Radar image noise reduction processing method based on deep learning

Abstract

The invention provides a radar image noise reduction processing method based on deep learning, and belongs to the technical field of machine learning, and the method comprises the steps: obtaining historical echo signal data, and determining a historical radar image data set based on the historical echo signal data; determining a training image set and a test image set based on the historical radar image data set, and constructing a noise reduction neural network model based on the training image set; evaluating the noise reduction neural network model based on the test image set, and performing second optimization on the noise reduction neural network model; and performing noise reduction processing on a real-time radar image data set determined by the real-time radar echo data based on the optimized noise reduction neural network model. The noise reduction learning ability of the model can be enhanced, the processing ability of the model to a complex noise environment can be improved, gradual fine adjustment of the model can be realized to form a performance closed loop, high-precision rapid processing of real-time radar data can be realized, the denoising effect of a radar image can be improved, and the noise reduction precision in the complex environment can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a radar image denoising processing method based on deep learning. Background Art

[0002] In the field of radar image processing, traditional noise reduction methods mainly include filtering-based algorithms (such as mean filtering, median filtering, and Wiener filtering) and statistical methods (such as Gaussian noise modeling). These methods smooth or remove noise in radar images by making assumptions about noise characteristics. However, these traditional methods have limitations when facing complex and changeable noise environments. It is difficult to effectively distinguish target signals from background noise, which can easily lead to target information loss or residual noise. At the same time, with the improvement of the resolution of modern radar systems and the complexity of application scenarios, the computational efficiency and adaptability of traditional methods can no longer meet the requirements of real-time and high precision.

[0003] Therefore, the present invention provides a radar image denoising method based on deep learning. Summary of the invention

[0004] The present invention provides a radar image denoising processing method based on deep learning. The method determines a historical radar image data set by analyzing historical echo data signals, constructs and optimizes a denoising neural network model, and performs denoising processing on a real-time radar image data set determined by real-time radar echo data through the optimized denoising neural network model. The denoising learning ability of the model can be enhanced, and the model's processing ability for complex noise environments can be improved. The model can be gradually refined to form a performance closed loop, and high-precision and rapid processing of real-time radar data can be achieved. The denoising effect of radar images is improved, and the denoising accuracy in complex environments is improved.

[0005] The present invention provides a radar image denoising processing method based on deep learning, comprising: 101: Acquire historical echo signal data, and determine a historical radar image data set based on the historical echo signal data; 102: Determine a training image set and a test image set based on the historical radar image data set, and build a denoising neural network model based on the training image set; 103: Evaluate the denoising neural network model based on the test image set, and perform a second optimization on the denoising neural network model; 104: Performing denoising on a real-time radar image data set determined by real-time radar echo data based on the optimized denoising neural network model.

[0006] According to a radar image denoising method based on deep learning provided by the present invention, historical echo signal data is obtained, and a historical radar image data set is determined based on the historical echo signal data, including: Acquire historical echo signal data of the radar within a historical specified time period, wherein the historical echo signal data includes the radar signal at each time point within the historical specified time period; Determine a segmented time window of a historical specified time period based on the pulse period, and divide the historical echo signal into a plurality of time echo signals based on the segmented time window; Based on the pulse compression algorithm, each time echo signal is converted into corresponding distance information, and the distance dimension information of each time echo signal is determined, wherein the distance dimension information includes the echo intensity of multiple distance units; Performing Fourier transform on the distance dimension data of each time echo signal to determine the Doppler frequency information of each distance dimension data; Determine a three-dimensional matrix of historical echo signal data based on all time echo signals, distance dimension information of all time echo signals, and Doppler frequency information of all distance units in the distance dimension information; Determine the two-dimensional matrix at each time point within the specified historical time period based on the time dimension of the three-dimensional matrix; Determine the historical range-Doppler image of each segmented time window based on the two-dimensional matrix of all time points in each segmented time window, and time-mark the historical range-Doppler image of each segmented time window; The historical range-Doppler images of all segmented time windows are sequentially combined based on the time tags to determine the historical radar image data set.

[0007] According to a radar image denoising method based on deep learning provided by the present invention, a training image set and a test image set are determined based on a historical radar image data set, including: Preprocessing the historical radar image dataset, wherein the preprocessing includes image quality inspection, image normalization, and image size adjustment; determining a target noise-free image for each historical range-Doppler image in the preprocessed radar image dataset; The division ratio of the historical radar image data set is determined, and the training image set and the test image set are determined based on the division ratio, the time tags of all historical range-Doppler images in the historical radar image set, and the target noise-free image.

