An adaptive X-ray dynamic image estimation method and system

Abstract

The present invention relates to a dynamic estimation method for observing the motion changes of the earth's magnetosphere in space science missions, and in particular to an adaptive X-ray dynamic image estimation method and system. The method comprises: taking the average of the X-ray images with integration times T1 and T3 as the starting X-ray image, performing motion estimation through the optical flow estimation model PWC-Net, and then simultaneously using an adaptive kernel density estimation method to capture the spatial evolution of pixel values ​​in actual observations, and using a learnable contribution graph to describe the correlation between different frames of the X-ray image, and ensuring the temporal continuity and consistency of the estimated X-ray image data through a time discriminator, thereby realizing the estimation of X-ray dynamic images at different times within the intermediate time T2. The present invention can more accurately reflect the real-time state changes of the magnetosphere, improve the accuracy and quality of X-ray dynamic image estimation, and can adapt to the requirements of different observation tasks.

Description

Technical Field

[0001] The present invention relates to a dynamic estimation method for observing the movement change of the earth's magnetosphere in a space science mission, and in particular to an adaptive X-ray dynamic image estimation method and system. Background Art

[0002] Observations of the Earth's magnetosphere play a vital role in space weather monitoring. The dynamics of the Earth's magnetic field directly affect the interaction between the solar wind and the Earth's atmosphere, which in turn affects critical infrastructure such as satellite communications, navigation systems, and power grids. The upcoming Solarwind Magnetospheric Ionosphere Explorer (SMILE) and Lunar Environment X-ray Imager (LEXI) missions aim to comprehensively observe the Earth's magnetosphere using the wide-angle Soft X-ray Imager (SXI) to gain a deeper understanding of the dynamic response of the Earth's magnetosphere to the impact of the solar wind.

[0003] When heavy ions in the solar wind collide with neutral atoms or molecules around the Earth, soft X-ray radiation is generated by solar wind charge exchange (SWCX). This constitutes the basic mechanism for SXI to observe the Earth's magnetosphere. However, in practice, the imaging process of SXI is affected by various factors, such as the soft X-ray background in outer space, instrument noise, and fluctuation noise, which leads to blurred magnetospheric structural features in 2D soft X-ray images and poses challenges for subsequent studies. The magnetospheric structural features can be highlighted by increasing the integration time of SXI. The integration time of SXI carried by SMILE and LEXI is set to 5 minutes and 10 minutes, respectively. However, this method of image quality at the expense of temporal resolution ignores the continuous changes in the magnetosphere itself, and the changes in the SXI observation position with the spacecraft (or the moon), such as Figure 1 As shown. Therefore, the SXI observations obtained by this method reflect the cumulative state of magnetospheric changes over the entire time period. Although this may have no effect on space weather events with larger time scales, it may ignore space events with small time scales and introduce discontinuities in the study of space event changes. Therefore, dynamically reconstructing magnetospheric changes based on accumulated SXI observations during the integration time is a necessary and scientifically important task.

[0004] SXI observation time resolution problem: SXI's integration time is set to 5 minutes and 10 minutes. This image quality method at the expense of time resolution ignores the continuous changes in the magnetosphere itself. This may lead to the introduction of discontinuities when studying the changes in space physics events and ignore burst events with small time scales. Therefore, it is necessary to find a way to obtain more accurate images without losing time resolution.

[0005] Utilization of the OMNI database: The OMNI database provides various solar wind parameters, magnetic field data, and particle measurement data, but how to combine these data with SXI observations to provide more accurate information on dynamic changes in the magnetosphere remains to be solved.

[0006] Motion estimation method: In order to estimate the dynamic changes of the magnetospheric X-ray images during the SXI integration time, an effective motion estimation method needs to be developed that can make comprehensive use of the solar wind parameters in the OMNI database and the SXI observation data. This requires combining deep learning techniques and image processing methods to obtain accurate estimates of magnetospheric changes.

[0007] Model validation issues: The proposed method needs to be verified in the experimental validation phase. Before the launch of the SMILE and LEXI missions, other models need to be used to replace the actual observations, which may introduce certain uncertainties. Therefore, the model selection and validation methods need to be carefully considered to ensure the reliability and effectiveness of the proposed method.

[0008] In summary, solving these problems requires interdisciplinary collaboration, including space physics, computer vision, image processing, and deep learning. By making full use of existing data and technical means, and continuously improving and innovating, we can better understand the dynamic response of the Earth's magnetosphere to the solar wind and provide more reliable data and predictions for space weather monitoring. Summary of the invention

[0009] The purpose of the present invention is to overcome the defects of the prior art and propose an adaptive X-ray dynamic image estimation method and system.

[0010] In order to achieve the above object, the present invention proposes an adaptive X-ray dynamic image estimation method for space science missions, comprising:

[0011] The X-ray images with integration times T1 and T3 are averaged as the starting X-ray image, and motion estimation is performed through the optical flow estimation model PWC-Net. At the same time, an adaptive kernel density estimation method is used to capture the spatial evolution of pixel values ​​in actual observations, and a learnable contribution graph is used to describe the correlation between different frames of the X-ray image. The temporal continuity and consistency of the estimated X-ray image data are ensured by the temporal discriminator, thereby realizing the estimation of X-ray dynamic images at different times within the intermediate time T2.

[0012] Preferably, the step of averaging the X-ray images with integration times T1 and T3 as the starting X-ray image comprises:

[0013] The X-ray images with integration times T1 and T3 are obtained by simulating the MHD model, and the X-ray images within the T2 time are adaptively and dynamically estimated with the set sampling interval, combined with the fluid dynamics motion information provided by the Jorgensen model; or

[0014] The actual X-ray images collected at integration times T1 and T3 are combined with the prior knowledge of magnetohydrodynamic motion provided by the MHD model to perform adaptive dynamic estimation of the X-ray images within the T2 time at the set sampling interval.

[0015] Preferably, the input of the optical flow estimation model PWC-Net is two X-ray images G0 and G1, and the output is the forward motion estimation field J from G0 to G1 0→1 , and the estimated reverse motion field J from G1 to G0 1→0 .

[0016] Preferably, the adaptive kernel density estimation method and the learnable contribution graph are implemented using U-Net, which includes an encoder, a decoder and a skip connection, wherein each processing unit contains 3×3 convolution and ReLU activation, the encoder uses average pooling to extract features, and the decoder upsampling layer uses bilinear interpolation to upsample the feature map in the encoder to a higher resolution, and a softmax layer is used as a constraint to ensure that the kernel weights are non-negative.

