Satellite image definition improving method based on elliptical Gaussian distribution

Abstract

The invention discloses a satellite image definition improving method based on elliptic Gaussian distribution, and relates to the technical field of remote sensing image processing, and the method comprises the steps: calculating and generating motion data according to collected physical imaging parameters, and constructing a reference motion vector based on the motion data, the motion data comprising a long-axis reference value and a reference fuzzy direction angle; a constraint space is constructed based on statistical characteristics of a real satellite image, and random degradation parameters are collected from the constraint space and comprise a long-axis disturbance item, a direction angle jitter item and a short-axis standard deviation; performing fusion processing on the reference motion vector and the random degradation parameter to generate a composite elliptical Gaussian degradation convolution kernel; according to the method, a long-axis reference value, a reference fuzzy direction angle, a random degradation parameter and alignment reference data are introduced in the construction process of a training sample pair set, so that the training sample pair better fits real satellite imaging in the aspects of fuzzy structure and space matching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and particularly relates to a satellite image definition enhancement method based on elliptical Gaussian distribution. BACKGROUND

[0002] With the increasing demand for satellite imaging tasks in fine ground monitoring, disaster analysis, urban change analysis and resource investigation, the role of satellite image definition in spatial target recognition and texture information analysis becomes increasingly critical. Imaging equipment usually needs to obtain continuous ground information under high-speed orbital motion, complex attitude disturbance and unstable atmospheric conditions, resulting in multi-factor coupling characteristics of motion blur and optical diffusion generated during imaging. Under such conditions, the processing methods that simply rely on fixed blur models, linear motion assumptions or static degradation kernels often cannot accurately describe the change pattern of real imaging degradation, making it difficult to support subsequent deep learning training tasks based on pseudo-satellite image data, thereby laying the demand background for proposing an image definition enhancement scheme that can reflect both the physical imaging process and the spatial statistical structure.

[0003] In the process of simulating imaging blur patterns based on physical imaging parameters, it is usually difficult to convert the combined effects of satellite orbit attitude angle, ground projection velocity, imaging scanning direction and exposure integration time into imaging motion characteristics that can stably describe the blur amplitude and blur direction, making the generated pseudo-satellite image data difficult to maintain correspondence with the real imaging process in terms of motion blur structure. When the shift of the blur direction angle, component noise and change of the motion scale are not fully expressed, the blur structure of the pseudo-satellite image data in the training sample pair set is prone to distortion, thereby affecting the learning ability of the subsequent reconstruction model for the real degradation pattern.

[0004] Secondly, in the process of constructing statistical degradation parameters required for degradation simulation, it is usually difficult to convert the edge gradient histogram, atmospheric disturbance standard deviation statistics and point spread function half-width value into a multi-dimensional constraint space that can be used for sampling, making the long-axis disturbance term, direction angle jitter term and short-axis standard deviation lack uniformity in statistical sources. In the absence of stable correlation between gradient section data, disturbance category data and diffusion interval data, the generated random degradation parameters are difficult to reflect the gradient distribution, disturbance structure and optical diffusion characteristics under real imaging conditions, resulting in insufficient blur diversity of the pseudo-satellite image data, thereby limiting the coverage ability of the training sample pair set for real blur conditions.

[0005] Further, in the process of establishing the spatial correspondence between the pseudo-satellite image data and the high-definition aerial image, slight pixel-level shifts often occur due to resampling operations, local texture differences, or image boundary processing, affecting the spatial consistency of the training sample pair set; if the coordinate difference values of the local regions in the pseudo-satellite image data lack effective screening, or the global alignment relationship cannot be stably estimated through the gradient intensity weight, spatial misalignment is likely to occur in the pixel-by-pixel pairing process, making it difficult for the deep learning network in the parameter iterative training stage to accurately model the true blur shape, spatial shift, and local changes, thereby affecting the efficiency of satellite image sharpness improvement. SUMMARY

[0006] To solve the above technical problems, the application provides a satellite image sharpness improvement method based on an elliptical Gaussian distribution, which comprises: S1, generating motion data according to the collected physical imaging parameters, and constructing a reference motion vector based on the motion data, wherein the motion data comprises a long-axis reference value and a reference blur direction angle; S2, constructing a constraint space based on the statistical characteristics of the real satellite image, and collecting random degradation parameters from the constraint space, wherein the random degradation parameters comprise a long-axis perturbation term, a direction angle jitter term, and a short-axis standard deviation; S3, fusing the reference motion vector and the random degradation parameters to generate a composite elliptical Gaussian degradation convolution kernel; S4, performing degradation simulation operation on the high-definition aerial image through the composite elliptical Gaussian degradation convolution kernel to generate pseudo-satellite image data, and performing pixel-level alignment between the pseudo-satellite image data and the corresponding high-definition aerial image to construct a training sample pair set.

[0007] Further, the physical imaging parameters comprise satellite orbit attitude angle, ground projection velocity, imaging scanning direction, and exposure integration time; and the statistical characteristics comprise edge gradient histogram of the real satellite image, atmospheric disturbance standard deviation statistics, and point spread function half-width value.

[0008] Further, the step of constructing the reference motion vector based on the motion data comprises: S11, performing parameter recording processing through the satellite orbit attitude angle, the ground projection velocity, the imaging scanning direction, and the exposure integration time to generate a physical imaging parameter record; S12, performing multiplication calculation based on the ground projection velocity and the exposure integration time in the physical imaging parameter record to generate a long-axis reference value; S13, performing angle analysis processing on the imaging scanning direction in the physical imaging parameter record to generate a reference blur direction angle; S14, performing vector construction operation according to the long-axis reference value and the reference blur direction angle to generate a reference motion vector.

[0009] Further, the logic of performing the angle resolving processing is: a1, performing a spatial direction decomposition processing by the imaging scanning direction to generate a direction decomposition quantity; a2, performing an angle offset estimation calculation according to the imaging scanning direction to generate an angle offset quantity; a3, performing a direction noise suppression calculation by the imaging scanning direction to generate a noise suppression quantity; a4, performing an angle fusion processing based on the direction decomposition quantity, the angle offset quantity and the noise suppression quantity to generate a reference ambiguous direction angle.

