A one-dimensional range image data cross-device utilization identification method

Abstract

The application discloses a one-dimensional range image data cross-device application identification method, comprising the following steps: collecting one-dimensional range image data of common observation targets by different radars, and constructing a data set; constructing a one-dimensional range image conversion network based on a CycleGAN model; performing interactive training on the constructed one-dimensional range image conversion network to obtain a trained one-dimensional range image conversion network; and converting and identifying the one-dimensional range image based on the trained one-dimensional range image conversion network. The application constructs a one-dimensional range image data conversion network by using a generative adversarial network, realizes mutual application and identification of one-dimensional range image data generated by two different radars on the same observation target, greatly improves the target identification efficiency, and provides help for accurate tracking of the target afterwards.

Description

Technical Field

[0001] This invention belongs to the field of data recognition and application technology, specifically relating to a method for cross-device application and recognition of one-dimensional distance image data. Background Technology

[0002] One-dimensional range imaging is an important component of radar imaging and recognition technology. It represents the distribution of a target model across the radar's radial range after a high-resolution radar transforms the target model into a set of independent scattering centers. One-dimensional range imaging is easy to implement and provides longitudinal structural information about the target, including the location, amplitude, and type of strong scattering points along the range direction—essential information for target identification. Therefore, it is widely used in many military and civilian fields, such as the identification of aerial targets, ground vehicle targets, and maritime ship targets.

[0003] With the deepening development of practical applications, the requirements for target recognition efficiency are also increasing. When using different radar devices as detection methods to observe the same target, the amount of scattering information collected by different radars will vary. Therefore, for the same target, the one-dimensional range images obtained by different radar devices will be different and cannot be mutually recognized, greatly reducing the efficiency of practical applications. Therefore, a one-dimensional range image conversion network between different radars can be constructed based on a certain method to achieve rapid identification of one-dimensional range image targets by different high-resolution radars.

[0004] Generative Adversarial Networks (GANs) are powerful data generation and transformation methods that can build rules based on existing information for training and learning, thereby constructing "new" information. For two different high-resolution radars, given a one-dimensional range profile of a target observed by one radar device, a one-dimensional range profile of the same target observed by the other radar device can be generated based on the GAN concept and latent rules. Therefore, GANs can be used to solve the problem of recognizing and utilizing one-dimensional range profiles from different devices. Summary of the Invention

[0005] To overcome the problems in the prior art, this invention proposes a method for cross-device recognition of one-dimensional distance image data.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] A method for cross-device recognition of one-dimensional distance image data includes the following steps:

[0008] Step 1. Collect one-dimensional range profile data of the common target from different radars and construct a dataset;

[0009] Step 2. Construct a one-dimensional distance image transformation network based on the CycleGAN model;

[0010] Step 3. Interactively train the constructed one-dimensional distance image transformation network based on the constructed dataset to obtain the trained one-dimensional distance image transformation network;

[0011] Step 4. Based on the trained one-dimensional distance image transformation network, perform transformation and recognition on the one-dimensional distance image.

[0012] Furthermore, the CycleGAN model comprises two GANs, each including a generator and a discriminator.

[0013] Furthermore, the generator includes an encoder, a converter, and a decoder, used to perform preprocessing, feature extraction, feature vector transformation, and output generated data on the input one-dimensional distance image data;

[0014] The encoder extracts features from the input data using a one-dimensional convolutional neural network; the converter uses six Reset modules to transform the feature vector by combining dissimilar features of the image data; the decoder uses a deconvolutional layer to restore low-level features from the feature vector, finally obtaining the generated one-dimensional distance image data.

[0015] Furthermore, the discriminator is used to discriminate the converted image and predict the probability that the generated image comes from the target image set.