[0008] According to a radar image denoising processing method based on deep learning provided by the present invention, a denoising neural network model is constructed based on a training image set, comprising: The input layer of the denoising neural network model receives all historical range-Doppler images in the training image set; Based on multiple convolutional layers of the denoising neural network model and the nonlinear activation function of each convolutional layer, spatial feature extraction and noise feature extraction are performed on each historical range-Doppler image received by the multi-input layer to determine the local feature vector of each historical range-Doppler image; Perform fine-grained feature extraction on each historical range-Doppler image in the training image set based on multiple deep convolutional layers of the denoising neural network model to determine a fine-grained feature vector for each historical range-Doppler image; Reconstructing the local feature vector and the fine-grained feature vector of each historical range-Doppler image in the training image set respectively using multiple deconvolution layers of the denoising neural network model to determine the predicted noise-free image of each historical range-Doppler image; Determining a comprehensive loss value based on a target noise-free image and a predicted noise-free image of a historical range-Doppler image in a training image set; The comprehensive loss value is compared with the preset loss threshold. If the comprehensive loss value is greater than or equal to the preset loss threshold, the denoising neural network model is first optimized based on the optimization algorithm until the comprehensive loss value is less than the preset loss threshold.

[0009] According to a radar image denoising method based on deep learning provided by the present invention, a comprehensive loss value is determined based on a target noise-free image of a historical range-Doppler image in a training image set and a predicted noise-free image, including: ; ; ; ; in, represents the comprehensive loss value, represents the loss value of the a-th historical distance-Doppler image in the training image set, They represent the maximum loss value and the minimum loss value of all historical distance-Doppler images in the training image set, respectively. Nu represents the number of historical distance-Doppler images in the training image set. represents the pixel loss value of the a-th historical range-Doppler image in the training image set, represents the frequency domain loss value of the a-th historical range-Doppler image in the training image set, represents the probability prediction value of the discriminator D that the a-th historical distance-Doppler image in the training image set belongs to the real image, express The expected value of , They represent the pixel value of the i-th pixel in the predicted noise-free image and the target noise-free image of the a-th historical range-Doppler image in the training image set, respectively; N1 represents the number of pixels in the predicted noise-free image and the target noise-free image of the historical range-Doppler image, represents the structural similarity value in the predicted noise-free image and the target noise-free image of the a-th historical range-Doppler image in the training image set, represents the weight of the structural similarity value, They represent the predicted noise-free image of the a-th historical range-Doppler image in the training image set and the k-th frequency component in the target noise-free image, respectively. They represent the phase of the kth frequency component in the predicted noise-free image and the target noise-free image of the ath historical range-Doppler image in the training image set, respectively. N2 represents the number of frequency components of the predicted noise-free image and the target noise-free image of the historical range-Doppler image. represents the weight of the frequency component, represents the weight of the phase of the frequency component, Represents the loss constant, with a value of 0.1.

[0010] According to a radar image denoising processing method based on deep learning provided by the present invention, a denoising neural network model is evaluated based on a test image set, and a second optimization is performed on the denoising neural network model, including: Input each historical range-Doppler image in the test set image into the first optimized denoising neural network model, and determine a predicted noise-free image for each historical range-Doppler image in the test set image based on an output result of the denoising neural network model; Compare the predicted noise-free images of all historical range-Doppler images in the test set images with the target noise-free images to determine the denoising effect of the denoising neural network model; Determine whether the denoising effect meets the preset denoising expectations. If not, adjust the model parameters of the denoising neural network model based on the denoising effect, and perform a second optimization on the denoising neural network model until the denoising effect meets the preset denoising expectations.

[0011] According to a radar image denoising method based on deep learning provided by the present invention, denoising is performed on a real-time radar image data set determined by real-time radar echo data based on an optimized denoising neural network model, comprising: Acquire real-time radar echo data, and determine a real-time radar image set based on the real-time radar echo data; The real-time radar image set is input into the second optimized denoising neural network model, and the denoising processing result of the real-time radar image set is determined based on the output result of the denoising neural network model.

[0012] According to a radar image denoising method based on deep learning provided by the present invention, a real-time radar image set includes a plurality of real-time range-Doppler images; The denoising result includes a predicted denoised image of each real-time range-Doppler image in the real-time radar image set.

