A face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation

Abstract

The application relates to the technical field of face image processing, and particularly discloses a face image super-resolution reconstruction method based on structure anchoring and edge frequency adaptive modulation, which comprises the following steps: after a low-resolution face image is acquired, a feature extraction unit is used to obtain multi-scale features, intermediate reconstruction features and deep or latent features; structure modulation parameters are generated based on face structure priori, and the deep or latent features are subjected to structure anchoring modulation to obtain structure anchoring features; edge response and frequency response are extracted under the constraint of the structure anchoring features, and a detail base under edge constraint is fused and generated; adaptive intensity refining parameters are generated according to the structure anchoring features, the edge frequency detail base and the intermediate reconstruction features, and the edge frequency details are selected, fused and injected; the scheme combines semantic structure position constraint, edge constraint high-frequency details and detail adaptive intensity injection into a continuous recovery chain, and can improve the recovery stability of structure-sensitive details in the low-resolution face image.

Description

Technical Field

[0001] This application relates to the field of face image processing technology, and specifically discloses a face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation. Background Technology

[0002] Facial images, as the most recognizable biometric images, are widely used in various civilian and security scenarios such as video surveillance, public security identity verification, mobile terminal facial recognition, natural human-computer interaction, smart park personnel management, smart construction site personnel attendance, long-distance image verification, and criminal investigation image tracing. They are the core input data for machine vision and intelligent analysis systems, and image quality directly determines the accuracy and stability of subsequent intelligent tasks such as face detection, key point localization, identity recognition, and behavior analysis.

[0003] In real-world scenarios, facial image acquisition is susceptible to multiple complex degradation factors. Specifically, long-distance imaging results in an extremely low percentage of effective facial pixels; limited camera hardware resolution, lens distortion, and sensor noise introduce inherent image distortion; movement of the subject and device jitter cause motion blur; video transmission and storage compression lead to loss of image details and blockiness; low-light environments such as nighttime, backlighting, and low light reduce the image signal-to-noise ratio; and issues such as masks, hats, glasses, hair occlusion, and pose shifts can cause missing local facial structures. These degradation problems, individually or in combination, result in low-resolution facial images that generally suffer from low resolution, weakened edge boundaries, missing skin texture and facial feature details, blurred local structures, and severe artifact interference. This significantly reduces the visual readability of facial images, causing subsequent intelligent analysis tasks to fail in feature extraction, drastically reduce recognition accuracy, and distort verification results, making it difficult to meet the high-precision and high-reliability requirements of practical applications.

[0004] Super-resolution facial image reconstruction technology is the core approach to solving the aforementioned problems. Its core principle is to accurately deduce and recover high-resolution facial images with complete details, accurate structure, and realistic texture from severely degraded low-resolution facial images using deep learning, signal processing, and other techniques. This effectively compensates for hardware imaging shortcomings, counteracts scene degradation interference, significantly improves the visual effect and image quality of facial images, and provides high-quality feature input for subsequent intelligent facial analysis tasks, thereby enhancing the stability and accuracy of the entire intelligent facial system. It has extremely high engineering application value and research significance.

[0005] Currently, the industry has developed several mature face super-resolution reconstruction technologies. The mainstream methods are mainly divided into five categories: reconstruction methods based on convolutional neural networks, reconstruction methods based on generative adversarial networks, reconstruction methods based on Transformer and attention mechanisms, structural reconstruction methods based on face priors, and reconstruction methods based on frequency domain and edge enhancement. Each method has its inherent technical advantages, but also significant limitations. Specifically, methods based on face structure priors can rely on prior information such as facial key points, face segmentation maps, and the distribution of facial features to constrain the overall facial contour and feature layout, correcting overall structural distortions in the reconstructed image. Methods based on frequency domain and edge enhancement can specifically enhance high-frequency information in the image, repairing high-frequency details such as facial edge contours and facial textures, and mitigating blurring and distortion. Methods based on attention mechanisms and feature modulation can adaptively select key feature regions in the image, enhance effective facial feature responses, suppress invalid background noise interference, and optimize feature fusion effects.

[0006] However, in demanding real-world scenarios such as high-magnification super-resolution reconstruction and complex hybrid degradation, reconstruction schemes relying solely on a single technical path have significant technical shortcomings. The deficiencies of various individual methods are further amplified, making it impossible to achieve high-precision and robust facial detail reconstruction. First, existing structural prior techniques mostly participate in reconstruction optimization indirectly through auxiliary image input, feature cascade embedding, and loss function supervision. They can only constrain the overall layout and macroscopic structure of the face, failing to provide precise positional constraints on fine-grained sensitive details such as eyeliner, lip line, tooth boundaries, beard texture, and fine facial contours. This easily leads to problems such as misalignment of facial features, texture shift, and local structural distortion, resulting in a mismatch between the reconstructed detail position and the real facial structure. Second, frequency domain enhancement and edge enhancement techniques lack precise constraints from structured positional priors. In areas where the facial structure is uncertain or severely degraded, they tend to over-amplify pseudo-textures and noisy textures, producing false details and edge jaggedness. In key boundary areas of real facial features, they suffer from insufficient enhancement and incomplete detail repair, ultimately resulting in messy textures, blurred edges, and poor visual realism in the reconstructed image.

[0007] Meanwhile, existing conventional attention mechanisms and feature modulation modules can only roughly determine the importance of regions, that is, they can only filter out the core area of ​​the face and the background area. They cannot accurately distinguish reliable real edges, effective high-frequency details and pseudo edges and noise pseudo details generated by degradation in the image. It is difficult to achieve targeted detail enhancement and noise suppression. They cannot answer the core question of "which edge frequency details are reliable and need to be enhanced, and which pseudo details need to be suppressed". Their adaptive modulation capability is seriously insufficient.

[0008] In summary, high-quality low-resolution face super-resolution reconstruction urgently needs to overcome three core technical challenges simultaneously: First, to achieve precise positioning of fine-grained facial details, ensuring that the restored details strictly conform to the actual structural position of the face; second, to restore edges and high-frequency details with reasonable physical morphology, eliminating the generation of false textures and ensuring the authenticity of details; and third, to achieve adaptive control of detail enhancement intensity, dynamically optimizing the enhancement intensity according to the degree of degradation and feature reliability of different regions of the image, taking into account both detail restoration and noise suppression.