[0010] Further, the step of constructing the constraint space based on the statistical characteristics of the real satellite image comprises: S21, performing a section gradient distribution determination based on the edge gradient histogram to generate gradient section data; S22, performing a disturbance amplitude classification processing according to the atmospheric disturbance standard deviation statistical quantity to generate disturbance category data; S23, performing a diffusion range interval division based on the point spread function half-width value to generate diffusion interval data; S24, performing a spatial domain layering processing by the gradient section data, the disturbance category data and the diffusion interval data to construct the constraint space.

[0011] Further, the step of generating the gradient section data is: S211, performing an amplitude density analysis based on each gradient amplitude section of the edge gradient histogram to generate a density mutation section record; S212, performing an amplitude section boundary determination according to the density mutation section record to obtain an amplitude boundary set; S213, performing a section division processing by the amplitude boundary set and the amplitude distribution of the edge gradient histogram to generate the gradient section data.

[0012] Further, the logic of performing the amplitude density analysis is: b1, performing a section gradient change rate analysis based on the amplitude section frequency change rate of the edge gradient histogram to obtain change rate fluctuation data; b2, performing an amplitude section stability determination according to the number of consecutive inversions of adjacent change rate signs in the change rate fluctuation data to obtain an unstable amplitude section set; b3, performing a mutation section screening processing by the unstable amplitude section set and the frequency gradient difference of the corresponding amplitude section in the edge gradient histogram to generate the density mutation section record.

[0013] Further, the generation logic of the composite elliptical Gaussian degradation convolution kernel is: S31, performing an amplitude direction superposition processing by the long axis reference value and the long axis disturbance term to generate a long axis fusion value; S32, performing angle correction processing based on the reference blur direction angle and the direction angle jitter term to obtain a direction fusion angle value; S33, performing axis parameter updating according to the short axis standard deviation and the correlation coefficient field to generate a short axis updated value; S34, performing convolution kernel parameter collection processing through the long axis fusion value, the direction fusion angle value and the short axis updated value to generate a composite elliptical Gaussian degradation convolution kernel.

[0014] Further, the logic for constructing the training sample pair is: c1, performing convolution processing through the composite elliptical Gaussian degradation convolution kernel and the high-definition aerial image to generate pseudo-satellite image data; c2, performing spatial coordinate correspondence analysis on the pseudo-satellite image data and the high-definition aerial image to obtain alignment reference data; c3, performing pixel-by-pixel pairing processing based on the alignment reference data to generate a training sample pair set.

[0015] Further, the step of performing spatial coordinate correspondence analysis is: c21, performing coordinate difference value calculation based on the corresponding pixel coordinates of the pseudo-satellite image data and the high-definition aerial image to generate difference segment data; c22, performing local consistency screening through the difference segment data to obtain candidate alignment unit data; c23, performing coordinate mapping confirmation processing according to the candidate alignment unit data to generate alignment reference data.

[0016] Compared with the prior art, the present application has the following advantages: The present application constructs imaging motion features based on physical imaging parameter records, long axis reference values and reference blur direction angles, and uses them to dynamically control the blur direction and blur scale in the subsequent degradation convolution kernel, so that the pseudo-satellite image data can be close to the spatial degradation characteristics of the real satellite image in the motion blur form, thereby establishing a blur structure mapping basis consistent with the real imaging environment in the training sample pair set; In addition, the present application obtains random degradation parameters from the constraint space composed of gradient segment data, disturbance category data and diffusion interval data, so that the long axis disturbance term, the direction angle jitter term and the short axis standard deviation are consistent in origin and physical consistency at the statistical level, so as to introduce controllable blur diversity in the composite elliptical Gaussian degradation convolution kernel, thereby improving the stability and rationality of the pseudo-satellite image data in the real blur state coverage.

[0017] Further, the application also performs spatial coordinate corresponding analysis, local consistency screening and alignment reference data generation processing on the pseudo-satellite image data and high-definition aerial image, so that the pixel-level corresponding relationship in the training sample pair set is established on the basis of global and local consistent spatial translation correction, thereby reducing the sample deviation caused by resampling error, window texture difference or local drift, improving the reliability and overall coordination of the training sample pair set in the spatial matching level, and further enhancing the effective modeling capability of the deep learning network on the difference between the pseudo-satellite image data and the high-definition aerial image in the parameter iterative training process.

[0018] Further, the application is based on the fuzzy structure consistency and spatial matching consistency of the training sample pair set in the parameter iterative training stage, so that the deep learning network can stably converge under the conditions of multi-type fuzzy form, different spatial position offset and multi-scale diffusion feature, and obtain a satellite image sharpening processing model capable of outputting high-fidelity reconstruction results under the matching error constraint, so that the image reconstruction task has higher reconstruction accuracy and stronger structure restoration capability when facing the motion blur, optical diffusion and local structure inconsistency of real satellite images. In summary, the application introduces the long axis reference value, the reference fuzzy direction angle, the random degradation parameter and the alignment reference data in the construction process of the training sample pair set, so that the training sample pair is more consistent with the real satellite imaging in terms of fuzzy structure and spatial matching. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0020] Figure 1 A flowchart of a satellite image sharpness enhancement method based on elliptical Gaussian distribution provided by the embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] Please refer to Figure 1As shown, the embodiment discloses a satellite image sharpness enhancement method based on elliptical Gaussian distribution, which comprises: S1, according to the collected physical imaging parameters, the motion data is calculated and generated, and the reference motion vector is constructed based on the motion data, wherein the motion data includes the long axis reference value and the reference blur direction angle; In one specific embodiment, for the basic physical imaging input parameters required for constructing the subsequent motion degradation model, the imaging physical quantities directly observed or retrieved in the satellite imaging process are structured and recorded.

