A data enhancement system for RD time-frequency data

Abstract

A data enhancement system for RD time-frequency data, wherein a data processing module receives RD time-frequency data, converts the RD time-frequency data into image data, performs grayscale processing on the image data, and generates an original grayscale image; a noise adjustment module uniformly adjusts the pixel values ​​belonging to the noise data part in the original grayscale image to generate a new grayscale image information matrix, and then fills the original target range pixel value matrix information back into the newly generated grayscale image information matrix to obtain a new grayscale image; a noise filtering module receives the original grayscale image, performs noise elimination processing, and obtains a grayscale image after eliminating periodic noise; a clustering splicing module generates a data-enhanced grayscale image, and finally restores the data-enhanced grayscale image to RD time-frequency data. The present invention utilizes limited RD time-frequency data to perform data enhancement and obtain more RD time-frequency data for deep learning.

Description

Technical Field

[0001] The invention belongs to the field of aerospace and image processing technology, and in particular to a data enhancement system for RD time-frequency data. Background Art

[0002] With the continuous development of radar technology, various radar technologies have been applied in various fields. The essence of its work is to determine the physical position of the target through radio waves. For example, in life, civil radar has a wide range of applications, such as supporting meteorological services, aviation and ship traffic control. Since the birth of radar technology, it has played an irreplaceable role in the military field, usually used to search, monitor and identify military targets.

[0003] Doppler radar is a common type of military radar. It is a radar that uses the Doppler effect to detect the position and relative speed of moving targets. When this radar emits a fixed-frequency pulse wave to scan the space, if it encounters a moving target, there will be a frequency difference between the frequency of the echo and the frequency of the transmitted wave, that is, the Doppler frequency. According to the size of the Doppler frequency, the radial relative motion speed of the target to the radar image can be measured; according to the time difference between the transmitted pulse and the received pulse, the distance of the target can be measured. This Doppler ranging radar has high ranging accuracy, strong anti-interference ability, good real-time performance, and can meet the needs of dynamic light intensity measurement. Although Doppler radar has many advantages, the processing of R (Range) D (Doppler) time-frequency data obtained by its drawing is an industry problem. Because if the RD time-frequency data we directly obtain is not processed, it is difficult to obtain useful data from it. And the RD time-frequency data is often full of a lot of interference data. It is a time-consuming and laborious task to find the real target from it, and the accuracy cannot be guaranteed.

[0004] With the continuous development of image processing technology and artificial intelligence technology, we know that training deep neural network models with a large amount of RD time-frequency data (training data) with accurate target annotation can effectively improve the speed and accuracy of obtaining targets from RD time-frequency data. However, if we can enhance the training data and improve the quality and quantity of the training data, it will further improve the speed and accuracy of obtaining targets from RD time-frequency data. Currently, common data enhancement technologies are often used to process image data, but there are relatively few data enhancements for RD time-frequency data. Therefore, it is very important to enhance the data of RD time-frequency data to improve the effectiveness and quantity of the data.

[0005] According to the literature, it is common to apply data enhancement methods to ordinary image data at home and abroad, but there is no data enhancement example for RD time-frequency data. Therefore, it is of great significance and practical value to develop a data enhancement technology for RD time-frequency data. Summary of the invention

[0006] The technical problem solved by the present invention is: to overcome the deficiencies of the prior art, to provide a data enhancement system for RD time-frequency data, which, relying on advanced data processing technology, can effectively improve the quality and quantity of RD time-frequency data, and remove or amplify interference data in the original data. The data-enhanced RD time-frequency data can be applied to deep learning technology to train a deep neural network for identifying target data, thereby providing strong technical support for quickly, accurately and automatically identifying target data in RD time-frequency data.