[0017] Preferably, the U-Net is trained using an Adamax optimizer, and the loss function is:

[0018] L=λ o L o +λ p L p +λ f L f +λ d L d

[0019] Among them, λ o ,λ p ,λ f ,λ d is the linear weighting weight, L o is the observation loss, L p is the perceptual loss, L f is the loss of simulating the effect of intermediate frame motion estimation, L d is the discriminator loss, which satisfies the following formulas:

[0020]

[0021] in, is a set of intermediate frames between the input images I0 and I1, i represents the frame number, N represents the total number of frames, and t i ∈[0,1], is the X-ray image observed within T time;

[0022]

[0023] Among them, φ represents the detail index after convolution;

[0024]

[0025] Among them, J is the optical flow field information estimated by PWC-Net, I i It indicates the estimated i-th frame X-ray image, I i+1 It indicates that the estimated i+1th frame of X-ray image is obtained, G i is the i-th frame X-ray image obtained by Jorgensen model simulation at the corresponding time, G i+1 is the i+1th frame of X-ray image obtained by Jorgensen model simulation at the corresponding time,

[0026]

[0027] Preferably, the temporal discriminator adopts a slow fusion model, the input is two consecutive frames obtained by adaptive estimation, and the output is the discrimination result; the processing process is: the two consecutive frames obtained by adaptive estimation are respectively sent to two independent processing streams, respectively simulating the context stream of the low-resolution image and the foveal stream for processing the high-resolution center crop; the two independent processing streams include alternating convolution, normalization and pooling layers, and then converge to the fully connected layer.

[0028] Preferably, the objective function of the slow fusion model training uses cross entropy and satisfies the following formula:

[0029]

[0030] in, is N frame samples, is N frames of negative samples, i represents the frame number, N represents the total number of frames, D t () represents the discriminator, and E() represents the expected value.

[0031] On the other hand, the present invention provides an adaptive X-ray dynamic image estimation system, comprising:

[0032] An acquisition input module is used to average the X-ray images with integration times T1 and T3 as the starting X-ray image;

[0033] Motion estimation module, used for motion estimation through the optical flow estimation model PWC-Net;

[0034] Adaptive kernel density estimation and contribution map estimation modules to capture the spatial evolution of pixel values ​​in actual observations and describe the correlation between different frames of X-ray images using a learnable contribution map; and

[0035] The time discriminator is used to ensure the temporal continuity and consistency of the estimated X-ray image data, thereby realizing the estimation of the X-ray dynamic images at different moments within the intermediate time T2.

[0036] Compared with the prior art, the advantages of the present invention are:

[0037] 1. Dynamic X-ray image adaptive estimation method: This application proposes a new method that can estimate the motion changes of the magnetosphere during the integration time. Compared with the traditional static estimation method, this dynamic method can more accurately reflect the real-time state changes of the magnetosphere and provide more favorable input for the subsequent dynamic three-dimensional reconstruction of the magnetosphere.

[0038] 2. Integrate actual observations with theoretical physics models: This application combines the solar wind parameter data observed by OMNI and the X-ray image data observed by SXI, and uses the X-ray images generated by the MHD model to provide prior motion knowledge for the adaptive estimation method. This integration can improve the accuracy and quality of X-ray image estimation.

[0039] 3. Adaptive kernel density estimation: In X-ray image processing, an adaptive kernel density estimation method is introduced to make full use of the information of SXI observation results and motion estimation, so as to better capture the spatial variation characteristics of the magnetosphere.

[0040] 4. Neural network-based model design: An innovative neural network model is proposed, which includes core parts such as motion estimation, adaptive kernel density estimation, learnable contribution map and time discriminator. This model can work together to provide accurate and dynamic estimates of magnetospheric changes.

[0041] 5. Flexibility for different time intervals: This method can be easily extended while considering period consistency, meeting the flexibility of time interval requirements for different application scenarios. This flexibility enables the method to adapt to different observation task requirements and provides greater convenience for practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the SXI observation position as the spacecraft (or the Moon) changes and the state of the magnetosphere boundary changes;

[0043] Figure 2 are the solar wind parameters observed from 00:00:00 UTC on May 14, 2009 to 22:4000 UTC on May 15, 2009;

[0044] Figure 3(a) shows the lunar trajectory in the time interval from 00:00:00 UTC on May 14, 2009 to 22:40:00 UTC on May 15, 2009;

[0045] Figure 3(b) is an ideal wide-field (40°×40°) soft x-ray image obtained based on the MHD simulation results;

[0046] Figure 4(a) and Figure 4(b) are the instantaneous frequency histograms when the maximum PSNR and SSIM values ​​appear within 5 minutes;

[0047] Figure 4(c) and Figure 4(d) are the corresponding PSNR and SSIM values ​​when the integration time is 10 minutes.

[0048] 5(a) and 5(b) are the PSNR and SSIM calculated between the corresponding linearly interpolated two-dimensional X-ray image and the original MHD two-dimensional X-ray image;

[0049] Figure 6 It is a dynamic X-ray image estimation framework;

[0050] Figure 7 It is a schematic diagram of the estimation of the bidirectional motion flow field;

[0051] FIG8(a) is a kernel-based X-ray estimation implementation process, FIG8(b) is a motion-based X-ray estimation implementation process, and FIG8(c) is an adaptive X-ray estimation implementation process;

[0052] Fig. 9 is a cycle consistency flow chart;

[0053] Fig.10 It is a slow fusion model architecture;

[0054] FIG11(a) and FIG11(b) are PSNR and SSIM calculated by the linear interpolation method and the adaptive dynamic x-ray image estimation method for the MHD model x-ray image, respectively; FIG11(c) is a PSNR histogram, and FIG11(d) is a histogram distribution of SSIM;

[0055] FIG12(a) and FIG12(b) are the statistical distributions of PSNR and SSIM of the linear interpolation method at time intervals of 5 minutes, 10 minutes, 15 minutes and 20 minutes; FIG12(c) and FIG12(d) are the statistical distributions of PSNR and SSIM of the adaptive dynamic X-ray estimation method at time intervals of 5 minutes, 10 minutes, 15 minutes and 20 minutes;

[0056] Figure 13(a) is the original two-dimensional X-ray image obtained from the MHD simulation; Figure 13(b) is the estimation result through linear interpolation; Figure 13(c) is the estimation result of the adaptive dynamic X-ray image estimation method. DETAILED DESCRIPTION