[0013] Compared with the prior art, the present invention has the following beneficial effects: By analyzing the historical echo data signal to determine the historical radar image data set, building and optimizing the denoising neural network model, and using the optimized denoising neural network model to perform denoising on the real-time radar image data set determined by the real-time radar echo data, the model's denoising learning ability can be enhanced, and the model's ability to handle complex noise environments can be improved. The model can be gradually refined to form a performance closed loop, and high-precision and rapid processing of real-time radar data can be achieved, thereby improving the denoising effect of radar images and the denoising accuracy in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 It is a flowchart of a radar image denoising method based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Embodiment 1: The embodiment of the present invention provides a radar image denoising method based on deep learning, such as Figure 1 As shown, including: 101: Acquire historical echo signal data, and determine a historical radar image data set based on the historical echo signal data; 102: Determine a training image set and a test image set based on the historical radar image data set, and build a denoising neural network model based on the training image set; 103: Evaluate the denoising neural network model based on the test image set, and perform a second optimization on the denoising neural network model; 104: Performing denoising on a real-time radar image data set determined by real-time radar echo data based on the optimized denoising neural network model.

[0018] In this embodiment, the historical echo signal data is the echo signal record collected by the radar in the past period of time, including the distance and speed information of the target object and the environmental noise. By processing these historical echo signal data, the corresponding historical radar image data set is generated.

[0019] In this embodiment, the historical radar image data set is divided into a training image set and a test image set. The training set is used for learning the model, and the test set is used to verify the model performance. Through deep learning, the training image set is used to train the neural network model so that it can learn how to extract target features from noisy images and remove noise.

[0020] In this embodiment, a test image set is used to evaluate the denoising effect of the initial neural network model to determine its adaptability and processing accuracy to different noise environments. A second round of optimization is performed on the model based on the evaluation results to improve the robustness and generalization ability of the model by adjusting the model architecture, hyperparameters or training strategies.

[0021] In this embodiment, after the model optimization is completed, the radar echo data received in real time is converted into a real-time radar image data set and input into the optimized denoising neural network model for real-time processing. The model can quickly generate clear denoised images, extract key features of the target, and remove environmental noise.

[0022] The beneficial effects of the above technical solution are as follows: by analyzing the historical echo data signal to determine the historical radar image data set, constructing and optimizing the denoising neural network model, and performing denoising processing on the real-time radar image data set determined by the real-time radar echo data through the optimized denoising neural network model, the denoising learning ability of the model can be enhanced, and the model's processing ability for complex noise environments can be improved. The model can be gradually refined to form a performance closed loop, and high-precision and rapid processing of real-time radar data can be achieved, the denoising effect of radar images can be improved, and the denoising accuracy in complex environments can be improved.

[0023] Embodiment 2: The embodiment of the present invention provides a radar image denoising processing method based on deep learning, which obtains historical echo signal data and determines a historical radar image data set based on the historical echo signal data, including: Acquire historical echo signal data of the radar within a historical specified time period, wherein the historical echo signal data includes the radar signal at each time point within the historical specified time period; Determine a segmented time window of a historical specified time period based on the pulse period, and divide the historical echo signal into a plurality of time echo signals based on the segmented time window; Based on the pulse compression algorithm, each time echo signal is converted into corresponding distance information, and the distance dimension information of each time echo signal is determined, wherein the distance dimension information includes the echo intensity of multiple distance units; Performing Fourier transform on the distance dimension data of each time echo signal to determine the Doppler frequency information of each distance dimension data; Determine a three-dimensional matrix of historical echo signal data based on all time echo signals, distance dimension information of all time echo signals, and Doppler frequency information of all distance units in the distance dimension information; Determine the two-dimensional matrix at each time point within the specified historical time period based on the time dimension of the three-dimensional matrix; Determine the historical range-Doppler image of each segmented time window based on the two-dimensional matrix of all time points in each segmented time window, and time-mark the historical range-Doppler image of each segmented time window; The historical range-Doppler images of all segmented time windows are sequentially combined based on the time tags to determine the historical radar image data set.

[0024] In this embodiment, the echo signal of the radar system is collected within a historical specified time period, and the radar signal at each time point within the time period is recorded. The echo signal reflects the characteristic information of the radar transmission signal reflected by the object, including the physical characteristics of the target object and the environment.