[0016] Furthermore, the model also includes a loss function, which comprises adversarial loss, cycle consistency loss, and identity mapping loss:

[0017] The overall loss function equation should be:

[0018] L(G,F,D X D Y ) = L adv (G X→Y D Y )+L adv (G Y→X D X )+λ cyc L cyc (G X→Y ,F Y→X )+λ id L id (G X→Y ,F Y→X )

[0019] Where, λ cyc and λ id To balance the parameters, L adv(G X→Y D Y L adv (G Y→X D X ) is an adversarial loss, L cyc (G X→Y ,F Y→X ) represents the loss due to circular consistency, L id (G X→Y ,F Y→X () represents the identity mapping loss.

[0020] Furthermore, the adversarial loss is:

[0021]

[0022] In the formula, x i y represents the i-th target in the radar 1 one-dimensional image dataset X within the total sample dataset D; i P represents the i-th target in the radar 2 one-dimensional image dataset Y within the total sample dataset D; X (x i P represents the probability distribution that the radar 1 one-dimensional image dataset X follows; Y (y i G represents the probability distribution that the radar 2 one-dimensional image dataset Y follows; X→Y (x i ) represents sample data x i Generate image data with the same distribution as the sample data Y using generator G; D Y (y i ) represents the actual data sample y i In discriminator D Y The rating value in the discriminator D ranges from [0,1]. The closer the rating value is to 1, the better the discriminator D becomes. Y The more genuine the sample, the better; logD Y (y i ) represents the discriminator D Y Determine the true data sample y i The probability of it being true; Discriminator D Y Determine the true data sample y i For true expectation; D Y (G X→Y (x i )) is the discriminator D Y For the generator G based on the input x i The generated image data is scored, with the score value ranging from [0,1]. The closer the score value is to 1, the higher the score. Y The more realistic the generated image data, the better; if the score is closer to 0, then D... YThe more fake the generated image data, the more likely it is to be; log(1-D) Y (G X→Y (x i ))) represents the discriminator D Y Determine the probability that the generated image data is fake; Discriminator D Y The expectation is to determine whether the generated image data is fake;

[0023]

[0024] In the formula, G Y→X (y i ) represents the sample data y i Generate image data that is identically distributed to sample X using generator G; D X (x i ) represents the actual data sample x i In discriminator D X The rating value in the discriminator D ranges from [0,1]. The closer the rating value is to 1, the better the discriminator D becomes. X The more genuine the sample, the better; log D X (x i ) represents the discriminator D X Determine the true data sample x i The probability of it being true; Discriminator D X Determine the true data sample x i For true expectation; D X (G Y→X (y i )) is the discriminator D X For the generator G based on the input y i The generated image data is scored, with the score value ranging from [0,1]. The closer the score value is to 1, the higher the score. X The more realistic the generated image data, the better; if the score is closer to 0, then D... X The more fake the generated image data, the more likely it is to be; log(1-D) X (G Y→X (y i ))) represents the discriminator D X Determine the probability that the generated image data is fake; Discriminator D X The expectation is to determine whether the generated image data is fake.

[0025] Furthermore, the cycle consistency loss is:

[0026]

[0027] In the formula, F Y→X (GX→Y (x i )) represents sample data x i Data is generated using generator G. Then, generator F continues to generate image data that is identically distributed to sample X. G X→Y (F Y→X (y i )) represents sample data y i Data is generated using generator F. Then, generator G continues to generate image data that is identically distributed to sample Y. Indicates the generation of data Compared with the original real data x i The expected gap; Indicates the generation of data Compared with the original real data y i The expected gap;

[0028] Furthermore, the identity mapping loss is:

[0029]

[0030] In the formula, F Y→X (x i ) represents sample data x i Image data generated by generator F; G X→Y (y i ) represents the sample data y i Image data generated by generator G; Indicates the generated data F Y→X (x i ) and the original input data x i The expected gap; Indicates the generation of data G X→Y (y i ) and the original input data y i The gap in expectations.