[0057] This application proposes an adaptive dynamic X-ray image estimation method based on two different types of observation data, making full use of OMNI solar wind parameters and SXI observation data. Specifically, we use the OMNI database to drive the magnetospheric fluid dynamics (MHD) simulation model to provide prior knowledge of the dynamic changes of the magnetosphere. Through the motion estimation neural network, we learn this prior knowledge and use the adaptive kernel density estimation method and the learnable contribution map to fuse the motion prior knowledge with the SXI observation results to estimate the dynamic changes of the magnetospheric X-ray image within the SXI integration time. The estimated continuous magnetospheric X-ray image changes are input into the time discriminator to ensure the temporal consistency of the motion. In the experimental verification stage, considering that the SMILE and LEXI missions have not yet been launched, we used two different magnetospheric fluid simulation models. The PPMLR-MHD model is considered to be the actual observation result of SXI and represents the most sophisticated magnetospheric fluid simulation model currently. On the other hand, the Jorgensen model is a simplified version of the MHD model, which is specifically designed to provide prior motion information. This equivalent alternative method has been widely used in space physics research.

[0058] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0059] Example 1

[0060] Embodiment 1 of the present invention proposes an adaptive X-ray dynamic image estimation method. The specific description is as follows:

[0061] 1. Analysis

[0062] 1. X-ray emissivity model

[0063] During the simulation, the analytical domain is divided into a grid of size M × N × L, where M, L, and N represent the number of grids in the X, Y, and Z directions, respectively. The volume emissivity is then estimated for each point in the grid.

[0064] P = a cx n H n sw <g>(eV·cm -3 ·s -1 )

[0065] where a cx 、n H 、n sw and <g>They represent the effective factor of interaction, the number density of hydrogen ions in the outer atmosphere, and the density and relative speed of solar wind ions. cx It depends on the charge transfer cross section (SWCX cross section), as well as the composition and density of high-valent ions in the solar wind. cx =1.0×10 -15 eV 2 , which is a reasonable estimate for the average solar wind (400 km / s). H It is provided by MHD simulation data and a reasonable approximation is selected. The general calculation formula is n H =n0(10(R E ) / r) 3 cm -3 . The average collision velocity is estimated from the plasma volume velocity and thermal velocity. The data volume of the three-dimensional magnetospheric x-ray emissivity in this paper is 128×128×128 voxels, corresponding to the X-axis range of -53.28~30Re in the GSM coordinate system, and the Y-axis and Z-axis range of -61.11~61.11Re. This is also the range of the region of interest (ROI) defined in this application.

[0066] Although there are obvious similarities between real observational data and physical simulation data, there are also differences. In order to simulate the difference between the two data sets, the actual observational data is replaced by more complex MHD model data, and the physical simulation is replaced by the Jorgensen radiation model. The Jorgensen model is based on the boundary between the magnetopause and the bow shock, and uses the boundary radiation model to simulate the MHD calculation results of simplified emissivity data. The boundary X-ray radiation formula can be expressed as:

[0067]

[0068] Where ref = 10R E is a constant when the solar wind conditions meet the following conditions: the number density of solar wind plasma N0 = 22.5 cm -3 , the x-direction component of the solar wind speed V x_0 =400km / s, components V in y and z directions y_0 =V z_0 = 0, the Z component of the interplanetary magnetic field B z_0 =5nT, and B in the x and y directions x_0 =B y_0 = 0. The most suitable parameter values ​​are shown below.

[0069] A1=3.2285×10 -5 eV -3 s -1 ,B=-1.7985×10 -5 eV -3 s -1

[0070] A2=1.3588×10 -5 eV -3 s -1 ,α=2.4908,β=-1.6458

[0071] In fact, the values ​​of parameters A1, B and A2 are related to N0×V x_0 The ratio is proportional to the value of , and the ratio is as follows.

[0072] times = N × V x ÷(22.5×400)

[0073] Therefore, for any solar wind plasma number density N and the x-direction component of the solar wind speed V x , the most suitable parameter values ​​are shown below.

[0074] A1=3.2285×10 -5 eV -3 s -1 ×times

[0075] B=-1.7985×10 -5 eV -3 s -1 ×times

[0076] A2=1.3588×10 -5 eV -3 s -1 ×times

[0077] α=2.4908,β=-1.6458

[0078] In this application, the simulation results of the PPMLR-MHD model will represent the actual state of the Earth's magnetosphere. At the same time, the use of the Jorgensen model as a theoretical framework provides prior information on the motion field and density field for dynamic estimation.

[0079] 2. OMNI solar wind parameters

[0080] The OMNI solar wind parameters used in this application are as follows: Figure 2 The complete time coverage range is from 00:00:00 UTC on May 14, 2009 to 22:40:00 UTC on May 15, 2009, with a total duration of 2800 minutes and a time resolution of 1 minute.

[0081] like Figure 2 The following are the solar wind parameters observed from 00:00:00 UTC on May 14, 2009 to 22:40:00 UTC on May 15, 2009. The panels show the solar wind proton density, the solar wind speed in three directions (V x ,V y , and V z )), interplanetary magnetic field strength in three directions (B x ,B y , and B z ), temperature (T).

[0082] It is necessary to obtain a complete solar wind parameter observation data set with a continuous period of 2800 minutes and a time resolution of 1 minute. However, the OMNI data set still has some missing data. To solve this problem, linear interpolation is used to provide a comprehensive input for subsequent simulations using the MHD model and the Jorgensen model.

[0083] Figure 2 The two red lines in correspond to the period from 01:43:00 on 15 May 2009 to 05:45:00 on 15 May 2009, during which there are 242 minutes of missing data in the OMNI dataset. We applied linear interpolation to ensure continuity of the simulation.

[0084] 3. 2-D X-ray image generation

[0085] The soft X-ray imager (SXI) is able to detect the Earth's magnetosphere because of the solar wind charge exchange (SWCX) that occurs when heavy ions in the solar wind collide with neutral atoms or molecules in outer space, thereby generating soft X-ray radiation. Solar wind charge exchange occurs when highly charged solar wind ions (such as O 7+ ) encounters a neutral atom (such as H) and captures an electron that remains in an excited state. This can be expressed as:

[0086] O 7+ +H→O 6+* +H + ,

[0087] When an ion decays to a lower energy state, it releases a photon, and SXI detects the resulting soft X-rays. SXI measures the integrated X-ray intensity along the viewing direction of a pixel. The path integral of the X-rays emitted from the charge-coupled device (CCD) at different distances in the line of sight is:

[0088]

[0089] P is the local volume energy rate (generated by the MHD model or Jorgensen). Figure 3(a) shows the lunar trajectory in the GSM coordinate system from 00:00:00 UTC on May 14, 2009 to 22:40:00 UTC on May 15, 2009, with a sampling interval of 1 minute. These coordinates are used as the observation positions of SXI, and the standardized observation direction is set to (8.0, 0, 0) Re. Figure 3(b) shows the MHD simulated X-ray emittance at 00:05:00 UTC on May 14, 2009, and the ideal soft X-ray image of the observer at the observation points S = (21.11, 59.40, 0.00) to L = (8.00, 0.00, 0.00).