[0025] In this embodiment, according to the pulse period of the radar, the historical specified time period is divided into several time windows, each window represents a segmented time range, and the historical echo signal is divided into multiple time echo signals according to these segmented time windows.

[0026] In this embodiment, a pulse compression algorithm is used to process each time echo signal, converting it from the time domain to the distance domain to obtain the distance dimension information of each echo signal. The distance dimension information represents the echo intensity at different distance units and is a direct reflection of the target distance and the reflected signal intensity.

[0027] In this embodiment, the distance dimension data of each time echo signal is Fourier transformed to extract the Doppler frequency information corresponding to each distance unit. The Doppler frequency reflects the speed characteristics of the target, and the position and motion state of the target can be determined in combination with the distance information.

[0028] In this embodiment, all time echo signals, their distance dimension information and Doppler frequency information of the distance unit are organized into a three-dimensional matrix, the dimensions of which are time, distance and Doppler frequency, respectively, which fully describes the radar observation information within a specified historical time period.

[0029] In this embodiment, based on the time dimension of the three-dimensional matrix, the data at each time point is extracted into a two-dimensional matrix, which represents the target distribution at that time point and its distance and speed information.

[0030] In this embodiment, the two-dimensional matrix of all time points in each segmented time window is synthesized to generate a historical range-Doppler image of the segmented time window. The range-Doppler image is a typical form of radar image, reflecting the distance and speed characteristics of the target. A time mark is attached to the image of each segmented time window to indicate the time sequence of the image.

[0031] In this embodiment, the horizontal axis of the range-Doppler image represents the Doppler frequency (target speed unit); the vertical axis represents the range unit (target range unit); and the color represents the echo intensity (the higher the intensity, the brighter the color, and a pseudo-color image can be used).

[0032] In this embodiment, based on the time stamp, the historical range-Doppler images of all segmented time windows are combined in chronological order to generate a complete historical radar image data set, which provides a basis for subsequent noise reduction and analysis.

[0033] In this embodiment, the pulse period represents the time interval of the radar transmitting signal and is used to determine the time window of signal sampling.

[0034] In this embodiment, the pulse compression algorithm means converting the radar echo from the time domain to the distance domain through signal processing technology to improve the distance resolution.

[0035] The beneficial effects of the above technical solution are as follows: acquiring historical echo signal data and determining the historical radar image data set based on the historical echo signal data can realize orderly and efficient organization of historical radar data, improve data integrity and accuracy, and provide a high-quality input data set for building a denoising neural network model.

[0036] Embodiment 3: The embodiment of the present invention provides a radar image denoising method based on deep learning, which determines a training image set and a test image set based on a historical radar image data set, including: Preprocessing the historical radar image dataset, wherein the preprocessing includes image quality inspection, image normalization, and image size adjustment; determining a target noise-free image for each historical range-Doppler image in the preprocessed radar image dataset; The division ratio of the historical radar image data set is determined, and the training image set and the test image set are determined based on the division ratio, the time tags of all historical range-Doppler images in the historical radar image set, and the target noise-free image.

[0037] In this embodiment, the image quality check means performing a quality assessment on each historical range-Doppler image in the historical radar image data set, eliminating invalid images caused by signal loss or abnormality, and ensuring the reliability and integrity of the data.

[0038] In this embodiment, image normalization means standardizing the pixel values ​​of the image (such as normalizing to the range of [0, 1] or [-1, 1]) to make the distribution of image data consistent, eliminate data deviations caused by different sampling dynamic ranges, and facilitate the convergence and performance improvement of the deep learning model.

[0039] In this embodiment, image resizing means uniformly adjusting the size of the image (such as cropping or interpolation processing) to a fixed size to facilitate the input and calculation of the neural network model while avoiding calculation problems caused by inconsistent sizes.

[0040] In this embodiment, for historical range-Doppler images, a target noise-free version of each image is extracted through prior knowledge or algorithm processing. The target noise-free image represents the ideal output after denoising and is used as "label" data for supervised learning to guide the denoising neural network model to learn how to remove noise and retain target features. These noise-free images can be generated by traditional filtering methods, expert annotation, or high-quality radar systems to ensure that they are as close as possible to the actual target reflection characteristics.

[0041] In this embodiment, the data is divided according to a common ratio (which can be 80% training set, 20% test set or 70% training set, 30% test set), the training set is used for model learning, and the test set is used for model verification, ensuring the independence of training and evaluation.

