An ultra-resolution face attribute lossless editing method and device and a readable storage medium

Abstract

The present application belongs to the technical field of face attribute fidelity editing, and relates to a super-resolution face attribute lossless editing method, device and readable storage medium; an original super-high resolution source image is input into a local perception network, a core topological anchor point set of a target face is output, a forward spatial affine transformation mapping operator is solved to perform affine clipping and normalized projection on a region of interest containing the target face in the original super-high resolution source image, and a local feature sub-manifold image is obtained; the local feature sub-manifold image is input into a generative adversarial network to output an evolved feature manifold image; a matrix inversion operation is performed on the forward spatial affine transformation mapping operator to obtain an inverse spatial mapping operator to perform topological inverse projection transformation on the evolved feature manifold image, and output a low-dimensional evolved image; the low-dimensional evolved image is enlarged and rotated to a physical coordinate system of the original super-high resolution source image by using a high-order resampling interpolation algorithm, and a high-dimensional registered evolved image is obtained.

Description

Technical Field

[0001] This invention relates to the field of facial attribute fidelity editing technology, and in particular to a super-resolution facial attribute non-destructive editing method, apparatus, and computer-readable storage medium. Background Technology

[0002] High-fidelity facial attribute editing requires precise editing of specific attributes (such as life cycle evolution, expression, and skin color) in real, ultra-high-resolution facial images while ensuring high fidelity and preserving non-target features. Essentially, this necessitates inverse semantic control capabilities—starting from real facial images and selectively adjusting target attributes. Early deep generative models could only generate entirely new images from random noise, lacking the ability to perform inverse semantic control on existing images, thus failing to locate and edit specific attributes of real faces.

[0003] With the continuous breakthroughs in multidimensional manifold learning theory in the field of computational photography, unsupervised feature decoupling and manifold synthesis technology based on generative adversarial networks has been gradually used for high-fidelity editing of facial attributes. By separating the attribute features (age, expression, etc.) of facial images from structural features (facial contour, facial feature layout, etc.), precise control at the semantic level can be achieved, thereby generating continuous and natural attribute transformation effects based on the continuous change law of facial attributes. For example, patent CN111985405B discloses a method and system for synthesizing facial age. It constructs a residual channel attention module at the bottleneck layer of a fully convolutional generator. Through global average pooling and max pooling, it aggregates spatial features and adaptively scales and recalibrates the latent features of facial images. Based on this, the generator employs a dual-branch decoding topology. One branch outputs the RGB image after age evolution, while the other branch regresses a continuous spatial attention mask matrix driven by the target age. This mask matrix aims to limit the injection region of nonlinear aging noise and attempts to preserve background information by weighted fusion of the source and synthesized images. Simultaneously, in the adversarial game stage, the method introduces a local discriminator targeting the geometric center of the face. The network uses a discriminator and a global discriminator for joint supervision to improve the realism of the generated facial texture. However, this approach gets caught in a perceptual-distortion game when generating macroscopic aging features. In order to forcibly fit the aging bone structure, the network over-smooths the high-frequency residual signals that represent identity recognition. This causes the reconstructed image to lose the original skin texture and present a severe plastic feel and artificial artifacts, which cannot meet the requirements of ultra-high resolution images for high-quality preservation of the original skin texture.Patent CN109509144B ​​discloses a job-related face aging method based on generative adversarial networks. Its core algorithm abandons a single generation path and instead employs a cascaded model combining a deep autoencoder network and a conditional generative adversarial network (CGAN). First, the deep autoencoder acts as an identity-preserving operator, compressing the youth source manifold into a low-dimensional latent space before decoding. Minimizing the L2 reconstruction loss forces the network to retain original personal features. Then, external prior semantic vectors (such as target age scalars and job category labels) are introduced as auxiliary input. The conditional generative adversarial network injects specific aging semantics into the latent space. Furthermore, to enhance the discriminative power of different conditions (such as different skin aging textures resulting from different jobs) within the feature manifold, a triplet ranking loss is introduced into the optimization objective. (Loss), which forces the distribution of features generated under the same conditions to be highly aggregated in the latent space, thereby realizing the recombination of aging features under multi-attribute constraints. However, this method relies on deep autoencoder networks to compress the source image to a low-dimensional latent space, resulting in a huge dissipation of high-resolution information. Since the feature map size of its decoding terminal is rigidly anchored, global downsampling and upsampling interpolation must be performed when processing ultra-high resolution input. This mechanism completely destroys the original high-precision pixel arrangement of the source image and introduces aliasing artifacts and truncation errors, making it completely lose the ability to perform lossless editing at the original pixel scale.

[0004] In summary, existing methods for high-fidelity facial attribute editing suffer from defects such as irreversible collapse of high-frequency micro-textures caused by global latent space inversion, physical resolution dimension locking, and topological collapse of non-target regions caused by global reconstruction. As a result, the reconstructed images exhibit details loss, structural distortion, and texture distortion, making it difficult to meet the requirements for high-fidelity facial attribute editing. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the defects of the existing high-fidelity facial attribute editing methods, such as irreversible collapse of high-frequency micro-textures caused by global latent space inversion, physical resolution dimension lock-up, and non-target region topological collapse caused by global reconstruction, resulting in the reconstruction of images with missing details, structural distortion and texture distortion, which make it difficult to meet the high-fidelity editing requirements of facial attributes.

[0006] To address the aforementioned technical problems, this invention provides a method for non-destructive editing of super-resolution facial attributes, comprising: The original ultra-high resolution source image is input into the local perception network, which outputs the core topological anchor point set of the target face. The objective is to minimize the difference between the core topological anchor point set after mapping by the forward spatial affine transformation mapping operator and the reference standard point set. The forward spatial affine transformation mapping operator is obtained by solving the problem. By using the forward spatial affine transformation mapping operator, the region of interest containing the target face in the original ultra-high resolution source image is affinely cropped and normalized projected to obtain a local feature submanifold image. The local feature submanifold image is input into the encoder in the adversarial generative network, and the output is the initial latent semantic coding matrix. After performing directional feature offset operation on the initial latent semantic coding matrix based on the topological evolutionary manifold direction of the facial attributes to be edited, it is input into the decoder in the adversarial generative network, and the output is the evolutionary feature manifold image. Perform matrix inversion on the forward spatial affine transformation mapping operator to obtain the inverse spatial mapping operator. Use the inverse spatial mapping operator to perform topological inverse projection transformation on the evolutionary feature manifold image to output a low-dimensional evolutionary image. The low-dimensional evolution image is magnified and rotated to the physical coordinate system of the original ultra-high resolution source image using a high-order resampling interpolation algorithm to obtain a high-dimensional registered evolution image.

