Radar data intelligent augmentation method and system based on probability diffusion model
Through the intelligent augmentation method of radar data based on the probability diffusion model, intelligent data augmentation of deep learning models in the field of radar detection is solved, and the generalization ability and robustness of the model are improved.
Patent Information
- Application Number
- CN202411969199.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-05-30
AI Technical Summary
Deep learning has sparse small sample problems in the application of radar detection, resulting in overfitting, poor robustness and weak generalization capabilities.
Using the intelligent augmentation method of radar data based on the probability diffusion model, a neural network based on the probability diffusion model is designed and trained to obtain intelligent augmentation data.
Large-scale intelligent data augmentation under the conditions of small sample data alleviates the small sample training problem of deep learning models during training, improves the generalization ability of deep learning networks, and reduces the overfitting problem.
Smart Images
Figure CN120071041A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of intelligent augmentation of radar data, and particularly relates to an intelligent augmentation method for radar data based on a probabilistic diffusion model. Background Art
[0002] In current artificial intelligence technologies, intelligent processing means represented by deep learning have achieved major breakthroughs. The premise for implementing the training of core network models by these cutting-edge technologies is the existence of a large number of effective training data samples corresponding to the problems to be solved, and the completion of the attribute annotation of the corresponding data samples. In the field of radar detection, in actual real scenarios, the number of effectively recorded data samples is very small, and due to the limitations of signal processing, the obtained data is also processed data, and the original data samples are rarely available. Therefore, the biggest contradiction for deep learning network models is the problem of sample completeness of training data. In order to apply artificial intelligence technologies to radar detection and its signal processing, the problems that must be solved are to rely on small-sample data to train the network model and effectively solve the robustness and generalization problems of the trained network model. Therefore, to solve the model design and training problems under small-sample data, it is necessary to break through the intelligent data augmentation technology relying on small samples. Therefore, there are currently sparse small-sample problems and the resulting overfitting, poor robustness, and weak generalization ability in the application of deep learning in the field of radar detection. Summary of the Invention
[0003] The purpose of the present invention is to provide an intelligent augmentation method and system for radar data based on a probabilistic diffusion model to solve the sparse small-sample problems and the resulting overfitting, poor robustness, and weak generalization ability in the current application of deep learning in the field of radar detection.
[0004] To achieve the above purpose, the present invention adopts the following technical solutions:
[0005] On the one hand, this specification provides an intelligent augmentation method for radar data based on a probabilistic diffusion model, including:
[0006] Step 102, performing Gaussian noise augmentation on the original radar image data to obtain a diffusion sample data set;
[0007] Step 104, designing and training a neural network based on a probabilistic diffusion model based on the diffusion sample data set to obtain a converged target neural network;
[0008] Step 106, obtaining intelligent augmented data based on the diffusion sample data and the target neural network.
[0009] On the other hand, this specification provides an intelligent augmentation system for radar data based on a probabilistic diffusion model, including:
[0010] Gaussian noise augmentation module, used to perform Gaussian noise augmentation on the original radar image data to obtain a diffusion sample data set;
[0011] Neural network determination module, used to design and train a neural network based on a probabilistic diffusion model based on the diffusion sample data set to obtain a converged target neural network;
[0012] Data intelligent augmentation module, used to obtain intelligent augmented data based on the diffusion sample data and the target neural network.
[0013] Based on the above technical solutions, this specification can achieve the following technical effects:
[0014] Based on the situation of obtaining less effective training data, this method designs a forward diffusion process and a reverse reconstruction process to perform large-scale intelligent data augmentation on a small number of effective samples, alleviates the small-sample training problem faced by the deep learning model during the training process, improves the generalization ability of the deep learning network, reduces the overfitting problem, and solves the sparse small-sample problem and the resulting overfitting, poor robustness, and weak generalization ability problems existing in the application process of current deep learning in the radar detection field. Brief Description of the Drawings
[0015] Figure 1 It is a schematic structural diagram of a radar data intelligent augmentation method based on a probabilistic diffusion model in an embodiment of the present invention.
[0016] Figure 2 It is a schematic structural diagram of a target neural network in an embodiment of the present invention.
[0017] Figure 3 It is a schematic structural diagram of a radar data intelligent augmentation system based on a probabilistic diffusion model in an embodiment of the present invention.