[0031] Furthermore, step 3 involves interactively training the constructed one-dimensional distance image transformation network based on the constructed dataset, including the following steps:

[0032] Step 3-1. Select the first i samples in the dataset as training samples and perform data augmentation preprocessing on the training samples;

[0033] Step 3-2. The generator uses a one-dimensional convolutional neural network. During generator training, data is read in pairs, with each sample (x) as an example. i ,y i For example, the specific steps are as follows:

[0034] Enter xi After being obtained by generator G Enter y i After being obtained by generator F

[0035] y is then generated by generator G. i Fake data x is then generated by generator F. i Fake data

[0036] After discriminator D X Generate the label x_f_dis. After discriminator D Y Generate the label y_f_dis;

[0037] Calculate the adversarial loss: L adv (G X→Y D Y ) and L adv (G Y→X D X );

[0038] y i After generator G, we get G(y) i ), x i After passing through generator F, we obtain F(x). i );

[0039] Calculate the cycle consistency loss L cyc (G X→Y ,F Y→X and identity mapping loss L id (G X→Y ,F Y→X );

[0040] Step 3-3. The discriminator uses a two-dimensional convolutional neural network to perform binary classification.

[0041] The training of the discriminator is x i and After discriminator D X Generate label matrices x_r_dis and x_f_dis respectively, y i and After discriminator D Y Generate the label matrices y_r_dis and y_f_dis respectively; calculate the loss function of the discriminator;

[0042] Steps 3-4. All network optimizers used are Adam;

[0043] Training used Adam stochastic gradient descent to optimize the total loss function, first optimizing the discriminator D. X and D Y Then optimize generators G and F:

[0044]

[0045] Furthermore, the data augmentation preprocessing operations include, but are not limited to, translation and rotation transformations.

[0046] Compared with the prior art, the present invention has the following technical effects:

[0047] This invention proposes a cross-device application identification method for one-dimensional range image data. It utilizes a generative adversarial network to construct a conversion network for one-dimensional range image data, enabling mutual application identification of one-dimensional range image data generated by two different radars for the same observed target. This greatly improves the efficiency of target identification and helps in the accurate tracking of the target. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0049] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0050] When the same target is detected and identified using different radar devices, the amount of scattering information acquired varies due to the different radar resolutions, resulting in differences in the obtained one-dimensional range profile data. Furthermore, cross-device application for identification is not possible. To enable the mutual application and identification of one-dimensional range profile data generated by different radars for a shared target and to improve target identification efficiency, [the following is proposed, referencing...] Figure 1 This embodiment provides a method for cross-device recognition of one-dimensional distance image data, including the following steps:

[0051] Step 1. Collect one-dimensional range profile data of the common target from different radars and construct a dataset;

[0052] For different radars, one-dimensional range profile data of common observation targets are collected to construct a dataset, with one-dimensional range profile data of the same target being used as a sample.

[0053] In a specific embodiment, one-dimensional range image data of a common target is collected for two radars with different resolutions, and a dataset (D, {X,Y}) is constructed. The one-dimensional range image data of the same target {one-dimensional image of radar 1, one-dimensional image of radar 2} is a sample.

[0054] The total sample contains m target samples, where the i-th target {x} i y i If} can be represented as a sample, then the one-dimensional image dataset of radar 1 in the total dataset D can be represented as X={x1,x2,…,x m The radar 2 one-dimensional image dataset is represented as Y = {y1, y2, ..., y}. m}

[0055] Step 2. Construct a one-dimensional distance image transformation network based on CycleGAN;

[0056] CycleGAN (Cycle Consistent Adversarial Networks) is a generative model that can be used for unsupervised training on non-parallel training data.

[0057] To adapt to different inputs and obtain different outputs, the constructed model contains two GANs. The first GAN can generate one-dimensional image data of radar 2 based on radar 1 one-dimensional image data, and the second GAN can generate one-dimensional image data of radar 1 based on radar 2 one-dimensional image data.