[0007] Preferably, the forward spatial affine transformation mapping operator Represented as: , in, The homogeneous coordinate matrix representing the set of core topological anchor points of the target face; Represents a reference standard point set; Denotes the square of the Frobenius norm; Represents the regularization coefficient; This represents the L1 regularization constraint, L2 regularization constraint, or low-rank constraint on the forward affine transformation mapping operator.

[0008] Preferably, the local feature submanifold image is represented as: , in, Represents the coordinates in the image of a local feature submanifold Pixels at that location; Indicates the region of interest containing the target face; Represents the coordinates in the original ultra-high resolution source image Pixels at that location; This represents the discrete sampling matching operator, where the homogeneous coordinates in the original ultra-high resolution source image... After positive spatial affine transformation mapping operator Transformed homogeneous coordinates in the local feature submanifold image When they overlap, ,on the contrary .

[0009] Preferably, the initial latent semantic coding matrix Represented as: , in, Indicates encoder; Represents a local feature submanifold image; The formula for performing directional feature shift operation on the initial latent semantic coding matrix is ​​as follows: , in, This represents the initial latent semantic encoding matrix after the directional feature offset operation; A scalar step size parameter representing the degree of evolution of the target facial attributes; Indicates the orientation of the topological evolution manifold of the facial attribute to be edited; Evolutionary feature manifold image Represented as: , in, This indicates the decoder.

[0010] Preferably, the inverse space mapping operator The acquisition process is as follows: Mapping operator for affine transformation of forward space By supplementing the information, a homogeneous space transformation matrix is ​​constructed. ; Alignment subspace transformation matrix Performing the standard matrix inversion operation yields ,extract The first two lines serve as the inverse space mapping operator. .

[0011] Preferably, the high-dimensional registration evolution image is represented as follows: , in, Represents coordinates in high-dimensional registration evolution images Pixels at that location; Represents coordinates in low-dimensional evolutionary images Pixels at that location; This represents a higher-order spatial resampling interpolation kernel function; This represents the inverse space mapping operator.

[0012] Preferably, after obtaining the high-dimensional registration evolution image, the process further includes: A two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation is used to perform convolution filtering on high-dimensional registration evolution images to extract low-frequency structural semantic components that characterize the evolution of macroscopic attributes. The original ultra-high resolution source image is convolved and filtered using a two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation to obtain a filtered image; the difference between the original ultra-high resolution source image and the filtered image is calculated to extract the high-frequency physical residual components that characterize the original absolute physical properties. After performing a matrix linear weighted summation operation on the low-frequency structural semantic components and the high-frequency physical residual components, the result is numerically truncated to reconstruct a frequency domain reconstructed image containing the evolutionary macroscopic structure and the original microscopic texture. Based on the two-dimensional geometric contour and anisotropic spatial gradient distribution of the target face in the original ultra-high resolution source image, an Alpha weight mask is constructed with the core facial region as the high-value retention area and the surrounding non-target edges exhibiting non-linear Gaussian smooth decay. By using an alpha weight mask, a pixel-level spatial gradient blending operation is performed on the frequency domain reconstructed image and the original super-resolution source image to obtain a super-resolution edited image.

[0013] Preferably, low-frequency structural semantic components Represented as: , in, Represents high-dimensional registration evolution images; This indicates that it has a preset bandwidth standard deviation. Two-dimensional spatial low-pass filter operator; Represents a discrete two-dimensional convolution operation; High-frequency physical residual components Represented as: , in, This represents the original ultra-high resolution source image; Frequency domain reconstructed image Represented as: , in, Indicates the numerical truncation operator; Indicates the texture preservation factor; The alpha weight mask is represented as follows: , in, Representing coordinates Weight mask at the location; Representing coordinates The square of the Euclidean distance to the boundary; Represents the two-dimensional geometric topological edge of the region of interest containing the target face; The standard deviation of the Gaussian attenuation rate at the nonlinear edges of the control mask represents the control mask's nonlinear edge smoothing attenuation rate. Super-resolution image editing Represented as: , in, Represents the Alpha weight mask; This represents a pixel-level multiplication operator.

[0014] The present invention also provides a super-resolution facial attribute non-destructive editing device, comprising: The forward mapping parameter acquisition module is used to input the original ultra-high resolution source image into the local perception network and output the core topological anchor point set of the target face. The forward spatial affine transformation mapping operator is obtained by solving the problem with the objective of minimizing the difference between the core topological anchor point set after mapping by the forward spatial affine transformation mapping operator and the reference standard point set. The region of interest processing module is used to perform affine cropping and normalized projection on the region of interest containing the target face in the original ultra-high resolution source image using the forward spatial affine transformation mapping operator to obtain a local feature submanifold image. The facial attribute evolution module is used to input the local feature submanifold image into the encoder in the adversarial generative network and output the initial latent semantic coding matrix. After performing directional feature offset operation on the initial latent semantic coding matrix based on the topological evolution manifold direction of the facial attributes to be edited, it is input into the decoder in the adversarial generative network and outputs the evolved feature manifold image. The inverse transformation module is used to perform matrix inversion on the forward spatial affine transformation mapping operator to obtain the inverse spatial mapping operator. The inverse spatial mapping operator is then used to perform topological inverse projection transformation on the evolutionary feature manifold image to output a low-dimensional evolutionary image. The evolutionary image acquisition module is used to enlarge and rotate the low-dimensional evolutionary image to the physical coordinate system of the original ultra-high resolution source image using a high-order resampling interpolation algorithm, so as to obtain a high-dimensional registered evolutionary image.

[0015] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described super-resolution facial attribute non-destructive editing method.

