A face aesthetic score method based on dual-domain cooperation

Abstract

The present application relates to a kind of based on double-domain collaborative human face aesthetic score method, the method includes: 1) face image and facial key point acquisition;2) geometric descriptor construction and coding;3) visual domain feature extraction and enhancement;4) double-domain semantic collaborative interaction;5) double-domain adaptive fusion;6) distribution expectation and residual correction score prediction.This method can improve the accuracy and stability of human face aesthetic score prediction.

Description

Technical Field

[0001] This invention relates to computer vision and artificial intelligence technologies, specifically a face aesthetic scoring method based on dual-domain collaboration. Background Technology

[0002] Facial aesthetic scoring aims to predict facial attractiveness scores that align with human subjective perception based on facial images. It is an important task in the field of facial analysis and can be applied to scenarios such as portrait enhancement, virtual makeup, and medical aesthetic analysis. In recent years, with the development of deep learning technology, automated facial aesthetic scoring has received widespread attention.

[0003] Existing facial aesthetic scoring methods mainly include those based on artificial geometric features and those based on deep learning. Methods based on artificial geometric features typically construct structural domain geometric features such as facial distance, proportion, and symmetry based on facial key points, and then combine them with regression models for score prediction. While these methods can utilize facial structural domain geometric features for analysis, their ability to represent visual information such as facial appearance, texture, and local details is limited. Deep learning-based methods typically use convolutional neural networks or Transformers to extract visual domain features from facial images and then predict facial aesthetic scores based on these visual domain features. These methods primarily rely on facial visual domain features and do not adequately utilize structural domain geometric features such as facial distance relationships, symmetry relationships, and proportion relationships.

[0004] Some facial aesthetic scoring methods attempt to incorporate facial key points into deep learning models, but they typically employ direct concatenation or one-way attention mechanisms to fuse visual and structural geometric features. The information interaction between these two types of features remains insufficient, hindering the effective fusion of visual and structural geometric features. Furthermore, facial aesthetic scoring is highly subjective; different raters may assign different scores to the same facial image. The scoring label for a facial image is usually represented by a subjective score distribution or the average of multiple subjective scores. Existing methods, if they only regress on a single score value, struggle to fully utilize the subjective differences contained in the score distribution. Therefore, how to effectively fuse visual and structural geometric features and fully utilize score distribution information for facial aesthetic score prediction remains a problem that existing facial aesthetic scoring methods need to address. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a face aesthetic scoring method based on dual-domain collaboration. This method can improve the accuracy and stability of face aesthetic score prediction.

[0006] The technical solution to achieve the objective of this invention is:

[0007] A face aesthetic scoring method based on dual-domain collaboration includes the following steps:

[0008] 1) Face Image and Facial Keypoint Acquisition: The face image to be scored is preprocessed to obtain a normalized face image, and a facial keypoint detection model is used to extract the set of facial keypoints. Specifically, let the face image to be scored be... ,in and Let the height and width of the face image to be scored be represented respectively. Then, the face image to be scored... Convert to an RGB three-channel image, and based on the input size required by the model, convert the face image to be scored. Scaling the size to Next, the scaled-down face image is subjected to pixel normalization. Pixel normalization is performed using a channel-based normalization method to obtain a normalized face image. The calculation formula is: ,in, Indicates pixel position, Indicates the color channel index. and They represent the first The mean and standard deviation of each color channel were calculated; the normalized face image was calculated using the MediaPipe Face Mesh facial landmark detection model. The facial key points are obtained by identifying key facial points in the eyebrows, eyes, nose, lips, facial contours, and iris area. ,in, Indicates the first Key facial features and They represent the first The x and y coordinates of each facial key point in the image coordinate system This represents the total number of facial key points, with a value of 478.

