A GPU-Accelerated Method for Simulating Wide-Squint Spotlight Replay SAR Targets

Abstract

The present invention discloses a method for simulating large squint spotlight SAR targets based on GPU acceleration, which includes the following steps: 1) A data redundancy removal algorithm based on range migration compensation; 2) Using a data adaptive block strategy to divide a larger data block and transfer the divided data into video memory; 3) Using the powerful parallel computing ability of the GPU to calculate the echo signal; 4) Asynchronous execution, with the help of streams for concurrent management, and transferring the calculated echo signal to memory; 5) Data stitching and restoration, restoring the data in memory to the position before range migration compensation to obtain the final echo data. The present invention not only significantly improves the target simulation efficiency and can effectively accelerate the simulation of large squint spotlight SAR targets, but also combines range migration compensation. After removing redundancy, it saves the space of memory and video memory, not only maintaining the high accuracy of the time-domain target simulation algorithm, but also significantly improving the operation efficiency of SAR target simulation in the case of large data volume.

Description

Technical Field

[0001] The invention relates to the technical field of signal processing, and in particular to a high-squint spotlight playback SAR target simulation method based on GPU acceleration. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing technology that has the ability to observe the earth around the clock and in all weather conditions. It originated in the 1950s, when the concept of "synthetic aperture" was first proposed by American scholar Carl Wiley. Subsequently, C. Sherwin and others developed the first X-band SAR system and obtained the first non-focused strip mode SAR image. Since then, SAR technology has begun to flourish.

[0003] Before each satellite is launched, it is an essential step to conduct a target simulation experiment. As the preliminary work for the performance verification of the SAR system, the successful launch of each satellite is inseparable from the contribution of target simulation. Target simulation can better determine the system parameters, and by simulating the echo data under various parameters, imaging and analysis can be performed to find the optimal parameters, and the system can be adjusted to the optimal state before the satellite is launched, effectively ensuring the stability of the SAR system. At present, the existing target simulation methods are generally divided into two categories: time domain target simulation and frequency domain target simulation. The time domain algorithm can obtain high-precision echo data, but the computational complexity is high and the efficiency is low. In the target simulation task of large scenes, the time cost is very high. Although the frequency domain algorithm can complete the target simulation with higher efficiency, its accuracy is lower than that of the time domain algorithm, and it is only suitable for target simulation under positive side view or small squint. When the squint angle increases, the coupling and range migration of the azimuth and range directions also increase, resulting in the inability of the traditional frequency domain algorithm to be used for target simulation under large squint conditions.

[0004] Currently, there is a synthetic aperture radar image simulation method for a ship scene disclosed in Chinese patent application publication number CN108594230A, and a SAR imaging simulation test system and method for a moving target under a sea background disclosed in Chinese patent application publication number CN108051789A. However, the above patented technologies are only applicable to ship scenes and are not applicable to target simulation under large squint conditions, so they cannot solve the above problems.

[0005] In summary, how to improve the computational efficiency of SAR target simulation in the high squint spotlight mode and ensure the superior properties of SAR echo signals is an urgent problem to be solved. Summary of the invention

[0006] The object of the present invention is to provide a large squint spotlight replay SAR target simulation method based on GPU acceleration to solve the above problems in view of the above deficiencies and defects of the prior art.

[0007] The technical problems solved by the present invention can be realized by the following technical solutions:

[0008] A large squint spotlight replay SAR target simulation method based on GPU acceleration includes the following steps:

[0009] 1) A data redundancy removal algorithm based on range migration compensation;

[0010] 2) Divide a relatively large data block by using a data adaptive block strategy and transfer the divided data into video memory;

[0011] 3) Use the powerful parallel computing ability of the GPU to calculate the echo signal;

[0012] 4) Asynchronous execution, perform concurrent management by means of streams, and transfer the calculated echo signal to memory;

[0013] 5) Data splicing and restoration, restore the data in memory to the position before range migration compensation to obtain the final echo data.

[0014] In a preferred embodiment of the present invention, the data redundancy removal algorithm in step 1) is as follows: First, determine the SAR geometric model and simulation parameters, and calculate the positions of valid data at each moment through range migration compensation to remove redundancy.