[0042] In this embodiment, the data is divided in chronological order in combination with time tags to ensure that the training set and the test set contain representative and diverse image samples to avoid overfitting.

[0043] In this embodiment, when dividing the training set and the test set, it is ensured that the target noise-free image corresponding to each range-Doppler image is also correctly divided to ensure data alignment for supervised learning.

[0044] The beneficial effects of the above technical solution are as follows: determining the training image set and the test image set based on the historical radar image data set can provide high-quality supervision signals, enhance the denoising learning ability of the model, and improve the robustness and accuracy of the denoising model in complex radar image processing tasks.

[0045] Embodiment 4: The embodiment of the present invention provides a radar image denoising processing method based on deep learning, which constructs a denoising neural network model based on a training image set, including: The input layer of the denoising neural network model receives all historical range-Doppler images in the training image set; Based on multiple convolutional layers of the denoising neural network model and the nonlinear activation function of each convolutional layer, spatial feature extraction and noise feature extraction are performed on each historical range-Doppler image received by the multi-input layer to determine the local feature vector of each historical range-Doppler image; Perform fine-grained feature extraction on each historical range-Doppler image in the training image set based on multiple deep convolutional layers of the denoising neural network model to determine a fine-grained feature vector for each historical range-Doppler image; Reconstructing the local feature vector and the fine-grained feature vector of each historical range-Doppler image in the training image set respectively using multiple deconvolution layers of the denoising neural network model to determine the predicted noise-free image of each historical range-Doppler image; Determining a comprehensive loss value based on a target noise-free image and a predicted noise-free image of a historical range-Doppler image in a training image set; The comprehensive loss value is compared with the preset loss threshold. If the comprehensive loss value is greater than or equal to the preset loss threshold, the denoising neural network model is first optimized based on the optimization algorithm until the comprehensive loss value is less than the preset loss threshold.

[0046] In this embodiment, the input layer of the denoising neural network model directly receives all historical range-Doppler images in the training image set, which contain target features and noise information. The function of the input layer is to take the original image as input and perform preliminary processing for subsequent convolutional layers to extract features.

[0047] In this embodiment, multiple convolutional layers of a denoising neural network are used to extract features from an image. The convolutional layer performs local operations on the image through a convolution kernel to capture spatial features (target distribution) and noise features. Non-linear activation functions (such as ReLU) introduce non-linear transformations to enhance the model's ability to express complex patterns and generate local feature vectors to describe the feature distribution in a small range in the image.

[0048] In this embodiment, multiple deep convolutional layers are used to further perform fine-grained feature extraction on the image. These deep convolutional layers have more convolution kernels and deeper levels, and can capture higher-level and more detailed features, generate fine-grained feature vectors, and further separate target and noise features.

[0049] In this embodiment, local feature vectors and fine-grained feature vectors are gradually decoded and reconstructed through multiple deconvolution layers to generate a predicted noise-free image. The deconvolution layer restores the low-dimensional feature vector to a high-dimensional image form and uses the extracted target features to restore the noise-free version of the image.

[0050] In this embodiment, if the comprehensive loss value is greater than or equal to the preset loss threshold, it indicates that the predictive ability of the model is insufficient and needs further optimization. The parameters of the model (such as convolution kernel weights) are adjusted through optimization algorithms (such as gradient descent or Adam algorithm) to continuously reduce the loss value until the comprehensive loss value is less than the preset threshold.

[0051] The beneficial effects of the above technical solution are as follows: constructing a denoising neural network model based on a training image set can enhance the model's ability to handle complex noise environments and significantly improve the denoising effect of radar images.

