Image processing method based on machine learning data completion algorithm

Abstract

The invention discloses an image processing method based on a machine learning data completion algorithm, and relates to the technical field of image processing. Based on a phase correlation sub-pixel registration technology, an image is aligned by using phase information of a frequency domain cross-power spectrum, and a non-rigid deformation field is constrained by combining substrate fracture distribution; the dislocation error of ultraviolet / infrared and visible light wave bands is obviously reduced; for a high-reflection area, the visible light overexposure weight is dynamically inhibited through an infrared penetrability characteristic, and characteristic conflicts of special materials such as gold foil and the like are solved; introducing an ancient pigment physicochemical characteristic database as a generation constraint, coordinating visible light texture and infrared underlying draft information in combination with a layered attention mechanism, and complementing colors conforming to mineral pigment spectral reflection characteristics in a fading area; spectral distortion caused by modern pigment differences in a traditional data driving method is avoided, so that a reconstructed color gamut strictly conforms to historical pigment chemical characteristics, and chromatic aberration deviation which can be distinguished by naked eyes is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image processing method based on a machine learning data completion algorithm. Background Technology

[0002] In the field of digital preservation of cultural relics, multispectral imaging technology has become a core tool for the analysis of ancient murals. This method can reveal underlying drafts, repair traces, and pigment composition information that are invisible to the naked eye by acquiring images in the ultraviolet, visible, and infrared bands. The current mainstream process relies on high-precision spectroscopic cameras to acquire multi-band data and combine them with registration algorithms to construct layered image sets. However, during implementation, physical damage such as cracking and peeling on the surface of the murals causes the imaging plane to become uneven, and there are non-rigid deformations between images in different bands. Existing registration methods based on feature points have difficulty extracting stable feature points when dealing with large areas of fading, while deformation field prediction schemes based on deep learning are limited by the problem of scarce samples.

[0003] In recent years, generative adversarial networks (GANs) have been widely used for color restoration in the reconstruction of faded areas. These methods establish a training set by collecting intact pigment samples from murals and learn the mapping relationship between local texture and color. However, in practice, the spectral reflectance characteristics of ancient mineral pigments are fundamentally different from those of modern chemical pigments, and the generated results are prone to spectral distortion. In addition, the personal brushstroke style of the mural artist (such as changes in brush pressure and direction) is difficult to capture effectively through data-driven methods. Some solutions attempt to introduce style transfer modules, but the lack of quantitative evaluation standards results in abrupt brushstrokes in the restored areas.

[0004] To address the challenges of multispectral data fusion, some newer solutions employ cross-modal attention mechanisms to align features across different bands. These models extract common features through a shared encoder and then separate band-specific information using a gating unit. However, when processing highly reflective regions (such as gold foil layers), the infrared and visible light bands conflict due to differences in reflectivity, causing attention weight allocation to fail. Other research has proposed physical guidance networks that use pigment spectral databases as prior knowledge to constrain the generation process. However, in practical applications, the complexity of ancient pigment chemical compositions and regional differences make it difficult to construct universally applicable constraint models. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides an image processing method based on machine learning data completion algorithm to solve the three major defects of existing methods, namely, multispectral registration misalignment, mineral pigment restoration distortion, and brushstroke style breakage, which restrict the high-quality restoration of ancient murals.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an image processing method based on a machine learning data completion algorithm, comprising: Step S1: Collect multispectral image sequences of the ancient murals, including ultraviolet, visible and infrared images; Step S2: Align the multi-band image through the phase correlation sub-pixel offset correction module. This module calculates the phase angle of the frequency domain cross power spectrum between bands and uses Gaussian surface fitting interpolation to achieve sub-pixel level registration. Step S3: Construct a mineral pigment spectral reflectance reconstruction network. Its input is a visible light band image and aligned multispectral data. The output layer is connected to an ancient pigment physicochemical property database and generates reflectance curves of the faded region through constraints. Step S4: Use a dynamic brushstroke generator to repair the damaged area. This generator generates an adversarial completion mask based on the local brushstroke direction field of the mural and applies a brushstroke trajectory continuity loss function. Step S5: Perform anisotropic feature fusion on the highly reflective areas and use the infrared band penetration features to suppress the weight of overexposed areas in the visible light band.