[0023] Specifically, the physical imaging parameters include satellite orbit attitude angle, ground projection speed, imaging scanning direction and exposure integration time; Among them, the satellite orbit attitude angle is output by the high-precision gyroscope and star sensor of the satellite attitude control system, which is used to describe the spatial attitude change of the satellite at the imaging moment, and is the key physical quantity reflecting the offset of the viewing axis.

[0024] The satellite orbit attitude angle includes roll angle, pitch angle and yaw angle, and there is a small disturbance at the imaging moment, which is °.

[0025] The ground projection speed is the speed of the satellite along the ground projection track direction at the imaging moment, which is measured by the orbit mechanics model combined with the global satellite navigation system, and can be used to describe the relative motion of the target ground object during imaging, which is m / s.

[0026] The imaging scanning direction is provided by the load scanning mechanism, which is used to describe the direction relationship of the sensor scanning line relative to the ground coordinate system during imaging, which is °.

[0027] It should be noted that the scanning direction is different in push-broom and swing-broom cameras, but it can be recorded as a single direction angle value.

[0028] The exposure integration time is set by the satellite load control system, which is basically constant in fixed imaging mode, but there is dynamic change in wide imaging or high-speed imaging conditions. This parameter is used to describe the cumulative time of the camera during single-pixel exposure, which is a linear factor for blur length calculation.

[0029] Specifically, the step of constructing the reference motion vector based on the motion data is: S11, through the satellite orbit attitude angle, the ground projection speed, the imaging scanning direction and the exposure integration time, the parameter recording process is performed, and the physical imaging parameter record is generated; In one preferred embodiment, in order to ensure the parameter consistency in the subsequent motion blur calculation process, the above four types of physical imaging parameters are generated according to the preset parameter field structure table to generate the physical imaging parameter record, and the parameter field structure table is shown in Table 1: Table 1: parameter field structure table

[0030] Wherein, the attitude angle vector adopts a three-dimensional column vector form, and is expressed as: ; In the formula, , , respectively represent the roll angle, the pitch angle and the yaw angle of the satellite at the collection moment; The generated physical imaging parameter record is expressed as: ; Exemplarily, assuming that the parameters collected at an imaging moment are: Satellite orbit attitude angle: ; ground projection speed: ; imaging scanning direction: ; exposure integration time: ; the generated physical imaging parameter record can be expressed as: ; It should be noted that: in order to avoid abnormal values caused by attitude jump or star sensor error from entering the degradation kernel calculation process, the following method is adopted for parameter validity verification in the embodiment: performing sliding window mean comparison on the set of satellite orbit attitude angles, and eliminating instantaneous spikes, in addition, the scanning direction and the exposure time are verified by running log interpolation, to ensure consistency, and the data after verification is taken as the final physical imaging parameter record.

[0031] S12, based on the ground projection speed and the exposure integration time in the physical imaging parameter record, performing multiplication calculation to generate a long axis reference value; Converting the exposure integration time into units to generate a converted exposure integration time , is the exposure integration time without unit conversion; By performing multiplication operation on the ground projection speed and the converted exposure integration time in the physical imaging parameter record, a long axis reference value representing the motion blur scale along the imaging scanning direction in the imaging process can be obtained, and is expressed as: ; In the formula, is the long axis reference value, is the ground projection speed; It should be noted that: the long axis reference value is used for subsequent construction of the long axis scale of the elliptical Gaussian degradation convolution kernel.

[0032] S13, perform angle resolving processing on the imaging scan direction in the physical imaging parameter record to generate a reference blur direction angle; Specifically, the logic for performing angle resolving processing is as follows: a1, perform spatial direction decomposition processing on the imaging scan direction to generate a direction decomposition quantity; The direction decomposition quantity includes a horizontal component and a vertical component; It should be understood that the imaging scan direction is a planar angle quantity, and the direction decomposition quantity used to describe the component structure of the scan direction can be generated by decomposition, which is expressed as: wherein, is the horizontal component of the imaging scan direction, is the vertical component of the imaging scan direction, is the imaging scan direction.

[0033] a2, perform angle offset estimation calculation on the imaging scan direction to generate an angle offset quantity; is expressed as: wherein, is the angle offset quantity, is an offset estimation coefficient, is the vertical component of the imaging scan direction, is an arctangent function; It should be noted that the angle offset quantity reflects the offset degree of the vertical component in the imaging scan direction to the blur direction, and the offset estimation coefficient is a direction stability feature obtained by statistically analyzing the variance of the angle change sequence of the imaging scan direction of the real satellite image, and preferably, The value range of is [0.01, 0.10].

[0034] a3, perform direction noise suppression calculation on the imaging scan direction to generate a noise suppression quantity; is expressed as: wherein, is the noise suppression quantity, is a noise suppression coefficient, is the horizontal component of the imaging scan direction; It should be noted that the noise suppression coefficient is obtained by statistically analyzing the edge gradient noise energy proportion of the real satellite image, and preferably, the noise suppression coefficient The value range of is [0.20, 0.35].

[0035] a4, perform angle fusion processing based on the direction decomposition quantity, the angle offset quantity, and the noise suppression quantity to generate a reference blur direction angle; ​​​is expressed as: ; In the formula, is a reference blur direction angle, is an arctangent function; It should be noted that the reference blur direction angle Comprehensive imaging scanning direction of the geometric component, offset and noise suppression factor, used to describe the imaging blur direction.

[0036] S14, according to the long axis reference value and reference blur direction angle to perform vector construction operation, generate reference motion vector; Based on the obtained long axis reference value And the reference blur direction angle , the reference motion vector is constructed; is expressed as: ; In the formula, is a reference motion vector; It should be noted that the reference motion vector is used to represent the blur direction and blur amplitude caused by imaging motion.

[0037] S2, based on the statistical characteristics of the real satellite image to construct the constraint space, and from the constraint space to collect random degradation parameters, the random degradation parameters include long axis disturbance term, direction angle jitter term and short axis standard deviation; In order to make the degradation simulation process can reflect the blur distribution pattern in the real imaging environment, through the statistical characteristics of the real satellite image to perform segmentation, classification and space layering construction, get the constraint space for sampling random degradation parameters; The constraint space is a multi-dimensional statistical structure, including gradient section data, disturbance category data and diffusion interval data; Based on the constraint space according to the set sampling strategy to collect random degradation parameters that meet the physical consistency, used for subsequent degradation kernel generation processing.