[0009] However, existing technical solutions are all single-module, single-dimensional optimization designs, which can only address specific problems such as structural distortion, blurred details, or feature selection, lacking a systematic reconstruction logic. Currently, no technical solution can organically combine accurate facial structure anchoring, accurate edge frequency detail modeling, and adaptive detail modulation, failing to form a continuous reconstruction chain of "structural localization - detail modeling - adaptive optimization" with sequential dependencies and closed-loop linkage. As a result, existing reconstruction algorithms generally suffer from structural misalignment, detail distortion, severe artifacts, and poor robustness in high-magnification and complex degradation scenarios, making it difficult to meet the needs of high-precision practical applications such as security monitoring and identity verification.

[0010] In view of this, the present invention provides a face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation to solve the above problems. Summary of the Invention

[0011] The purpose of this invention is to propose a new technical solution for reliably restoring sensitive details of facial structure by utilizing structural anchoring, edge frequency detail modeling, and adaptive modulation in synergy.

[0012] To achieve the above objectives, the present invention provides the following basic solution: A face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation includes the following steps: Step S1, Image Acquisition and Preprocessing: Acquire the low-resolution face image to be reconstructed, preprocess the low-resolution face image, and input it into the face image super-resolution reconstruction network; Step S2, Multi-scale feature extraction: Multi-scale feature extraction is performed through the feature extraction unit to obtain intermediate reconstructed features and deep features or latent features; Step S3, latent space semantic structure anchoring: Generate structural prior features based on the deep features or latent features, convert the structural prior features into structural modulation parameters, and perform structural anchoring modulation on the deep features or latent features; Step S4, Edge Constraint Detail Generation: Based on the structural anchoring feature extraction and fusion of edge response and frequency response, an edge frequency detail basis constrained by the face structure position is generated. The edge frequency detail basis is used to characterize the candidate detail morphology of the face structure sensitive region. Step S5, Adaptive intensity refinement: Generate adaptive modulation parameters based on the structural anchoring features, the edge frequency detail basis, and the intermediate reconstruction features; select, fuse, and inject the edge frequency detail basis to obtain detail modulation features; Step S6, Decoding, Reconstruction and Image Output: Fuse structural anchoring features, detail modulation features and reconstruction features from the decoding stage, and generate a high-resolution face image through the reconstruction unit.

[0013] Based on the above steps, this invention further defines the following aspects regarding latent space semantic structure anchoring, edge constraint detail generation, adaptive intensity refinement, and training methods: Furthermore, the structural prior features include facial landmark heatmaps, facial semantic parsing maps, facial component region maps, facial contour maps, structural attention maps, structural features output by the facial structure prediction branch, or combinations of at least two of the above features.

[0014] Furthermore, the structural modulation parameters include at least one of affine modulation parameters, spatial attention weights, channel attention weights, region gating weights, or normalized modulation parameters.

[0015] Furthermore, when the structural modulation parameters are selected as affine modulation parameters, the affine modulation parameters include scaling parameter γ and translation parameter β, and the structural anchoring modulation is achieved by performing at least one of scaling, translation, element-wise multiplication and addition, gated fusion or normalization modulation on the deep features or latent features.

[0016] Furthermore, the edge-constrained high-frequency detail base includes high-frequency sub-band features, horizontal edge features, vertical edge features, gradient magnitude features, directional edge features, frequency band selection features, or a fusion result of at least two of the above features, used to characterize the directional boundary, high-frequency texture, or local detail morphology of at least one structurally sensitive region among eye contour, eyebrow, nose wing, lip line, tooth boundary, beard, and facial contour.

[0017] Furthermore, the high-frequency response is obtained through wavelet transform, high-pass filtering, Fourier transform, learnable frequency filtering, frequency domain convolution, directional high-frequency extraction, or at least two of the above methods; the edge response is obtained through gradient operator, Sobel operator, Scharr operator, Laplacian operator, directional convolution, edge detection network, or at least two of the above methods.

[0018] Furthermore, the adaptive modulation parameters include at least one of the following: a frequency band selection gate, a detail reliability gate, an output injection gate, a spatial attention weight, a channel attention weight, or a region adaptive weight. The frequency band selection gate is used to control the weights of different high-frequency sub-bands or directional frequency components participating in the reconstruction. The detail reliability gate is used to determine the credibility of the edge frequency detail basis at the current spatial location or channel dimension. The output injection gate is used to control the intensity of the modulated detail features injected into the decoded reconstruction features.

[0019] Furthermore, the adaptive modulation parameters enhance at least one structurally sensitive detail among the eye contour, eyebrows, nose wings, lip lines, tooth boundaries, beard, and facial contour, based on the structural reliability or degradation degree of different facial regions, and suppress pseudo-high-frequency responses in smooth areas or structurally uncertain areas such as cheeks and forehead.

[0020] Furthermore, the face image super-resolution reconstruction network adopts any one or any combination of encoder-decoder structure, convolutional neural network structure, residual network structure, Transformer structure, U-Net structure or image restoration network structure. When training the face image super-resolution reconstruction network, pixel reconstruction loss, structural prior supervision loss, edge loss, frequency loss, perceptual loss, identity preservation loss or a combination of at least two of the above losses are used.

[0021] Corresponding to the above method steps, the face image super-resolution reconstruction network of the present invention mainly includes the following functional modules: The image acquisition module is used to acquire low-resolution face images to be reconstructed. The feature extraction module is used to perform multi-scale feature extraction on the low-resolution face image to obtain intermediate reconstruction features and deep features or latent features. The latent space semantic structure anchoring module is used to generate semantic structure prior features, convert the structure prior features into structure modulation parameters, and perform structure anchoring modulation on the deep features or latent features according to the structure modulation parameters. The edge constraint detail generation module is used to extract edge response and frequency response under structural anchoring feature constraints, and fuse them to generate edge frequency detail basis; An adaptive intensity refinement module is used to generate adaptive modulation parameters based on the structural anchoring features, the edge frequency detail basis, and the intermediate reconstruction features, and to select, fuse, and inject the edge frequency detail basis using the adaptive modulation parameters; The image reconstruction output module is used to generate high-resolution face images based on structural anchoring features, detail modulation features, and reconstruction features from the decoding stage.