[0090] Figure 3(a) depicts the lunar trajectory in the time interval from 00:00:00 UTC on May 14, 2009 to 22:40:00 UTC on May 15, 2009, with a sampling interval of 1 minute and an angular distance of 24.34° between the start and end points. The cyan scatter points are calculated using the magnetopause model of Shue et al. (1997). (b) is an ideal wide-field (40°×40°) soft X-ray image obtained based on the MHD simulation results. The image resolution is 0.2°×0.2° per pixel, the full size is 201×201 pixels, and the integration time is set to 1 minute. The black dashed box represents the field of view size of SMILE, and the white dashed box represents the field of view size of LEXI.

[0091] 4. 2-D X-ray Image Analysis

[0092] The integrated 2D X-ray images over a period of time (T) provide a comprehensive representation of the magnetic field variations over T minutes. To deeply investigate the similarity between the 2D X-ray images over T minutes and a specific moment in the integration time (with a sampling interval of 1 minute), we performed an exhaustive comparative analysis. We compared the 2D X-ray images accumulated over T minutes with the 2D X-ray images taken every minute over the integration time T using the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) as evaluation metrics. PSNR is a commonly used objective image evaluation metric that quantifies the degree of difference in pixel values ​​between two images.

[0093]

[0094] Where X is the image to be compared, Y is the reference image, L is the maximum value of the reference image, and the dimensions m and n are the height and width of the image. In general, PSNR values ​​are usually divided into four ranges: greater than 40dB, indicating that the image quality is comparable to the original image; between 30-40dB, indicating that the distortion is within an acceptable range; between 20-30dB, indicating that the image quality is relatively poor; and below 20dB, indicating severe image distortion. SSIM measures the structural similarity of two images, usually of the same size, or detects the degree of distortion of the image:

[0095]

[0096] Among them, μ X and are the mean and variance of X respectively; μ Y and are the mean and variance of Y respectively; σ XY is the covariance of X and Y; C1=(K1L) 2 and C2=(K2L) 2 It is a constant used to maintain stability. Generally, K1=0.01, K2=0.03, and L is the maximum value in the data. The value range of SSIM is generally between -1 and 1. The closer it is to 1, the higher the similarity.

[0097] The 2800 minutes were divided into 560 5-minute intervals and 280 10-minute time periods. Each time interval was integrated and the resulting 2D x-ray images were averaged. Subsequently, the 2D x-ray average image was compared with the 2D x-ray image per minute and the corresponding PSNR and SSIM were calculated. Finally, the moments corresponding to the maximum values ​​of PSNR and SSIM in each time interval were determined by statistical analysis. Figure 4 shows the statistical occurrence of the maximum values ​​of PSNR and SSIM in the 5-minute and 10-minute integration times, indicating the corresponding moments in the integration period.

[0098] By observing Figure 4, it is clear that the average 2D X-ray image produced by integration over the entire time span has obvious similarities with the 2D X-ray image at each specific moment. More specifically, the average 2D X-ray image is significantly similar to the image captured at the intermediate moments, as indicated by the statistical results when the PSNR and SSIM reach their peaks. It is worth noting that in (a) and (b), the frequency of the third minute significantly exceeds that of the other moments, while in (c) and (d), the frequency of the fifth and sixth minutes is significantly elevated. Therefore, based on probabilistic statistical inference, it is not only reasonable but also appropriate to use the initial estimate of the average 2D X-ray image over the entire integration period as a representative of the 2D X-ray image at the intermediate moments.

[0099] 5. Linear interpolation analysis

[0100] Suppose two adjacent frames are I0 and I1, and their time interval is T. To generate a new image I at the intermediate time t t , which is between I0 and I1, can use the linear interpolation formula:

[0101]

[0102] in, is the interpolation weight parameter, indicating I t Relative to the position of I0 and I1 in time, when t = 0, I t =I0; when t = T, I t =I1. In other cases, I t is the linear weighted combination of I0 and I1 according to the time weight. With the conclusion of the 2-D X-Ray image analysis, we choose to average the integral images of 10 minutes and use them as the estimated images at the intermediate moment. Then, we use the linear interpolation method to estimate the X-ray images at other time points based on the estimated X-ray images. The estimated X-ray images (a total of 2780 images, from May 14, 2009, 00:10:00 UTC to May 15, 2009, 22:30:00) are compared with the actual MHD X-Ray images at the corresponding time, and the results of the mean square error (MSE) and structural similarity index (SSIM) are shown in Figure 5(a) and Figure 5(b). Figure 5(a) and Figure 5(b) correspond to the PSNR and SSIM calculated between the 2D X-ray image obtained by linear interpolation and the original MHD 2D X-ray image, respectively.

[0103] When the solar wind parameters change slowly and tend to be stable, linear interpolation presents excellent results; however, once the solar wind parameters change significantly, the quality of the linear interpolation results decreases significantly. This phenomenon has an adverse impact on space weather forecasts, especially because space weather forecasts pay more attention to the evolution of X-ray images under drastic changes in solar wind parameters, which may trigger new space environmental events. Given that the dynamic estimation of X-ray images under drastic changes in solar wind parameters is of great scientific significance, this application proposes an innovative method to improve the accurate dynamic estimation of X-ray images under the condition of significant changes in solar wind parameters, providing more reliable and accurate information for space weather forecasts.

[0104] 2. Main Methods

[0105] An adaptive dynamic X-ray image estimation method based on motion estimation and kernel density estimation. Its specific implementation process includes motion estimation network, density estimation network, contribution maps, cycle consistency and spatiotemporal discriminator.