[0018] Figure 4 It is a schematic diagram of an electronic device of the present invention. Detailed Description of the Embodiments
[0019] The following further elaborates on the present invention in detail in conjunction with the drawings and specific embodiments. According to the following description and the claims, the advantages and features of the present invention will be clearer. It should be noted that the drawings are all in a very simplified form and all use non-precise scales, only for the purpose of facilitating and clearly assisting in explaining the purpose of the embodiments of the present invention.
[0020] It should be noted that, in order to clearly illustrate the content of the present invention, the present invention specifically gives multiple embodiments to further illustrate different implementation manners of the present invention. Among them, the multiple embodiments are listed rather than exhaustive. In addition, for the sake of brevity of description, the content already mentioned in the previous embodiments is often omitted in the subsequent embodiments. Therefore, the content not mentioned in the subsequent embodiments can be correspondingly referred to the previous embodiments.
[0021] Example 1
[0022] Please refer to Figure 1 , Figure 1 The intelligent augmentation method for radar data based on the probability diffusion model provided in this example is shown below.
[0023] In this example, the method includes:
[0024] Step 102: Perform Gaussian noise augmentation on the original radar image data to obtain a diffusion sample data set;
[0025] In this example, one implementation of Step 102 is:
[0026] Step 202: Determine the number of diffusion steps and the initial diffusion rate;
[0027] Step 204: Based on the number of diffusion steps and the initial diffusion rate, gradually add Gaussian noise to the original radar image data to obtain a number of diffusion sample data corresponding to each original radar image;
[0028] In this example, one implementation of Step 204 is:
[0029] Step 302: Based on the number of diffusion steps and the initial diffusion rate, determine the diffusion rate corresponding to each diffusion step;
[0030] Step 304: Based on the diffusion rate and Gaussian noise, obtain the current Gaussian noise to be added at the current diffusion step;
[0031] Step 306: Add the current Gaussian noise to the diffusion sample data corresponding to the previous diffusion step to obtain the diffusion sample data corresponding to the current diffusion step.
[0032] In this example, another implementation of Step 204 is:
[0033] Step 402: Based on the initial diffusion rate, determine the diffusion coefficient;
[0034] Step 404: Based on the diffusion rate and Gaussian noise, obtain the Gaussian noise distribution;
[0035] Step 406: Multiply the diffusion sample data corresponding to the previous diffusion step by the diffusion coefficient and then add the Gaussian noise distribution to obtain the diffusion sample data corresponding to the current diffusion step. Step 206: Based on the diffusion sample data of all original radar images, obtain a diffusion sample data set.
[0036] In this example, the mean of the Gaussian noise distribution is 0, and the variance is the square root of the diffusion rate.
[0037] Step 104: Design a neural network based on the probabilistic diffusion model and train it based on the diffusion sample dataset to obtain a converged target neural network.
[0038] In this embodiment, the target neural network is composed of several layers of UNet networks, and the probabilistic diffusion model is implemented in the form of an encoder-decoder. The diffusion sample dataset is added as the residual input for each layer in each network.
[0039] In this embodiment, the activation function of the target neural network is the SILU function.
[0040] Step 106: Obtain intelligent augmented data based on the diffusion sample data and the target neural network.
[0041] In this embodiment, one implementation of Step 106 is as follows:
[0042] Step 502: Input the diffusion sample data corresponding to the current diffusion step into the target neural network to obtain augmented data.
[0043] Step 504: Based on the Gaussian white noise variance, standard Gaussian white noise, and augmented data, obtain the intelligent augmented data corresponding to the previous diffusion step.
[0044] Specifically, in the first step, assume that the radar image dataset to be augmented is \(X = \{x 0 ,x 1 , \cdots, x n \}\), where \(n\) is the number of samples in the dataset to be augmented. Perform a Gaussian noise diffusion process on the radar image dataset \(X\) to be augmented, set the diffusion step as \(T\) and the diffusion rate as \(\beta\) to obtain the diffusion dataset:
[0045]
[0046]
[0047] M
[0048]
[0049] where, is the dataset diffused from \(x 0 as the original image, is the dataset diffused from \(x 1 as the original image, is the dataset diffused from \(x n as the original image, \(t\) is the step variable, and its value range is \([0, T - 1]\). Therefore, the dataset is expanded from \(n\) to \(nT\).