[0022] Based on the above technical solution, the beneficial effects of the present invention are as follows: 1. Compared with the prior art, the purpose of this invention is to solve the problems of inaccurate structural position, blurred edge shape, unbalanced local detail response, and easy amplification of pseudo texture in existing low-resolution face image reconstruction. This invention does not simply improve image sharpness, but determines the detail recovery position through structural anchoring, characterizes the available local detail shape through edge frequency detail basis, and controls the detail injection intensity through adaptive modulation, thereby achieving structure-dependent detail recovery.

[0023] 2. Compared with existing technologies, this method abandons the traditional approach of relying solely on loss supervision and feature stitching to indirectly constrain facial structure. Instead, it innovatively introduces a structural anchoring mechanism: generating facial key points and semantic parsing from deep features, and transforming them into modulation parameters such as affine scaling and spatial gating that directly act on latent features. This locks the positions of facial features and contours at a global level, ensuring that the fine-grained details such as eyeliner, lip line, and nose wings that are subsequently restored strictly conform to the real facial structure. This thoroughly improves the defects of facial feature misalignment, texture misalignment, and local structural distortion under high-magnification reconstruction.

[0024] 3. Compared with existing technologies, an edge frequency detail basis is constructed, which integrates wavelet or high-pass frequency domain high-frequency features with Sobel / Scharr spatial edge features to generate only candidate details that are attached to the boundaries of the face structure. All high-frequency and edge extractions are constrained by structural anchoring features, avoiding blindly amplifying noise and compressing artifacts without structural constraints. It distinguishes between real facial feature boundaries and false noise textures, and reconstructs images with clean edges and realistic textures without burrs.

[0025] 4. Compared with other technologies, the three-layer adaptive modulation gating can dynamically adjust the enhancement intensity of details according to the facial region: enhance high-frequency edge details in structurally reliable areas such as eyebrows, lip lines, teeth, and facial contours; and automatically suppress pseudo-high-frequency responses in smooth or structurally blurred areas such as cheeks and forehead, achieving "enhancement where it should be enhanced and suppression where it should be suppressed", adaptively adapting to different degrees of degradation and different facial regions, and greatly improving robustness.

[0026] 5. In summary, as Figure 1 As shown, compared with the representative advanced methods listed in the table, the method of the present invention achieves the best results in all six commonly used evaluation indicators, including PSNR, SSIM, and LPIPS, and demonstrates comprehensive advantages over other representative methods in terms of pixel fidelity, structural consistency, and perceptual quality. Figure 6 The visualization results further demonstrate that the present invention can recover clearer and more stable local details in structurally sensitive areas such as eye contours, pupil boundaries, nose wings, lip lines, and facial contours, while reducing boundary blurring, structural misalignment, and pseudo-textures. (See Table 2 and...) Figure 7As shown, each key design exhibits a clear and progressive quality improvement in the aforementioned structurally sensitive areas. This demonstrates that the present invention forms a continuous recovery mechanism with a priori dependence and synergistic effect through structural anchoring, edge frequency detail generation, and adaptive intensity refinement, rather than a simple superposition of existing structural a priori, high-frequency enhancement, and feature modulation methods. This reflects the originality, necessity, and effectiveness of the technical design of the present invention. Attached Figure Description

[0027] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments are briefly described below.

[0028] Figure 1 A flowchart illustrating the face image super-resolution reconstruction method of the present invention is shown; Figure 2 This diagram illustrates the overall structure and core module layout of the face image super-resolution reconstruction network of the present invention. Figure 3 A schematic diagram of the latent space semantic structure anchoring module of the present invention is shown; Figure 4 A schematic diagram of the edge constraint detail generation module of the present invention is shown; Figure 5 A schematic diagram of the adaptive strength refinement module of the present invention is shown; Figure 6 The following diagrams show the results of face super-resolution reconstruction and local details of the method of this invention and representative advanced methods. Figure 7 The key functional modules of the present invention and the effectiveness of loss supervision are shown in the diagram, along with the distribution of local reconstruction quality gain. Detailed Implementation

[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0030] The following combination Figures 1 to 7 The specific embodiments, core module structure, and technical effects of the present invention are described below. A face image super-resolution reconstruction method based on structure anchoring and edge frequency adaptive modulation includes the following steps: Step S1, acquiring a low-resolution face image to be reconstructed, preprocessing the low-resolution face image, and inputting it into a face image super-resolution reconstruction network.

[0031] First, the low-resolution face image is a face region image from a surveillance video frame, mobile terminal image, long-distance acquisition image, compressed transmission image, drone inspection image, smart construction site acquisition image, smart park acquisition image, or low-bandwidth transmission image. The low-resolution face image is then cropped, normalized, aligned, pre-upsampled, or otherwise pre-processed to obtain the input image to be reconstructed.

[0032] Regarding the face image super-resolution reconstruction network: The face image super-resolution reconstruction network adopts any one or any combination of encoder-decoder structure, convolutional neural network structure, residual network structure, Transformer structure, U-Net structure or image restoration network structure. When training the face image super-resolution reconstruction network, pixel reconstruction loss, structural prior supervision loss, edge loss, frequency loss, perceptual loss, identity preservation loss or a combination of at least two of the above losses are used.

[0033] The training data includes low-resolution training images and their corresponding high-resolution target images. The low-resolution training images are input into the reconstruction network to obtain super-resolution output. The pixel reconstruction loss, structural supervision loss, high-frequency loss, and edge loss are calculated using the high-resolution target images to achieve a joint collaborative supervision mechanism. The network parameters are then updated through backpropagation.

[0034] The total loss can be summarized as follows: In the formula, Represents the total loss function; This represents the pixel reconstruction loss, used to constrain the consistency between the super-resolution reconstructed image and the high-resolution target image in pixel space; The structural prior supervision loss is used to constrain the consistency between the face key region heatmap predicted by the structural anchoring module and the corresponding structural label. This represents the high-frequency consistency loss, used to constrain the consistency between the super-resolution reconstructed image and the high-resolution target image in the directional high-frequency subband; This represents the edge consistency loss, used to constrain the consistency of the spatial edge response between the super-resolution reconstructed image and the high-resolution target image; , and These represent the non-negative weighting coefficients of the corresponding loss terms.