[0008] In a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, step S2 involves performing phase-correlation sub-pixel offset correction. First, integer displacement is derived in the frequency domain, and then sub-pixel compensation is obtained by Gaussian surface fitting to estimate the band. With band The translation amounts between them specifically include: Perform a two-dimensional Fourier transform on the input image: , , in, For the first Band spectrum, Represents the two-dimensional discrete Fourier transform operator. For the first Band spectrum, For frequency domain coordinates, and The two bands are respectively located in spatial coordinates. Pixel intensity below; Normalized cross-power spectrum writing: , in, For the normalized cross-power spectrum, For conjugate spectrum, Represents the modulus of a complex number. It serves as a stabilizing factor to suppress the denominator from approaching zero at low-energy spectral points; The stability factor is defined as: , in, This is an empirical proportionality coefficient. This indicates the entire frequency domain. The arithmetic mean of the sampling points is taken. Calculate the inverse transform to obtain , in, For spatial domain related responses, For inverse Fourier operators; Peak coordinates are Pixel displacement estimation, the formula is: , in, Integer pixel displacement estimation; Take the logarithm of the nine points in the peak neighborhood: , in, For logarithmic response, Index for relative displacement; Derivation of subpixel compensation based on the analytical expression of a bivariate quadratic surface: , , and These are the horizontal and vertical subpixel compensation amounts, respectively. but in, For the final translation estimate; according to right Perform a second B-spline resampling to obtain the registered image. , Indicates by displacement vector The resampled registered image has the same unit of intensity as the input pixel.

[0009] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the phase correlation sub-pixel offset correction module in step S2 further includes: Constructing multi-scale pyramid decomposition layers to handle non-rigid deformation; Local structural tensor constraints are introduced at each scale layer, and the smoothness of the deformation field is adjusted according to the elastic coefficient of the mural substrate material.

[0010] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the mineral pigment spectral reflectance reconstruction network comprises: The encoder employs a multi-branch residual structure to extract visible light texture features and multispectral band features respectively. The decoder injects a layered attention gate mechanism during the decoding stage, activating the pigment generation path based on the underlying draft information revealed by the infrared band.

[0011] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the dynamic brushstroke generator comprises: Pressure-speed mapping model: By analyzing the rate of change of the brush stroke width, the physical brush movement parameters during creation are deduced. Trajectory prediction unit: A dual-stream LSTM network is used to model the changes in horizontal pen stroke force and the vertical turning angle respectively.

[0012] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the calculation method of the local structure tensor constraint is as follows: Extract the microscopic crack distribution map of the mural substrate material; obtain the substrate crack distribution map based on the mural microscopic imaging, and dynamically adjust the deformation field regularization intensity according to the crack density field to reduce the deformation degree of freedom in the crack-dense area.

[0013] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the hierarchical attention gate mechanism is executed in the infrared band feature channel: Generate a heatmap of mutual information between infrared and visible light features; The interference of visible light characteristics in the gold foil-covered area is suppressed by using heatmap weights.

[0014] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the pen stroke trajectory continuity loss function includes: Curvature consistency measure: Calculates the standard deviation of the curvature change between the completed area and the original stroke; Pressure gradient constraint: Forces the pressure values ​​of adjacent stroke nodes to conform to the attenuation distribution model.

[0015] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, in step S4, the standard deviation of curvature change is defined and incorporated into the stroke trajectory continuity loss function, specifically as follows: For the original pen stroke node sequence With complete stroke node sequence Perform a spline resampling, maintaining equal arc length intervals; Original trajectory number The node discrete curvature is: , in, The angle between adjacent line segments. Denotes the Euclidean norm. The angle between adjacent line segments. The local average arc length, Original curvature; complete trajectory curvature Using the same formula, we can derive: , This indicates the completion of node curvature. To complete the angle between adjacent line segments at the nth node of the stroke, This represents the local average arc length at node 𝑖; Define the curvature difference sequence and its mean as: , in, For the two trajectories in the first... The curvature difference of the nodes, Its arithmetic mean The total number of nodes after spline resampling is dimensionless. Define the standard deviation of curvature variation as: , in, The standard deviation of the curvature difference measures the overall uniformity of bending. Construct the loss term: , ,in, For curvature uniformity loss, For adaptive weights, This is the baseline coefficient for curvature loss. The original curvature average, This represents the original maximum curvature.