[0009] 2) Geometric descriptor construction and encoding: based on facial key point set Three types of geometric descriptors—distance descriptors, symmetry descriptors, and scale descriptors—are constructed, and the geometric features of the structural domain are obtained by encoding the distance descriptors, symmetry descriptors, and scale descriptors respectively. Specifically:

[0010] 2-1) Select several groups of key points from the set of facial key points, calculate the Euclidean distance between each group of key points, and normalize it using the facial scale factor to obtain the distance descriptor. Specifically: Let the first... The and the first The two-dimensional coordinates of the facial key points are as follows: and The normalized distance between the two key points The calculation formula is: ,in Represents the face scale factor;

[0011] 2-2) Select several sets of corresponding key points from the symmetrical regions of the face, calculate the mirror error in the horizontal direction and the positional error in the vertical direction for each set of key points, and normalize them using the facial scale factor to obtain the symmetry descriptor. The calculation formula is: and ,in Indicates the position of the midline of the face;

[0012] 2-3) Select several sets of distances that can reflect the relative scale of the facial region, calculate the ratio between the two distances in each set, and obtain the scale descriptor. , No. proportional features corresponding to group distance The calculation formula is: ,in, and These represent the facial distances used as the numerator and denominator when calculating this proportional characteristic. This represents a constant used to prevent the denominator from being zero;

[0013] 2-4) Obtaining the distance descriptor Symmetric descriptors and scale descriptor Then, the three types of geometric descriptors are encoded respectively, specifically as follows: Let Indicates the first Geometry-like descriptors, where For distance, symmetry, and scale descriptors respectively, first, each type of geometric descriptor is normalized. Then, each normalized descriptor is transformed through a feature mapping operation. We obtain a feature vector and then embed the corresponding descriptor identifier into it to get the descriptor feature sequence. The calculation formula is as follows:

[0014] ,

[0015] in, Indicates the number of this type of descriptor. This indicates a normalization operation. Indicates the first Feature mapping operations for class descriptors Then, set for each type of descriptor. Learnable query vectors These query vectors are automatically updated during model training and are used to derive features from descriptor sequences. Select and aggregate relevant geometric information, specifically, using As a query feature As key and value features, the importance of each descriptor to different query vectors is calculated through multi-head cross-attention, and weighted aggregation is performed according to the corresponding weights. Then, the aggregation result is residually connected with the query vector, and then processed by a feedforward network to finally obtain the first... Aggregated features corresponding to geometry descriptors The calculation formula is as follows:

[0016] ,

[0017] ,

[0018] in, This indicates a multi-headed cross-attention operation. Indicates a feedforward network. Indicates the first The type embedding of the geometric descriptor is performed; finally, the three types of aggregated features are concatenated and normalized along the feature sequence dimension to obtain the structural domain geometric features. The calculation formula is:

[0019]

[0020] in Indicates a splicing operation;

[0021] 3) Visual Domain Feature Extraction and Enhancement: The normalized face image is input into the visual feature extraction network to obtain initial visual domain features. After processing by the visual aesthetic feature enhancement module, residual fusion is performed with the initial features to obtain the visual domain features. Specifically, the normalized face image is... Input the Swin Transformer visual domain feature extraction network and output the initial visual domain features. The calculation formula is: ,in This represents the pre-training parameters of the Swin Transformer visual feature extraction network; it represents the initial visual domain features. The input visual aesthetic feature enhancement module acquires local visual information through convolutional operations at different receptive fields and generates visual enhancement weights. Based on visual enhancement weights, the initial visual domain features are... Weighting is performed to obtain enhanced visual domain features. ,Will and The features are multiplied, and the fused features are flattened and normalized to obtain the visual domain features. The calculation formula is: , in, , Indicates the flattening operation. This represents an element-wise multiplication operation;

[0022] 4) Dual-domain semantic collaborative interaction: Utilizing bidirectional cross-attention interaction between visual domain features and structural domain geometric features, structurally perceptual visual features and visually perceptual structural features are obtained respectively; subsequently, through residual fusion and attention pooling, global visual domain features and global structural domain geometric features are obtained. Specifically: first, the visual domain features... Normalization and linear mapping are performed to obtain visual query features. Visual key features Visual value features The calculation formula is: ,