[0058] Each GAN includes one generator model and one discriminator model. Therefore, the CycleGAN model has a total of four models: two generators (G and F) and two discriminators (D). X and D Y To ensure accurate conversion of the input and output corresponding to the same target when two different radars switch between multiple targets, i.e., to avoid all X being mapped to the same Y.

[0059] A generator G is used to implement the mapping from X to Y, which is represented by G in the formula. X→Y The generator F implements the mapping from Y to X, which is represented as F in the formula. Y→X The generator includes an encoder, a converter, and a decoder, used to perform preprocessing, feature extraction, feature vector transformation, and output generated data from the input one-dimensional distance image data (X or Y). or ).

[0060] The encoder uses a one-dimensional convolutional neural network to extract features from the input data (X or Y).

[0061] The converter uses six Reset modules to transform feature vectors in the X (or Y) domain into feature vectors in the Y (or X) domain by combining dissimilar features of the image data. Each Reset module consists of two convolutional layers, enabling the transformation to be completed while preserving the features of the original input.

[0062] The decoder uses decovolution layers to recover low-level features from the feature vectors, ultimately producing one-dimensional distance image data. or ).

[0063] Discriminator (D) X and D Y The discriminator (D) is used to distinguish the converted image and predict the probability that the generated image comes from the target image set. X It encourages the generator F to transform Y into X, and the discriminator D... Y It encourages the generator G to transform X into Y.

[0064] The CycleGAN model also includes loss functions, which are key to the transformation network. These include adversarial loss, cycle-consistency loss, and identity-mapping loss.

[0065] Adversarial loss measures the difference between the transformed one-dimensional range image and the target's one-dimensional range image. The smaller the difference between the two, the smaller the adversarial loss function. To make the transformed features indistinguishable from the target, the generator should minimize the difference between the output and the input.

[0066]

[0067] In the formula, x i y represents the i-th target in the radar 1 one-dimensional image dataset X within the total sample dataset D; i P represents the i-th target in the radar 2 one-dimensional image dataset Y within the total sample dataset D; X (x i P represents the probability distribution that the radar 1 one-dimensional image dataset X follows; Y (y i G represents the probability distribution that the radar 2 one-dimensional image dataset Y follows; X→Y (x i ) represents sample data x i Generate image data with the same distribution as the sample data Y using generator G; D Y (y i ) represents the actual data sample y i In discriminator D Y The rating value in the discriminator D ranges from [0,1]. The closer the rating value is to 1, the better the discriminator D becomes. Y The more genuine the sample, the better; log D Y (y i ) represents the discriminator DY Determine the true data sample y i The probability of it being true; Discriminator D Y Determine the true data sample y i For true expectation; D Y (G X→Y (x i )) is the discriminator D Y For the generator G based on the input x i The generated image data is scored, with the score value ranging from [0,1]. The closer the score value is to 1, the higher the score. Y The more realistic the generated image data, the better; if the score is closer to 0, then D... Y The more fake the generated image data, the more likely it is to be; log(1-D) Y (G X→Y (x i ))) represents the discriminator D Y Determine the probability that the generated image data is fake; Discriminator D Y The expectation is to determine whether the generated image data is fake.

[0068] Focusing only on adversarial losses L adv (G X→Y D Y It cannot guarantee the preservation of the original feature information of X, nor can it guarantee the consistency of input and output features, because the generator may tend to convert all X into a one-dimensional distance image data. Therefore, it is also necessary to calculate the adversarial loss L. adv (G Y→X D X ).

[0069]

[0070] In the formula, G Y→X (y i ) represents the sample data y i Generate image data that is identically distributed to sample X using generator G; D X (x i ) represents the actual data sample x i In discriminator D X The rating value in the discriminator D ranges from [0,1]. The closer the rating value is to 1, the better the discriminator D becomes. X The more genuine the sample, the better; log D X (x i ) represents the discriminator D X Determine the true data sample x i The probability of it being true; Discriminator D X Determine the true data sample xi For true expectation; D X (G Y→X (y i )) is the discriminator D X For the generator G based on the input y i The generated image data is scored, with the score value ranging from [0,1]. The closer the score value is to 1, the higher the score. X The more realistic the generated image data, the better; if the score is closer to 0, then D... X The more fake the generated image data, the more likely it is to be; log(1-D) X (G Y→X (y i ))) represents the discriminator D X Determine the probability that the generated image data is fake; Discriminator D X The expectation is to determine whether the generated image data is fake.