[0035] The structural prior supervision loss is used to constrain the consistency between structural prior features and real face structural information. The edge loss is used to constrain the consistency between the super-resolution output image and the real high-resolution image in edge response. The frequency loss is used to constrain the consistency between the super-resolution output image and the real high-resolution image in high-frequency or frequency domain response.

[0036] Regarding the training phase, pairs of low-resolution and high-resolution face images can be used as training samples. Low-resolution face images can be obtained through real-world acquisition or by simulating downsampling, blurring, compression, noise perturbation, or composite degradation of high-resolution images. During the inference phase, the system does not require real high-resolution images as input; it only needs to input the low-resolution face image to be processed to output the corresponding high-resolution face image.

[0037] Regarding the inference phase, the inference phase only requires inputting the low-resolution face image to be reconstructed into the trained network to obtain the super-resolution face image. High-resolution target images, structural labels, and various training losses are not used as inference inputs. The information required for structural anchoring and detail modulation is automatically generated by the network based on the internal features of the input image.

[0038] Step S2: Multi-scale feature extraction is performed on the low-resolution face image by the feature extraction unit to obtain intermediate reconstruction features and at least one deep feature or latent feature.

[0039] Specifically, multi-scale feature extraction is performed on the input image using an encoder or other feature extraction network. Shallow features mainly contain low-level texture and local edge information, intermediate reconstructed features retain a certain spatial resolution and semantic information, and deep features or latent features have a larger receptive field and stronger facial structure representation ability.

[0040] Step S3: Generate structural prior features related to the facial structure location based on the deep features or latent features, convert the structural prior features into structural modulation parameters, and use the structural modulation parameters to perform structural anchoring modulation on the deep features or latent features to obtain structural anchoring features.

[0041] The semantic structure prior features include facial landmark heatmaps, facial semantic parsing maps, facial component region maps, facial contour maps, structural attention maps, structural features output by the facial structure prediction branch, or a combination of at least two of the above features.

[0042] The structural modulation parameters include at least one of affine modulation parameters, spatial attention weights, channel attention weights, region gating weights, or normalized modulation parameters. When the structural modulation parameters are selected as affine modulation parameters, the affine modulation parameters include scaling parameters γ and translation parameters β. The structural anchoring modulation is achieved by performing at least one of scaling, translation, element-wise multiplication and addition, gating fusion, or normalized modulation on the deep features or latent features.

[0043] Regarding semantic structure anchoring modulation: Structure anchoring modulation is applied to the deep feature layer or latent feature layer of the reconstructed network, and the edge frequency detail basis and the adaptive modulation parameters are applied to the latent feature layer, at least one intermediate decoding layer, or a combination of the above layers.

[0044] Step S4: Based on the structural anchoring features, the edge response and frequency response in the intermediate reconstruction features are extracted and fused to generate an edge frequency detail basis constrained by the face structure position. The edge frequency detail basis is used to characterize the candidate detail morphology of the face structure sensitive region.

[0045] In step S4, the frequency response is obtained through wavelet transform, high-pass filtering, Fourier transform, learnable frequency filtering, frequency domain convolution, directional high-frequency extraction, or at least two of the above methods; in step S4, the edge response is obtained through gradient operator, Sobel operator, Scharr operator, Laplacian operator, directional convolution, edge detection network, or at least two of the above methods.

[0046] Step S5: Generate adaptive modulation parameters based on the structural anchoring features, the edge frequency detail bases, and the intermediate reconstruction features. Then, use the adaptive modulation parameters to select, fuse, and inject the edge frequency detail bases to obtain detail modulation features. In step S5, the adaptive modulation parameters are used to control at least one of the frequency band selection, detail reliability, and output injection strength, so that the edge frequency detail bases are enhanced in the structurally reliable region and suppressed in the smooth region or the structurally uncertain region.

[0047] In step S5, the adaptive modulation parameters include at least one of a frequency band selection gate, a detail reliability gate, an output injection gate, a spatial attention weight, a channel attention weight, or a region adaptive weight. The frequency band selection gate is used to control the weights of different high-frequency sub-bands or directional frequency components participating in the reconstruction. The detail reliability gate is used to determine the credibility of the edge frequency detail basis at the current spatial location or channel dimension. The output injection gate is used to control the intensity of the modulated detail features injected into the decoded reconstruction features.

[0048] The adaptive intensity refinement parameters enhance at least one structurally sensitive detail in the eye contour, eyebrows, nose, lip line, tooth boundaries, beard, and facial contour, based on the structural reliability or degradation degree of different facial regions, while suppressing pseudo-high-frequency responses in smooth areas or structurally uncertain areas such as cheeks and forehead.

[0049] The edge frequency constrained detail base includes high-frequency subband features, horizontal edge features, vertical edge features, gradient magnitude features, directional edge features, frequency band selection features, or a fusion of at least two of the above features, used to characterize the directional boundaries, high-frequency textures, or local detail morphology of at least one structurally sensitive region among eye contours, eyebrows, nose wings, lip lines, tooth boundaries, beards, and facial contours.

[0050] Step S6: The structural anchoring features, the detail modulation features, and the reconstruction features from the decoding stage are fused together, and a high-resolution face image is generated through the decoding and reconstruction unit.

[0051] The following describes the specific implementation of each step in conjunction with its functional modules. Its main design includes: The system comprises the following modules: an image acquisition module for acquiring a low-resolution face image to be reconstructed; a feature extraction module for performing multi-scale feature extraction on the low-resolution face image to obtain intermediate reconstruction features and deep or latent features; a latent space semantic structure anchoring module for generating structural prior features, converting the structural prior features into structural modulation parameters, and performing structural anchoring modulation on the deep or latent features according to the structural modulation parameters; an edge frequency detail basis generation module for extracting edge responses and frequency responses under the constraints of structural anchoring features, and fusing them to generate edge frequency detail basis; an adaptive modulation module for generating adaptive modulation parameters based on the structural anchoring features, the edge frequency detail basis, and the intermediate reconstruction features, and using the adaptive modulation parameters to select, fuse, and inject the edge frequency detail basis; and a reconstruction output module for generating a high-resolution face image based on the structural anchoring features, detail modulation features, and reconstruction features from the decoding stage.