[0007] The technical solution of the present invention is:

[0008] A data enhancement system for RD time-frequency data, comprising: a data processing module, a noise adjustment module, a noise filtering module and a clustering splicing module;

[0009] Data processing module: receiving RD time-frequency data, converting the RD time-frequency data into image data, performing grayscale processing on the image data, generating a corresponding grayscale image as the original grayscale image, and transmitting it to the noise adjustment module and the noise filtering module;

[0010] Noise adjustment module: receives the original grayscale image transmitted by the data processing module, uniformly adjusts the pixel values ​​of the noise data part of the original grayscale image, enlarges or reduces the pixel values ​​of the noise area in the original grayscale image according to the use needs to obtain a new grayscale image, and then transmits the new grayscale image to the clustering and splicing module;

[0011] Specifically, the following steps are performed: according to the range information of the target data in the RD time-frequency data, the pixel value data matrix information corresponding to the target data is extracted and saved, and then the pixel values ​​other than this part of the target data are multiplied by the interference intensity coefficient to form a new grayscale image information matrix, and the pixel value data corresponding to the previously extracted target data is backfilled into the newly generated grayscale image information matrix, thereby obtaining a new grayscale image and transmitting it to the clustering splicing module;

[0012] Noise filtering module: receives the original grayscale image transmitted by the data processing module, performs noise elimination, retains the target data points in the original grayscale image, obtains the grayscale image after eliminating periodic noise, and transmits it to the clustering splicing module;

[0013] Clustering and splicing module: Receive the new grayscale image transmitted by the noise adjustment module and the grayscale image after eliminating periodic noise transmitted by the noise filtering module, generate a data-enhanced grayscale image, and finally restore the data-enhanced grayscale image to RD time-frequency data to obtain the final result.

[0014] Optionally, the interference intensity coefficient in the noise adjustment module is greater than 1 when amplifying the noise, and is less than 1 otherwise.

[0015] Optionally, the method for identifying the noise data period in the original grayscale image in the noise filtering module is specifically:

[0016] 1) First, an adaptive algorithm is used to obtain the noise period;

[0017] 2) Setting the value of the noise period to L according to the image characteristics;

[0018] 3) Using the initial 3×3 pixel elimination template, the noise data in the grayscale image is subtracted and eliminated with L as the noise period value, and then the pixel values ​​in the result image are summed to obtain the pixel cumulative sum;

[0019] 4) Perform conditional judgment on the pixel accumulation and if the condition is met, the optimal cycle length is obtained and used as the cycle of the noise data; otherwise, go to step 5);

[0020] 5) Adjust the period length L to be longer, repeat steps 2) to 4) again until the optimal period length is obtained, and take an image of the periodic noise area as the elimination template.

[0021] Optionally, the method for condition determination in step 4) is specifically:

[0022] If the accumulated sum of the resulting pixel values ​​is less than the threshold, the condition is determined to be met, otherwise it is determined that the condition is not met.

[0023] Optionally, the threshold value in step 4) ranges from 10 to 20.

[0024] Optionally, the clustering and stitching module generates a data-enhanced grayscale image, specifically:

[0025] The processed grayscale images are clustered and divided by a clustering algorithm, and the grouping of the grayscale images is determined according to the clustering results. The grouped grayscale images are randomly spliced ​​within or across groups to generate data-enhanced grayscale images.

[0026] Optionally, the noise filtering module performs noise elimination processing, specifically:

[0027] The adaptive algorithm first identifies the period of the noise data in the original grayscale image, then takes an image of a periodic noise area as an elimination template, and subtracts the noise data of the remaining periodic noise areas in the original grayscale image in turn according to the elimination template to eliminate the noise.

[0028] The advantages of the present invention compared with the prior art are:

[0029] (1) The data enhancement technology for RD time-frequency data provided by the present invention provides a powerful solution for data processing and enhancement of RD time-frequency data. In the face of poor data quality, strong interference, small quantity, etc., it can maximize the quality and quantity of RD time-frequency data and achieve the purpose of data enhancement. In the application scenario of deep learning technology, it can play a typical role and assist deep neural networks to quickly, accurately and automatically identify target data in RD time-frequency data.

[0030] (2) The present invention designs a noise adjustment method for grayscale images, which can efficiently and reliably adjust the interference intensity of the background and other shapes of interference data in the grayscale image, and enhance the typical feature of interference intensity while ensuring that the annotation content (if any) in the data is retained, thereby generating RD time-frequency data with weak interference and RD time-frequency data with strong interference (if the RD time-frequency data contains original standard information, the original annotation information of the data is retained). These data can be directly applied to deep learning model training.