[0050] Next, take as an example to explain the meaning of the expansion. To Except for the first item of the subscript being different, the rest of the data sets are the same. In the following, unless otherwise specified, the expanded data set will be taken as an example for analysis and introduction.
[0051] The expansion process of 0 is as follows: Given the original image x 00 , Gaussian noise is gradually added to it. The noise addition process lasts for T times, generating a series of noisy images {x 01 , x 0(T-1)},..., x 0(t-1)} to achieve the purpose of damaging the image. During the process of adding noise from any t-th image x 0t in the series of noisy images to x t , the variance of the added noise is β t , which is the diffusion rate and is a value that gradually increases with the increase of the diffusion steps. The relationship between the diffusion rate β
[0052]
[0053] and the data set is as follows: t where ε t is the Gaussian noise of the standard normal distribution, and its distribution is ε
[0054] ~N(0, I), and I is the identity matrix.
[0055]
[0056] where q(x 0t |x 0(t-1) ) is defined as the noise addition process, which can be regarded as multiplying the coefficient 0(t-1) on the basis of x in the previous state and then adding a Gaussian noise distribution with a mean of 0 and a variance of .
[0057] In the second step, design a neural network ε θ based on the probabilistic diffusion model. The network input is the original data x 0 , the noisy images {x 00 , x 01 ,..., x 0(T-1)}, the number of steps t, and the diffusion rate β t , and the output is the network prediction value The specific update formula is:
[0058]
[0059] where Δ is the gradient descent update function, and ε t is the standard Gaussian noise. is the network prediction output value, x 0 is the original image, is the step variable, is defined as:
[0060]
[0061] where T is the total number of augmented steps, β t +α t = 1, that is
[0062] Update the neural network ε according to the Δ formula θ , until the network converges.
[0063] The neural network ε θ adopts the basic structure based on UNet and implements the probability diffusion model in the way of encoder-decoder. Convert the original data x 0 into single-channel image data as the original input, design a T-layer UNet network, and add residual inputs {x 01 , K, x 0(T-1)} in each layer. The activation function selects the SILU function, and the formula is as follows:
[0064]
[0065] where e is the natural logarithm, and the neural network structure of the overall probability diffusion model is as Figure 2 .
[0066] where T is the total number of augmented steps, downsampling is deceleration sampling and the number of channels of the subsequent feature maps changes from 20 to 2T. Its function is to increase the number of feature layers and facilitate deep feature extraction. The pooling layer adopts the maximum pooling method, and its function is to highlight obvious features and smooth insignificant features. The input noisy image {x 00 , x 01 , K, x 0(T-1)} is used in each layer to increase the residual connection and facilitate the training of the network model. The input of each layer is the multi-channel feature map processed by the previous layer, and the output is the multi-channel feature map after convolution processing.
[0067] In the third step, use the neural network ε θ to generate the intelligent augmented data, and the formula is as follows:
[0068]
[0069] where x t-1 is the augmented data at the (t - 1)-th step of inference, α t = 1 - β t , x t is the diffusion data at the t-th step, ε θFor a trained neural network, the input is x t , the number of diffusion steps t, σ t is the variance of Gaussian white noise, and z is standard Gaussian white noise.
[0070] In summary, based on obtaining a small amount of effective training data, by designing the forward diffusion process and the reverse reconstruction process, this method performs large-scale intelligent data augmentation on a small number of effective samples, alleviates the small-sample training problem faced by deep learning models during the training process, improves the generalization ability of deep learning networks, reduces the overfitting problem, and solves the sparse small-sample problem and the resulting overfitting, poor robustness, and weak generalization ability problems existing in the current application process of deep learning in the field of radar detection.
[0071] Embodiment 2
[0072] Please refer to Figure 3 , Figure 3 shown is the intelligent radar data augmentation system based on the probabilistic diffusion model provided by this embodiment.
[0073] In this embodiment, the system includes:
[0074] A Gaussian noise augmentation module, used to perform Gaussian noise augmentation on the original radar image data to obtain a diffusion sample data set;
[0075] A neural network determination module, used to design and train a neural network based on the probabilistic diffusion model based on the diffusion sample data set to obtain a converged target neural network;
[0076] A data intelligent augmentation module, used to obtain intelligent augmented data based on the diffusion sample data and the target neural network.