[0052] For ease of explanation, this embodiment will refer to the latent space semantic structure anchoring module as LSGA, the edge constraint detail generation module as EAHC, the adaptive intensity refinement module as ARDR, and the frequency edge detail loss supervision used in the training phase as FEDS.

[0053] like Figure 2As shown, this embodiment employs an encoder-decoder multi-scale reconstruction network. Low-resolution face images undergo shallow feature extraction and multi-scale encoding to obtain latent layer deep features, which are then sequentially processed through a latent space semantic structure anchoring module, an edge constraint detail generation module, and an adaptive intensity refinement module to complete structure anchoring, candidate detail generation, and reliable detail selection and injection. Edge constraint detail generation and adaptive intensity refinement modules are further implemented in decoding layers 3 and 2 to progressively restore local details with increasing resolution. Encoding layers E3, E2, and E1 are fused with features from decoding layers 3, 2, and 1 via skip connections, respectively, ultimately generating a high-resolution face image via the reconstruction output unit. Frequency edge detail loss supervision only constrains the frequency and edge responses of the reconstruction results during the training phase and does not participate in the forward processing during the inference phase.

[0054] like Figure 3 As shown, the latent space semantic structure anchoring module is used to determine the locations where facial details should be restored. For severely degraded low-resolution facial images, structural cues such as the corners of the eyes, the wings of the nose, the lip lines, and the facial contours may be compressed, blurred, or lost. If only pixel regression or ordinary attention mechanisms are relied upon, the model is prone to generating local details with positional shifts.

[0055] In one implementation, the structural anchoring module receives deep or latent features and uses a structural prediction branch to generate a keypoint heatmap or semantic structure map. Subsequently, modulation parameters are generated through one or more convolutional layers, normalization layers, attention layers, or a multilayer perceptron. The modulation parameters may include scaling parameters γ and translation parameters β, or spatial attention weights, channel attention weights, or region gating weights.

[0056] Specifically: the semantic structure anchoring module inputs latent layer deep features The processing procedure involves obtaining facial geometric and semantic features through a lightweight structure extractor and predicting structural response maps of one or more key facial regions. These structural response maps receive structural supervision during the training phase, but the structural information is not limited to auxiliary branches; it is further used to generate scaling parameters. Translation parameters Alternatively, other equivalent gating modulation parameters can be used to affinely modulate the backbone features of the latent layer. Preferably, only a portion of the channels are modulated, while the remaining channels retain their original semantic information, thus avoiding complete structural prior coverage of the backbone representation. The structural modulation relationship can be summarized as follows: , in, This represents the structural anchoring characteristics obtained after structural anchoring modulation. This represents the latent deep features of the input semantic structure anchoring module; This represents the scaling parameters generated based on the structural response map; This represents the translation parameters generated based on the structural response diagram; element-wise multiplication The latent space semantic structure anchoring module stabilizes the positions and contours of facial features before entering the decoding stage, providing a global structural reference for subsequent edge frequency detail restoration. This differs from schemes that merely stitch keypoint heatmaps into the network, calculate keypoint losses, or use them only as auxiliary information for the discriminator. Structure anchoring modulation obtains structural anchoring features through scaling, translation, element-wise multiplication and addition, gated fusion, or normalization modulation of deep features. This approach enables subsequent decoding to restore details with a relatively stable facial structure reference.

[0057] In alternative implementations, structural prior features can be generated by an external face landmark detector, a face parsing network, a 3D face estimation module, a semantic segmentation module, or by structural branches learned end-to-end within a super-resolution network.

[0058] like Figure 4 As shown, the edge frequency detail basis generation module is used to provide local detail sources. Facial structurally sensitive details typically possess both edge and high-frequency attributes. Using only frequency domain enhancement may not guarantee that details adhere to the boundaries of the real facial structure, while using only edge enhancement may ignore frequency differences in detail texture. Therefore, this embodiment uses both edge response and frequency response to construct candidate detail representations.

[0059] In one implementation, the system performs frequency decomposition on intermediate reconstructed features or structural anchoring features to obtain low-frequency features, high-frequency features, or multiple directional sub-band features. Frequency decomposition methods may include Haar wavelet transform, discrete wavelet transform, Fourier transform, high-pass filtering, learnable frequency filtering, or convolutional high-frequency extraction.

[0060] Simultaneously, the system extracts edge responses from intermediate reconstructed features, structural anchoring features, or intermediate reconstructed images. Edge response methods can include Scharr operator, Sobel operator, Laplacian operator, gradient magnitude calculation, directional convolution, or learnable edge detection networks.

[0061] Subsequently, high-frequency sub-band features, directional edge features, gradient magnitude features, or combinations thereof are fused to generate an edge frequency detail basis. This detail basis is not used directly as the output texture, but rather serves as a candidate detail source for subsequent adaptive modulation modules.

[0062] Specifically, the edge frequency detail basis generation module takes the latent layer or intermediate decoding layer features as input and first obtains feature branches for constructing detail basis and bypass features for subsequent output control through channel transformation. Directional high-frequency decomposition is performed on the detail basis branches, preferentially obtaining high-frequency responses in the LH, HL, and HH directions. Simultaneously, edge operators are used to extract horizontal and vertical edge responses, forming edge amplitude features. The directional high-frequency responses and spatial edge responses together constitute candidate detail basis.

[0063] In a preferred embodiment, directional high-frequency decomposition employs three-way Haar wavelet detail features, and edge extraction uses the Scharr operator; resulting in: ,in, Represents the edge frequency detail basis; , and These represent the three directional high-frequency detail features obtained from Haar wavelet decomposition, where... Indicates low-pass filtering. This indicates a high-pass filter, with the two subscript letters corresponding to two spatial directions; The text indicates that the edge response features extracted using the Scharr operator are represented; the curly braces indicate that the three directional high-frequency detail features and the edge response features together constitute the edge frequency detail basis. The edge frequency detail basis generation module is used to provide interpretable candidate detail sources, but does not directly assume that all high-frequency responses should be enhanced.