[0016] As a preferred embodiment of the image processing method based on machine learning data completion algorithm described in this invention, the anisotropic feature fusion further includes: A penetrating mask for infrared band feature maps is constructed, which is dynamically adjusted based on pixel-level reflection intensity. A weighted summation operation is used in the fusion layer, where the weighting coefficient of the overexposed region in the visible light band is inversely proportional to the infrared mask value; After fusion, edge-preserving filtering is performed to smooth feature transitions.

[0017] The beneficial effects of this invention are as follows: This invention breaks through the bottleneck of multispectral data fusion in the restoration of ancient murals by deeply coupling multispectral physical characteristics with machine learning generation: Based on phase-correlated subpixel registration technology, it uses the phase information of frequency domain cross power spectrum to align images, and combines the distribution of substrate cracks to constrain the non-rigid deformation field, which significantly reduces the misalignment error between ultraviolet / infrared and visible light bands; for highly reflective areas, it dynamically suppresses the overexposure weight of visible light through infrared penetrability characteristics, and resolves the feature conflicts of special materials such as gold foil; this design enables high-precision fusion of multispectral data in spatial and feature dimensions, providing reliable input for reconstruction.

[0018] This invention introduces an ancient pigment physicochemical properties database as a generation constraint, and combines a layered attention mechanism to coordinate visible light texture and infrared underlying draft information to fill in colors in faded areas that conform to the spectral reflectance characteristics of mineral pigments; it avoids the spectral distortion caused by differences in modern pigments in traditional data-driven methods, so that the reconstructed color gamut strictly conforms to the chemical characteristics of historical pigments, and significantly reduces color deviations that can be perceived by the naked eye. Furthermore, the brushstroke patterns of ancient artisans were modeled through dynamic inversion: pressure-velocity parameters were inferred from the rate of change of brushstroke width, and the brushstroke trajectory was predicted using a dual-stream LSTM, while applying constraints on curvature consistency and pressure gradient; this ensured that the completed brushstrokes maintained stylistic consistency with the original murals in details such as bending angle and pressure attenuation, avoiding the abrupt brushstroke problems caused by mechanical generation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the image processing method based on the machine learning data completion algorithm in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figure 1 This embodiment provides an image processing method based on a machine learning data completion algorithm, including the following steps: Step S1: Collect multispectral image sequences of the ancient murals, including ultraviolet, visible and infrared images; Step S2: Align the multi-band image through the phase correlation sub-pixel offset correction module. This module calculates the phase angle of the frequency domain cross power spectrum between bands and uses Gaussian surface fitting interpolation to achieve sub-pixel level registration. The phase-correlation subpixel offset correction module in step S2 further includes: Constructing multi-scale pyramid decomposition layers to handle non-rigid deformation; Local structural tensor constraints are introduced at each scale layer, and the smoothness of the deformation field is adjusted according to the elastic coefficient of the mural substrate material. The calculation method for local structure tensor constraints is as follows: Microscopic crack distribution map of mural substrate material; Based on the microscopic imaging of murals, the distribution map of substrate cracks is obtained, and the deformation field regularization intensity is dynamically adjusted according to the crack density field to reduce the deformation degree of freedom in the crack-dense area. In step S2, phase-correlation subpixel offset correction is performed. First, the integer displacement is derived in the frequency domain, and then subpixel compensation is obtained by fitting a Gaussian surface to estimate the band. With band The translation amounts between them specifically include: Perform a two-dimensional Fourier transform on the input image: , , in, For the first Band spectrum, Represents the two-dimensional discrete Fourier transform operator. For the first Band spectrum, For frequency domain coordinates, and The two bands are respectively located in spatial coordinates. Pixel intensity below; Normalized cross-power spectrum writing: , in, For the normalized cross-power spectrum, For conjugate spectrum, Represents the modulus of a complex number. It serves as a stabilizing factor to suppress the denominator from approaching zero at low-energy spectral points; The stability factor is defined as: , in, The empirical proportionality coefficient (taken as follows) ), This indicates the entire frequency domain. The arithmetic mean of the sampling points is taken. Calculate the inverse transform to obtain , in, For spatial domain related responses, For inverse Fourier operators; Peak coordinates are Pixel displacement estimation, the formula is: , in, Integer pixel displacement estimation; Take the logarithm of the nine points in the peak neighborhood: , in, For logarithmic response, Index for relative displacement; Derivation of subpixel compensation based on the analytical expression of a bivariate quadratic surface: , , and These are the horizontal and vertical subpixel compensation amounts, respectively. but in, For the final translation estimate; according to right Perform a second B-spline resampling to obtain the registered image. , Indicates by displacement vector The resampled registered image has the same unit of intensity as the input pixel; Specifically, frequency domain normalization reduces amplitude interference caused by