 distance is used to measure the similarity of attribute parameters between capsules, and the formula is as follows:

[0020]

[0021] In the formula, The vector corresponds to four types of attribute information of capsule attribute parameter M1: amplitude, azimuth, pitch angle and shape; The vector corresponds to the four types of attribute information of the capsule attribute parameter M2, namely, amplitude, azimuth, pitch angle and shape, and t indicates that the sum is performed on the first four elements.

[0022] Based on the above content, the present invention further discloses a computer device, including: a memory for storing a computer program; and a processor for implementing any of the above methods when executing the computer program.

[0023] Based on the above content, the present invention further discloses a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, any of the methods described above is implemented.

[0024] Based on the above technical solution, the beneficial effects of the present invention are:

[0025] 1) The present invention provides a multi-level capsule fusion network (multi-Level Capsule Fusion Network), which can realize feature extraction of attributes at different levels. Multi-layer fusion capsules contain feature information at different levels at the same time, which can better extract the physical attributes corresponding to the target features, enhance the interpretability of the extracted features, and improve the network's ability to recognize targets;

[0026] 2) The present invention provides an attribute attention reconstruction module, which expands the distinction between the attribute parameters of the target attribute scattering center. The attribute attention reconstruction module reconfigures the advanced capsule four-dimensional vector through a two-layer fully connected self-attention mechanism, which enhances the difference between targets and greatly reduces the number of parameters in the network;

[0027] 3) The present invention combines a multi-level capsule fusion network with an attribute scattering center, and uses the multi-level capsule fusion network to extract the deep representation of four attribute parameters: amplitude, azimuth, pitch angle, and shape. The interpretability of the deep attributes is verified through the attribute attention reconstruction module. In addition, a new capsule form of similarity measurement between attribute parameters is defined to realize the training of the multi-level capsule fusion network. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of the structure of a multi-level capsule fusion network in a SAR target recognition method based on a multi-level capsule fusion network in an embodiment;

[0029] Figure 2 is a schematic diagram of the structure of a Basic block in an embodiment;

[0030] Figure 3 is a feature extraction method based on capsule network in one embodiment;

[0031] Figure 4is a network back propagation path in one embodiment;

[0032] Figure 5 is a schematic diagram of dynamic routing in one embodiment;

[0033] Figure 6 is a schematic diagram of the structure of an attribute attention reconstruction module in one embodiment;

[0034] Figure 7 Schematic diagram of the process of an attribute attention reconstruction module under directional perturbation in an embodiment. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] This embodiment provides a SAR target recognition method based on a multi-level capsule fusion network, and proposes a multi-level capsule fusion network, such as Figure 1 As shown in Figure 1, the multi-level capsule fusion network includes a feature extraction module, an initial capsule, a classification capsule, and an attribute attention reconstruction module. The feature extraction module combines the multi-level capsule fusion network with the attribute parameters of the target attribute scattering center to extract the attribute features corresponding to the target and construct them in the form of capsules. The network adopts a multi-level capsule fusion structure to extract attribute features at different levels. Capsule block 1 and capsule block 2 are respectively connected through two basic blocks (Basic block structure, which is a double-layer convolutional neural network with a residual structure, such as Figure 2 As shown in the figure, feature extraction is performed on the two capsules, which contain detail information. Capsule block 3 and capsule block 4 are respectively extracted through two Basic block structures and three Basic block structures, and the two capsules contain significant features and semantic information. Multi-level fusion capsules constitute initial capsules and contain feature information at different levels. At the same time, the embodiment of the present invention defines a new capsule form of similarity index between attribute parameters to realize the training of multi-level capsule fusion network. Secondly, the dynamic routing method is used to continuously iteratively update the weight distribution between the initial capsule and the classification capsule, so that the network can better capture the spatial structural relationship between capsules and provide a more robust feature representation; finally, after dynamic routing, the attribute attention reconstruction module (AttributeAttentionReconstructionModule, AARM) is introduced to expand the differences between target attribute parameters, enhance the separability between similar targets, and significantly reduce the number of participants in the network. By adding perturbations to different dimensions of the capsule and reconstructing them, the results further prove the physical interpretability of the proposed features. The specific structure and function of each module will be described in detail in the subsequent embodiments.