[0064] like Figure 5 As shown, the adaptive intensity refinement module is used to determine which details are reliable and should be enhanced. Different facial regions have different structural determinism and detail requirements. For example, areas such as the eyes, lip lines, tooth boundaries, beards, and facial contours usually require stronger edge frequency detail compensation; while relatively smooth areas or structurally uncertain areas such as cheeks and foreheads should avoid excessive enhancement of pseudo-high frequencies.

[0065] In one implementation, the adaptive modulation module generates three levels of modulation weights based on high-frequency directional subbands, edge constraint details, and local response features: band selection weights, detail reliability weights, and backbone injection weights. The band selection weights are used to select different high-frequency subbands or directional frequency components; the detail reliability weights are used to determine the reliability of the detail basis in the current region; and the output injection weights are used to control the intensity of the modulated detail features injected into the decoded features.

[0066] Adaptive intensity refinement can be achieved through Sigmoid gating, attention mechanisms, element-wise multiplication, channel-wise weighting, spatial weighting, residual injection, and convolutional fusion after feature concatenation. The modulated detail features are then fused with structural anchoring features or decoding features and input into the subsequent decoding layer.

[0067] Specifically: The adaptive response modulation module takes into account the current layer backbone features, directional high-frequency subbands, edge responses, candidate detail bases, and bypass features. The processing includes at least the following three types of controls: Directional subband selection gate: controls the retention ratio of different directional high-frequency subbands for different local regions; Edge detail modulation gate: combines the directional high-frequency response and spatial edge response to determine whether the candidate high frequencies have a credible edge dependency relationship; Adaptive intensity injection gate: controls which local responses in the bypass features can be converted into detail residuals and enter the main recovery path. The selected details are injected into the backbone features as residuals, and a learnable scaling factor s can be set to control the overall injection intensity. , in, This represents the backbone features of the current layer input; This represents the candidate detail features obtained after directional subband selection and edge detail modulation; This represents the gate weights generated by the adaptive strength injection gate; Represents the activation function of the Gaussian error linear unit; This represents element-wise multiplication; Indicates the use of channel mapping and output projection. Convolution operation; This represents a learnable scaling factor that controls the overall injection strength of the detailed residuals; " indicates scalar multiplication; This represents the detail modulation features obtained after detail residual injection. The adaptive intensity refinement module preserves directional details in locations with stable edge references, such as eyelids, lip lines, nose wings, and facial contours, while suppressing unreliable high frequencies caused by noise, ringing, or interpolation in smooth areas like cheeks and forehead, reducing the risk of pseudo-textures.

[0068] After completing the detailed description of the network structure and each key module, the following uses quantitative comparison, ablation experiments and visualization analysis to verify the overall reconstruction performance of the method of the present invention and the role of each key design.

[0069] To verify the overall performance of the method of this invention in the face image super-resolution reconstruction task, and the role of the latent space semantic structure anchoring module, edge constraint detail generation module, adaptive intensity refinement module, and frequency edge detail loss supervision in the reconstruction results, this embodiment conducts ×8 face image super-resolution reconstruction experiments on a representative test set, and uses Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Learned Perceptual Patch Similarity (LPIPS) as evaluation metrics. Higher PSNR and SSIM values ​​indicate higher pixel consistency and structural similarity between the reconstructed image and the high-resolution target image; lower LPIPS values ​​indicate smaller perceptual differences between the reconstructed image and the high-resolution target image.

[0070] To evaluate the overall reconstruction performance of the method of the present invention, the method of the present invention is compared with the Bicubic interpolation method and representative methods from previous years such as DIC, NLSN, KDFSR, CTCNet, SFMNet, MPSL, WFEN, SADT, and PortraitSR. The specific quantitative results are shown in Table 1.

[0071] Table 1. Comparison of quantitative indicators with representative advanced methods over the years

[0072] As shown in Table 1, under the representative methods listed in the table and the evaluation conditions adopted in this embodiment, the method of the present invention achieves PSNR, SSIM and LPIPS of 28.049, 0.8418 and 0.1333 on the Helen test set, and PSNR, SSIM and LPIPS of 28.055, 0.8209 and 0.1220 on the CelebA test set, respectively. All six evaluation indicators achieve the best results in the table.

[0073] Compared to PortraitSR, the method of this invention improves PSNR and SSIM by 0.343 and 0.0134, respectively, and reduces LPIPS by 0.0115 on the Helen test set; on the CelebA test set, it improves PSNR and SSIM by 0.147 and 0.0053, respectively. LPIPS on the CelebA test set is 0.0014 lower than the second-best NLSN in the table. These results demonstrate that the method of this invention can improve pixel fidelity and structural consistency while further enhancing the perceptual quality of reconstructed face images, thus validating the effectiveness of the continuous recovery chain comprised of structural position constraints, edge frequency detail modeling, and adaptive detail refinement.

[0074] To further demonstrate the reconstruction effect of the method of the present invention, Figure 6The paper presents a low-resolution input image, a reconstructed image using a representative advanced method, a reconstructed image using the method of this invention, and a high-resolution target image, with local magnification of the eye region.

[0075] like Figure 6 As shown, compared with low-resolution input images and the listed representative advanced methods, the method of the present invention achieves clearer and more stable reconstructed details in the eye contour, pupil boundary, and adjacent structural regions, while reducing local boundary blurring and structural misalignment. This visualization result corroborates the quantitative evaluation results in Table 1, demonstrating that the method of the present invention not only improves the overall evaluation index but also enhances the local restoration quality of facial structurally sensitive areas.

[0076] To verify the contribution of each key design element to the reconstruction performance, a reconstruction network without the latent space semantic structure anchoring module (LSGA), edge constraint detail generation module (EAHC), adaptive intensity refinement module (ARDR), and frequency edge detail loss supervision (FEDS) was used as the baseline network, and ablation experiments were conducted while keeping other experimental conditions consistent. FEDS was only used during the network training phase to constrain the frequency and edge responses of the reconstruction results and was not used as input during the inference phase. Specific results are shown in Table 2.