[0052] Embodiment 5: The embodiment of the present invention provides a radar image denoising method based on deep learning, which determines a comprehensive loss value based on a target noise-free image of a historical range-Doppler image in a training image set and a predicted noise-free image, including: ; ; ; ; in, represents the comprehensive loss value, represents the loss value of the a-th historical distance-Doppler image in the training image set, They represent the maximum loss value and the minimum loss value of all historical distance-Doppler images in the training image set, respectively. Nu represents the number of historical distance-Doppler images in the training image set. represents the pixel loss value of the a-th historical range-Doppler image in the training image set, represents the frequency domain loss value of the a-th historical range-Doppler image in the training image set, represents the probability prediction value of the discriminator D that the a-th historical distance-Doppler image in the training image set belongs to the real image, express The expected value of , They represent the pixel value of the i-th pixel in the predicted noise-free image and the target noise-free image of the a-th historical range-Doppler image in the training image set, respectively; N1 represents the number of pixels in the predicted noise-free image and the target noise-free image of the historical range-Doppler image, represents the structural similarity value in the predicted noise-free image and the target noise-free image of the a-th historical range-Doppler image in the training image set, represents the weight of the structural similarity value, They represent the predicted noise-free image of the a-th historical range-Doppler image in the training image set and the k-th frequency component in the target noise-free image, respectively. They represent the phase of the kth frequency component in the predicted noise-free image and the target noise-free image of the ath historical range-Doppler image in the training image set, respectively. N2 represents the number of frequency components of the predicted noise-free image and the target noise-free image of the historical range-Doppler image. represents the weight of the frequency component, represents the weight of the phase of the frequency component, Represents the loss constant, with a value of 0.1.

[0053] In this embodiment, The value range is 0-1. Taking the logarithm emphasizes the small probability prediction values ​​(such as close to 0) the penalty for loss will be greater if Too low and the loss value will quickly become high.

[0054] In this embodiment, represents the adversarial loss value of the a-th range-Doppler image in the training image set.

[0055] The beneficial effects of the above technical solution are as follows: by determining the comprehensive loss value based on the target noise-free image of the historical range-Doppler image in the training image set and the predicted noise-free image, the high precision and robustness of the denoising result can be improved, and the denoising neural network model can generate high-quality predicted noise-free images.

[0056] Embodiment 6: The embodiment of the present invention provides a radar image denoising processing method based on deep learning, which evaluates a denoising neural network model based on a test image set and performs a second optimization on the denoising neural network model, including: Input each historical range-Doppler image in the test set image into the first optimized denoising neural network model, and determine a predicted noise-free image for each historical range-Doppler image in the test set image based on an output result of the denoising neural network model; Compare the predicted noise-free images of all historical range-Doppler images in the test set images with the target noise-free images to determine the denoising effect of the denoising neural network model; Determine whether the denoising effect meets the preset denoising expectations. If not, adjust the model parameters of the denoising neural network model based on the denoising effect, and perform a second optimization on the denoising neural network model until the denoising effect meets the preset denoising expectations.

[0057] In this embodiment, each historical range-Doppler image in the test set is taken as input and passed into the denoising neural network model that has undergone the first round of optimization. The model extracts features based on the input image and generates a corresponding predicted noise-free image based on the parameters optimized by previous training, that is, the denoising output of the model for the test data.

[0058] In this embodiment, the predicted noise-free image output by the model is compared one by one with the target noise-free image in the test set. The target noise-free image represents the ideal denoising result, while the predicted noise-free image is the actual denoising result of the model. By comparing the two, the performance of the model in the denoising task is evaluated.

[0059] In this embodiment, the denoising effect evaluation can use quantitative indicators (such as mean square error MSE, signal-to-noise ratio SNR, or structural similarity SSIM) to evaluate the overall denoising effect of the test set, quantify the denoising ability of the model, and determine whether the model accurately restores the target characteristics while effectively suppressing noise.

[0060] In this embodiment, the denoising effect is compared with the preset denoising expectation, which is the lower limit of the performance indicator set according to the actual application requirements (such as MSE is less than a certain value, SNR reaches a certain threshold, etc.). If the denoising effect of the model is lower than this expectation, it means that the current model performance is not enough and needs further optimization.

[0061] In this embodiment, if the denoising effect does not meet expectations, the model parameters (such as convolution kernel weights, learning rate, etc.) are adjusted and the denoising neural network model is optimized for a second round in combination with the optimization algorithm. The optimization process will further reduce the loss value and improve the denoising effect of the model on the test set. This process is iterative until the model performance reaches or exceeds the preset denoising expectations, thereby ensuring the reliability and accuracy of the model in practical applications.

[0062] The beneficial effects of the above technical solution are as follows: by evaluating the denoising neural network model based on the test image set and performing a second optimization on the denoising neural network model, the denoising effect can be accurately quantified, and the model can be gradually refined to form a performance closed loop, thereby improving the denoising accuracy in complex environments.