[0023] in, This represents a linear mapping operation applied to the geometric features of the structural domain. Normalization and linear mapping are performed to obtain structural query features. Structural bond features and structural value characteristics The calculation formula is: In the interaction direction of the visual query structure, with As a query, with and Using these as keys and values ​​respectively, we obtain structure-aware visual features. In the synergistic direction from structure to vision, with As a query, with and As keys and values, we obtain visual perceptual structural features. The structure-aware visual features are residually fused with the original visual domain features, and the visually-aware structural features are residually fused with the original structural domain geometric features. The resulting visual domain features and structural domain geometric features are then input into attention pooling layers. Normalized weights are first obtained through two layers of linear mapping and the Softmax function. and The calculation formula is as follows:

[0024]

[0025]

[0026] in, Let represent the hyperbolic tangent activation function; then, based on the attention weights... and The global visual domain features are obtained by weighted summation of the visual domain features and structural domain geometric features after interaction. and global structural domain geometric features The calculation formula is as follows: and ,in, and They represent the first Visual domain features and structural domain geometric features;

[0027] 5) Dual-domain adaptive fusion: Fusion weights are generated based on the global visual domain features and the global structural domain geometric features. Adaptive weighted fusion of the global visual domain features and the global structural domain geometric features is then performed based on these fusion weights to obtain dual-domain fused features. Specifically, the global visual domain features are... and global structural domain geometric features The concatenation is performed and input into the gating weight generation unit to obtain the fused weight. The calculation formula is as follows:

[0028] ,

[0029] in, Represents the activation function, based on Global visual domain features and global structural domain geometric features Adaptive weighted fusion is performed to obtain dual-domain fusion features. The calculation formula is: ;

[0030] 6) Expected Distribution and Residual Correction Score Prediction: Based on the dual-domain fusion features, predict the probability distribution and residual correction value of the face aesthetic score, and add the expected score corresponding to the probability distribution to the residual correction value to obtain the final face aesthetic prediction score. Specifically: First, the dual-domain fusion features... Input the rating distribution prediction layer and normalize it using the Softmax function to obtain the rating probability distribution. The calculation formula is: ,in, This represents the linear mapping operation of the score distribution prediction layer. This indicates the number of preset rating intervals. Indicates that the predicted score belongs to the first The probability of each rating interval, and satisfying Subsequently, the predicted probabilities are weighted and summed based on the center value of each scoring interval to obtain the expected score distribution. The calculation formula is: ,in Indicates the first The center value of each scoring interval; simultaneously, the residual correction branch is used to fuse features from the two domains. The predicted scoring bias is determined and its range is limited by the hyperbolic tangent function, then multiplied by the residual range control coefficient. The residual correction value is obtained. The calculation formula is: ,in, This represents a multilayer perceptron in the residual correction branch. The residual range control coefficient is used to obtain the final facial aesthetic score. .

[0031] This technical solution first preprocesses the face image to be rated and extracts facial key points. Then, based on the facial key points, distance descriptors, symmetry descriptors, and proportion descriptors are constructed and encoded to obtain structural domain geometric features. Simultaneously, visual domain features are extracted and enhanced from the preprocessed face image. Next, through the dual-domain semantic collaborative interaction between the visual and structural domains, the visual domain features and structural domain geometric features engage in bidirectional information exchange. The resulting visual domain features and structural domain geometric features are then adaptively fused to obtain dual-domain fused features. Finally, the probability distribution of the face aesthetic score and the residual correction value are predicted based on the dual-domain fused features. The expected score corresponding to the probability distribution is added to the residual correction value to obtain the final face aesthetic score.

[0032] This technical solution utilizes facial visual features in face images and geometric structural features constructed based on facial key points. Through bidirectional information interaction and adaptive fusion between the visual and structural domains, and combined with score distribution and residual correction, it predicts facial aesthetic scores to improve the accuracy and stability of facial aesthetic score prediction.