[0016] The super-resolution non-destructive facial attribute editing method provided in this application has the following beneficial effects: 1. First, the original ultra-high resolution source image is input into a local perceptual network to extract the core topological anchor point set of the target face. The forward spatial affine transformation mapping operator is then solved with the objective of minimizing the difference between the anchor points and the reference standard point set, achieving accurate localization and geometric normalization of the facial region of interest (ROI) and avoiding redundant interference introduced by global processing. Subsequently, this forward mapping operator is used to perform affine cropping and normalized projection on the ROI in the original ultra-high resolution image, obtaining a local feature submanifold image. Only this local image is input into the encoder of the generative adversarial network to obtain the initial latent semantic coding matrix, without performing global latent space inversion on the entire ultra-high resolution image, thus avoiding the irreversible collapse of high-frequency microtextures caused by global compression. Afterwards, the base... The initial latent code is directionally offset along the direction of the facial attribute topological evolution manifold, and then the decoder outputs the evolutionary feature manifold image. Attribute editing is only completed in the local normalized space, without changing the structure and texture of non-target regions. Then, by inverting the forward affine transformation matrix to obtain the inverse space mapping operator, a topological inverse projection transformation is performed on the evolutionary feature manifold image to accurately map the edited local features back to the original image coordinate space, avoiding problems such as topological collapse and structural distortion in non-target regions caused by global reconstruction. Finally, a high-order resampling interpolation algorithm is used to restore the inverse-projected low-dimensional evolutionary image to the original ultra-high resolution physical coordinate system, achieving sub-pixel level detail reconstruction and texture fidelity, effectively compensating for defects such as detail loss and texture distortion. 2. After obtaining the high-dimensional registered evolution image, the final ultra-high-fidelity facial editing is achieved through frequency domain separation reconstruction and facial adaptive mask fusion. A two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation is used to perform convolution filtering on the high-dimensional registered evolution image and the original ultra-high resolution source image, respectively. This accurately separates the low-frequency structural semantic components of macroscopic facial attribute changes in the evolution image from the high-frequency physical residual components of skin texture, pores, and hair strands in the original image, which are not editable. By performing linear weighted summation and numerical truncation on the low-frequency semantic components and the high-frequency physical residual components, the edited facial structure can be strictly integrated with the real microscopic texture of the original image. This preserves the effectiveness of facial attribute editing while completely retaining the original high-frequency details, thus avoiding texture distortion, detail blurring, and texture loss from the root. Subsequently, based on the two-dimensional geometric contour and anisotropic spatial gradient distribution of the target face, an Alpha weight mask is constructed with the core facial region as the high-value retention area and the non-target region exhibiting Gaussian smooth decay. This mask is used to perform pixel-level gradual weighted fusion of the frequency domain reconstructed image and the original ultra-high resolution source image, which can achieve a seamless transition between the edited and non-edited regions, avoiding harsh edges, structural misalignment, and distortion of non-target regions. The final output is a super-resolution edited image that combines natural editing, structural fidelity, texture realism, and spatial coherence. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 Flowchart of the super-resolution facial attribute non-destructive editing method provided in this application; Figure 2 The original ultra-high resolution source image provided for the embodiments of this application; Figure 3 Evolutionary feature manifold images provided in embodiments of this application; Figure 4 Super-resolution edited images provided in the embodiments of this application; Figure 5 A schematic diagram of the super-resolution facial attribute non-destructive editing device provided in this application. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Please see Figure 1 , Figure 1 The diagram shows a flowchart of the super-resolution non-destructive facial attribute editing method provided in this application, which includes steps S10 to S50: S10: Transfer the original ultra-high resolution source image Input a local perceptual network and output a set of core topological anchor points for the target face. ( (representing the nth anchor point), with the objective of minimizing the difference between the core topological anchor point set and the reference standard point set after mapping by the forward affine transformation mapping operator, the forward affine transformation mapping operator is obtained by solving.

[0020] Specifically, the local perception network is based on the facial key point detection algorithm to extract core topological anchor points containing multi-dimensional spatial features such as the corners of the eyes, the tip of the nose, and the corners of the mouth in a complex global field of view, thereby obtaining a set of core topological anchor points of the target face.

[0021] It should be noted that the local perceptual network in this application is a pre-trained local perceptual network, and its training process includes steps 1-1 to 1-4: Step 1-1: Construct a training dataset and perform pose normalization, illumination equalization, and data augmentation on the samples in the training dataset. The training dataset contains multiple sets of labeled ultra-high resolution face image samples. Each set of samples is labeled with the pixel coordinates of key topological anchor points such as the corners of the eyes, the tip of the nose, and the corners of the mouth of the target face. Data augmentation includes operations such as random rotation, horizontal flipping, random cropping, brightness perturbation, and contrast perturbation to improve the network's robustness to faces with different poses, illuminations, and expressions.

[0022] Steps 1-2: Construct a local perceptual network model; the local perceptual network model is based on a lightweight feature extraction backbone (such as MobileNet, ResNet-18 or improved HRNet), extracts local facial features under different receptive fields through a multi-scale feature pyramid structure, and then connects to a keypoint regression head to directly output the two-dimensional coordinates of a preset number of core topological anchor points of the target face.

[0023] Steps 1-3: Define the network loss function; the loss function includes two parts: keypoint coordinate regression loss and topology constraint loss. The keypoint coordinate regression loss uses mean squared error loss (MSE) or smoothed L1 loss to measure the deviation between the predicted anchor point coordinates and the labeled true values. The topology constraint loss introduces facial structure geometric constraints (such as the proportions of key structures such as eye distance, nose distance, and mouth width) to regularize the relative positional relationship between predicted anchor points, so as to ensure the topological consistency of the output anchor point set.

[0024] Steps 1-4: Use the Adam optimizer with learning rate decay to train the network end-to-end. During training, use the ground truth of the labeled anchor points as supervision signals to iteratively optimize the network parameters until the loss function converges or the preset number of iterations is reached, and the trained local perceptron network is obtained.

[0025] Furthermore, the forward affine transformation mapping operator Represented as: , in, The homogeneous coordinate matrix representing the set of core topological anchor points of the target face; Represents a reference standard point set; Denotes the square of the Frobenius norm; Represents the regularization coefficient; This represents a constraint on the forward affine transformation mapping operator; optionally, this constraint can be an L1 regularization constraint, an L2 regularization constraint, or a low-rank constraint, with regularization coefficients... The value range is [0.01, 0.1], with an optimal value of 0.05. This can be achieved by adjusting the regularization coefficient. The value of can strike a balance between alignment accuracy and deformation smoothness.