[0106] 1. X-ray image interpolation framework

[0107] We average the X-ray images of the integration time T as the initial observation estimate of the X-ray image at the intermediate time. Our goal is to obtain the X-ray image at a certain time t (0≤t≤T) in the middle. Make an estimate, generate The process is mainly divided into three parts: motion estimation, adaptive kernel density estimation and contribution map estimation. If we define the adaptive X-ray estimator as Γ, the entire X-ray image estimation process can be expressed as,

[0108]

[0109] Among them, Z t←0 and Z t←1 are contribution maps, whose size is consistent with the X-ray image. For each pixel p, Zt ←0 (p)∈[0,1] represents the relationship between I0 and I1 t I0 and I1 are the initial estimates of the X-ray image mentioned above. N = (1-t) Zt ←0 +tZt ←1 is the normalization factor. t→0 and J t→1 yes The detailed framework is as follows: Figure 6 As shown,

[0110] The motion estimation module learns key prior knowledge about the magnetospheric motion changes from a series of 2D X-ray images simulated by the Jorgensen model. The adaptive kernel density estimation module more effectively captures the spatial evolution of pixel values ​​in actual observations. The learnable contribution map is used to describe the correlation between different frames of the X-ray image, while the temporal discriminator ensures the continuity and consistency of the estimated X-ray image data in time.

[0111] 2. Adaptive X-ray Image Estimator

[0112] 1) Bidirectional motion flow field estimation

[0113] In this application, the optical flow estimation model PWC-Net is used as the motion estimation network. Since it is extremely difficult to perform motion estimation in the absence of true value supervision, the pre-trained PWC-Net is used to initialize the motion estimation network to accelerate the model training process and further improve its generalization performance. Figure 7 Shown is a schematic diagram of bidirectional motion flow field estimation.

[0114] The Jorgensen model is integrated in continuous time to obtain two X-ray images (G0 and G1) and input them into PWC-Net; the estimated forward motion field from G0 to G1 is called J 0→1 , the estimated reverse motion field J from G1 to G0 1→0 ,

[0115] J 0→1 =PWC-Net(G0→G1)

[0116] J 1→0 =PWC-Net(G1→G0)

[0117] J 0→1 and J 1→0 It is essentially a two-dimensional vector field, and each two-dimensional vector represents the horizontal and vertical displacement of a pixel in the image.

[0118] 2) Adaptive Kernel Density Estimation

[0119] The idea of ​​traditional kernel density estimation is to use a kernel function to weight the area around each observation value, and only consider the pixels to be estimated that are related to the pixels within the convolution kernel window. This method will no longer be applicable when the object or camera undergoes a large movement change. Deformed convolution is used, which adds the offset vector Δl = (U, V) to the classic convolution.

[0120]

[0121] Where L is the set of all pointing points (i, j) in the two-dimensional vector field J estimated by PWC-Net, and length(L) is the number of all points in the set. k is the kernel weight matrix, the size is 1×L, U k and V k Represents the horizontal and vertical displacement of the k-th pixel L(k) pointing to point (i, j). In order to make the estimation more accurate and the convolution operation more convenient, expand each pixel L(k), that is, perform a kernel convolution operation with a window size of M×N on L(k), which helps the adaptive kernel density estimation method explore a wider area. Perform the following operations.

[0122]

[0123] Therefore, Γ(I,J) is defined as the value for each pixel of the X-ray image to be estimated. Perform the above estimation operation. It is worth noting that I represents the previous frame I0 and I1 of the frame to be estimated, and J represents the forward and reverse bidirectional motion field calculated by PWC-net. The principle of adaptive kernel density estimation is shown in Figure 8(a), Figure 8(b), and Figure 8(c), which generally describe the implementation process of adaptive kernel density estimation. Figure 8(a) is the kernel-based X-ray estimation implementation process, Figure 8(b) is the motion-based X-ray estimation implementation process, and Figure 8(c) is the adaptive X-ray estimation implementation process; the purple part in each figure represents the generation of the target pixel, while the other colors represent reference points from different sources.

[0124] 3) Learnable contribution maps

[0125] In order to improve the rationality and accuracy of the estimation, this paper introduces an innovative method, which is to learn the contribution map through a U-Net network. The contribution map plays a key role in the estimation process of X-ray images, which reflects the correlation between the X-ray image to be estimated and the previous and next frames. Specifically, the X-ray radiation intensity of the X-ray to be estimated may be generated by the SWCX process in the previous frame, or by the SWCX process in the next frame, or even by both the previous and next frames at the same time.

[0126] The generation of the contribution map mainly relies on the gradient information of the motion field generated by PWC-net. The gradient is a local measure of the change in motion in the motion field, which tells us which areas in the image have motion. The magnitude and direction of the gradient provide an estimate of the speed and direction of motion. In the gradient field, large gradients usually correspond to significant motion or the location of the edge of an object in the image. A preliminary estimate of the contribution map can be obtained by normalizing or weighting the gradient field. In this process, areas with larger gradients are considered to contribute more to the optical flow field estimation, while areas with smaller gradients have less impact on the estimation. The preliminary estimated contribution map can be further optimized by a deep learning model (U-net) to more accurately reflect the contribution of the previous and next frames to the X-ray image estimation process.

[0127] Z=U-Net(▽J,θ)

[0128] Among them, ▽J is the gradient of the motion field information J calculated by PWC-Net, and θ represents the parameters used by the U-Net network learning. The application of this method in X-ray image estimation provides more comprehensive information for the deep learning model, enabling it to more accurately capture the relationship between X-ray images and previous and next frames, thereby improving the performance of X-ray image estimation.

[0129] 3. Cycle consistency

[0130] The adaptive X-ray image method proposed in this application requires X-ray image data within three time intervals T. The time interval can be set freely (5 minutes for SMILE and 10 minutes for LEXI), and the sampling interval is 1 minute, which is determined by the minimum sampling rate of OMNI data. The three time intervals are called T1, T2, and T3. The integrated X-ray images within T1 and T3 are averaged and used as the starting X-ray images I0 and I1. Here, we summarize the inputs and final outputs of our method. The inputs include the simulated SXI observations I0 and I1 (the X-ray images obtained by integrating the MHD model are used in this paper), the X-ray images G1...G obtained by integrating the Jorgensen model with a time interval of 1 minute from I0 to I1, and the X-ray images G2, G3, G4, G5, G6, G7, G8, G9, G10, G11, G12, G13, G14, G15, G16, G17, G18, G19, G20, G219, G211, G221, G23, G24, G25, G26, G27, G28, G29, G30, G311, G322, G33, G34, G35, G36, G37, G38, G39, G40, G41, G42, G43, G44, G45, G46, G47, G48, G49, G50, G51, G52, G53, G54, G55, G56, G57, G58, G59, G60, G61 2N The output is all X-ray images with a 1-minute interval within the T2 time period. Therefore, the method of the present application can be extended to any time interval T (greater than 3 minutes), thereby meeting the application requirements under different situation tasks.