[0038] Specifically, the statistical characteristics include the edge gradient histogram of the real satellite image, the atmospheric disturbance standard deviation statistics and the point spread function half width value; The edge gradient histogram is based on the use of linear array remote sensing camera to collect real satellite image, through the extraction of edge gradient value in fixed size window and statistical generation; The atmospheric disturbance standard deviation statistics is based on the continuous sampling data of cross scanning direction of spaceborne multispectral radiometer, by calculating the fluctuation value of adjacent pixel brightness difference; The point spread function half width value is based on the measurement results of the imaging load optical calibration device in the laboratory calibration process of point light source imaging response.

[0039] Specifically, the steps for constructing a constrained space based on the statistical features of real satellite imagery include: S21, perform segment gradient distribution determination based on edge gradient histogram, and generate gradient segment data; In real satellite imagery, edge gradient histograms can reflect the changes in texture intensity in different regions. In order to transform the gradient distribution into discrete segments suitable for statistical sampling, the edge gradient histogram is segmented to generate gradient segment data.

[0040] Specifically, the steps for generating gradient segment data are as follows: S211, perform amplitude density analysis on each gradient amplitude segment based on the edge gradient histogram to generate density abrupt change segment records; Specifically, the logic for performing amplitude density analysis is as follows: b1, based on the amplitude segment frequency change rate of the edge gradient histogram, perform segment gradient change rate analysis to obtain change rate fluctuation data; The frequency of each gradient amplitude segment is counted, and the rate of change of frequency between adjacent amplitude segments is calculated, expressed as: ; In the formula, The edge gradient histogram of the th Frequency of each amplitude range The edge gradient histogram of the th Frequency of each amplitude range For the first The rate of change of frequency in the amplitude range; Based on multiple groups Construct a set of frequency change rates This data is then output as fluctuation data of the rate of change.

[0041] It should be noted that the frequency of a certain gradient amplitude range is the number of pixels whose pixel gradient falls within that amplitude range.

[0042] b2, based on the number of consecutive reversals of adjacent rate of change signs in the rate of change fluctuation data, perform a stability determination of the amplitude segment to obtain a set of unstable amplitude segments; In a specific embodiment, when performing gradient change stability determination based on the number of sign reversals of the rate of change, it is necessary to retrieve a preset threshold for the number of consecutive reversals. When the number of consecutive reversals is greater than or equal to the threshold, it is marked as an unstable amplitude segment, and an unstable amplitude segment set is constructed based on multiple unstable amplitude segments. It should be noted that: the continuous inversion number threshold is set according to the statistical analysis result of the real satellite image, the sign set of the change rate of each amplitude section is counted in the edge gradient histogram of a large number of samples, the occurrence number of continuous sign inversion is recorded, and the continuous inversion number threshold with a higher occurrence frequency is extracted as the value range of the threshold; the continuous inversion number threshold obtained by this method is a small positive integer; Preferably, the continuous inversion number threshold value range can be between 1 and 3.

[0043] b3, by executing the mutation section screening processing on the unstable amplitude section set and the frequency gradient difference value of the corresponding amplitude section in the edge gradient histogram, generating a density mutation section record; In the formula, is the frequency gradient difference value of the amplitude section, is the frequency of adjacent amplitude sections; When a certain amplitude section belongs to the unstable amplitude section set at the same time, and the frequency gradient difference value reaches the preset frequency gradient difference value, the amplitude section is marked as a density mutation section, and written into the density mutation section record.

[0044] Preferably, the preset frequency gradient difference value is 1% of the total number of histogram pixels; S212, performing amplitude section boundary determination according to the density mutation section record to obtain an amplitude boundary set; In this step, in order to divide the continuous gradient amplitude space into discrete sections with statistical representation, it is necessary to perform amplitude section boundary determination according to the generated density mutation section record; Specifically, each density mutation section contained in the density mutation section record represents the most significant boundary position of the gradient change in the gradient histogram, and the amplitude section boundary can be determined accordingly. The amplitude boundary set is obtained based on a plurality of amplitude section boundaries; Exemplarily, after performing amplitude density analysis on the edge gradient histogram, it is assumed that the obtained density mutation section record contains three amplitude sections: It should be noted that: the amplitude axis of the edge gradient histogram is divided into a plurality of continuous amplitude sections, 4 can be understood as the 4th amplitude section, and the section width is set to 0.05, then: The amplitude range of the 4th section is about [0.15, 0.20], and the amplitude section boundary is taken as 0.20; The amplitude range of the 9th section is about [0.40, 0.45], and the amplitude section boundary is taken as 0.45; The amplitude range of the 12th section is about [0.55, 0.60], and the amplitude section boundary is taken as 0.60;​​​​​ The obtained amplitude section boundaries are sorted in ascending order to generate an amplitude boundary set, denoted as: ; Since real satellite images have strong complexity in edge structure and texture distribution, the boundary determination method based on abrupt section can effectively depict the structural change of gradient distribution, so that the generated gradient section data is more consistent with the edge response characteristics of real images.

[0045] S213, performing section division processing on the amplitude boundary set and the amplitude distribution of the edge gradient histogram to generate gradient section data; For each divided amplitude section, based on the frequency of each amplitude section in the corresponding edge gradient histogram, the proportion of the frequency of each amplitude section to the total frequency of the amplitude section is calculated, and these proportion values are sorted in ascending order of amplitude to form a structured set, forming an amplitude section density structure. The amplitude section density structure is used to represent the distribution pattern of the gradient density inside the amplitude section.

[0046] Further, the amplitude range, frequency and corresponding amplitude section density structure of each amplitude section are combined into a structured record as a component unit of the gradient section data, and finally the gradient section data is generated; It should be noted that the amplitude range is illustrated in step S212 and is not described in detail.