[0026] It should be noted that the reference standard point set is obtained based on a standardized facial alignment set specification. Specifically, a standard facial key point template with a preset dimension is constructed based on facial anatomy and facial alignment specifications. This template includes core topological anchor points such as the corners of the eyes, eyebrows, tip of the nose, alar of the nose, corners of the mouth, and chin. The relative positions and proportional relationships between each anchor point are determined based on the statistical average results of a large sample of faces, forming a fixed standard facial topological structure. Then, a standardized facial coordinate system is set with the midpoint of the line connecting the centers of the eyes as the origin, the horizontal direction as the x-axis, and the vertical direction as the y-axis. Based on key dimensions such as interocular distance and facial height, the standard facial topological structure is scaled to ensure that faces of different poses and sizes can be mapped to the same standard space. Based on the standard facial topological structure and the normalized coordinate system, a reference standard point set containing a preset number of reference standard points is generated.

[0027] This application employs the least squares method to optimize and solve for a positive spatial affine transformation mapping operator that can eliminate spatial scale differences and rotational distortion. Then, using the positive spatial affine transformation mapping operator as a physical criterion, affine cropping and normalized projection are performed on the original ultra-high resolution source image to obtain a centered, scale-normalized feature submanifold image with redundant background removed, providing a standardized data foundation for subsequent semantic inference.

[0028] S20: Using the forward spatial affine transformation mapping operator, the region of interest containing the target face in the original ultra-high resolution source image is affinely cropped and normalized projected to obtain a local feature submanifold image.

[0029] Furthermore, the local feature submanifold image is represented as: , in, Represents the coordinates in the image of a local feature submanifold Pixels at that location; Indicates the region of interest containing the target face; Represents the coordinates in the original ultra-high resolution source image Pixels at that location; This represents the discrete sampling matching operator, where the homogeneous coordinates in the original ultra-high resolution source image... After positive spatial affine transformation mapping operator Transformed homogeneous coordinates in the local feature submanifold image When they overlap, ,on the contrary .

[0030] Specifically, the process of obtaining the region of interest (ROI) of the target face in the original ultra-high resolution source image is as follows: based on the core topological anchor point set of the target face and combined with a predefined standard 3D face model, the PnP algorithm or EPnP algorithm is used to solve the 3D spatial pose parameter vector of the target face in the camera coordinate system. Here, 'a' represents the global yaw parameter, 'b' represents the pitch attitude parameter, and 'c' represents the roll attitude parameter, thus realizing the estimation of three-dimensional attitude from two-dimensional image; using the bounding box of the face corresponding to the three-dimensional attitude parameters as a reference, it is expanded in all directions by a preset ratio to include key areas such as facial contour, ears, and forehead, thus obtaining the expanded bounding box; the expanded bounding box is projected back onto the two-dimensional image plane to obtain the initial rectangular bounding box; finally, the initial rectangular bounding box is corrected to obtain the region of interest of the target face.

[0031] S30: Input the local feature submanifold image into the encoder in the adversarial generative network to output the initial latent semantic coding matrix; perform directional feature offset operation on the initial latent semantic coding matrix based on the topological evolutionary manifold direction of the facial attributes to be edited, and input it into the decoder in the adversarial generative network to output the evolutionary feature manifold image.

[0032] Furthermore, the initial latent semantic encoding matrix Represented as: , in, Indicates encoder; This represents a local feature submanifold image.

[0033] The formula for performing directional feature shift operation on the initial latent semantic coding matrix is ​​as follows: , in, This represents the initial latent semantic encoding matrix after the directional feature offset operation; A scalar step size parameter representing the degree of evolution of the target facial attributes; This represents the orientation of the topological evolution manifold of the facial attributes to be edited; in some embodiments of this application, The value range of is [-5, 5]. At that time, the initial latent semantic encoding does not shift, and the reconstructed native face is directly output. As facial features evolve towards aging, the larger the value, the greater the span of age increase. At that time, facial features are reverse-engineered to look younger, through adjustments... This allows for continuous, smooth, and lossless editing of age attributes across the target manifold direction.

[0034] For example, to obtain the topological evolution manifold direction of the facial attribute to be edited, aging... First, a predetermined number of initial latent semantic encoding matrix samples are randomly sampled from the latent space of the pre-trained generative network. Then, a pre-trained age classification attribute evaluation network is used to score the images generated from the initial latent semantic encoding matrix samples, classifying them into a youth feature set and an aging feature set. Next, the initial latent semantic encoding matrix samples from both feature sets are used as training data to train a linear support vector machine classifier to find the linear hyperplane that maximizes the distinction between the two feature sets. The orthogonal normal vector of this linear hyperplane in the latent space is used as the direction of the topological evolution manifold for the aging facial attribute to be edited. .

[0035] Evolutionary feature manifold image Represented as: , in, This indicates the decoder.

[0036] It should be noted that the hidden layer weights of the adversarial generative network have been pre-adaptively fine-tuned using a massive dataset of facial images of specific ethnicities, thereby forcing the recalibration of the pre-spatial feature distribution and correcting the tendency of topological distortion during cross-ethnic evolution.

[0037] For example, the facial attribute to be edited can be the life cycle. The model strictly follows the objective life cycle topological evolution law of the target race. Along the set semantic boundary of the target attribute (such as towards the aging feature), it performs continuous and non-linear directional offset operation on the initial latent semantic encoding. The decoder then reconstructs the offset features. The output not only has the macroscopic physical structure of the target evolution, but its physical resolution also strictly maintains the normalized standard of the initial evolution feature manifold.

[0038] S40: Perform matrix inversion on the forward spatial affine transformation mapping operator to obtain the inverse spatial mapping operator. Use the inverse spatial mapping operator to perform topological inverse projection transformation on the evolutionary feature manifold image to output a low-dimensional evolutionary image.

[0039] Furthermore, the inverse space mapping operator The acquisition process is as follows: Mapping operator for affine transformation of forward space By supplementing the information, a homogeneous space transformation matrix is ​​constructed. ; Alignment subspace transformation matrix Performing the standard matrix inversion operation yields ,extract The first two lines serve as the inverse space mapping operator. .

[0040] For example, the affine transformation operator for a 2×3 positive space. Fill in the bottom row [0,0,1] to construct a 3×3 homogeneous space transformation matrix. After that, Performing the inversion of the standard matrix yields ,Will The first two lines serve as the inverse space mapping operator. .

[0041] S50: The low-dimensional evolution image is magnified and rotated to the physical coordinate system of the original ultra-high resolution source image using a high-order resampling interpolation algorithm to obtain a high-dimensional registered evolution image.

[0042] Furthermore, the high-dimensional registration evolution image is represented as: , in, Represents coordinates in high-dimensional registration evolution images Pixels at that location; Represents coordinates in low-dimensional evolutionary images Pixels at that location; This represents a higher-order spatial resampling interpolation kernel function; This represents the inverse space mapping operator.