brightness differences in each band, and the stability factor is adaptively adjusted to eliminate low-energy noise amplification; the phase response is a single-peak pulse in the spatial domain, and the Gaussian surface is linearly solvable in the logarithmic domain, directly providing sub-pixel fine displacement and avoiding uncertain convergence of iterative algorithms; the two-level estimation of integers and sub-pixels achieves a balance between computational load and precision, and can maintain high alignment quality for slight non-rigid stretching of ancient murals in multiple bands; Step S3: Construct a mineral pigment spectral reflectance reconstruction network. Its input is a visible light band image and aligned multispectral data. The output layer is connected to an ancient pigment physicochemical property database and generates reflectance curves of the faded region through constraints. The mineral pigment spectral reflectance reconstruction network includes: The encoder employs a multi-branch residual structure to extract visible light texture features and multispectral band features respectively. The decoder injects a layered attention gate mechanism during the decoding stage, which activates the pigment generation path based on the underlying draft information revealed by the infrared band. The hierarchical attention gate mechanism is executed during the decoder phase: First, multi-scale pooling is performed on the infrared band feature map to generate a feature importance weight map; Secondly, the weight map and the visible light feature map are multiplied channel by channel by the gating unit to dynamically suppress the visible light feature interference in highly reflective areas; Finally, a normalization operation is applied before outputting the fused feature map to ensure that the feature values ​​are within a reasonable range. The hierarchical attention gate mechanism is executed in the infrared band feature channel: Generate a heatmap of mutual information between infrared and visible light features; Suppressing visible light feature interference in the gold foil-covered area by using heatmap weighting; Step S4: Use a dynamic brushstroke generator to repair the damaged area. This generator generates an adversarial completion mask based on the local brushstroke direction field of the mural and applies a brushstroke trajectory continuity loss function. The dynamic stroke generator includes: Pressure-speed mapping model: By analyzing the rate of change of the brush stroke width, the physical brush movement parameters during creation are deduced. Trajectory prediction unit: A dual-stream LSTM network is used to model the changes in horizontal pen stroke force and the vertical turning angle respectively; The pen stroke trajectory continuity loss function includes: Curvature consistency measure: Calculates the standard deviation of the curvature change between the completed area and the original stroke; Pressure gradient constraint: forces the pressure values ​​of adjacent stroke nodes to conform to the attenuation distribution model; The pen stroke trajectory continuity loss function further includes: Velocity-acceleration coupling term: Based on the pen stroke parameters output by the dual-stream LSTM network, calculate the velocity change rate and acceleration matching degree between adjacent nodes; Directional consistency constraint: Through directional field gradient analysis, ensure that the deviation of the tangent angle between the completed stroke and the original stroke at the turning point does not exceed the set threshold; In step S4, the standard deviation of curvature change is defined and incorporated into the stroke trajectory continuity loss function to measure the difference in geometric curvature between the completed trajectory and the original stroke. Specifically: For the original pen stroke node sequence With complete stroke node sequence Perform a spline resampling, maintaining equal arc length intervals; Original trajectory number The node discrete curvature is: , in, The angle between adjacent line segments. Denotes the Euclidean norm. The angle between adjacent line segments. The local average arc length, Original curvature; complete trajectory curvature Using the same formula, we can derive: , This indicates the completion of node curvature. To complete the angle between adjacent line segments at the nth node of the stroke, This represents the local average arc length at node 𝑖; Define the curvature difference sequence and its mean as: , in, For the two trajectories in the first... The curvature difference of the nodes, Its arithmetic mean The total number of nodes after spline resampling is dimensionless. Define the standard deviation of curvature variation as: , in, The standard deviation of the curvature difference measures the overall uniformity of bending. Construct the loss term: , ,in, For curvature uniformity loss, For adaptive weights, This is the baseline coefficient for curvature loss (empirical value 0.8). The original curvature average, The original maximum curvature; Specifically, curvature, composed of node angles and local arc lengths, reflects the degree of bending of the brushstroke at a microscopic level. The variance of the difference between the curvature and the original curvature measures the geometric consistency of the completion process, effectively suppressing harsh lines and excessive smoothing; adaptive weighting... The constraint of complex curve segments is dynamically amplified based on the original curvature distribution of the mural, making the loss more sensitive to the high bending areas of the brushstrokes; Step S5: Perform anisotropic feature fusion on the highly reflective areas and use the infrared band penetration features to suppress the weight of overexposed areas in the visible light band. Anisotropic feature fusion further includes: A penetrating mask for infrared band feature maps is constructed, which is dynamically adjusted based on pixel-level reflection intensity. A weighted summation operation is used in the fusion layer, where the weighting coefficient of the overexposed region in the visible light band is inversely proportional to the infrared mask value; After fusion, edge-preserving filtering is performed to smooth feature transitions.