[0047] S22, performing disturbance amplitude classification processing according to the atmospheric disturbance standard deviation statistic to generate disturbance category data; Real satellite images are easily affected by atmospheric disturbances during imaging, and this disturbance will be manifested as random fluctuations of different intensities in the image edge structure and brightness changes; Based on the atmospheric disturbance standard deviation statistic obtained by a large number of real images, the disturbance amplitude can be classified to construct different disturbance levels.

[0048] In specific implementation, the atmospheric disturbance standard deviation statistic is subjected to disturbance amplitude classification processing according to a preset classification rule, and the disturbance standard deviation value is assigned to a number of disturbance level intervals to obtain disturbance category data. The disturbance category data describes the disturbance amplitude range corresponding to different atmospheric disturbance intensity intervals.

[0049] S23, performing diffusion range interval division based on the point spread function half-width value to generate diffusion interval data; The half-width value of the point spread function (PSF) can describe the diffusion range when the optical system is imaged, and is an important physical index of imaging clarity and blurring degree; In order to make the simulated degradation convolution kernel be able to map the diffusion characteristics of the real imaging system, interval division of the diffusion range needs to be performed based on the half-width value. In the implementation process, the statistical point spread function half-width values are partitioned according to numerical values, divided into multiple diffusion range intervals, and diffusion interval data is formed; The diffusion interval data describes the diffusion scale that the optical system may produce under different imaging conditions, and provides sampling basis for the short-axis standard deviation in the random degradation parameter.

[0050] The diffusion interval data established in this way can ensure that the short-axis diffusion characteristics in the degradation simulation are consistent with the performance of the real satellite image in the optical diffusion layer S24, performing spatial domain layering processing through the gradient section data, the disturbance category data and the diffusion interval data, and constructing a constraint space; In one specific embodiment, the gradient structure described by the gradient section data is the first spatial dimension, the atmospheric disturbance level described by the disturbance category data is the second spatial dimension, and the diffusion scale interval described by the diffusion interval data is the third spatial dimension; According to the combination of the three types of data in three spatial dimensions, the combination rule of "gradient section data x disturbance category data x diffusion interval data" is used to construct a multi-dimensional space with gradient-disturbance-diffusion triple constraints, and the multi-dimensional space is output as a constraint space; Based on the constraint space, the degradation parameters of each spatial dimension can be collected as random degradation parameters, and the collection logic is: The preset long-axis disturbance item is extracted from the first spatial dimension of the constraint space, the preset direction angle jitter item is extracted from the second spatial dimension of the constraint space, and the preset short-axis standard deviation is extracted from the third spatial dimension of the constraint space.

[0051] It should be noted that: the long-axis disturbance item is obtained by statistically analyzing the edge spread width along the imaging motion direction in the real satellite image. Specifically, the edges of a large number of real satellite images are scanned row by row, and the gradient diffusion width along the imaging direction is measured, so that the long-axis disturbance item directly used for degradation simulation can be obtained, and the unit is m; The direction angle jitter item is obtained by statistically analyzing the main direction angle fluctuation of the local structure in the real satellite image. Specifically, the structure direction of multiple local regions of the image is analyzed, the main direction angle of each region and the change amplitude relative to the adjacent region are measured, so that the angle fluctuation representing the direction instability is obtained, and the unit is °; The generating step of the short-axis standard deviation is specifically: selecting a region with uniform texture and gentle brightness change in a plurality of real satellite images, using a point source or an approximate point-like spot commonly used in imaging instrument calibration, recording the corresponding horizontal width when the brightness of the point-like feature in the horizontal direction is reduced to half of the peak value by analyzing the brightness diffusion form of the point-like feature in the horizontal direction. The horizontal width values obtained from a plurality of images are sorted to form a set of horizontal diffusion width data that can be directly used for degradation simulation. The horizontal diffusion width data is calculated by standard deviation, and the generated result is the short-axis standard deviation.

[0052] Exemplarily, it is assumed that the first dimension of the space, i.e., the gradient section data, contains three gradient sections, and the gradient section table is as shown in Table 2. Table 2: Gradient section table

[0053] It is assumed that the second dimension of the space, i.e., the disturbance category data, contains three disturbance levels, and the disturbance level table is as shown in Table 3. Table 3: Disturbance level table

[0054] It is assumed that the third dimension of the space, i.e., the diffusion interval data, contains three diffusion scale intervals divided based on the half-width value of the point spread function, and the diffusion scale interval table is as shown in Table 4. Table 4: Diffusion scale interval table

[0055] Through the combination rule of "gradient section data x disturbance category data x diffusion interval data", the following can be generated: space level units Each "unit" corresponds to a blurred state that may occur in the real world. For example, a unit can represent: gradient section G3, disturbance level T2, and diffusion scale E1. It should be noted that the specific space level units are not only 27, but can be more than 27 or less than 27 according to different dimension levels. The above content is only for example to facilitate understanding It is assumed that the random falls in the space level unit (G3, T2, E1) of the constraint space. Collecting random degradation parameters, i.e., extracting the long-axis disturbance item from the first dimension of the constraint space G3, extracting the direction angle jitter item from the second dimension of the constraint space T2, and extracting the short-axis standard deviation from the third dimension of the constraint space E1.

[0056] S3, fusing the reference motion vector and the random degradation parameter to generate a composite elliptical Gaussian degradation convolution kernel. Specifically, the generation logic of the composite elliptical Gaussian degradation convolution kernel is as follows: S31, performing amplitude superposition processing on the long axis reference value and the long axis disturbance term to generate a long axis fusion value; The generation formula of the long axis fusion value is as follows: In the formula, is the unconstrained long axis fusion value, is the long axis disturbance term; It should be noted that the amplitude superposition processing reflects the joint action of the physical motion blur scale and the statistical blur scale in actual imaging The lower limit constraint processing is performed on the unconstrained long axis fusion value to obtain the long axis fusion value; is expressed as: is the long axis fusion value, is the minimum acceptable long axis scale set according to the real image blur width statistics, to prevent the degradation kernel from degrading to a very small scale, is selected and as the long axis fusion value for output.