[0077] Optionally, the Gaussian noise augmentation module includes:
[0078] A parameter determination sub-module, used to determine the number of diffusion steps and the initial diffusion rate;
[0079] A noise addition sub-module, used to gradually add Gaussian noise to the original radar image data based on the number of diffusion steps and the initial diffusion rate to obtain a number of diffusion sample data corresponding to each original radar image;
[0080] A diffusion data acquisition sub-module, used to obtain a diffusion sample data set based on the diffusion sample data of all original radar images.
[0081] Optionally, the noise addition sub-module may include:
[0082] A diffusion rate determination unit, used to determine the diffusion rate corresponding to each diffusion step based on the number of diffusion steps and the initial diffusion rate;
[0083] A Gaussian noise determination unit for obtaining the current Gaussian noise to be added for the current diffusion step based on the diffusion rate and Gaussian noise;
[0084] A noise addition unit for adding the current Gaussian noise to the diffusion sample data corresponding to the previous diffusion step to obtain the diffusion sample data corresponding to the current diffusion step.
[0085] Optionally, the noise addition sub-module may include:
[0086] A diffusion coefficient determination unit for determining the diffusion coefficient based on the initial diffusion rate;
[0087] A noise distribution determination unit for obtaining the Gaussian noise distribution based on the diffusion rate and Gaussian noise;
[0088] A Gaussian noise addition unit for multiplying the diffusion sample data corresponding to the previous diffusion step by the diffusion coefficient and then adding the Gaussian noise distribution to obtain the diffusion sample data corresponding to the current diffusion step.
[0089] Optionally, the mean of the Gaussian noise distribution is 0 and the variance is the square root of the diffusion rate.
[0090] Optionally, the target neural network is composed of several layers of UNet networks, and the probability diffusion model is implemented in an encoder-decoder manner, and the diffusion sample data set is added as the residual input for each layer in each layer of the network.
[0091] Optionally, the activation function of the target neural network is the SILU function.
[0092] Optionally, the data intelligent augmentation module includes:
[0093] A neural network output unit for inputting the diffusion sample data corresponding to the current diffusion step into the target neural network to obtain augmented data;
[0094] A data intelligent augmentation unit for obtaining the intelligent augmented data corresponding to the previous diffusion step based on the Gaussian white noise variance, standard Gaussian white noise, and augmented data.
[0095] Based on this, on the basis of obtaining less effective training data, the system designs a forward diffusion process and a reverse reconstruction process to perform large-scale intelligent data augmentation on a small number of effective samples, alleviates the small sample training problem faced by the deep learning model during training, improves the generalization ability of the deep learning network, reduces the overfitting problem, and solves the sparse small sample problem and its resulting overfitting, poor robustness, and weak generalization ability problems existing in the current application of deep learning in the radar detection field.
[0096] Embodiment 3
[0097] Please refer toFigure 4 , this embodiment provides an electronic device, which includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it, forming a radar data intelligent augmentation method based on a probability diffusion model at the logical level. Of course, in addition to the software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. That is, the execution entity of the following processing flow is not limited to each logical unit, and can also be hardware or a logic device.
[0098] The network interface, processor, and memory can be interconnected through a bus system. The above bus can be divided into an address bus, a data bus, a control bus, etc.
[0099] The memory is used to store programs. Specifically, the program can include program code, and the above program code includes computer operation instructions. The memory can include a read-only memory and a random access memory, and provides instructions and data to the processor.
[0100] The processor is used to execute the program stored in the above memory, and specifically execute:
[0101] Step 102, perform Gaussian noise augmentation on the original radar image data to obtain a diffusion sample data set;
[0102] Step 104, design a neural network based on a probability diffusion model and train it based on the diffusion sample data set to obtain a converged target neural network;
[0103] Step 106, obtain intelligent augmented data based on the diffusion sample data and the target neural network.
[0104] The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed through the integrated logic circuit of the processor's hardware or instructions in software form.
[0105] Based on the same inventive concept, this embodiment of the specification also provides a computer-readable storage medium. The above computer-readable storage medium stores one or more programs. When the above one or more programs are executed by an electronic device including multiple application programs, the above electronic device is caused to execute Figure 1 - Figure 2 The corresponding embodiment provides a radar data intelligent augmentation method based on a probability diffusion model.