[0131] 4. Time Discriminator

[0132] Time Discriminator (D t ) is used to process X-ray images of consecutive frames to ensure that the generated X-ray images have a smooth transition in time, thereby improving the continuity and consistency of the generated X-ray image data in time. The input data of the temporal discriminator is the time series data of multiple consecutive frames generated by the adaptive X-ray image estimator. In order to better capture the information of the time dimension, the temporal discriminator D t A slow fusion model is used. Unlike the method of directly using four convolutional layers for convolution, a hierarchical structure is introduced to process the input data. As the data is slowly passed through the network, a gradual fusion method is used to fuse consecutive frames in pairs until all frames are fused into one. Fig.10 Shown is the slow fusion model architecture. Two consecutive frames obtained by adaptive estimation are fed into two independent processing streams: a context stream that simulates the low-resolution image and a foveal stream that processes the high-resolution center crop.

[0133] Two consecutive frames or feature maps are fused in groups until the number is reduced to 1. The last layer is a fully connected node layer, which is used to generate the final discriminant result. Discriminator D t The judgment basis is learned from 2800 X-ray images generated by the Jorgensen model. t The objective function of the training uses the cross entropy which is most commonly used in adversarial training. The specific form is as follows:

[0134]

[0135] In one embodiment, 2000 sets of consecutive N frames of X-ray images are randomly selected from 2800 images as And use label smoothing technology to add 2000 groups Randomly generate label values ​​of 0.9-1. At the same time, in order to improve the discriminator D t The generalization performance of , and then randomly select 1000 groups of N frames of disordered X-ray images as negative samples Participation Discriminator D t training.

[0136] 5. Training and implementation details

[0137] This application designs four loss functions for different aspects. The final loss function is a linear combination of the four terms, λ o ,λ p ,λ f ,λ d are the linearly weighted weights.

[0138] L=λ o L o +λ p L p +λ f L f +λ d L d

[0139] Assume input images I0 and I1, a set of intermediate frames between them where t i ∈[0,1]. Estimate of the intermediate frame Observation loss L o Intermediate frames designed to constrain the estimate The sum of the X-ray images observed by the actual SXI within T time Keep it as consistent as possible. The specific calculation formula is as follows:

[0140]

[0141] Perceptual loss L p .Even if we use L1 loss to simulate the estimation error of the intermediate frame X-ray image, it may still cause the estimated image to be blurred. Therefore, perceptual loss is used to preserve details, making the estimated X-ray image clearer;. Specifically, the perceptual loss L p Defined as:

[0142]

[0143] Where φ represents a commonly used deep learning detail acquisition indicator. Flow loss L f Simulate the effect of motion estimation of intermediate frames:

[0144]

[0145] J is the optical flow field information estimated using PWC-Net. i Represents the estimated X-ray image, G i is the X-ray image obtained by Jorgensen simulation at the corresponding time. f The significance of is to make the motion field estimation of the intermediate estimation frame consistent with the motion estimation calculated by the Jorgensen model. The adaptive X-ray image estimator is regarded as a generator G, whose output is the estimated X-ray image Discriminator loss L d The role is to improve Continuity and consistency in time, ensuring Has a smooth transition in time. d is defined as:

[0146]

[0147] In one embodiment, the selected linear combination of loss function weights is: λ0=0.7, λ1=0.1, λ2=0.8, λ3=0.6, which is summarized based on training experience.

[0148] During model training, continuous-time integral observations are used. And the constraints of training the model through the simulation results of the Jorgensen model driven by OMNI data. These two types of data are available in actual tasks, so this method is applicable to actual soft X-ray observation tasks in the problem of dynamic X-ray image estimation. The training method belongs to the self-supervised learning method, that is, each dynamic estimation of a group of N X-ray images needs to be retrained based on the preprocessing model, which highlights the adaptability of this method and expands the scope of application of the model.

[0149] In the present invention, PWC-Net is used to obtain the motion field change information J between X-ray images at 1 minute time intervals (total time 2800 minutes) generated by Jorgensen model simulation.

[0150] Regarding adaptive kernel density estimation and learnable contribution maps, a fully convolutional neural network is designed to estimate the kernel weights W k (m,n) and contribution maps to estimate the parameters θ. The specific network structure uses U-Net, which consists of an encoder, a decoder, and a skip connection. Each processing unit basically contains 3×3 convolution and ReLU activation. For the encoder part, average pooling is used to extract features. The decoder upsampling layer uses bilinear interpolation for upsampling to upsample the feature map in the encoder to a higher resolution. For the kernel weight W k In the case where (m,n) must be non-negative, a softmax layer is used as a constraint, and a sigmoid activation function is used to satisfy Z∈[0,1]. The Adamax optimizer is used for iterative updates of parameters during the entire training process.

[0151] In the process of estimating dynamic X-ray images with a time interval of N, the training epoch is set to 4000 and the batch size is 16. The initial learning rate of the Adamax optimizer is set to 0.0002 and the Adam motion is set to 0.5. The learning rate of the first 2000 epochs remains unchanged. In the next 2000 epochs, if the loss function L does not decrease in 5 iterations, we reduce the learning rate by 0.2 times until the learning rate drops to 0 or the maximum number of training times (2000) is reached, and then the training is stopped.

[0152] 3. Experiment

[0153] 1. Introduction of experimental data

[0154] The experimental data time range is from May 14, 2009, 12:10:00 UTC, to May 16, 2009, 10:30:00 UTC, a total of 2780 minutes. The MHD model and the Jorgensen model and the corresponding solar wind parameters are used to simulate the dynamic three-dimensional magnetospheric X-ray rate with a time interval of 1 minute. The X-ray image simulation technology is used to generate continuous dynamic MHD model X-ray images and Jorgensen model X-ray images with the corresponding time and lunar orbit position. Continuous dynamic Jorgensen model X-ray images It is used as input to provide prior motion information for the adaptive X-ray image dynamic estimation method. The continuous dynamic MHD model X-ray images will also be accumulated at a time interval of T = 10 minutes as SXI simulation observation results (I0 and I1). At the same time, 2780 copies of MHD model X-ray images will also be used as Ground_turth for comparison.