[0063] Embodiment 7: The embodiment of the present invention provides a radar image denoising method based on deep learning, which performs denoising on a real-time radar image data set determined by real-time radar echo data based on an optimized denoising neural network model, including: Acquire real-time radar echo data, and determine a real-time radar image set based on the real-time radar echo data; The real-time radar image set is input into the second optimized denoising neural network model, and the denoising processing result of the real-time radar image set is determined based on the output result of the denoising neural network model.

[0064] In this embodiment, the radar device continuously receives real-time echo signals from the target area. These echo data are converted into visual two-dimensional range-Doppler images after signal processing, which reflect the distance and speed of the target object and the noise distribution in the surrounding environment in real time.

[0065] In this embodiment, the continuously acquired radar echo data are divided according to the radar pulse period to generate a real-time radar image set.

[0066] In this embodiment, the real-time radar image set is input into the denoising neural network model that has completed the second round of optimization.

[0067] In this embodiment, the model processes each input real-time range-Doppler image one by one to generate corresponding predicted noise-free images, which remove environmental noise and retain target features, and finally outputs the denoising results of the real-time radar image set.

[0068] The beneficial effect of the above technical solution is: based on the optimized denoising neural network model, the real-time radar image data set determined by the real-time radar echo data is subjected to denoising processing, which can improve the denoising adaptability of the radar echo data in dynamic scenes.

[0069] Embodiment 8: The embodiment of the present invention provides a radar image denoising method based on deep learning, wherein a real-time radar image set includes a plurality of real-time range-Doppler images; The denoising result includes a predicted denoised image of each real-time range-Doppler image in the real-time radar image set.

[0070] In this embodiment, the denoising neural network model processes each real-time range-Doppler image in the real-time radar image set one by one, and outputs the corresponding predicted denoised image. The predicted denoised image retains the distance and speed characteristics of the target object, while removing environmental noise, interference signals and system errors, making the radar image clearer and more accurate. Finally, the denoising processing result is composed of all the predicted denoised images of the real-time radar image set, providing a complete denoising view.

[0071] The beneficial effects of the above technical solution are as follows: determining the real-time radar image set and the denoising processing results can ensure the accuracy and continuity of the denoising process and output a high-quality predicted denoised image set.

[0072] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0073] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A radar image denoising method based on deep learning, characterized in that: include: 101: Acquire historical echo signal data, and determine a historical radar image data set based on the historical echo signal data; 102: Determine a training image set and a test image set based on a historical radar image data set, Determining a comprehensive loss value based on a target noise-free image of a historical range-Doppler image in a training image set and a predicted noise-free image determined based on a denoising neural network model; Compare the comprehensive loss value with the preset loss threshold, and if the comprehensive loss value is greater than or equal to the preset loss threshold, perform a first optimization on the denoising neural network model based on the optimization algorithm until the comprehensive loss value is less than the preset loss threshold; 103: Evaluate the denoising neural network model based on the test image set, and perform a second optimization on the denoising neural network model; 104: Performing denoising on a real-time radar image data set determined by real-time radar echo data based on the optimized denoising neural network model.

2. The radar image denoising method based on deep learning according to claim 1, characterized in that: Obtain historical echo signal data, and determine a historical radar image data set based on the historical echo signal data, including: Acquire historical echo signal data of the radar within a historical specified time period, wherein the historical echo signal data includes the radar signal at each time point within the historical specified time period; Determine a segmented time window of a historical specified time period based on the pulse period, and divide the historical echo signal into a plurality of time echo signals based on the segmented time window; Based on the pulse compression algorithm, each time echo signal is converted into corresponding distance information, and the distance dimension information of each time echo signal is determined, wherein the distance dimension information includes the echo intensity of multiple distance units; Performing Fourier transform on the distance dimension data of each time echo signal to determine the Doppler frequency information of each distance dimension data; Determine a three-dimensional matrix of historical echo signal data based on all time echo signals, distance dimension information of all time echo signals, and Doppler frequency information of all distance units in the distance dimension information; Determine the two-dimensional matrix at each time point within the specified historical time period based on the time dimension of the three-dimensional matrix; Determine the historical range-Doppler image of each segmented time window based on the two-dimensional matrix of all time points in each segmented time window, and time-mark the historical range-Doppler image of each segmented time window; The historical range-Doppler images of all segmented time windows are sequentially combined based on the time tags to determine the historical radar image data set.