[0077] Wherein, Baseline represents the baseline network; “+LSGA”, “+LSGA+EAHC”, “+LSGA+EAHC+ARDR”, and “+LSGA+EAHC+ARDR+FEDS” constitute the experimental chain of progressively added key designs; “+EAHC+FEDS” is an auxiliary combination experiment without LSGA, used to observe the combined effect of edge constraint detail generation and frequency edge detail loss supervision, as shown in Table 2 below: Table 2 Ablation results of key functional modules of the present invention

[0078] LSGA represents the latent space semantic structure anchoring module, EAHC represents the edge-constrained detail generation module, ARDR represents the adaptive intensity refinement module, and FEDS represents the frequency edge detail loss supervision module. Table 2 shows that after adding LSGA to the baseline network, the PSNR and SSIM on the CelebA test set increased from 27.7890 and 0.8119 to 27.8708 and 0.8145, respectively, while LPIPS decreased from 0.1311 to 0.1293; on the Helen test set, the PSNR and SSIM increased from 27.7802 and 0.8293 to 27.9097 and 0.8338, respectively, while LPIPS decreased from 0.1612 to 0.1509. These results demonstrate that latent space semantic structure anchoring can provide a stable reference for the facial structure location in subsequent detail recovery and improve the overall reconstruction quality.

[0079] Table 1 shows the performance comparison of the proposed method with representative state-of-the-art (SOTA) face super-resolution methods such as Bicubic, DIC, NLSN, KDFSR, CTCNet, SFMNet, MPSL, WFEN, SADT, and PortraitSR on the Helen and CelebA datasets using ×8 reconstruction. The results show that, under the listed methods and the same evaluation metrics, the proposed method achieves the best results in all six metrics (PSNR, SSIM, and LPIPS) on both datasets. This demonstrates that the proposed method, through the collaborative processing of structural anchoring, edge frequency detail generation, and adaptive intensity refinement, can simultaneously improve pixel consistency, structural similarity, and perceptual quality of the reconstructed image.

[0080] Table 2 shows the ×8 ablation experiment results of the key functional modules and loss supervision of this invention. By sequentially introducing the Latent Space Semantic Structure Anchoring (LSGA), Edge Constrained Details Generation (EAHC), Adaptive Intensity Refinement (ARDR), and Frequency Edge Details Loss Supervision (FEDS) into the baseline network, the effects of each key design on the reconstruction performance of the CelebA and Helen datasets are verified. Experimental results show that each design improves the reconstruction results from aspects such as face structure location constraints, edge frequency candidate detail generation, reliable detail selection and injection, and frequency edge supervision during the training phase. The complete scheme achieves a significant overall performance improvement compared to the baseline network.

[0081] After further adding EAHC to LSGA, the PSNR and SSIM on the CelebA test set improved from 27.8708 and 0.8145 to 27.9575 and 0.8174, respectively, while LPIPS decreased from 0.1293 to 0.1266; on the Helen test set, the PSNR and SSIM improved from 27.9097 and 0.8338 to 27.9328 and 0.8351, respectively, while LPIPS decreased from 0.1509 to 0.1453. These results demonstrate that extracting and fusing edge response and frequency response under structural anchoring feature constraints can provide more reliable candidate details for structurally sensitive regions such as the eyes, nose, lips, and facial contours.

[0082] After adding ARDR to LSGA and EAHC, the PSNR, SSIM, and LPIPS on the CelebA test set reached 28.0523, 0.8206, and 0.1231, respectively; and on the Helen test set, they reached 28.1189, 0.8413, and 0.1375, respectively. Compared to the configuration with LSGA and EAHC, the PSNR and SSIM on the CelebA test set increased by 0.0948 and 0.0032, respectively, while LPIPS decreased by 0.0035; and the PSNR and SSIM on the Helen test set increased by 0.1861 and 0.0062, respectively, while LPIPS decreased by 0.0078. These results demonstrate that orientation selection, reliability assessment, and injection strength control for candidate details can further improve the structural consistency and perceptual quality of the reconstructed image.

[0083] After further adding FEDS to the complete network structure, the PSNR, SSIM, and LPIPS on the CelebA test set changed from 28.0523, 0.8206, and 0.1231 to 28.0624, 0.8209, and 0.1228, respectively. PSNR and SSIM improved by 0.0101 and 0.0003, respectively, while LPIPS decreased by 0.0003. On the Helen test set, the PSNR, SSIM, and LPIPS changed from 28.1189, 0.8413, and 0.1375 to 28.1369, 0.8423, and 0.1325, respectively. PSNR and SSIM improved by 0.0180 and 0.0010, respectively, while LPIPS decreased by 0.0050. These results demonstrate that FEDS can further improve the pixel fidelity, structural consistency, and perceptual quality of the reconstructed images.

[0084] Compared to the baseline network, the complete method improves PSNR and SSIM by 0.2734 and 0.0090, respectively, and reduces LPIPS by 0.0083 on the CelebA test set; on the Helen test set, it improves PSNR and SSIM by 0.3567 and 0.0130, respectively, and reduces LPIPS by 0.0287. These results demonstrate that LSGA, EAHC, ARDR, and FEDS contribute to structural location constraints, candidate detail generation, adaptive detail control, and training loss supervision, respectively, and collectively improve the performance of face image super-resolution reconstruction through a synergistic relationship of prior and subsequent dependencies.

[0085] To further analyze the impact of key design features on the reconstruction quality of faces at different spatial locations, Figure 7 The baseline network and the reconstruction results after progressively adding LSGA, EAHC, ARDR and FEDS are presented, and the spatial quality variation between adjacent experimental configurations is represented by local PSNR and SSIM gain distributions.

[0086] like Figure 7 As shown in the figure, the baseline network and the reconstruction results after progressively adding LSGA, EAHC, ARDR, and FEDS are displayed sequentially. Local PSNR and SSIM gain maps are used to represent the spatial quality changes between adjacent experimental configurations. All gain maps are uniformly normalized to [0, 1]. The closer the color is to red, the greater the local reconstruction quality gain at the corresponding facial spatial location; the closer the color is to blue, the smaller the gain. Local gains are mainly distributed in structurally sensitive areas such as the eyes, nose, lips, and facial contours, indicating that the key functional modules of this invention can progressively improve the recovery of local details at facial structural locations. The cumulative gain map on the far right shows the overall local quality improvement of the complete method compared to the baseline network.