[0043] In the embodiments of this application, the higher-order spatial resampling interpolation kernel function The standard mathematical expression for the bicubic interpolation kernel function is: , in, It is an unknown. This represents the deformation control parameter, typically taking values ​​of -0.5 or -0.75, as described in the embodiments of this application. The optimal value is -0.5.

[0044] It should be noted that when performing interpolation calculations on two-dimensional images, the two-dimensional interpolation kernel function... Using a one-dimensional high-order space resampling interpolation kernel function The orthogonal product form is used for calculation, i.e. The target pixel value in the high-dimensional registration evolution image is determined by its inverse mapping coordinates in the low-dimensional evolution image. The 16 source pixels in the neighborhood are obtained by weighted summation using the two-dimensional interpolation kernel function described above, thereby achieving sub-pixel-level smooth reconstruction.

[0045] Specifically, by reading and parsing the forward spatial affine transformation mapping operator, and through high-precision internal matrix inversion operations, the inverse spatial mapping parameters for spatial restoration are accurately derived. Based on these inverse parameters, a high-order topological inverse projection transformation is performed on the normalized evolving manifold generated by the deep network. In this process, in order to effectively suppress pixel jaggedness and feature aliasing that may be caused by cross-dimensional mapping, a high-order resampling interpolation algorithm is introduced to ensure that the low-dimensional generated manifold can be smoothly enlarged and generate accurate geometric rotation without topological distortion. Finally, it achieves a perfect pixel-level alignment with the absolute physical coordinate system of the original ultra-high resolution source image, generating a high-dimensional registered evolving image with macroscopic spatial consistency.

[0046] Furthermore, after obtaining the high-dimensional registered evolutionary image, in order to completely avoid the collapse of high-frequency physical details caused by the global reconstruction of the deep network, this application realizes the orthogonal separation of the macroscopic bone structure and the microscopic skin structure through the frequency domain analysis mechanism, and then realizes the pixel-level lossless integration of the evolutionary manifold and the high-dimensional original background environment through frequency domain feature recombination and multi-level mask adaptive mosaicking.

[0047] Specifically, after obtaining the high-dimensional registration evolution image, steps 2-1 to 2-5 are also included: Step 2-1: Use a two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation to perform convolution filtering on the high-dimensional registration evolution image to extract low-frequency structural semantic components that characterize the evolution of macroscopic attributes.

[0048] Specifically, low-frequency structural semantic components Represented as: , in, This represents high-dimensional registration evolution images; This indicates that it has a preset bandwidth standard deviation. Two-dimensional spatial low-pass filter operator; This represents a discrete two-dimensional convolution operation.

[0049] Specifically, it has a preset bandwidth standard deviation. The two-dimensional spatial low-pass filter operator, whose impulse response characteristics closely match the differences in human visual perception of structure and texture, can filter out false high-frequency noise forcibly introduced by deep generative networks when performing convolutional filtering on high-dimensional registered evolutionary images. It can independently extract low-frequency structural semantic components representing the evolution of macroscopic attributes (such as soft tissue indentation and large-scale muscle deformation). Optionally, a preset bandwidth standard deviation is used. The value range is [5,25]. This range is based on the physical scale of the face in the original ultra-high resolution source image and is used to accurately separate the macroscopic bone structure and the microscopic skin texture.

[0050] Step 2-2: Perform convolution filtering on the original ultra-high resolution source image using a two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation to obtain a filtered image; calculate the difference between the original ultra-high resolution source image and the filtered image to extract the high-frequency physical residual components that characterize the original absolute physical properties.

[0051] Specifically, high-frequency physical residual components Represented as: , in, This represents the original ultra-high resolution source image.

[0052] Specifically, a low-pass filter of the same frequency band is performed on the original ultra-high resolution source image, and the mathematical difference between the original pixel array and the low-frequency array is calculated, thereby orthogonally extracting the high-frequency physical residual tensor that is completely uncontaminated by the deep generative network and contains the original pore distribution pattern and the optical reflectivity of the real dermis.

[0053] Steps 2-3: After performing a matrix linear weighted summation operation on the low-frequency structural semantic components and the high-frequency physical residual components, the operation results are numerically truncated to reconstruct a frequency domain reconstructed image containing the evolved macroscopic structure and the original microscopic texture.

[0054] Specifically, frequency domain reconstructed image Represented as: , in, Indicates the numerical truncation operator; This represents the texture preservation factor.

[0055] It should be noted that the numerical truncation operator The value range is [0, 255]. Because after the low-frequency structural semantic components and high-frequency physical residual components undergo matrix weighted summation, some pixel values ​​may overflow due to the injection of high-frequency noise (less than 0 or greater than 255), causing abnormal pixel patches in the final image. This truncation operator forces all pixel values ​​to be constrained to the standard 8-bit RGB color space's physically valid range (i.e., [0, 255]), thus ensuring that the image after high-frequency micro-texture reconstruction can be normally parsed and rendered in the physical coordinate system; texture preservation coefficient. The value range is [0.3, 0.7]. A value exceeding 0.7 will result in overly sharp facial textures and excessively pronounced grooves after facial evolution. A value less than 0.3 will result in blurred skin texture after facial evolution. In the embodiments of this application, The optimal value for is 0.5, at which point the best smooth natural evolution effect can be achieved.

[0056] Steps 2-4: Based on the two-dimensional geometric contour and anisotropic spatial gradient distribution of the target face in the original ultra-high resolution source image, construct an Alpha weight mask with the core facial region as the high-value retention area and the surrounding non-target edges exhibiting non-linear Gaussian smooth decay.

[0057] Specifically, the Alpha weight mask is represented as: , in, Representing coordinates Weight mask at the location; Representing coordinates The square of the Euclidean distance to the boundary; Represents the two-dimensional geometric topological edge of the region of interest containing the target face; The standard deviation of Gaussian represents the rate of smooth attenuation at the nonlinear edges of the control mask.

[0058] It should be noted that the Gaussian standard deviation of the nonlinear edge smoothing attenuation rate of the control mask is... Based on the adaptive determination of the target face scale, if the eye-to-eye center distance in the core topological anchor point set is... ,but This ensures that the width of the feathered edge can dynamically scale with the image resolution, achieving a seamless transition of spatial gradient blending.

[0059] Steps 2-5: Use the Alpha weight mask to perform pixel-level spatial gradient blending operations on the frequency domain reconstructed image and the original super-resolution source image to obtain the super-resolution edited image.