[0057] S32, performing angle correction processing based on the reference blur direction angle and the direction angle jitter term to obtain a direction fusion angle value; In the formula, is the non-normalized direction fusion angle value, is the direction angle jitter term; It should be noted that the superposition of this step is used to simulate the random disturbance of the scanning direction angle in the real imaging process; The normalization processing is performed on the non-normalized direction fusion angle value to generate the direction fusion angle value; is expressed as: In the formula, is the direction fusion angle value, and the value range is [0, ), is the angle normalization processing performed on the non-normalized direction fusion angle value so that the output is within the interval [0, ).

[0058] S33, performing axis parameter updating according to the short axis standard deviation and the correlation coefficient field to generate a short axis update value; is expressed as: In the formula, is the unconstrained short axis update value,​​​​​ is a short axis standard deviation, is a correlation coefficient field, used to describe the correlation degree between the major axis direction and the minor axis direction of the elliptical Gaussian kernel; It should be noted that the axial parameter update of this step reflects the change trend of the short axis scale under the joint action of atmospheric diffusion structure and optical correlation; correlation coefficient field obtained by performing covariance statistical processing on the two-dimensional local blur shape of the typical edge region in the real satellite image; Specifically, based on the edge region in the high-resolution satellite image which has uniform texture and stable contour structure, the local gray gradient field of the region is extracted, and the horizontal component and the vertical component of the gradient field in the horizontal direction and the vertical direction are calculated to generate a local gradient covariance matrix. Based on the local gradient covariance in the matrix and the variance of each direction corresponding to the local gradient covariance, the direction coupling degree is calculated to serve as the correlation coefficient field. is expressed as: ; In the formula, is a horizontal component of the local gradient, is a vertical component of the local gradient, is a local gradient covariance of the horizontal component and the vertical component, , is a variance of the horizontal component and the vertical component; unconstrained short axis update value performing lower limit constraint to generate a short axis update value; is expressed as: ; In the formula, is a short axis update value, is a minimum acceptable short axis scale set according to the diffusion width statistics of the real image, used to avoid the short axis scale from being reduced to a physically unreasonable minimum value.

[0059] S34, performing convolution kernel parameter collection processing on the long axis fusion value, the direction fusion angle value and the short axis update value to generate a composite elliptical Gaussian degradation convolution kernel; is expressed as: ; is a composite elliptical Gaussian degradation convolution, indicates that the collection processing of constructing an elliptical Gaussian function is performed with the long axis fusion value, the direction fusion angle value and the short axis update value as inputs.

[0060] It should be noted that before the long axis fusion value, the direction fusion angle value and the short axis update value enter the composite elliptical Gaussian degradation convolution kernel for updating, they all need to be converted based on the imaging ground resolution to be expressed in pixel scale in the convolution kernel. The conversion logic is: Ground Sampling Distance, GSD, of satellite imagery, all parameters in meters Divide by GSD to convert to pixel units.

[0061] S4, performing degradation simulation operation on the high-definition aerial image by the composite elliptical Gaussian degradation convolution kernel to generate pseudo-satellite image data, and pixel-level aligning the pseudo-satellite image data with the corresponding high-definition aerial image to construct a set of training sample pairs; Based on the composite elliptical Gaussian degradation convolution kernel obtained in step S3, performing degradation simulation operation on the high-definition aerial image as the true value to generate pseudo-satellite image data with real blur morphology, and constructing training sample pairs that can be used for deep learning network training by performing spatial coordinate corresponding analysis and pixel-by-pixel pairing processing on the pseudo-satellite image data and the high-definition aerial image.

[0062] Specifically, the logic for constructing the training sample pairs is: c1, generating pseudo-satellite image data by performing convolution processing on the high-definition aerial image through the composite elliptical Gaussian degradation convolution kernel; Obtaining the high-definition aerial image after resolution unification, denoted as: ; Wherein, and are discrete pixel coordinate indexes, and the unit is pixel, in order to ensure that the degradation simulation process is computable on the pixel scale, if there is a difference between the spatial resolution of the original high-definition aerial image and the ground resolution of the target satellite image, then before entering this step, perform resampling processing on the high-definition aerial image to make its pixel size consistent with the pixel size of the target satellite image, which can use bilinear interpolation or cubic interpolation method, and will not be described in detail; Denote the composite elliptical Gaussian degradation convolution kernel obtained in step S3 as: ; Wherein, is the discrete pixel displacement index inside the convolution kernel, and the size of the convolution kernel is , and are positive integers; Performing two-dimensional discrete convolution operation on the high-definition aerial image and the composite elliptical Gaussian degradation convolution kernel to generate pseudo-satellite image data, denoted as: ; In the formula, is the gray value of the pseudo-satellite image data at the pixel coordinate ; It should be noted that: The pixel coordinate To maintain spatial consistency between pseudosatellite imagery data and high-resolution aerial imagery, this embodiment employs a mirror-fill strategy for image edges during convolution, i.e., when... When the image exceeds the original image range, the pixels at the boundary are mirrored about the boundary, and the expanded pixel values ​​are used as the input to the convolution, thereby avoiding the edge brightness decay caused by zero padding. After the convolution operation is completed, pseudosatellite image data with the same size as the high-resolution aerial image is obtained. The pseudo-satellite image data maintains the same spatial resolution as the satellite image, while its blurring morphology is controlled by a composite elliptical Gaussian degenerate convolution kernel to simulate motion blur and optical diffusion in the real satellite imaging process.