[0037] Specifically, the input feature map is extracted through four extraction layers, where the Basic block structure is as follows: Figure 2 As the network depth increases, the input feature map will lose information during transmission. The Basic Block structure contains two 3×3 convolutional layers and jump connections. The stride of the two convolutional layers is 1, and its output results contain the initial input information, which can maintain the integrity of the information and alleviate the problem of gradient disappearance or explosion in deep networks.

[0038] Next, the extracted features are constructed into capsule blocks. Capsule network-based feature extraction encapsulates features of the same attribute together to form a capsule, which can accurately separate and extract attribute features at different levels. This is different from the CNN method, which mixes various attribute features together during the feature extraction process, destroying the inherent physical meaning of the attribute features. Figure 3 The feature extraction method based on capsule network is shown. Different shapes represent different attributes, and different colors represent different features. The extracted features correspond to the attribute parameters of the attribute scattering center, including four types of attributes: amplitude, azimuth, pitch angle, and shape. These four dimensions are selected after reconstruction and verification to obtain the parameters of the four effective attribute scattering centers.

[0039] Finally, the four-layer capsule blocks are stacked to form an initial capsule. The initial capsule contains multiple layers of attribute features. Its BasicBlock layer can be directly connected to the initial capsule. The network back propagation path is as follows: Figure 4 The solid line represents forward propagation, and the dotted line represents backward propagation.

[0040] The multi-layer fusion capsule network proposed in the embodiment of the present invention adopts cross-layer connection, and the initial capsule layer can be directly connected to the Basic Block layer, which is not affected by the network depth, significantly reduces the loss propagation distance, and learns the attribute characteristics of different layers.

[0041] The dynamic routing designed in one embodiment is an iterative mechanism that dynamically assigns input information to capsule units, which helps improve the accuracy of the model in estimating and classifying the target pose. The dynamic routing mechanism is used to pass the output information of the low-level capsules in the initial capsule to the high-level capsules in the classification capsule, and the weight distribution between the initial capsule and the classification capsule is continuously updated iteratively. The principle diagram of dynamic routing is shown in the figure. Figure 5 As shown in Equation 1, the low-level capsule u i After weighted summation, we get s j The value process is:

[0042]

[0043] u i represents the i-th capsule in the low-level capsule; Wij represents the weight matrix of the connection between the i-th capsule in the low-level capsule and the j-th capsule in the high-level capsule. Then the classification capsule is obtained by nonlinear transformation through the Squashing function. The allocation of input information of the capsule unit and the adjustment of posture estimation are achieved by continuously iteratively updating the weights. This enables the network to more accurately capture the posture, relationship and representation of the target, thereby improving the recognition performance of the model. ji is the coupling coefficient, defined as shown in Equation 2, b ji represents the prior probability of the jth capsule in the lower-level capsule relative to the ith capsule in the higher-level capsule, b jk It represents the prior probability of the j-th capsule in the lower-level capsule to the k-th capsule in the higher-level capsule.

[0044]

[0045] By applying the Squashing function, the capsule network can retain the relative direction information of the input vector and normalize the length to better capture the posture relationship of the target. The definition is shown in Equation 3, where v j is the output vector of the Squashing function, s j is the input vector, ∥s j ∥ is the vector s j The Euclidean length of .