[0060] Specifically, super-resolution image editing Represented as: , in, Represents the Alpha weight mask; This represents a pixel-level multiplication operator.

[0061] It should be noted that in the high-resolution facial attribute lossless editing scenario of this application, the Alpha weight mask is a two-dimensional floating-point matrix with the same size as the source image. The weight value at each coordinate point in the mask is in the range of [0,1]. Specifically, the weight value in the core facial editing area is close to 1, indicating that the pixels in this area are completely reconstructed using the frequency domain to reconstruct the evolutionary features of the image; the weight value in non-target areas (background, clothing) far from the core of the face is close to 0, indicating that the original pixels of the original ultra-high resolution source image are completely preserved losslessly; the weight value in the facial contour boundary area presents a floating-point value transitioning from 0 to 1, for example, 0.5, indicating that the new and old pixels are gradually mixed according to the proportional space, thereby achieving visual seamless stitching.

[0062] Specifically, frequency domain feature reconstruction and multi-level mask mosaicking, as the final rendering and spatial compositing, are responsible for breaking down the visual disconnect between the local evolutionary manifold and the global ultra-high-definition native background, achieving seamless pixel-level physical integration between the two. First, at the frequency domain level, linear reconstruction and numerical truncation are performed on the low-frequency structural semantic tensor and the high-frequency physical residual tensor to reconstruct a fusion feature tensor that combines both evolved bone structure and absolutely native skin texture. Then, the two-dimensional geometric contours and anisotropic spatial gradient distribution of the target face are analyzed in depth, adaptively generating a multi-level spatial weight mask tensor with the core facial evolutionary region as the high-value retention area and a non-linear Gaussian smooth decay towards the outer non-target edges. Using this dynamic mask tensor as a flexible physical constraint boundary for the mixing process, high-precision spatial gradient mixing operations are performed between the reconstructed high-dimensional fusion features and the original complex environmental background. The final output super-resolution edited image not only achieves high-fidelity evolution of target attributes in the facial region but also ensures low-loss and physical-level preservation of microscopic hair topology, clothing texture, and complex background environment at the field of view edges.

[0063] The effectiveness of the super-resolution non-destructive facial attribute editing method provided in this application is verified through specific examples below: The software and hardware programming environment based on this embodiment is as follows: operating system Ubuntu 20.04 (or Windows 10 / 11), deep learning framework using PyTorch 1.10 or above, image processing and matrix operations relying on OpenCV and NumPy libraries, feature point detection using Dlib library, and hardware using NVIDIA GPUs (such as RTX 30 series or equivalent computing power devices) for core computing acceleration.

[0064] This embodiment uses a high-resolution human face image with an actual resolution of 3456×5184 pixels as input, and takes the age feature transformation attribute editing as an example. Figure 2 The image shown is the original ultra-high resolution source image provided in an embodiment of this application. The specific operation steps are as follows: 1. Reading of ultra-high resolution source images and extraction of facial topological feature points: Reading high-resolution source images with a resolution of 3456×5184, calling 68 facial key point detection algorithms in the Dlib library, accurately extracting the coordinates of the center of the eyes, the tip of the nose and the corners of the mouth of the target face, and using them as core topological anchor points. Based on the above anchor point parameters, the region of interest (ROI) containing the core facial features is delineated in the source image.

[0065] 2. Spatial decoupling and mapping operator persistence based on affine transformation: To remove the fixed physical resolution limitation of the deep generative model, the following spatial dimensionality reduction and mapping parameter extraction operations are performed: (1) Based on the above defined ROI boundary and standard face alignment geometry, a 2×3 positive spatial affine mapping transformation operator is derived and calculated; (2) To ensure the consistency of physical location during later restoration, the operator is serialized into an independent binary file using the NumPy library and persistently stored locally; (3) Perform an affine transformation on the source image using the forward spatial affine mapping transformation operator, extract and output a center-aligned face image with a size of strictly 1024×1024 pixels, denoted as the normalized submanifold image.

[0066] 3. Latent space attribute derivation of specific ethnic manifolds in a vertical domain: The aforementioned 1024×1024 pixel normalized submanifold image is input into a deep encoder and StyleGAN generative network model. This model has been pre-adaptively fine-tuned using a massive amount of facial images (more than 20,000 samples) to correct the distribution bias of traditional models for aging characteristics of other baseline populations. The network evolves in the latent space according to the target age scalar, outputting an evolved image with the aging bone structure and skin aging characteristics of the target population, maintaining a resolution of 1024×1024 pixels. Figure 3 The image shown is an evolutionary feature manifold image provided in an embodiment of this application.

[0067] 4. Topological inverse registration mapping of the evolutionary image: The program reads the previously saved binary forward spatial affine transformation operator and calculates its inverse spatial mapping operator. Using the inverse spatial mapping operator, the generated aging image is subjected to inverse spatial mapping, which undistortedly enlarges, rotates, and precisely registers it back to the coordinate system position of the 3456×5184 pixel high-resolution source image, resulting in a spatially registered image.

[0068] 5. Adaptive decoupling extraction of high and low frequency features in the spatial frequency domain of images: To address the loss of high-frequency details (pores, skin texture) caused by direct replacement, the following frequency domain separation operation is performed: (1) Set the standard deviation parameter of the two-dimensional Gaussian filter kernel as follows: ; (2) Perform Gaussian filtering on the registered image to filter out the generated false details and extract only the low-frequency semantic components that characterize macro-aging structures (such as soft tissue collapse of the cheeks and deep wrinkles); (3) Perform Gaussian filtering with the same parameters on the high-definition source image synchronously, and subtract the filtering result from the original image matrix to extract the absolute high-fidelity residual component containing native pores, real skin color and micro-texture.

[0069] 6. Frequency domain feature orthogonal recombination and multi-order mask feathering mosaicking: To achieve natural evolution of facial attributes and lossless processing of non-target regions, the following fusion operations are performed: (1) The residuals of low-frequency aging bone phase and high-frequency original skin phase are linearly summed by matrix and numerical clipping is performed to ensure that the pixel values ​​are normalized in the interval [0, 255] to obtain the frequency domain reconstructed image; (2) Calculate and construct an Alpha weighted dynamic mask that covers the core facial region and decays non-linearly with a Gaussian weight towards the edge; (3) Using the Alpha mixing algorithm, the frequency domain reconstructed image is seamlessly embedded back into the high-resolution source image according to the weight mask, and the final edited image is output, such as... Figure 4 The image shown is a super-resolution edited image provided in an embodiment of this application. This image not only achieves natural facial aging and skin texture fidelity, but also maintains its original resolution of 3456×5184 pixels, and achieves pixel-level lossless rendering of hair topology and complex background outside the boundary.