[0063] c2 performs spatial coordinate correspondence analysis on pseudosatellite image data and high-resolution aerial image data to obtain alignment reference data; In one specific embodiment, in order to build reliable training sample pairs at the pixel level, it is necessary to establish a stable spatial coordinate correspondence between pseudosatellite image data and high-resolution aerial imagery. Since pseudosatellite imagery data is obtained by convolutional degradation of high-resolution aerial imagery, the two should ideally maintain a one-to-one correspondence on the pixel grid. However, in engineering implementation, there may be slight pixel-level offsets due to resampling, cropping, or external preprocessing. Therefore, this step generates alignment reference data through spatial coordinate correspondence analysis to describe the overall or local pixel displacement between the two images; Specifically, the steps for performing spatial coordinate correspondence analysis are as follows: c21, calculates the coordinate difference based on the corresponding pixel coordinates of pseudosatellite imagery data and high-resolution aerial imagery, and generates difference segment data; In pseudosatellite imagery and high-resolution aerial imagery, a set of sampling regions with stable texture structures is selected. These sampling regions can be divided according to a fixed-interval grid, for example, extracting a small block of region every few pixels. The center pixel coordinates of each small block of region are denoted as follows in the high-resolution aerial imagery: ; Mark the center pixel coordinates of the corresponding small region in the pseudosatellite image data: ; in, Index for the sampling region; Ideally, if the two images have no additional geometric offset, then: ; To detect and quantify potential offsets, this embodiment calculates the coordinate difference for each sampling region, expressed as: ; In the formula, the transverse difference value of the first sampling region, the longitudinal difference value of the first sampling region; the transverse difference value of the first sampling region, the longitudinal difference value of the first sampling region; the coordinate difference values of all the sampling regions are recorded according to the region index to form difference segment data, denoted as: In the formula, is the difference segment data, is the total number of sampling regions, and the difference segment data is used to describe the pixel displacement of the pseudo-satellite image data relative to the high-definition aerial image at different spatial positions.

[0064] c22, performing local consistency screening on the difference segment data to obtain candidate alignment unit data; Based on the fact that some of the sampling regions may be located in a low-texture region of the image or exist noise interference, directly using all the difference segment data for displacement estimation may lead to a large error bias. Therefore, the difference segment data is subjected to local consistency screening in this step to eliminate unstable regions and retain regions with consistent displacement characteristics to generate candidate alignment unit data.

[0065] Specifically, a local analysis window centered on the spatial position of the sampling region is constructed, and a plurality of spatially adjacent sampling regions are grouped into the same local window. For the difference segment data in each local window, a statistical quantity is calculated and denoted as: In the formula, , is the average displacement amount in the horizontal direction and the vertical direction within the window, , is the dispersion degree in the horizontal direction and the vertical direction, is the number of sampling regions within the window; When the and in a certain local window are both less than the dispersion threshold value preset based on historical experimental data, it is considered that the coordinate difference values in the local window have good consistency. The local window is recorded as a stable window, and the displacement statistical result corresponding to the local window is recorded as a candidate alignment unit. On the contrary, the local window is regarded as an unstable region and is not recorded as a candidate alignment unit. ​​​​​​​

[0066] The average displacement information corresponding to all local windows satisfying the consistency condition is recorded as: ; In the formula, is the candidate alignment unit data, is the stable window set screened by local consistency c23, performing coordinate mapping confirmation processing according to the candidate alignment unit data, generating alignment reference data; In one specific embodiment, in order to establish a unified spatial alignment relationship between the entire pseudo-satellite image data and the high-definition aerial image, a representative global alignment displacement is calculated based on the candidate alignment unit data, and the displacement is output as the alignment reference data; Specifically, the average displacement of each candidate alignment unit data set is weighted and counted to obtain the global alignment displacement , and the calculation formula is: ; ; ; In the formula, is the weight of the local window , which is set based on the gradient strength in the local window to improve the contribution of the texture-rich area in displacement estimation; Specifically, the horizontal gradient strength and the vertical gradient strength in the local window are extracted to calculate the average gradient strength of the local window , which is represented as: ; In the formula, is the average gradient strength, , is the horizontal gradient strength and the vertical gradient strength of the pixel point , and is the number of sampling regions; The average gradient strength is normalized, and the normalized result is output as the weight of the local window , and the calculation formula is: ; The global alignment displacement is output as the alignment reference data, which is used to describe the overall translation relationship of the pseudo-satellite image data on the pixel grid relative to the high-definition aerial image.

[0067] c3, based on the alignment reference data, performing pixel-by-pixel pairing processing to generate a set of training sample pairs; In one specific embodiment, the alignment reference data is given according to alignment reference data , the pixel coordinates of the pseudo-satellite image data are translated and corrected, and a one-to-one correspondence is established between the corrected pseudo-satellite image data and the high-definition aerial image at the pixel coordinate level; Specifically, for each pixel coordinate in the pseudo-satellite image data , the corresponding pixel coordinate in the high-definition aerial image is calculated according to the alignment reference data , which is expressed as: ; It should be noted that when the calculated is not an integer, the pixel value at the corresponding position in the high-definition aerial image is obtained by a bilinear interpolation method, denoted as: ; The pixel value of the pseudo-satellite image data at is combined into a pair of training sample records, and all pixel coordinates in the full image range that meet the boundary conditions are paired and processed as described above, and a set of training sample pairs is constructed, denoted as: ; In the formula, is the set of training sample pairs, is the set of pixel coordinates that are still in the valid area of the image after the alignment displacement transformation; S5, based on the set of training sample pairs, the parameter iterative training of the deep learning network is performed to generate a satellite image sharpening processing model; In one specific embodiment, the training sample pairs generated by the above method are used to train the model by taking the pseudo-satellite image data as the input image and the high-definition aerial image as the target image to generate a satellite image sharpening processing model; Specifically, the steps of generating a satellite image sharpening processing model are: S51, the training sample pairs are divided into a sharpening processing training set and a sharpening processing verification set, and the training sample pairs include pseudo-satellite image data and corresponding high-definition aerial images; S52, a deep learning network is constructed, the pseudo-satellite image data in the sharpening processing training set is taken as the input of the deep learning network, the corresponding high-definition aerial image is taken as the output of the deep learning network, the parameter iterative training of the deep learning network is performed, and an initial satellite image sharpening processing network is obtained; It should be noted that the parameter iterative training includes but is not limited to forward calculation, difference measurement calculation, and gradient update processing, and the deep learning network includes but is not limited to an attention-enhanced network or a Transformer image reconstruction network; S53, verifying the initial satellite image sharpening network by the verification set to verify the initial satellite image sharpening network, and outputting the initial satellite image sharpening network with a verification error less than or equal to a preset matching error threshold as a pre-constructed satellite image sharpening model.