[0025] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An image processing method based on a machine learning data completion algorithm, characterized in that, include, Step S1: Collect multispectral image sequences of the ancient murals, including ultraviolet, visible and infrared images; Step S2: Align the multi-band image through the phase correlation sub-pixel offset correction module. This module calculates the phase angle of the frequency domain cross power spectrum between bands and uses Gaussian surface fitting interpolation to achieve sub-pixel level registration. Step S3: Construct a mineral pigment spectral reflectance reconstruction network. Its input is a visible light band image and aligned multispectral data. The output layer is connected to an ancient pigment physicochemical property database and generates reflectance curves of the faded region through constraints. Step S4: Use a dynamic brushstroke generator to repair the damaged area. This generator generates an adversarial completion mask based on the local brushstroke direction field of the mural and applies a brushstroke trajectory continuity loss function. Step S5: Perform anisotropic feature fusion on the highly reflective areas and use the infrared band penetration features to suppress the weight of overexposed areas in the visible light band.

2. The image processing method based on machine learning data completion algorithm as described in claim 1, characterized in that, In step S2, phase-correlation subpixel offset correction is performed. First, the integer displacement is derived in the frequency domain, and then subpixel compensation is obtained by fitting a Gaussian surface to estimate the band. With band The translation amounts between them specifically include: Perform a two-dimensional Fourier transform on the input image: ， ， in, For the first Band spectrum, This represents the two-dimensional discrete Fourier transform operator. For the first Band spectrum, For frequency domain coordinates, and The two bands are respectively located in spatial coordinates. Pixel intensity below; Normalized cross-power spectrum writing: ， in, For the normalized cross-power spectrum, For conjugate spectrum, Represents the modulus of a complex number. It serves as a stabilizing factor to suppress the denominator from approaching zero at low-energy spectral points; The stability factor is defined as: ， in, This is an empirical proportionality coefficient. This indicates the entire frequency domain. The arithmetic mean of the sampling points is taken. Calculate the inverse transform to obtain , in, For spatial domain related responses, For inverse Fourier operators; Peak coordinates are Pixel displacement estimation, the formula is: ， in, Integer pixel displacement estimation; Take the logarithm of the nine points in the peak neighborhood: ， in, For logarithmic response, Index for relative displacement; Derivation of subpixel compensation based on the analytical expression of a bivariate quadratic surface: ， ， and These are the horizontal and vertical subpixel compensation amounts, respectively. but in, For the final translation estimate; according to right Perform a second B-spline resampling to obtain the registered image. , Indicates by displacement vector The resampled registered image has the same unit of intensity as the input pixel.