[0087] Summary Table 1 Figure 6 Table 2 and Figure 7 As shown, this invention determines the location of detail restoration through latent space semantic structure anchoring, generates candidate details constrained by structural location through edge-constrained detail generation, selects, fuses, and injects candidate details through adaptive intensity refinement, and utilizes frequency edge detail loss to supervise the high-frequency and edge responses of the reconstruction results during the training phase. The above technical design can stably improve the restoration quality of facial structure-sensitive regions on different test sets and reduce the risk of structural misalignment, edge blurring, and pseudo-high-frequency responses.

[0088] To further demonstrate the reconstruction effect of the method of the present invention, Figure 6 (a) is a low-resolution input image. Figure 6 (b) is a reconstructed image from a representative state-of-the-art method. Figure 6 (c) is the reconstructed image of the method of the present invention. Figure 6(d) shows the high-resolution target image. The figure also includes a magnified view of the eye region. Compared to low-resolution input images and the listed representative advanced methods, the method of this invention achieves clearer and more stable details in the eye contour, pupil boundary, and adjacent structural regions, while reducing local boundary blurring and structural misalignment.

[0089] In summary, quantitative evaluation, visual comparison, and ablation experiments all demonstrate that the method of this invention has comprehensive performance advantages in terms of pixel fidelity, structural consistency, and perceptual quality. In particular, it can improve the detail clarity and positional stability of structurally sensitive areas such as eye contours, pupil boundaries, nose ala, lip lines, and facial contours, while reducing structural misalignment, edge blurring, and pseudo-high-frequency responses. These effects do not stem from a simple superposition of structural priors, high-frequency enhancement, and feature modulation, but rather from the collaborative processing that seamlessly connects latent space semantic structure anchoring, edge-constrained detail generation, and adaptive intensity refinement. This forms a unique technical concept that differs from existing methods that rely on single structural constraints or single detail enhancements.

[0090] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions, simple modifications, or other equivalent alterations made by those skilled in the art without departing from the technical concept of the present invention should be subject to the scope of protection defined in the claims.

Claims

1. A face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation, characterized in that, Includes the following steps: Step S1, Image Acquisition and Preprocessing: Acquire the low-resolution face image to be reconstructed, preprocess the low-resolution face image, and input it into the face image super-resolution reconstruction network; Step S2, Multi-scale feature extraction: Multi-scale feature extraction is performed through the feature extraction unit to obtain intermediate reconstructed features and deep features or latent features; Step S3, latent space semantic structure anchoring: Generate structural prior features based on the deep features or latent features, convert the structural prior features into structural modulation parameters, and perform structural anchoring modulation on the deep features or latent features; Step S4, Edge Constraint Detail Generation: Based on the structural anchoring feature extraction and fusion of edge response and frequency response, an edge frequency detail basis constrained by the face structure position is generated. The edge frequency detail basis is used to characterize the candidate detail morphology of the face structure sensitive region. Step S5, Adaptive intensity refinement: Generate adaptive modulation parameters based on the structural anchoring features, the edge frequency detail basis, and the intermediate reconstruction features; select, fuse, and inject the edge frequency detail basis to obtain detail modulation features; Step S6, Decoding, Reconstruction and Image Output: The structural anchoring features, detail modulation features and reconstruction features from the decoding stage are fused and a high-resolution face image is generated through the decoding and reconstruction unit.

2. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 1, characterized in that, In step S5, the adaptive modulation parameters are used to control at least one of frequency band selection, detail reliability, and output injection strength, so that the edge frequency detail basis is enhanced in the structurally reliable region and suppressed in the smooth region or the structurally uncertain region.

3. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 2, characterized in that, The structural prior features include facial landmark heatmaps, facial semantic parsing maps, facial component region maps, facial contour maps, structural attention maps, structural features output by the facial structure prediction branch, or combinations of at least two of the above features.

4. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 3, characterized in that, The structural modulation parameters include at least one of affine modulation parameters, spatial attention weights, channel attention weights, region gating weights, or normalized modulation parameters.

5. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 4, characterized in that, When the structure modulation parameters are selected as affine modulation parameters, the affine modulation parameters include scaling parameter γ and translation parameter β. The structure anchoring modulation is achieved by performing at least one of scaling, translation, element-wise multiplication and addition, gated fusion or normalization modulation on the deep features or latent features.

6. A face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 4 or 5, characterized in that, The edge frequency detail base includes high-frequency sub-band features, horizontal edge features, vertical edge features, gradient amplitude features, directional edge features, frequency band selection features, or a fusion result of at least two of the above features, used to characterize the directional boundary, high-frequency texture, or local detail morphology of at least one structurally sensitive region among eye contour, eyebrow, nose wing, lip line, tooth boundary, beard, and facial contour.

7. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 6, characterized in that, In step S4, the frequency response is obtained through wavelet transform, high-pass filtering, Fourier transform, learnable frequency filtering, frequency domain convolution, directional high-frequency extraction, or at least two of the above methods; in step S4, the edge response is obtained through gradient operator, Sobel operator, Scharr operator, Laplacian operator, directional convolution, edge detection network, or at least two of the above methods.

8. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 7, characterized in that, In step S5, the adaptive modulation parameters include at least one of a frequency band selection gate, a detail reliability gate, an output injection gate, a spatial attention weight, a channel attention weight, or a region adaptive weight. The frequency band selection gate is used to control the weights of different high-frequency sub-bands or directional frequency components participating in the reconstruction. The detail reliability gate is used to determine the credibility of the edge frequency detail basis at the current spatial location or channel dimension. The output injection gate is used to control the intensity of the modulated detail features injected into the decoded reconstruction features.

9. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 8, characterized in that, The adaptive modulation parameters enhance at least one structurally sensitive detail in the eye contour, eyebrows, nose, lip line, tooth boundaries, beard, and facial contour, based on the structural reliability or degradation degree of different facial regions, and suppress pseudo-high-frequency responses in smooth areas or structurally uncertain areas such as cheeks and forehead.

10. The face image super-resolution reconstruction method based on structural anchoring and edge frequency adaptive modulation according to claim 9, characterized in that, The face image super-resolution reconstruction network adopts any one or any combination of encoder-decoder structure, convolutional neural network structure, residual network structure, Transformer structure, U-Net structure or image restoration network structure. When training the face image super-resolution reconstruction network, pixel reconstruction loss, structural prior supervision loss, edge loss, frequency loss, perceptual loss, identity preservation loss or a combination of at least two of the above losses are used.

Images