[0070] Based on the super-resolution non-destructive editing method for facial attributes provided in the above embodiments, this application also provides a super-resolution non-destructive editing device for facial attributes, such as... Figure 5 As shown, the device specifically includes: The forward mapping parameter acquisition module 10 is used to input the original ultra-high resolution source image into the local perception network and output the core topological anchor point set of the target face. The forward spatial affine transformation mapping operator is obtained by solving the problem with the objective of minimizing the difference between the core topological anchor point set after mapping by the forward spatial affine transformation mapping operator and the reference standard point set.

[0071] The region of interest processing module 20 is used to perform affine cropping and normalized projection on the region of interest containing the target face in the original ultra-high resolution source image using a forward spatial affine transformation mapping operator to obtain a local feature submanifold image.

[0072] The facial attribute evolution module 30 is used to input the local feature submanifold image into the encoder in the adversarial generative network and output the initial latent semantic coding matrix; after performing directional feature offset operation on the initial latent semantic coding matrix based on the topological evolution manifold direction of the facial attributes to be edited, it is input into the decoder in the adversarial generative network and output the evolved feature manifold image.

[0073] The inverse transformation module 40 is used to perform matrix inversion on the forward spatial affine transformation mapping operator to obtain the inverse spatial mapping operator. The inverse spatial mapping operator is then used to perform topological inverse projection transformation on the evolutionary feature manifold image to output a low-dimensional evolutionary image.

[0074] The evolutionary image acquisition module 50 is used to enlarge and rotate the low-dimensional evolutionary image to the physical coordinate system of the original ultra-high resolution source image using a high-order resampling interpolation algorithm to obtain a high-dimensional registered evolutionary image.

[0075] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described super-resolution facial attribute non-destructive editing method.

[0076] The proposed super-resolution facial attribute lossless editing method utilizes parameterized operators for cross-dimensional manifold anchoring, completely breaking the inherent physical resolution topology lock of deep generative models. This scheme achieves lossless input and output at the original pixel level for ultra-high resolution images, significantly improving the spatial accuracy of cross-dimensional attribute editing. Simultaneously, it innovatively performs orthogonal decoupling and recombination of the semantic evolution manifold and the high-frequency residuals of the source image at the frequency domain level. This mechanism avoids the "perceptual-distortion" collapse caused by global inversion, achieving fidelity in the original epidermal microtexture and physical optical properties. Furthermore, it constructs a multi-order dynamic mask space adaptive fusion mechanism constrained by complex boundary topology, effectively mitigating edge crosstalk and artifacts caused by cross-dimensional reprojection, achieving in-situ lossless stripping of target attribute evolution and ultra-high-definition complex backgrounds. This method is lightweight and... Frequency domain matrix recombination and topological registration replace computationally intensive global nonlinear pixel reconstruction. This not only fundamentally eliminates memory overflow and accuracy degradation caused by ultra-large tensor operations, but also significantly enhances the robustness and generalization ability of the algorithm in medical-grade image processing. In addition, by independently training and fine-tuning parameters on ultra-large-scale vertical ethnic datasets, the realism of facial evolution generation for specific ethnic groups is significantly improved. It should also be noted that the generation operator of this application corrects the domain offset bias of the original model, not only accurately replicating the real aging topological patterns of the target ethnic group, but also effectively preventing identity distortion caused by cross-ethnic feature crosstalk.

[0077] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0079] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for super-resolution non-destructive editing of facial attributes, characterized in that, include: The original ultra-high resolution source image is input into the local perception network, which outputs the core topological anchor point set of the target face. The objective is to minimize the difference between the core topological anchor point set after mapping by the forward spatial affine transformation mapping operator and the reference standard point set. The forward spatial affine transformation mapping operator is obtained by solving the problem. By using the forward spatial affine transformation mapping operator, the region of interest containing the target face in the original ultra-high resolution source image is affinely cropped and normalized projected to obtain a local feature submanifold image. The local feature submanifold image is input into the encoder of the adversarial generative network (GCN) to output an initial latent semantic encoding matrix. After performing a directional feature shift operation on the initial latent semantic encoding matrix based on the topological evolutionary manifold orientation of the facial attribute to be edited, it is input into the decoder of the GCN to output an evolutionary feature manifold image. The facial attribute to be edited is aging, and the steps for obtaining its topological evolutionary manifold orientation include: A predetermined number of initial latent semantic encoding matrix samples are obtained by randomly sampling in the latent space of the pre-trained generative network; The images generated from the initial latent semantic encoding matrix samples are scored using a pre-trained age classification attribute evaluation network, and then classified into a young feature set and an aging feature set. Using initial latent semantic encoding matrix samples from two feature sets as training data, a linear support vector machine classifier is trained to find the linear hyperplane that maximizes the distinction between the two feature sets. The orthogonal normal vector of this linear hyperplane in the latent space is used as the direction of the topological evolution manifold for the aging facial attribute to be edited. ; The formula for performing directional feature shift operation on the initial latent semantic coding matrix is ​​as follows: ， in, This represents the initial latent semantic encoding matrix after the directional feature offset operation; This represents the initial latent semantic encoding matrix; A scalar step size parameter representing the degree of evolution of the target facial attributes; Indicates the orientation of the topological evolution manifold of the facial attribute to be edited; Perform matrix inversion on the forward spatial affine transformation mapping operator to obtain the inverse spatial mapping operator. Use the inverse spatial mapping operator to perform topological inverse projection transformation on the evolutionary feature manifold image to output a low-dimensional evolutionary image. The low-dimensional evolution image is magnified and rotated to the physical coordinate system of the original ultra-high resolution source image using a high-order resampling interpolation algorithm to obtain a high-dimensional registered evolution image.

2. The super-resolution facial attribute non-destructive editing method according to claim 1, characterized in that, Forward spatial affine transformation mapping operator Represented as: ， in, The homogeneous coordinate matrix representing the set of core topological anchor points of the target face; Represents a reference standard point set; Denotes the square of the Frobenius norm; Represents the regularization coefficient; This represents the L1 regularization constraint, L2 regularization constraint, or low-rank constraint on the forward affine transformation mapping operator.