[0046]

[0047] In order to effectively train the multi-level capsule fusion network, we define the L2 distance between capsules of the target attribute scattering center attribute parameters, which enables the network to be trained to learn to capture the features that are critical to distinguishing different target attributes and strengthen the network's ability to understand the key attributes of the input data. For a given capsule attribute parameter M1 and M2, the distance between them is shown in Formula 4. The vector corresponds to four types of attribute information of capsule attribute parameter M1: amplitude, azimuth, pitch angle and shape; The vector corresponds to the four types of attribute information of the capsule attribute parameter M2, namely, amplitude, azimuth, pitch angle and shape, and t indicates that the sum is performed on the first four elements.

[0048]

[0049] At present, most capsule networks use three layers of fully connected layers as reconstruction networks, which have strong model representation capabilities and can learn more complex features and patterns. However, the number of parameters in the three-layer fully connected layer is too large, which is prone to overfitting and affects the generalization performance of the network. To this end, the embodiment of the present invention proposes an attribute attention reconstruction module (AttributeAttention Reconstruction Module, AARM), the structure of the self-attention module is as follows Figure 6 As shown in the figure, the attribute attention reconstruction module consists of a self-attention module and two layers of fully connected connections, which can reduce the number of parameters while expanding the distinction between the attribute parameters of the target attribute scattering center and enhancing the distinguishability between similar targets. The self-attention mechanism in the attribute attention reconstruction module focuses on different dimensions of the input information and weights them to highlight the distinguishability of features in the same dimension. Corresponding to the target attribute parameters, the distinguishability of attribute parameters between different targets is enhanced.

[0050] Self-attention module processing process: First, use point convolution to perform three nonlinear transformations on the input classification capsule to generate query Q vector, key K vector and value V vector respectively; then calculate the dot product of the query Q vector and the key K vector to obtain the similarity matrix S; then scale the similarity matrix, and apply the Softmax function to convert the similarity into the attention weight matrix A; finally, multiply the value V by the attention weight matrix A to obtain a classification capsule with high discriminability, which enhances the separability between similar targets.

[0051] The features proposed based on the multi-level capsule fusion network method correspond to the attribute parameters of the target attribute scattering center, including four valid attribute parameters: amplitude, azimuth, pitch angle and shape. To verify this conclusion, we will use the attribute attention reconstruction module to reconstruct the capsule-based attribute parameters, and add perturbations to each attribute parameter separately during the entire reconstruction process. The perturbation range is [-0.25, 0.25] and the interval is 0.05. Adding perturbations to the attributes of each dimension will produce changes in the corresponding physical attributes. The attribute attention reconstruction module is as follows Figure 7 As shown in the figure. When perturbations are added to the dimension where the magnitude information is located, the scattering intensity of the SAR image will change dramatically. When perturbations are made to the dimension where the directional information is located, the overall direction of the SAR image will change. When the pitch angle changes, the number of points in the target scattering center will also change, and the highlights of the resulting SAR image will also change. When we add perturbations to the shape information, the structural information of the SAR image will also change to varying degrees.

[0052] The above are only preferred embodiments of the embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