[0071] To further standardize the mapping and prevent overfitting of generators G and F, the cycle consistency loss includes forward cycle loss and backward cycle loss.

[0072] The meaning of forward loop:

[0073] The meaning of a backward loop:

[0074]

[0075] In the formula, F Y→X (G X→Y (x i )) represents sample data x i Data is generated using generator G. Then, generator F continues to generate image data that is identically distributed to sample X. G X→Y (F Y→X (y i )) represents sample data y i Data is generated using generator F. Then, generator G continues to generate image data that is identically distributed to sample Y. Indicates the generation of data Compared with the original real data x i The expected gap; Indicates the generation of data Compared with the original real data y i The gap in expectations.

[0076] Identity mapping loss measures the difference between the one-dimensional range profile data obtained by inputting the one-dimensional range profile data X from radar 1 into generator F and X, and vice versa. Theoretically, it should not change the original input data as much as possible, that is, the smaller the difference, the better.

[0077]

[0078] In the formula, F Y→X (x i ) represents sample data x i Image data generated by generator F; G X→Y (y i ) represents the sample data y i Image data generated by generator G; Indicates the generated data F Y→X (x i ) and the original input data x i The expected gap; Indicates the generation of data G X→Y (y i ) and the original input data y i The gap in expectations.

[0079] Therefore, the overall loss function equation should be:

[0080] L(G,F,D X D Y ) = L adv (G X→Y D Y )+L adv (G Y→X D X )+λ cyc L cyc (G X→Y ,F Y→X )+λ id L id (G X→Y ,F Y→X )

[0081] Where, λ cyc and λ id To weigh the parameters.

[0082] Step 3. Interactively train the constructed one-dimensional distance-image transformation network model based on the constructed dataset to obtain the trained one-dimensional distance-image transformation network;

[0083] Step 3-1. Select the first i samples from the dataset as training samples, i.e., (x1, y1), (x2, y2), ..., (x... i ,y iThe training samples are preprocessed with data augmentation techniques, including translation and rotation transformations, to ensure that the extracted features are translation and rotation invariant. (x) i+1 ,y i+1 ),(x i+2 ,y i+2 ),…,(x m ,y m () as test samples.

[0084] Step 3-2. The generator uses a one-dimensional convolutional neural network, including an input layer, convolutional layers, activation functions, pooling layers, fully connected layers, and an output layer.

[0085] The input layer is used to normalize the input data (X or Y) to reduce the impact of amplitude sensitivity on target recognition.

[0086] The purpose of the convolutional layer is to extract features from the one-dimensional distance image of the target. The convolutional layer takes the output feature vector of the previous layer as input and enters the convolutional layer, performs convolution operation with the convolutional kernel of each channel, and then passes through the activation function to obtain the output feature vector.

[0087] The activation function uses a gated linear unit (GLU) mechanism, which selectively propagates information based on the previous layer state to determine which features should be focused on and which should be ignored.

[0088] Pooling layers are used to remove redundant information in feature extraction. They divide the input feature vector into multiple connected, non-overlapping vector blocks of the same length, and obtain the output feature vector by averaging or taking the maximum value of each vector block.

[0089] Fully connected layers are used to map the extracted feature vectors to the output layer.

[0090] The output layer outputs the generated data, resulting in ( or ).