[0068] The above embodiments are only used to illustrate the technical method of the present application but not limit the present application. Although the present application is explained in detail with reference to the preferred embodiments, it should be understood by those ordinary skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A method for improving the definition of satellite images based on elliptical Gaussian distribution, characterized in that, The method comprises: S1, calculating generated motion data according to the collected physical imaging parameters, and constructing a reference motion vector based on the motion data, wherein the motion data comprises a long axis reference value and a reference blur direction angle; S2, constructing a constraint space based on the statistical characteristics of the real satellite image, and collecting random degradation parameters from the constraint space, wherein the random degradation parameters comprise a long axis disturbance term, a direction angle jitter term and a short axis standard deviation; S3, fusing the reference motion vector and the random degradation parameters to generate a composite elliptical Gaussian degradation convolution kernel; S4, performing degradation simulation operation on the high-definition aerial image through the composite elliptical Gaussian degradation convolution kernel to generate pseudo-satellite image data, and performing pixel-level alignment on the pseudo-satellite image data and the corresponding high-definition aerial image to construct a training sample pair set.

2. The method of claim 1, wherein the method is based on an elliptical Gaussian distribution. The physical imaging parameters comprise satellite orbit attitude angle, ground projection velocity, imaging scanning direction and exposure integration time; and the statistical characteristics comprise edge gradient histogram of the real satellite image, atmospheric disturbance standard deviation statistics and point spread function half-width value.

3. The method of claim 2, wherein the method is based on an elliptical Gaussian distribution. The step of constructing the reference motion vector based on the motion data comprises: S11, performing parameter recording processing through the satellite orbit attitude angle, the ground projection velocity, the imaging scanning direction and the exposure integration time to generate a physical imaging parameter record; S12, performing multiplication calculation based on the ground projection velocity and the exposure integration time in the physical imaging parameter record to generate the long axis reference value; S13, performing angle analysis processing on the imaging scanning direction in the physical imaging parameter record to generate the reference blur direction angle; S14, performing vector construction operation according to the long axis reference value and the reference blur direction angle to generate the reference motion vector.

4. The method of claim 3, wherein the method is based on an elliptical Gaussian distribution. The logic of performing the angle analysis processing comprises: a1, performing spatial direction decomposition processing through the imaging scanning direction to generate a direction decomposition quantity; a2, performing angle offset estimation calculation according to the imaging scanning direction to generate an angle offset quantity; a3, performing direction noise suppression calculation through the imaging scanning direction to generate a noise suppression quantity; a4, performing angle fusion processing based on the direction decomposition quantity, the angle offset quantity and the noise suppression quantity to generate the reference blur direction angle.

5. The method of claim 2, wherein the method is based on an elliptical Gaussian distribution. The step of constructing the constraint space based on the statistical characteristics of the real satellite image comprises: S21, performing section gradient distribution determination based on the edge gradient histogram to generate gradient section data; S22, performing disturbance amplitude classification processing according to the atmospheric disturbance standard deviation statistics to generate disturbance category data; S23, performing diffusion range interval division based on the point spread function half-width value to generate diffusion interval data; S24, performing spatial domain layering processing through the gradient section data, the disturbance category data and the diffusion interval data to construct the constraint space.

6. The method of claim 5, wherein the method is based on an elliptical Gaussian distribution. The step of generating the gradient section data comprises: S211, performing amplitude density analysis based on each gradient amplitude section of the edge gradient histogram to generate a density mutation section record; S212, performing amplitude section boundary determination according to the density mutation section record to obtain an amplitude boundary set; S213, performing section division processing through the amplitude boundary set and the amplitude distribution of the edge gradient histogram to generate the gradient section data.

7. The method of claim 6, wherein the method is based on an elliptical Gaussian distribution. The logic of performing the amplitude density analysis comprises: b1, perform segment gradient change rate analysis based on the amplitude segment frequency change rate of the edge gradient histogram to obtain change rate fluctuation data; b2, perform amplitude segment stability determination according to the number of consecutive inversions of adjacent change rate signs in the change rate fluctuation data to obtain an unstable amplitude segment set; b3, perform mutation segment screening processing by the unstable amplitude segment set and the frequency gradient difference of the corresponding amplitude segment in the edge gradient histogram to generate density mutation segment records.

8. The method of claim 1, wherein the method is based on an elliptical Gaussian distribution. The generation logic of the composite elliptical Gaussian degradation convolution kernel is: S31, perform amplitude superposition processing by the long axis reference value and the long axis perturbation term to generate a long axis fusion value; S32, perform angle correction processing based on the reference blur direction angle and the direction angle jitter term to obtain a direction fusion angle value; S33, perform axis parameter updating according to the short axis standard deviation and the correlation coefficient field to generate a short axis updated value; S34, perform convolution kernel parameter collection processing by the long axis fusion value, the direction fusion angle value, and the short axis updated value to generate a composite elliptical Gaussian degradation convolution kernel.

9. The method of claim 1, wherein the method is based on an elliptical Gaussian distribution. The logic for constructing training sample pairs is: c1, perform convolution processing by the composite elliptical Gaussian degradation convolution kernel and the high-definition aerial image to generate pseudo-satellite image data; c2, perform spatial coordinate correspondence analysis on the pseudo-satellite image data and the high-definition aerial image to obtain alignment reference data; c3, perform pixel-by-pixel pairing processing based on the alignment reference data to generate a training sample pair set.

10. The method of claim 9, wherein the method is based on an elliptical Gaussian distribution. The steps for performing spatial coordinate correspondence analysis are: c21, perform coordinate difference value calculation based on the corresponding pixel coordinates of the pseudo-satellite image data and the high-definition aerial image to generate difference value segment data; c22, perform local consistency screening by the difference value segment data to obtain candidate alignment unit data; c23, perform coordinate mapping confirmation processing according to the candidate alignment unit data to generate alignment reference data.

Images