1. A SAR target recognition method based on a multi-level capsule fusion network, characterized in that: The steps include: Acquire SAR images of the target area; The SAR image of the target area is input into a multi-level capsule fusion network, which includes a feature extraction module, an initial capsule, a classification capsule and an attribute attention reconstruction module, and a dynamic routing mechanism is used to continuously iterate and update the weight distribution between the initial capsule and the classification capsule. The feature extraction module extracts features through four extraction layers respectively and constructs them in the form of capsules to form an initial capsule; The output information of the initial capsule is passed to the classification capsule through a dynamic routing mechanism. The features include four types of attribute information: amplitude, azimuth, pitch angle, and shape. The attribute attention reconstruction module consists of a self-attention module and two fully connected layers connected in sequence, which is used to reconstruct the 10 4-dimensional vectors output by the classification capsule to form a reconstructed image. Specifically, The four extraction layers include a first extraction layer, a second extraction layer, a third extraction layer, and a fourth extraction layer, wherein the first extraction layer, the second extraction layer, the third extraction layer, and the fourth extraction layer all use a Basic block structure, wherein the first extraction layer, the second extraction layer, and the third extraction layer all include 2 Basic block structures, and the fourth extraction layer includes 3 Basic block structures; the Basic Block structure includes two 3x3 convolutional layers and a jump connection, wherein the stride of the two convolutional layers is 1, and a ReLU activation function is used; The self-attention module is used to use point convolution to perform three nonlinear transformations on the input classification capsule to generate a query Q vector, a key K vector and a value V vector respectively, perform a dot product operation on the query Q vector and the key K vector to obtain a similarity matrix and scale the similarity matrix, and obtain an attention weight matrix through a Softmax function; multiply the value V vector by the attention weight matrix to obtain a classification capsule with high discrimination; the first fully connected layer of the two fully connected layers uses a ReLU activation function, and the second fully connected layer uses a sigmoid activation function; During the training process of the multi-level capsule fusion network, the L2 distance is used to measure the similarity of attribute parameters between capsules. The formula is as follows: In the formula, The vector corresponds to four types of attribute information of capsule attribute parameter M1: amplitude, azimuth, pitch angle and shape; The vector corresponds to the four types of attribute information of the capsule attribute parameter M2, namely, amplitude, azimuth, pitch angle and shape, and t indicates that the sum is performed on the first four elements.

2. A SAR target recognition device based on a multi-level capsule fusion network, characterized in that: include: Get module and multi-level capsule fusion network, where, The acquisition module is used to acquire a SAR image of the target area; The multi-level capsule fusion network is used to input the SAR image of the target area into the multi-level capsule fusion network. The multi-level capsule fusion network includes a feature extraction module, an initial capsule, a classification capsule and an attribute attention reconstruction module. The weight distribution between the initial capsule and the classification capsule is continuously updated iteratively by using a dynamic routing mechanism. The feature extraction module performs feature extraction through four extraction layers respectively and constructs the initial capsule in the form of capsules; the output information of the initial capsule is transmitted to the classification capsule through the dynamic routing mechanism, and the feature includes four types of attribute information: amplitude, azimuth, pitch angle and shape; the attribute attention reconstruction module is composed of a self-attention module and two fully connected layers connected in sequence, and is used to reconstruct 10 4-dimensional vectors output by the classification capsule to form a reconstructed image. Specifically, The four extraction layers include a first extraction layer, a second extraction layer, a third extraction layer, and a fourth extraction layer, wherein the first extraction layer, the second extraction layer, the third extraction layer, and the fourth extraction layer all use a Basic block structure, wherein the first extraction layer, the second extraction layer, and the third extraction layer all include 2 Basic block structures, and the fourth extraction layer includes 3 Basic block structures; the Basic Block structure includes two 3x3 convolutional layers and a jump connection, wherein the stride of the two convolutional layers is 1, and a ReLU activation function is used; The self-attention module is used to use point convolution to perform three nonlinear transformations on the input classification capsule to generate a query Q vector, a key K vector and a value V vector respectively, perform a dot product operation on the query Q vector and the key K vector to obtain a similarity matrix and scale the similarity matrix, and obtain an attention weight matrix through a Softmax function; multiply the value V vector by the attention weight matrix to obtain a classification capsule with high discrimination; the first fully connected layer of the two fully connected layers uses a ReLU activation function, and the second fully connected layer uses a sigmoid activation function; During the training process of the multi-level capsule fusion network, the L2 distance is used to measure the similarity of attribute parameters between capsules. The formula is as follows: In the formula, The vector corresponds to four types of attribute information of capsule attribute parameter M1: amplitude, azimuth, pitch angle and shape; The vector corresponds to the four types of attribute information of the capsule attribute parameter M2, namely, amplitude, azimuth, pitch angle and shape, and t indicates that the sum is performed on the first four elements.

3. A computer device, characterized in that: The device comprises: a memory for storing a computer program; A processor, configured to implement the method according to claim 1 when executing the computer program.

4. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to claim 1 is implemented.