[0091] During generator training, data is read in pairs, with each pair containing a sample (x) i ,y i For example, the specific steps are as follows:

[0092] Enter x i After being obtained by generator G Enter y i After being obtained by generator F

[0093] y is then generated by generator G. i Fake data x is then generated by generator F.i Fake data

[0094] After discriminator D X Generate the label x_f_dis. After discriminator D Y Generate the label y_f_dis;

[0095] Calculate the adversarial loss: L adv (G X→Y D Y ) and L adv (G Y→X D X );

[0096] y i After generator G, we get G(y) i ), x i After passing through generator F, we obtain F(x). i );

[0097] Calculate the cycle consistency loss L cyc (G X→Y ,F Y→X and identity mapping loss L id (G X→Y ,F Y→X ).

[0098] Step 3-3. The discriminator uses a two-dimensional convolutional neural network to perform binary classification.

[0099] The discriminator extracts features from the image and then determines whether the extracted features belong to a specific category by adding a convolutional layer that produces a one-dimensional output. It finds the optimal decision boundary between real data and transformed features by maximizing adversarial loss, while generating deceptive features by minimizing cycle consistency loss and identity mapping loss.

[0100] The training of the discriminator is x i and After discriminator D X Generate label matrices x_r_dis and x_f_dis respectively, y i and After discriminator D Y Generate label matrices y_r_dis and y_f_dis respectively; calculate the loss function of the discriminator.

[0101] Steps 3-4. All network optimizers used are Adam.

[0102] The training used Adam stochastic gradient descent. First, 100 epochs of learning were performed with a small learning rate, followed by another 100 epochs of learning with the learning rate gradually decreasing. The model was updated after each sample was processed.

[0103] To optimize the total loss function, first optimize the discriminator D. X and D Y Then optimize generators G and F:

[0104]

[0105] Step 4. Based on the trained one-dimensional distance image transformation network, perform transformation and recognition on the one-dimensional distance image.

[0106] Based on the trained one-dimensional range image transformation network, when the newly observed one-dimensional range image of radar 1 is known, the newly observed one-dimensional range image of radar 1 is transformed into a one-dimensional range image of radar 2, which is used as a target database data sample for radar 2 to identify. Similarly, when the newly observed one-dimensional range image of radar 2 is known, the newly observed one-dimensional range image of radar 2 is transformed into a one-dimensional range image of radar 1, which is used as a target database data sample for radar 1 to identify.

[0107] This invention is based on GAN theory. After knowing the one-dimensional range image data of a radar for observing a target, a transformation network is trained using CycleGAN to generate one-dimensional range image data of another radar for observing the same target, thereby realizing the cross-device application and recognition of one-dimensional range image data.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-device recognition of one-dimensional distance image data, characterized in that, Includes the following steps: Step 1. Collect one-dimensional range profile data of the common target from different radars and construct a dataset; Step 2. Based on the CycleGAN model, construct a one-dimensional distance-image transformation network; the model also includes loss functions, which include adversarial loss, cycle consistency loss, and identity mapping loss: The overall loss function equation should be: in, and To balance the parameters, , For counter-loss, For cycle consistency loss, This is the loss due to identity mapping. Step 3. Interactively train the constructed one-dimensional distance image transformation network based on the constructed dataset to obtain the trained one-dimensional distance image transformation network; Step 4. Based on the trained one-dimensional distance image transformation network, transform and recognize the one-dimensional distance image data.

2. The method for cross-device recognition of one-dimensional distance image data according to claim 1, characterized in that, The CycleGAN model contains two GANs, each of which includes a generator and a discriminator.

3. The method for cross-device recognition of one-dimensional distance image data according to claim 2, characterized in that, The generator includes an encoder, a converter, and a decoder, used to perform preprocessing, feature extraction, feature vector transformation, and output generated data on the input one-dimensional distance image data. The encoder extracts features from the input data using a one-dimensional convolutional neural network; the converter uses six Reset modules to transform the feature vector by combining dissimilar features of the image data; the decoder uses a deconvolutional layer to restore low-level features from the feature vector, finally obtaining the generated one-dimensional distance image data.