[0015] In a preferred embodiment of the present invention, in step 1), calculate the valid region of the echo in the CPU, and then calculate the offset of the range direction data at each azimuth moment according to range migration. When transferring it to the video memory for operation, it can be equivalent to the target simulation calculation without the influence of range migration.

[0016] In a preferred embodiment of the present invention, in step 1), let t r be the range time axis at time t a ,

[0017]

[0018] where F r is the range sampling rate, N r is the number of range sampling points, i represents the serial number of the range sampling point. According to formula (1), it can be known that the absolute value of t r should be less than half of the pulse width. Therefore, the two boundary values i near , i far

[0019]

[0020] Among them, R near (t a ) indicates t a The closest distance between the radar and the scene at any moment, R far (t a ) indicates t a The maximum distance from the radar to the scene, R c represents the slant distance from the center of the scene, c represents the speed of light, T r represents the pulse width,

[0021] Let N a is the number of sampling points in the azimuth direction. According to the time of each azimuth moment, the distance Range that needs to be compensated can be calculated, that is, t a The distance of the moment data offset, assuming θ is the oblique angle, t0 is the center moment,

[0022]

[0023] From formula (3), we can calculate t a The size of the pixel offset Pix is:

[0024] Pix= Range / ΔR (4)

[0025] Among them, ΔR represents the pixel interval, and the near boundary i of the data is near and far boundary i far After shifting Pix respectively, the shifted boundary i can be obtained near '、i far '.

[0026]

[0027] After eliminating it through distance migration compensation, the computing efficiency will be significantly improved and the occupancy of video memory will be reduced.

[0028] In a preferred embodiment of the present invention, the data adaptive partitioning in step 2) is as follows: obtaining the available video memory capacity in the system, calculating the size of the data block that can be stored at one time, that is, obtaining the number of blocks required.

[0029] In a preferred embodiment of the present invention, in step 2), the number of available GPUs in the system and the capacity of all available video memory are obtained, and then, based on the total amount of video memory, the size of the data block that can be stored at one time is calculated, that is, the number of blocks that need to be divided is obtained, and the memory and video memory are allocated, and the echo simulation calculations are performed in sequence. Each time a part of the calculation is completed, the block of data is transferred from the video memory to the internal memory and the video memory is released.

[0030] In a preferred embodiment of the present invention, in step 2), let Gpumem be the available video memory capacity of the system. According to the size of the video memory, the size of the data block that can be processed at one time can be calculated. Let col be the maximum number of columns that can be processed at one time.

[0031] col = Gpumem / (N r · size) (6)

[0032] where size is the size of a single data, and N r ·size is the data size of each column. According to formula (6) and the number of azimuth samplings N a , the number of blocks N that need to be segmented can be calculated.

[0033]

[0034] represents rounding up. After segmentation, video memory allocation is performed according to the segmentation result, and then the time-domain target simulation algorithm in the GPU is calculated sequentially.

[0035] In a preferred embodiment of the present invention, in step 3), using the powerful parallel ability of the GPU, the echo data at each azimuth time is calculated.

[0036] In a preferred embodiment of the present invention, in step 3), when calculating the echo data in the GPU, the expression of the echo signal is:

[0037]

[0038]

[0039] In the above formulas (8) and (9), v represents the flight speed of the carrier aircraft, λ represents the wavelength, x represents the abscissa of the target point, y represents the ordinate of the target point, and K r coordinate represents the range chirp rate.

[0040] In a preferred embodiment of the present invention, the asynchronous execution in step 4) is as follows: with the help of streams for concurrent management, after the range-direction data calculation at each azimuth time is completed, the data of this column is transferred to the memory, without affecting other threads to continue the calculation of this data block.

[0041] In a preferred embodiment of the present invention, the data splicing and restoration in step 5) are as follows: in the memory, every time a part of the data is transferred from the video memory, it is stored according to the corresponding time on the azimuth time axis. After storage, the data is padded with zeros to generate a data block before range migration compensation. At this time, this data block is the echo data after range migration compensation.