3. The radar image denoising method based on deep learning according to claim 2, characterized in that: The training image set and the test image set are determined based on the historical radar image dataset, including: Preprocessing the historical radar image dataset, wherein the preprocessing includes image quality inspection, image normalization, and image size adjustment; determining a target noise-free image for each historical range-Doppler image in the preprocessed radar image dataset; The division ratio of the historical radar image data set is determined, and the training image set and the test image set are determined based on the division ratio, the time tags of all historical range-Doppler images in the historical radar image set, and the target noise-free image.

4. The radar image denoising method based on deep learning according to claim 3, characterized in that: Determine the predicted noise-free image based on the denoising neural network model, including: The input layer of the denoising neural network model receives all historical range-Doppler images in the training image set; Based on multiple convolutional layers of the denoising neural network model and the nonlinear activation function of each convolutional layer, spatial feature extraction and noise feature extraction are performed on each historical range-Doppler image received by the multi-input layer to determine the local feature vector of each historical range-Doppler image; Perform fine-grained feature extraction on each historical range-Doppler image in the training image set based on multiple deep convolutional layers of the denoising neural network model to determine a fine-grained feature vector for each historical range-Doppler image; The local feature vectors and fine-grained feature vectors of each historical range-Doppler image in the training image set are reconstructed using multiple deconvolution layers of the denoising neural network model to determine the predicted noise-free image of each historical range-Doppler image.

5. The radar image denoising method based on deep learning according to claim 4, characterized in that: The comprehensive loss value is determined based on the target noise-free image and the predicted noise-free image of the historical range-Doppler image in the training image set. include: ； ； ； ； in, represents the comprehensive loss value, represents the loss value of the a-th historical distance-Doppler image in the training image set, They represent the maximum loss value and the minimum loss value of all historical distance-Doppler images in the training image set, respectively. Nu represents the number of historical distance-Doppler images in the training image set. represents the pixel loss value of the a-th historical range-Doppler image in the training image set, represents the frequency domain loss value of the a-th historical range-Doppler image in the training image set, represents the probability prediction value of the discriminator D that the a-th historical distance-Doppler image in the training image set belongs to the real image, express The expected value of , They represent the pixel value of the i-th pixel in the predicted noise-free image and the target noise-free image of the a-th historical range-Doppler image in the training image set, respectively; N1 represents the number of pixels in the predicted noise-free image and the target noise-free image of the historical range-Doppler image, represents the structural similarity value in the predicted noise-free image and the target noise-free image of the a-th historical range-Doppler image in the training image set, represents the weight of the structural similarity value, They represent the predicted noise-free image of the a-th historical range-Doppler image in the training image set and the k-th frequency component in the target noise-free image, respectively. They represent the phase of the kth frequency component in the predicted noise-free image and the target noise-free image of the ath historical range-Doppler image in the training image set, respectively. N2 represents the number of frequency components of the predicted noise-free image and the target noise-free image of the historical range-Doppler image. represents the weight of the frequency component, represents the weight of the phase of the frequency component, Represents the loss constant, with a value of 0.

1.

6. The radar image denoising method based on deep learning according to claim 4, characterized in that: The comprehensive loss value is determined based on the target noise-free image and the predicted noise-free image of the historical range-Doppler image in the training image set, including: Input each historical range-Doppler image in the test set image into the first optimized denoising neural network model, and determine a predicted noise-free image for each historical range-Doppler image in the test set image based on an output result of the denoising neural network model; Compare the predicted noise-free images of all historical range-Doppler images in the test set images with the target noise-free images to determine the denoising effect of the denoising neural network model; Determine whether the denoising effect meets the preset denoising expectations. If not, adjust the model parameters of the denoising neural network model based on the denoising effect, and perform a second optimization on the denoising neural network model until the denoising effect meets the preset denoising expectations.

7. The radar image denoising method based on deep learning according to claim 1, characterized in that: Based on the optimized denoising neural network model, the real-time radar image data set determined by the real-time radar echo data is subjected to denoising, including: Acquire real-time radar echo data, and determine a real-time radar image set based on the real-time radar echo data; The real-time radar image set is input into the second optimized denoising neural network model, and the denoising processing result of the real-time radar image set is determined based on the output result of the denoising neural network model.

8. The radar image denoising method based on deep learning according to claim 1, characterized in that: The real-time radar image set includes multiple real-time range-Doppler images; The denoising result includes a predicted denoised image of each real-time range-Doppler image in the real-time radar image set.

Images