4. The method for cross-device recognition of one-dimensional distance image data according to claim 2, characterized in that, The discriminator is used to distinguish the converted image and predict the probability that the generated image comes from the target image set.

5. The method for cross-device recognition of one-dimensional distance image data according to claim 1, characterized in that, The adversarial loss is: In the formula, Represents the total sample dataset D One-dimensional image dataset of radar 1 X The first in i One goal; Represents the total sample dataset D China Radar 2 1D Image Dataset Y The first in i One goal; Let X represent the probability distribution that the one-dimensional image dataset X of radar 1 follows; Represents the radar 2 one-dimensional image dataset Y The probability distribution it follows; Representing sample data Through generator G Generation and sample data Y Image data with the same distribution; Represents real data samples In the discriminator The rating value in the system ranges from [0,1]. Discriminator Determine the true data sample The probability of it being true; Discriminator Determine the true data sample For true expectations; It is a discriminator For generator G According to the input The generated image data is scored, and the score value ranges from [0,1]. Discriminator Determine the probability that the generated image data is fake; Discriminator The expectation is to determine whether the generated image data is fake; In the formula, Representing sample data Through generator G Generation and Samples X Image data with the same distribution; Represents real data samples In the discriminator The rating value in the system ranges from [0,1]. Discriminator Determine the true data sample The probability of it being true; Discriminator Determine the true data sample For true expectations; It is a discriminator For generator G According to the input The generated image data is scored, and the score value ranges from [0,1]. Discriminator Determine the probability that the generated image data is fake; Discriminator The expectation is to determine whether the generated image data is fake.

6. The method for cross-device recognition of one-dimensional distance image data according to claim 1, characterized in that, The cycle consistency loss is: In the formula, Representing sample data Through generator G Generate data Then, continue through the generator. F Generate image data that is identically distributed to sample X. ; Representing sample data Through generator F Generate data Then, continue through the generator. G Generate image data that is identically distributed to sample Y. ; Indicates the generation of data Compared with the original real data The expected gap; Indicates the generation of data Compared with the original real data The gap in expectations.

7. The method for cross-device recognition of one-dimensional distance image data according to claim 1, characterized in that, The identity mapping loss is: In the formula, Representing sample data Through generator F The generated image data; Representing sample data Through generator G The generated image data; Indicates the generation of data Compared with the original input data The expected gap; Indicates the generation of data Compared with the original input data The gap in expectations.

8. The method for cross-device recognition of one-dimensional distance image data according to claim 1, characterized in that, Step 3 involves interactively training the constructed one-dimensional distance image transformation network based on the constructed dataset, including the following steps: Step 3-1. Select the previous data in the dataset. i One sample is used as the training sample, and data augmentation preprocessing is performed on the training sample; Step 3-2. The generator uses a one-dimensional convolutional neural network. During generator training, data is read in pairs, with samples... For example, the specific steps are as follows: enter After generator G get ,enter After generator F get ; After generator G Regeneration Fake data , After generator F Regeneration Fake data ; After discriminator D X Generate the label x_f_dis. After discriminator D Y Generate the label y_f_dis; Calculate adversarial loss: and ; After generator G get , After generator F get ; Calculate cycle consistency loss and identity mapping loss ; Step 3-3. The discriminator uses a two-dimensional convolutional neural network to perform binary classification. The training of the discriminator is and After discriminator D X Generate label matrices x_r_dis and x_f_dis respectively. and After discriminator D Y Generate the label matrices y_r_dis and y_f_dis respectively; calculate the loss function of the discriminator; Steps 3-4. All network optimizers used are Adam; Training used Adam stochastic gradient descent to optimize the total loss function, starting with optimizing the discriminator. D X and D Y Further optimize the generator G and F : 。 9. The method for cross-device recognition of one-dimensional distance image data according to claim 8, characterized in that, The data augmentation preprocessing operations include, but are not limited to, translation and rotation transformations.

Images