[0155] 2. Dynamic X-ray Image Estimation Results

[0156] The adaptive X-ray image dynamic estimation method of the present application is used to estimate the X-ray images corresponding to 2780 minutes. Specifically, the time interval is set to 10 minutes and the adaptive X-ray image dynamic estimation method is used to perform 278 independent X-ray image estimations, thereby obtaining a 2780-minute dynamic X-ray image. PSNR and SSIM were used to evaluate the estimation results. The distribution of PSNR and SSIM of the 2780-minute X-ray image is shown in Figures 11(a), 11(b), 11(c) and 11(d). Figures 11(a) and 11(b) respectively represent the PSNR and SSIM calculated by the linear interpolation method (black) and the adaptive dynamic x-ray image estimation method (red) for the MHD model x-ray image (Ground Truth). Subsequently, Figures 11(c) and 11(d) show the histogram distribution of PSNR and SSIM over the entire time period. In these histograms, the blue bars represent the results obtained by the linear interpolation method, and the orange bars represent the results obtained by the adaptive dynamic x-ray image estimation method.

[0157] Compared with the linear interpolation method, the adaptive X-ray image estimation method proposed in this paper has significant advantages in both PSNR and SSIM. This conclusion is drawn from the trend that the red line in Figure 11 (a) and Figure 11 (b) continues to exceed the black line. In particular, the advantage is more obvious in terms of PSNR, while in SSIM, the adaptive X-ray image estimation method is significantly different from the linear interpolation method only when the solar wind changes drastically. However, from the numerical distribution of PSNR and SSIM in Figure 11 (c) and Figure 11 (d), it can be seen that the PSNR values ​​of the adaptive X-ray image estimation method are mainly distributed above 30dB, and most of them are concentrated above 40dB. Similarly, most of the SSIM values ​​remain above 0.95. In contrast, the PSNR distribution range of the linear interpolation method is wider, which indirectly reflects the accuracy and stability of the adaptive X-ray image estimation method.

[0158] In summary, two conclusions can be drawn: First, when the solar wind parameters change gradually and the magnetosphere remains relatively stable, both the adaptive X-ray image estimation method and the linear interpolation method perform well in terms of PSNR and SSIM. Second, during periods of significant changes in solar wind parameters, the interaction between the magnetosphere and energetic particles in the solar wind leads to changes in morphology. At this time, the PSNR and SSIM values ​​of the linear interpolation method drop significantly, while the adaptive X-ray image estimation method continues to show robustness. Typically, rapid changes in magnetospheric morphology indicate the occurrence of space weather events, demonstrating the ability of the proposed method to accurately estimate the state of the magnetosphere during such events. As shown in Table 1:

[0159] Table 1 Statistical distribution of PSNR and SSIM results

[0160]

[0161] In Table 1, PSNR is divided into four groups: less than 20dB, 20-30dB, 30-40dB, and greater than 40dB. For SSIM, we divide it into four groups: greater than 0.98, 0.9-0.98, 0.8-0.9, and less than 0.8. At the same time, the number and proportion of the linear interpolation method and the adaptive X-ray image estimation method in the PSNR and SSIM range are counted. The results show that the PSNR calculated by the adaptive dynamic X-ray image estimation method is mainly concentrated in the two ranges of 30-40dB and greater than 40dB, while SSIM is mainly concentrated in the two groups of greater than 0.98 and 0.9-0.98, which reflects the superiority and stability of the method in this paper. In contrast, the values ​​of PSNR and SSIM of the linear interpolation method are lower and the distribution is more scattered, reflecting the instability of the method.

[0162] 3. Analysis of dynamic X-ray image results at different integration times

[0163] The time interval T has a certain influence on both the linear interpolation method and the adaptive X-ray image estimation method in this paper. First, the time interval T will affect the accuracy of the estimation of the initial state of the magnetosphere. Second, the longer the time interval T, the more obvious the magnetosphere features from the SXI observation image will be. We analyzed the linear interpolation results and adaptive X-ray image estimation results with time intervals of 5 minutes, 10 minutes, 15 minutes and 20 minutes, and their PSNR and SSIM statistical distribution curves are shown in Figures 12(a), 12(b), 12(c) and 12(d). Figures 12(a) and 12(b) show the statistical distribution of PSNR and SSIM of the linear interpolation method at time intervals of 5 minutes, 10 minutes, 15 minutes and 20 minutes. Figures 12(c) and 12(d) show the statistical distribution of PSNR and SSIM of the adaptive dynamic X-ray estimation method at time intervals of 5 minutes, 10 minutes, 15 minutes and 20 minutes.

[0164] As shown in Figure 12, the linear interpolation method is more significantly affected by the time interval T than the adaptive X-ray image estimation method. As the time interval T increases, the PSNR and SSIM corresponding to the linear interpolation results and the adaptive X-ray image estimation results both show a certain degree of decline. Specifically, the distribution curves of Figures 12(a) and 12(b) shift significantly to the left with the increase of the time interval T, which represents a significant decrease in PSNR and SSIM. In addition, there is a huge left shift from T = 10 minutes (red curve) to T = 15 minutes (yellow curve) in Figure 12(a), which reflects from the side that the linear interpolation method is not suitable for interpolation estimation tasks with a time interval greater than 10 minutes. On the contrary, Figures 12(c) and 12(d) also show a certain degree of left shift. This shows that although the adaptive X-ray image estimation method proposed in this paper is also affected by the time interval T, the degree of influence is relatively low. Specifically, when the time interval is 20 minutes, the PSNR of the adaptive X-ray image estimation method can be maintained above 30dB, and the SSIM can also be maintained above 0.95.

[0165] 4. Estimation of detailed analysis of X-ray images

[0166] The edge information of X-ray images plays a vital role in the three-dimensional reconstruction of the magnetosphere. For example, the maximum position of the X-ray image is used to reconstruct the three-dimensional magnetopause; and the bow shock wave in the outer layer directly reflects the specific changes in the interaction between the magnetosphere and the high-energy particles carried by the solar wind. Figure 13 (a), Figure 13 (b), and Figure 13 (c) show the changes in the maximum and bow shock waveforms of the original MHD X-ray image, linear interpolation results, and adaptive estimation results from May 14, 2009, 00:10:00 UTC to May 16, 2009, 10:30:00 UTC (10 minutes). Figure 13 (a) shows the original two-dimensional X-ray image obtained from the MHD simulation. Figure 13 (b) illustrates the estimation result by linear interpolation as an approximation of the original image. Figure 13 (c) shows the estimation result of the adaptive dynamic X-ray image estimation method proposed in this paper. In addition, each row maintains a consistent color range for comparison.