[0106] Those skilled in the art should understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this specification can take the form of a computer program product implemented on one or more computer-readable storage media containing computer-usable program code.
[0107] In addition, for the specific implementation of the above system, since it is basically similar to the method implementation, the description is relatively simple. For related parts, please refer to the corresponding description in the method implementation part. Moreover, it should be noted that in each module of the system of this application, the components are logically divided according to the functions to be realized. However, this application is not limited thereto, and the components can be re-divided or combined as needed.
[0108] The embodiments in this specification are all described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments.
[0109] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order than in the embodiments and still achieve the desired results. Additionally, in the processes depicted in the figures, it is not necessarily required to show a specific or sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0110] The above are only the embodiments of this application and are not intended to limit this application. For those skilled in the art, various changes and modifications can be made to this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this application shall be included within the scope of the claims of this application.
Claims
1. The radar data intelligent augmentation method based on probability diffusion model is characterized by: include: Perform Gaussian noise expansion on the original radar image data to obtain a diffusion sample data set; Based on the diffusion sample data set, a neural network based on the probability diffusion model is designed and trained to obtain a converged target neural network; Based on the diffuse sample data and the target neural network, intelligent augmented data is obtained.
2. The method according to claim 1, characterized in that The Gaussian noise expansion of the original radar image data to obtain a diffusion sample data set includes: Determine the number of diffusion steps and initial diffusion rate; Based on the diffusion steps and the initial diffusion rate, Gaussian noise is gradually added to the original radar image data to obtain a number of diffusion sample data corresponding to each original radar image; Based on the diffusion sample data of all original radar images, a diffusion sample data set is obtained.
3. The method according to claim 2, characterized in that The step of gradually adding Gaussian noise to the original radar image data based on the diffusion step number and the initial diffusion rate to obtain a number of diffusion sample data corresponding to each original radar image includes: Based on the number of diffusion steps and the initial diffusion rate, determining the diffusion rate corresponding to each diffusion step; Based on the diffusion rate and Gaussian noise, the current Gaussian noise that should be added to the current diffusion step is obtained; The current Gaussian noise is added to the diffusion sample data corresponding to the previous diffusion step number to obtain the diffusion sample data corresponding to the current diffusion step number.
4. The method according to claim 2, characterized in that: The step of gradually adding Gaussian noise to the original radar image data based on the diffusion step number and the initial diffusion rate to obtain a number of diffusion sample data corresponding to each original radar image includes: Based on the initial diffusion rate, the diffusion coefficient is determined; Based on the diffusion rate and Gaussian noise, a Gaussian noise distribution is obtained; The diffusion sample data corresponding to the previous diffusion step number is multiplied by the diffusion coefficient and then added to the Gaussian noise distribution to obtain the diffusion sample data corresponding to the current diffusion step number.
5. The method according to claim 4, characterized in that The mean of the Gaussian noise distribution is 0, and the variance is the square root of the diffusion rate.
6. The method according to claim 1, characterized in that The target neural network is a UNet network with several layers, and the probability diffusion model is implemented in a codec manner, and the diffusion sample data set is added to each layer of the network as the residual input of each layer.
7. The method according to claim 6, characterized in that The activation function of the target neural network is a SILU function.
8. The method according to claim 1, characterized in that The step of obtaining intelligent augmented data based on the diffusion sample data and the target neural network includes: Input the diffusion sample data corresponding to the current diffusion step number into the target neural network to obtain augmented data; Based on the Gaussian white noise variance, standard Gaussian white noise and augmented data, the intelligent augmented data corresponding to the previous diffusion step is obtained.
9. The radar data intelligent augmentation system based on probability diffusion model is characterized by: include: Gaussian noise expansion module, used to perform Gaussian noise expansion on the original radar image data to obtain a diffusion sample data set; A neural network determination module is used to design and train a neural network based on a probability diffusion model based on a diffusion sample data set to obtain a converged target neural network; The data intelligent augmentation module is used to obtain intelligent augmented data based on the diffusion sample data and the target neural network.
10. An electronic device, characterized in that: include: processor; and a memory arranged to store computer executable instructions which, when executed, cause the processor to perform the steps of the method as claimed in any one of claims 1 to 8.