[0031] (3) The present invention designs an image periodic noise adaptive filtering algorithm, which can accurately and effectively eliminate periodic noise images. For noise images with uncertain periods, the algorithm can perform periodic interference adaptation. Without input, the periodic interference in the image can be automatically identified and eliminated. Compared with the traditional method of manually confirming the period and then eliminating noise, the human resource cost is greatly saved.

[0032] (4) The method designed in the present invention can enhance the RD time-frequency data according to the needs of users, including the removal and amplification of interference data, the increase of data volume, the classification of data, etc., so as to be suitable for different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a system block diagram of the present invention;

[0034] Figure 2 The data processing module workflow diagram of the present invention;

[0035] Figure 3 This is a working diagram of the noise adjustment module of the present invention;

[0036] Figure 4 This is a working diagram of the noise filtering module of the present invention;

[0037] Figure 5 This is a workflow diagram of the clustering and splicing module of the present invention. DETAILED DESCRIPTION

[0038] like Figure 1 As shown, a data enhancement system for RD time-frequency data includes: a data processing module, a noise adjustment module, a noise filtering module and a clustering splicing module.

[0039] Data processing module: receives RD time-frequency data, converts the RD time-frequency data into image data, performs grayscale processing on the image data, generates a corresponding grayscale image as the original grayscale image, and transmits it to the noise adjustment module and the noise filtering module; the grayscale image contains interference of various shapes including bright targets, dark backgrounds, and bright spots.

[0040] Noise adjustment module: receives the original grayscale image transmitted by the data processing module, uniformly adjusts the pixel values ​​belonging to the noise data part of the original grayscale image, enlarges or reduces the pixel values ​​of the noise area in the original grayscale image according to usage needs to obtain a new grayscale image, and then transmits the new grayscale image to the clustering and splicing module.

[0041] The method for uniformly adjusting the pixel values ​​belonging to the noise data part in the original grayscale image by the noise adjustment module is specifically as follows: uniformly adjusting the pixel values ​​belonging to the noise data part in the original grayscale image, extracting and saving the pixel value data matrix information corresponding to the target data according to the range information of the target data in the RD time-frequency data, and then multiplying the pixel values ​​​​except this part of the target data with the interference intensity coefficient (if the interference needs to be amplified, the coefficient is set to be greater than 1, otherwise the coefficient is set to be less than 1), thereby forming a new grayscale image information matrix and filling the pixel value data corresponding to the previously extracted target data into the newly generated grayscale image information matrix, thereby obtaining a new grayscale image and transmitting it to the clustering and splicing module.

[0042] Noise filtering module: Receives the original grayscale image transmitted by the data processing module, and automatically identifies the period of the noise data in the original grayscale image through an adaptive algorithm. Then, an image of a periodic noise area is taken as an elimination template. According to the elimination template, the noise data of the remaining periodic noise areas in the original grayscale image are subtracted in sequence to eliminate the noise, retain the target data points in the original grayscale image, obtain the grayscale image after eliminating the periodic noise, and transmit it to the clustering and splicing module.

[0043] The method for identifying the noise data period in the original grayscale image in the noise filtering module is as follows:

[0044] 1) First, an adaptive algorithm is used to obtain the noise period, that is, a 3×3 pixel adaptive template is used;

[0045] 2) Setting the value of the noise period to L according to the image characteristics;

[0046] 3) Using the initial 3×3 pixel elimination template, the noise data in the grayscale image is subtracted and eliminated with L as the noise period value, and then the pixel values ​​in the result image are summed;

[0047] 4) If the periodic noise filtering effect is obvious, the cumulative sum of the resulting pixel values ​​is less than the threshold (the best case is that only white real target points exist), then the optimal period length is determined as the period of the noise data; otherwise, go to step 5); the threshold value range is 10 to 20;

[0048] 5) If it is not within the threshold, the period length L is readjusted to be longer, and steps 2) to 4) are repeated again until the optimal period length is obtained, and an image of a periodic noise area is taken as the elimination template.

[0049] Clustering and splicing module: Receive the new grayscale image transmitted by the noise adjustment module and the grayscale image after eliminating periodic noise transmitted by the noise filtering module, cluster the processed grayscale image through the clustering algorithm, determine the grouping of the grayscale image according to the clustering result, and randomly splice the grouped grayscale images within or across groups, so as to generate a large number of data-enhanced grayscale images, and finally restore them to RD time-frequency data to obtain the final result.