[0167] Compared with the original MHD X-ray image, the adaptive estimation results show significant consistency in both the maximum shock shape and the bow shock shape, while the linear interpolation results show obvious differences. Specifically, the 5 minutes shown in Figure 13 captures the shape change cycle after the interaction between the solar wind and the magnetosphere. In the first and third columns, the deformation process of the maximum position of the magnetosphere and the bow shock in the 2D X-ray image can be seen when interacting with the solar wind. In contrast, linear interpolation cannot capture the significant changes in the maximum position of the magnetosphere and the bow shock. The PSNR and SSIM results of the linear interpolation and adaptive dynamic X-ray image estimation methods in Figure 13 compared with the original MHD X-ray image are shown in Table 2.

[0168] Table 2 PSNR and SSIM corresponding to the X-ray images in Figure 13

[0169]

[0170] This highlights the considerable advantage of the adaptive estimation method in accurately capturing the details of X-ray motion. This advantage is attributed to the X-ray images simulated using the OMNI solar wind parameters and the Jorgensen model. These inputs provide prior information that improves the accuracy of the adaptive estimation method in capturing the details of X-ray image motion.

[0171] This method has unique advantages in accurately obtaining boundary and detail information by capturing subtle changes in the interaction between the solar wind and the magnetosphere, laying a solid foundation for a deep understanding of the evolution of the magnetosphere structure. This not only helps to fully understand the magnetosphere dynamics, but also provides important data support for in-depth research on the relationship between solar activity and the Earth's magnetosphere. This method will have a profound impact on the scientific understanding of the Earth's space environment and its applications.

[0172] Example 2

[0173] Embodiment 2 of the present invention provides an adaptive X-ray dynamic image estimation system, which is implemented based on the method of embodiment 1. The system includes:

[0174] An acquisition input module is used to average the X-ray images with integration times T1 and T3 as the starting X-ray image;

[0175] Motion estimation module, used for motion estimation through the optical flow estimation model PWC-Net;

[0176] Adaptive kernel density estimation and contribution map estimation modules to capture the spatial evolution of pixel values ​​in actual observations and describe the correlation between different frames of X-ray images using learnable contribution maps;

[0177] The time discriminator is used to ensure the temporal continuity and consistency of the estimated X-ray image data, thereby realizing the estimation of the X-ray dynamic images at different moments within the intermediate time T2.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.< / g> < / g>

Claims

1. An adaptive X-ray dynamic image estimation method for space science missions, comprising: The X-ray images with integration times T1 and T3 are averaged as the starting X-ray image, and motion estimation is performed through the optical flow estimation model PWC-Net. At the same time, an adaptive kernel density estimation method is used to capture the spatial evolution of pixel values ​​in actual observations, and a learnable contribution graph is used to describe the correlation between different frames of the X-ray image. The temporal continuity and consistency of the estimated X-ray image data are ensured by the temporal discriminator, thereby realizing the estimation of X-ray dynamic images at different times within the intermediate time T2. The method of averaging the X-ray images with integration times T1 and T3 as the starting X-ray image comprises: The X-ray images with integration times T1 and T3 are obtained by simulating the MHD model, and the X-ray images within the T2 time are adaptively and dynamically estimated with the set sampling interval, combined with the fluid dynamics motion information provided by the Jorgensen model; or The actual X-ray images at integration time T1 and T3 are collected, and combined with the prior knowledge of magnetohydrodynamic motion provided by the MHD model, the X-ray images within time T2 are adaptively and dynamically estimated at the set sampling interval; Regarding adaptive kernel density estimation and learnable contribution maps, a fully convolutional neural network is designed to estimate the kernel weights W k (m,n) and the parameter θ estimated by the contribution maps. The specific network structure adopts U-Net, which includes an encoder, a decoder and a skip connection. Each processing unit contains 3×3 convolution and ReLU activation. The encoder uses average pooling to extract features. The decoder upsampling layer uses bilinear interpolation to upsample the feature map in the encoder to a higher resolution. The softmax layer is used as a constraint to ensure that the kernel weight is non-negative. The temporal discriminator adopts a slow fusion model, the input is two consecutive frames obtained by adaptive estimation, and the output is the discrimination result; the processing process is: the two consecutive frames obtained by adaptive estimation are sent to two independent processing streams respectively, which respectively simulate the context stream of low-resolution images and the foveal stream for processing high-resolution center cropping; the two independent processing streams include alternating convolution, normalization and pooling layers, and then converge to a fully connected layer.

2. The adaptive X-ray dynamic image estimation method according to claim 1, characterized in that: The input of the optical flow estimation model PWC-Net is two X-ray images G0 and G1, and the output is the forward motion estimation field J from G0 to G1. 0→1 , and the estimated reverse motion field J from G1 to G0 1→0 .

3. The adaptive X-ray dynamic image estimation method according to claim 1, characterized in that: The U-Net training adopts Adamax optimizer, and the loss function is: L=λ o L o +λ p L p +λ f L f +λ d L d Among them, λ o ,λ p ,λ f ,λ d is the linear weighting weight, L o is the observation loss, L p is the perceptual loss, L f is the loss of simulating the effect of intermediate frame motion estimation, L d is the discriminator loss, which satisfies the following formulas: in, is a set of intermediate frames between the input images I0 and I1, i represents the frame number, N represents the total number of frames, and t i ∈[0,1], is the X-ray image observed within T time; Among them, φ represents the detail index after convolution; Among them, J is the optical flow field information estimated by PWC-Net, I i It indicates the estimated i-th frame X-ray image, I i+1 It indicates that the estimated i+1th frame of X-ray image is obtained, G i is the i-th frame X-ray image obtained by Jorgensen model simulation at the corresponding time, G i+1 is the i+1th frame of X-ray image obtained by Jorgensen model simulation at the corresponding time, 4. The adaptive X-ray dynamic image estimation method according to claim 1, characterized in that: The objective function of the slow fusion model training uses cross entropy and satisfies the following formula: in, is N frame samples, is N frames of negative samples, i represents the frame number, N represents the total number of frames, D t () represents the discriminator, and E() represents the expected value.

5. A system based on the adaptive X-ray dynamic image estimation method of claim 1, characterized in that: include: An acquisition input module is used to average the X-ray images with integration times T1 and T3 as the starting X-ray image; Motion estimation module, used for motion estimation through the optical flow estimation model PWC-Net; Adaptive kernel density estimation and contribution map estimation modules to capture the spatial evolution of pixel values ​​in actual observations and describe the correlation between different frames of X-ray images using learnable contribution maps; and The time discriminator is used to ensure the temporal continuity and consistency of the estimated X-ray image data, thereby realizing the estimation of the X-ray dynamic images at different moments within the intermediate time T2.

Images