3. The super-resolution facial attribute non-destructive editing method according to claim 1, characterized in that, The local feature submanifold image is represented as: ， in, Represents the coordinates in the image of a local feature submanifold Pixels at that location; Indicates the region of interest containing the target face; Represents the coordinates in the original ultra-high resolution source image Pixels at that location; This represents the discrete sampling matching operator, where the homogeneous coordinates in the original ultra-high resolution source image... After positive spatial affine transformation mapping operator The transformed homogeneous coordinates in the local feature submanifold image When they overlap, ,on the contrary .

4. The super-resolution facial attribute non-destructive editing method according to claim 1, characterized in that, Initial latent semantic encoding matrix Represented as: ， in, Indicates encoder; Represents a local feature submanifold image; Evolutionary feature manifold image Represented as: ， in, This indicates the decoder.

5. The super-resolution facial attribute non-destructive editing method according to claim 1, characterized in that, Inverse space mapping operator The acquisition process is as follows: Mapping operator for affine transformation of forward space By supplementing the information, a homogeneous space transformation matrix is ​​constructed. ; Alignment subspace transformation matrix Performing the standard matrix inversion operation yields ,extract The first two lines serve as the inverse space mapping operator. .

6. The super-resolution facial attribute non-destructive editing method according to claim 1, characterized in that, The high-dimensional registration evolution image is represented as: ， in, Represents coordinates in high-dimensional registration evolution images Pixels at that location; Represents coordinates in low-dimensional evolutionary images Pixels at that location; This represents a higher-order spatial resampling interpolation kernel function; This represents the inverse space mapping operator.

7. The super-resolution facial attribute non-destructive editing method according to claim 1, characterized in that, After obtaining the high-dimensional registration evolution image, the following is also included: A two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation is used to perform convolution filtering on high-dimensional registration evolution images to extract low-frequency structural semantic components that characterize the evolution of macroscopic attributes. The original ultra-high resolution source image is convolved and filtered using a two-dimensional spatial low-pass filter operator with a preset bandwidth standard deviation to obtain a filtered image; the difference between the original ultra-high resolution source image and the filtered image is calculated to extract the high-frequency physical residual components that characterize the original absolute physical properties. After performing a matrix linear weighted summation operation on the low-frequency structural semantic components and the high-frequency physical residual components, the result is numerically truncated to reconstruct a frequency domain reconstructed image containing the evolutionary macroscopic structure and the original microscopic texture. Based on the two-dimensional geometric contour and anisotropic spatial gradient distribution of the target face in the original ultra-high resolution source image, an Alpha weight mask is constructed with the core facial region as the high-value retention area and the surrounding non-target edges exhibiting non-linear Gaussian smooth decay. By using an alpha weight mask, a pixel-level spatial gradient blending operation is performed on the frequency domain reconstructed image and the original super-resolution source image to obtain a super-resolution edited image.

8. The super-resolution facial attribute non-destructive editing method according to claim 7, characterized in that, Low-frequency structural semantic components Represented as: ， in, Represents high-dimensional registration evolution images; This indicates that it has a preset bandwidth standard deviation. Two-dimensional spatial low-pass filter operator; Represents a discrete two-dimensional convolution operation; High-frequency physical residual components Represented as: ， in, This represents the original ultra-high resolution source image; Frequency domain reconstructed image Represented as: ， in, Indicates the numerical truncation operator; Indicates the texture preservation factor; The alpha weight mask is represented as follows: ， in, Representing coordinates Weight mask at the location; Representing coordinates The square of the Euclidean distance to the boundary; Represents the two-dimensional geometric topological edge of the region of interest containing the target face; The standard deviation of the Gaussian attenuation rate at the nonlinear edges of the control mask represents the control mask's nonlinear edge smoothing attenuation rate. Super-resolution image editing Represented as: ， in, Represents the Alpha weight mask; This represents a pixel-level multiplication operator.

9. A super-resolution facial attribute non-destructive editing device, characterized in that, include: The forward mapping parameter acquisition module is used to input the original ultra-high resolution source image into the local perception network and output the core topological anchor point set of the target face. The forward spatial affine transformation mapping operator is obtained by solving the problem with the objective of minimizing the difference between the core topological anchor point set after mapping by the forward spatial affine transformation mapping operator and the reference standard point set. The region of interest processing module is used to perform affine cropping and normalized projection on the region of interest containing the target face in the original ultra-high resolution source image using the forward spatial affine transformation mapping operator to obtain a local feature submanifold image. The facial attribute evolution module is used to input the local feature submanifold image into the encoder of the adversarial generative network (GCN) and output an initial latent semantic encoding matrix. Based on the topological evolution manifold direction of the facial attribute to be edited, the initial latent semantic encoding matrix undergoes a directional feature shift operation, which is then input into the decoder of the GCN to output an evolved feature manifold image. The facial attribute to be edited is aging, and the steps for obtaining its topological evolution manifold direction include: A predetermined number of initial latent semantic encoding matrix samples are obtained by randomly sampling in the latent space of the pre-trained generative network; The images generated from the initial latent semantic encoding matrix samples are scored using a pre-trained age classification attribute evaluation network, and then classified into a young feature set and an aging feature set. Using initial latent semantic encoding matrix samples from two feature sets as training data, a linear support vector machine classifier is trained to find the linear hyperplane that maximizes the distinction between the two feature sets. The orthogonal normal vector of this linear hyperplane in the latent space is used as the direction of the topological evolution manifold for the aging facial attribute to be edited. ; The formula for performing directional feature shift operation on the initial latent semantic coding matrix is ​​as follows: ， in, This represents the initial latent semantic encoding matrix after the directional feature offset operation; This represents the initial latent semantic encoding matrix; A scalar step size parameter representing the degree of evolution of the target facial attributes; Indicates the orientation of the topological evolution manifold of the facial attribute to be edited; The inverse transformation module is used to perform matrix inversion on the forward spatial affine transformation mapping operator to obtain the inverse spatial mapping operator. The inverse spatial mapping operator is then used to perform topological inverse projection transformation on the evolutionary feature manifold image to output a low-dimensional evolutionary image. The evolutionary image acquisition module is used to enlarge and rotate the low-dimensional evolutionary image to the physical coordinate system of the original ultra-high resolution source image using a high-order resampling interpolation algorithm, so as to obtain a high-dimensional registered evolutionary image.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the super-resolution facial attribute non-destructive editing method according to any one of claims 1 to 8.

Images