[0042] Due to the above technical solution, the present invention is accelerated based on the GPU, combines the range migration compensation technology to remove redundancy. The method based on GPU acceleration not only significantly improves the target simulation efficiency, but also can effectively accelerate the simulation of large squint spotlight SAR targets. And by combining range migration compensation and removing redundancy, the memory and video memory space are saved. It not only maintains the high accuracy of the time-domain target simulation algorithm, but also significantly improves the operation efficiency of SAR target simulation in the case of large data volume. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0044] Figure 1 It is the overall flowchart of the present invention.

[0045] Figure 2 It is the amplitude diagram of the echo data before range migration compensation.

[0046] Figure 3 It is the amplitude diagram of the echo data after range migration compensation.

[0047] Figure 4 It is the imaging comparison diagram of the echo data of 5 point targets before using the target simulation acceleration algorithm by using the backprojection algorithm of the present invention.

[0048] Figure 5 It is the imaging comparison diagram of the echo data of 5 point targets after using the target simulation acceleration algorithm by using the backprojection algorithm of the present invention.

[0049] Figure 6 It is the point target analysis of target 3 before acceleration.

[0050] Figure 7 It is the point target analysis of target 3 after acceleration. DETAILED DESCRIPTION OF THE INVENTION

[0051] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below.

[0052] See Figures 1 to 7 As shown, a method for simulating large squint spotlight SAR targets accelerated based on GPU (graphics processing unit) includes the following steps:

[0053] 1) Data redundancy removal algorithm based on range migration compensation. Preferably, first determine the SAR geometric model and simulation parameters, and through range migration compensation, calculate the positions of effective data at each moment and remove the redundancy.

[0054] To improve the efficiency of the algorithm, calculate the effective area of the echo in the CPU, and then calculate the offset of the range data at each azimuth moment according to the range migration. When passing it into the video memory for operation, it can be equivalent to the target simulation calculation without the influence of range migration.

[0055] Let t r be the range time axis at time t a moment.

[0056]

[0057] Among them, F r is the range sampling rate, N r is the number of range sampling points, and i represents the serial number of the range sampling point. According to formula (1), it can be known that the absolute value of t r should be less than half of the pulse width. Therefore, the two boundary values i near and i far

[0058]

[0059] Among them, R near (t a ) represents the closest distance from the radar to the scene at time t a moment, R far (t a ) represents the farthest distance from the radar to the scene at time t a moment, R c represents the slant range of the scene center, c represents the speed of light, and T r represents the pulse width.

[0060] Let N a be the number of azimuth sampling points. According to the time at each azimuth moment, the range Range that needs to be compensated can be calculated, that is, the distance by which the data at time t a moment is offset.

[0061]

[0062] From formula (3), the size Pix of the pixel offset at time t a can be calculated as:

[0063] Pix = Range / ΔR (4)

[0064] Among them, ΔR represents the pixel interval, and the near boundary i nearand the far boundary i far After shifting Pix respectively, the shifted boundary i can be obtained near ’ and i far ’

[0065]

[0066] After eliminating it through range migration compensation, the calculation efficiency will be significantly improved, and the memory usage will also be reduced

[0067] Use the data adaptive block strategy to divide larger data blocks and transfer the divided data into the video memory. Preferably, the data adaptive block is: obtain the available video memory capacity in the system, calculate the size of the data block that can be stored at one time, that is, obtain the number of blocks to be divided

[0068] Preferably, obtain the available GPU number and all available video memory capacities in the system, and then, according to the total amount of video memory, calculate the size of the data block that can be stored at one time, that is, obtain the number of blocks to be divided. Allocate memory and video memory, and perform the calculation of echo simulation in turn. After each part of the calculation is completed, transfer the data of this block from the video memory to the memory and release the video memory

[0069] Let Gpumem be the available video memory capacity of the system. According to the size of the video memory, the size of the data block that can be processed at one time can be calculated. Let col be the maximum number of columns that can be processed at one time

[0070] col = Gpumem / (N r · size) (6)

[0071] where size is the size of a single data, and N r · size is the data size of each column. According to formula (6) and the azimuth sampling number N a , the number of blocks to be divided N can be calculated

[0072]

[0073] represents rounding up. After the block division is completed, allocate video memory according to the block division result, and then perform the calculation of the time-domain target simulation algorithm in the GPU in turn

[0074] 3) Use the powerful parallel ability of the GPU to calculate the echo signal. Preferably, calculate the echo data at each azimuth moment of the SAR, calculate the echo data in the GPU, and the expression of the echo signal is

[0075]

[0076]

[0077] In the above formulas (8) and (9), v represents the flight speed of the carrier aircraft, λ represents the wavelength, x represents the abscissa of the target point, y represents the ordinate of the target point, and K r The coordinate represents the range chirp rate.