[0050] like Figure 2 As shown, the workflow of the data processing module is:

[0051] 1. Prompt the user to input RD time-frequency data, read it, and determine the basic information of the data;

[0052] 2. Perform pulse compression, convert the input RD time-frequency data into a vector, and perform matched filtering, that is, autocorrelate with its own transmitted signal. Then convert the algorithm result of the matched filter into a matrix, perform a fast Fourier transform (FFT) operation on each row, and extract the Doppler frequency of the target. Finally, generate the corresponding two-dimensional image data based on the conversion result, with the horizontal axis representing time and the vertical axis representing frequency;

[0053] 3. Perform grayscale processing on the image data generated in step 2, that is, operate the pixel matrix of the image data so that the RGB value of each pixel in the image data is equal, and generate a corresponding grayscale image;

[0054] 4. Store the generated grayscale image information data to prepare data for other modules. At this point, it indicates that the RD time-frequency data has been input and exit the data processing module.

[0055] like Figure 3 As shown, the workflow of the noise adjustment module is:

[0056] 1. Read the grayscale image converted by the data processing module;

[0057] 2. According to the target data range information, the pixel value data matrix information corresponding to the part is cached;

[0058] 3. Adjust the interference intensity coefficient as needed. If the coefficient is greater than 1, the interference effect is amplified; if the coefficient is less than 1, the interference effect is weakened;

[0059] 4. According to the interference intensity coefficient determined in step 3, multiply the pixel value of the grayscale image by the interference intensity coefficient;

[0060] 5. Fill back the pixel value matrix information within the cached target range to obtain new grayscale image data after changing the interference intensity;

[0061] 6. The grayscale image information of the changed interference intensity is stored to prepare data for other modules. At this point, it indicates that the RD time-frequency data has completed the change of interference intensity and exits the noise adjustment module.

[0062] like Figure 4 As shown in the figure, the workflow of the noise filtering module is:

[0063] 1. Read the grayscale image converted by the data processing module;

[0064] 2. Adopt an adaptive algorithm to obtain the noise cycle. Specifically, adopt a 3×3 pixel adaptive template, set the initial value L of the noise cycle according to the characteristics of this type of image, and perform periodic noise subtraction to eliminate the periodic noise for the period length L. Then sum the pixel values ​​in the result image. If the periodic noise filtering effect is obvious, the cumulative sum of the result pixel values ​​should be small (the best case is that only white real target points exist). Then set a threshold for the cumulative sum. If it is not within this threshold, readjust the period length L. Repeat the above process again until the optimal period length is obtained.

[0065] 3. Perform noise filtering according to the period length determined in step 2. Specifically, take an image of a periodic noise area as an elimination template, subtract it from the remaining periodic noise areas of the original image in sequence according to the period length, and finally eliminate the periodic noise;

[0066] 4. Store the grayscale image information that eliminates periodic noise and prepare data for other modules. At this point, it indicates that the RD time-frequency data has completed the change of interference intensity and exits the noise filtering module.

[0067] like Figure 5 As shown in the figure, the workflow of the clustering splicing module is:

[0068] 1. Read the grayscale image data after noise adjustment or noise filtering;

[0069] 2. The user inputs the required number of groups of RD time-frequency data (grayscale image) to determine the number of categories that need to be clustered, recorded as n;

[0070] 3. Randomly select n image data objects as the initial clustering centers;

[0071] 4. Calculate the distance between the remaining image data objects and the n initial cluster centers, and assign the image data to the cluster that is closest to it;

[0072] 5. Adjust and recalculate the center of the new category;

[0073] 6. Calculate the clustering criterion function E to determine whether the convergence condition is met. If not, return to step 4. If it is, continue to step 7.

[0074] 7. According to the clustering results, determine the grouping of RD time-frequency data (grayscale image);

[0075] 8. Data splicing within or across groups, that is, splicing multiple grayscale images into one. This process can be repeated randomly to effectively increase the number of generated grayscale images;

[0076] 9. According to the splicing result of the grayscale image, restore it to RD time-frequency data. This indicates that the data enhancement operation of the RD time-frequency data has been completed, and exit the clustering splicing module.