3. The image processing method based on machine learning data completion algorithm as described in claim 1, characterized in that, The phase-correlation subpixel offset correction module in step S2 further includes: Constructing multi-scale pyramid decomposition layers to handle non-rigid deformation; Local structural tensor constraints are introduced at each scale layer, and the smoothness of the deformation field is adjusted according to the elastic coefficient of the mural substrate material.

4. The image processing method based on machine learning data completion algorithm as described in claim 1, characterized in that, The mineral pigment spectral reflectance reconstruction network includes: The encoder employs a multi-branch residual structure to extract visible light texture features and multispectral band features respectively. The decoder injects a layered attention gate mechanism during the decoding stage, activating the pigment generation path based on the underlying draft information revealed by the infrared band.

5. The image processing method based on machine learning data completion algorithm as described in claim 1, characterized in that, The dynamic stroke generator includes: Pressure-speed mapping model: By analyzing the rate of change of the brush stroke width, the physical brush movement parameters during creation are deduced. Trajectory prediction unit: A dual-stream LSTM network is used to model the changes in horizontal pen stroke force and the vertical turning angle respectively.

6. The image processing method based on machine learning data completion algorithm as described in claim 3, characterized in that, The calculation method for the local structure tensor constraint is as follows: Extract the microscopic crack distribution map of the mural substrate material; obtain the substrate crack distribution map based on the mural microscopic imaging, and dynamically adjust the deformation field regularization intensity according to the crack density field to reduce the deformation degree of freedom in the crack-dense area.

7. The image processing method based on machine learning data completion algorithm as described in claim 4, characterized in that, The hierarchical attention gating mechanism is executed in the infrared band feature channel: Generate a heatmap of mutual information between infrared and visible light features; The interference of visible light characteristics in the gold foil-covered area is suppressed by using heatmap weights.

8. The image processing method based on machine learning data completion algorithm as described in claim 1, characterized in that, The continuity loss function for the pen stroke trajectory includes: Curvature consistency measure: Calculates the standard deviation of the curvature change between the completed area and the original stroke; Pressure gradient constraint: Forces the pressure values ​​of adjacent stroke nodes to conform to the attenuation distribution model.

9. The image processing method based on machine learning data completion algorithm as described in claim 8, characterized in that, In step S4, the standard deviation of curvature change is defined and incorporated into the continuity loss function of the pen stroke trajectory, specifically as follows: For the original stroke node sequence With complete stroke node sequence Perform a spline resampling, maintaining equal arc length intervals; Original trajectory number The node discrete curvature is: ， in, The angle between adjacent line segments. Denotes the Euclidean norm. The angle between adjacent line segments. The local average arc length, Original curvature; complete trajectory curvature Using the same formula, we can derive: , This indicates the completion of node curvature. To complete the angle between adjacent line segments at the nth node of the stroke, This represents the local average arc length at node 𝑖; Define the curvature difference sequence and its mean as: ， in, For the two trajectories in the first... The curvature difference of the nodes, Its arithmetic mean The total number of nodes after spline resampling is dimensionless. Define the standard deviation of curvature variation as: ， in, The standard deviation of the curvature difference measures the overall uniformity of bending. Construct the loss term: , ,in, For curvature uniformity loss, For adaptive weights, This is the baseline coefficient for curvature loss. The original curvature average, This represents the original maximum curvature.

10. The image processing method based on machine learning data completion algorithm as described in claim 1, characterized in that, The anisotropic feature fusion further includes: A penetrating mask for infrared band feature maps is constructed, which is dynamically adjusted based on pixel-level reflection intensity. A weighted summation operation is used in the fusion layer, where the weighting coefficient of the overexposed region in the visible light band is inversely proportional to the infrared mask value; After fusion, edge-preserving filtering is performed to smooth feature transitions.

Images