[0078] 4) Asynchronous execution, with concurrent management through streams, and the calculated echo signals are transmitted to the memory; preferably, the asynchronous execution is: with concurrent management through streams, after the data calculation in the range direction at each azimuth time is completed, this column of data is transmitted to the memory, without affecting other threads to continue the calculation of this data block. The calculations and data copies within the same stream are carried out sequentially, and the calculations within one stream and the data transmission of another stream can be carried out simultaneously, enabling the execution units and memory control units in the GPU to work simultaneously, improving resource utilization and further enhancing the efficiency of the algorithm. In the actual algorithm implementation, after each thread finishes calculating the data at one azimuth time, the data is transmitted to the memory, and after the transmission is completed, the calculation of the data at the next azimuth time continues. The transmission and calculation processes of each block of echo data do not need to wait and can continue with the transmission and calculation of the next data block.

[0079] 5) Data stitching and restoration, restoring the data in the memory to the position before range migration compensation to obtain the final echo data. Preferably, the data stitching and restoration is: in the memory, every time a part of the data is fetched from the video memory, it is stored according to the corresponding time on the azimuth time axis. After storage, the data is filled with zeros to generate a data block before range migration compensation. At this time, this data block is the echo data after range migration compensation. Further, after each block of data is calculated, the data needs to be fetched from the video memory into the memory. Since the influence caused by range migration is compensated during the optimization of the calculation amount, it is necessary to restore it in the CPU, compensate the data back to the position affected by the squint angle again, and restore all the divided data into a complete piece of data; in the memory, every time a part of the data is fetched from the video memory, it is stored according to the corresponding time on the azimuth time axis. After storage, the data is filled with zeros to generate a data block of a certain size. At this time, this data block is the echo data after range migration compensation; according to the calculated compensation value, each azimuth time in this matrix is restored to the position before range migration compensation in the range direction, that is, the inverse process of range migration compensation in step 1). In this way, the target simulation data under large squint angles is restored, that is, the final echo data.

[0080] The technical solution of the present invention will be further described in detail below in conjunction with specific embodiments

[0081] Embodiment 1

[0082] Figure 2 、 3 are the amplitude diagrams of the echo data before and after range migration compensation. It can be clearly observed that after redundancy removal, the number of calculation units is significantly reduced.Figure 4 , 5 To use the back-projection algorithm, the echo data of 5 point targets before and after using the target simulation acceleration algorithm are imaged and compared. Figure 4 and Figure 5 There is almost no difference, indicating that the precision of the present invention is relatively high. Figure 6 Analysis of the point target of target 3 before acceleration. Figure 7 Analysis of the point target of target 3 after acceleration. It can be found that the effects of the point targets are all good and there is no obvious change.

[0083] The various indexes of the point target are shown in Table 1.

[0084] Table 1

[0085]

[0086] As can be seen from Table 1, the present invention is equivalent in range and azimuth to the traditional time-domain echo simulation algorithm, and there is no obvious difference, indicating that there is no loss of precision in the present invention.

[0087] The comparison of the calculation efficiency is shown in Table 2.

[0088] Table 2

[0089]

[0090] As can be seen from Table 2, the operation efficiency of the present invention is significantly improved, and the efficiency has been significantly improved compared with the algorithms without acceleration and after traditional acceleration.