[0077] The present invention can significantly enhance the RD time-frequency data, and can efficiently and quickly improve the quality and quantity of this type of data. Through adaptive filtering, noise adjustment, clustering and splicing, etc., interference data can be effectively eliminated or amplified, and a large amount of effective RD time-frequency data can be generated. It provides strong technical support for obtaining target data in RD time-frequency data based on artificial intelligence methods. Therefore, it is of great significance to apply this technology to engineering practice.

[0078] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

[0079] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A data enhancement system for RD time-frequency data, characterized in that: include: Data processing module, noise adjustment module, noise filtering module and clustering splicing module; Data processing module: receiving RD time-frequency data, converting the RD time-frequency data into image data, performing grayscale processing on the image data, generating a corresponding grayscale image as the original grayscale image, and transmitting it to the noise adjustment module and the noise filtering module; Noise adjustment module: receives the original grayscale image transmitted by the data processing module, uniformly adjusts the pixel values ​​of the noise data part of the original grayscale image, enlarges or reduces the pixel values ​​of the noise area in the original grayscale image according to the use needs to obtain a new grayscale image, and then transmits the new grayscale image to the clustering and splicing module; Noise filtering module: receives the original grayscale image transmitted by the data processing module, performs noise elimination, retains the target data points in the original grayscale image, obtains the grayscale image after eliminating periodic noise, and transmits it to the clustering splicing module; Clustering and splicing module: Receive the new grayscale image transmitted by the noise adjustment module and the grayscale image after eliminating periodic noise transmitted by the noise filtering module, generate a data-enhanced grayscale image, and finally restore the data-enhanced grayscale image to RD time-frequency data to obtain the final result; The method for identifying the noise data period in the original grayscale image in the noise filtering module is as follows: 1) First, an adaptive algorithm is used to obtain the noise period; 2) Setting the value of the noise period to L according to the image characteristics; 3) Using the initial elimination template, the noise data in the grayscale image is subtracted and eliminated with L as the noise period value, and then the pixel values ​​in the result image are summed to obtain the pixel cumulative sum; 4) Perform conditional judgment on the pixel accumulation and if the condition is met, the optimal cycle length is obtained and used as the cycle of the noise data; otherwise, go to step 5); 5) Adjust the period length L to be longer, repeat steps 2) to 4) again until the optimal period length is obtained, and take an image of the periodic noise area as the elimination template; The noise filtering module performs noise elimination processing, specifically: The adaptive algorithm first identifies the period of the noise data in the original grayscale image, then takes an image of a periodic noise area as an elimination template, and subtracts the noise data of the remaining periodic noise areas in the original grayscale image in turn according to the elimination template to eliminate the noise.

2. According to claim 1, a data enhancement system for RD time-frequency data is characterized in that: The method in which the noise adjustment module uniformly adjusts the pixel values ​​belonging to the noise data part in the original grayscale image is as follows: according to the range information of the target data in the RD time-frequency data, the pixel value data matrix information corresponding to the target data is extracted and saved, and then the pixel values ​​​​except this part of the target data are multiplied by the interference intensity coefficient to form a new grayscale image information matrix and the pixel value data corresponding to the previously extracted target data is filled into the newly generated grayscale image information matrix, thereby obtaining a new grayscale image and transmitting it to the clustering splicing module.

3. The data enhancement system for RD time-frequency data according to claim 2, characterized in that: The interference intensity coefficient in the noise adjustment module is greater than 1 when the noise is amplified, and is less than 1 otherwise.

4. The data enhancement system for RD time-frequency data according to claim 3, characterized in that: The method for condition determination in step 4) is specifically as follows: If the accumulated sum of the resulting pixel values ​​is less than the threshold, the condition is determined to be met, otherwise it is determined that the condition is not met.

5. The data enhancement system for RD time-frequency data according to claim 4, characterized in that: The threshold value in step 4) ranges from 10 to 20.

6. A data enhancement system for RD time-frequency data according to any one of claims 2 to 5, characterized in that: The clustering and stitching module generates data-enhanced grayscale images, specifically: The processed grayscale images are clustered and divided by a clustering algorithm, and the grayscale images are grouped according to the clustering results. The grouped grayscale images are randomly spliced ​​within or across groups to generate data-enhanced grayscale images.

Images