[0091] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for simulating large squint spotlight replay SAR targets based on GPU acceleration, characterized in that, It includes the following steps: 1) A data redundancy removal algorithm based on range migration compensation; 2) Using a data adaptive block strategy to divide a larger data block and transfer the divided data into video memory; 3) Using the powerful parallel ability of the GPU to calculate echo signals; 4) Asynchronous execution, with the help of streams for concurrent management, and transferring the calculated echo signals to memory; 5) Data splicing and restoration, restoring the data in memory to the position before range migration compensation to obtain the final echo data; The data redundancy removal algorithm in step 1) is as follows: First, determine the SAR geometric model and simulation parameters, calculate the position of valid data at each moment through range migration compensation, and remove the redundancy; In step 1), calculate the effective area of the echo in the CPU, and then calculate the offset of the range direction data at each azimuth moment according to range migration. When transferring it to the video memory for operation, it is equivalent to the target simulation calculation without the influence of range migration; In step 1), let t r be the distance time axis at time t a , Among them, F r is the range sampling rate, N r is the number of range sampling points, i represents the serial number of the range sampling point. According to formula (1), the absolute value of t r should be less than half of the pulse width. Therefore, for each azimuth time t, the two boundary values of the effective range, i near and i far Among them, R near (t a ) indicates t a The closest distance between the radar and the scene at any moment, R far (t a ) indicates t a The maximum distance from the radar to the scene, R c represents the slant distance from the center of the scene, c represents the speed of light, T r represents the pulse width, Let N a be the number of azimuth sampling points. According to the time at each azimuth moment, calculate the range Range to be compensated, that is, the distance by which the data at time t a is offset. Let θ be the squint angle and t0 be the central moment, Calculate t according to formula (3). a The size Pix of pixel offset is as follows: Pix = Range / ΔR (4) where ΔR represents the pixel interval, and after shifting the near boundary i near and the far boundary i far of the data by Pix respectively, the shifted boundaries i near ’ and i far ’ are obtained. After eliminating it through range migration compensation, the calculation efficiency will be significantly improved, and the occupation of video memory will also be reduced.

2. The method for simulating large squint spotlight replay SAR targets based on GPU acceleration according to claim 1, characterized in that, The data adaptive block in step 2) is as follows: Obtain the available video memory capacity in the system, calculate the size of the data block that can be stored at one time, that is, obtain the number of blocks that need to be divided.

3. The method for simulating large squint spotlight replay SAR targets based on GPU acceleration according to claim 2, characterized in that, In step 2), obtain the number of available GPUs and all available video memory capacities in the system. Then, according to the total amount of video memory, calculate the size of the data block that can be stored at one time, that is, obtain the number of blocks that need to be divided, allocate memory and video memory, and perform the calculation of echo simulation in sequence. After each part of the calculation is completed, transfer the block data from the video memory to memory and release the video memory.

4. The method for simulating large squint spotlight replay SAR targets based on GPU acceleration according to claim 3, characterized in that, In step 2), let Gpumem be the available video memory capacity of the system. Calculate the size of the data block that can be processed at one time according to the video memory size. Let col be the maximum number of columns that can be processed at one time, and col = Gpumem / (N r ·size) (6) where size is the size of a single data, N r · size is the data size of each column. According to formula (6) and the number of azimuth samples N a , the number of blocks N to be divided can be calculated Represents rounding up. After the block division is completed, allocate video memory according to the block division result, and then perform the calculation of the time domain target simulation algorithm in the GPU in sequence.

5. The method for simulating large squint spotlight replay SAR targets based on GPU acceleration according to claim 1, characterized in that, In step 3), use the powerful parallel ability of the GPU to calculate the echo data at each azimuth moment of the SAR.

6. The method for simulating large squint spotlight replay SAR targets based on GPU acceleration according to claim 5, characterized in that, In step 3), calculate the echo data in the GPU, and the expression of the echo signal is: In the above formulas (8) and (9), v represents the flight speed of the carrier aircraft, λ represents the wavelength, x represents the abscissa of the target point, y represents the ordinate of the target point, and K r The coordinate represents the range modulation frequency.

7. A large squint spotlight replay SAR target simulation method based on GPU acceleration as claimed in claim 1, wherein, The asynchronous execution in step 4) is as follows: With the help of streams for concurrent management, after the calculation of the range direction data at each azimuth moment is completed, transfer the column data to memory without affecting other threads to continue the calculation of the data block.

8. A large squint spotlight replay SAR target simulation method based on GPU acceleration as claimed in claim 1, wherein, The data splicing and restoration in step 5) is as follows: In memory, every time a part of the data is transferred from the video memory, store it at the moment corresponding to the azimuth time axis. After the storage is completed, fill the data with zeros to generate a data block before range migration compensation. At this time, this data block is the echo data after range migration compensation.

Images