[0033] This method can improve the accuracy and stability of facial aesthetic score prediction. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the method used in the embodiment.

[0035] Figure 2 This is a schematic diagram illustrating the construction and encoding process of the structural domain geometry descriptor in the embodiment;

[0036] Figure 3 This is a schematic diagram of the dual-domain semantic collaborative interaction process in the embodiment;

[0037] Figure 4 This is the result of facial image and facial key point region segmentation in the embodiment;

[0038] Figure 5 This is a structurally-aware visual feature response map in the embodiment. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0040] Example:

[0041] Reference Figure 1 A face aesthetic scoring method based on dual-domain collaboration includes the following steps:

[0042] 1) Face Image and Facial Keypoint Acquisition: The face image to be scored is preprocessed to obtain a normalized face image, and a facial keypoint detection model is used to extract the set of facial keypoints. Specifically, let the face image to be scored be... ,in and Let the height and width of the face image to be scored be represented respectively. Then, the face image to be scored... Convert to an RGB three-channel image, and based on the input size required by the model, convert the face image to be scored. Scaling the size to Next, the scaled-down face image is subjected to pixel normalization. Pixel normalization is performed using a channel-based normalization method to obtain a normalized face image. The calculation formula is: ,in, Indicates pixel position, Indicates the color channel index. and They represent the first The mean and standard deviation of each color channel were calculated; the normalized face image was calculated using the MediaPipe Face Mesh facial landmark detection model. The facial key points are obtained by identifying key facial points in the eyebrows, eyes, nose, lips, facial contours, and iris area. ,in, Indicates the first Key facial features and They represent the first The x and y coordinates of each facial key point in the image coordinate system This represents the total number of facial key points, with a value of 478.

[0043] 2) Geometric descriptor construction and encoding: based on facial key point set Three types of geometric descriptors—distance descriptors, symmetry descriptors, and scale descriptors—are constructed, and the geometric features of the structural domain are obtained by encoding the distance descriptors, symmetry descriptors, and scale descriptors respectively. Specifically:

[0044] 2-1) Select several groups of key points from the set of facial key points, calculate the Euclidean distance between each group of key points, and normalize it using the facial scale factor to obtain the distance descriptor. Specifically: Let the first... The and the first The two-dimensional coordinates of the facial key points are as follows: and The normalized distance between the two key points The calculation formula is: ,in Represents the face scale factor;

[0045] 2-2) Select several sets of corresponding key points from the symmetrical regions of the face, calculate the mirror error in the horizontal direction and the positional error in the vertical direction for each set of key points, and normalize them using the facial scale factor to obtain the symmetry descriptor. The calculation formula is: and ,in Indicates the position of the midline of the face;

[0046] 2-3) Select several sets of distances that can reflect the relative scale of the facial region, calculate the ratio between the two distances in each set, and obtain the scale descriptor. , No. proportional features corresponding to group distance The calculation formula is: ,in, and These represent the facial distances used as the numerator and denominator when calculating this proportional characteristic. This represents a constant used to prevent the denominator from being zero;

[0047] 2-4) Obtaining the distance descriptor Symmetric descriptors and scale descriptor Then, the three types of geometric descriptors are encoded respectively, specifically as follows: Let Indicates the first Geometry-like descriptors, where For distance, symmetry, and scale descriptors respectively, first, each type of geometric descriptor is normalized. Then, each normalized descriptor is transformed through a feature mapping operation. We obtain a feature vector and then embed the corresponding descriptor identifier into it to get the descriptor feature sequence. The calculation formula is as follows:

[0048] ,

[0049] in, Indicates the number of this type of descriptor. This indicates a normalization operation. Indicates the first Feature mapping operations for class descriptors Then, set for each type of descriptor. Learnable query vectors These query vectors are automatically updated during model training and are used to derive features from descriptor sequences. Select and aggregate relevant geometric information, specifically, using As a query feature As key and value features, the importance of each descriptor to different query vectors is calculated through multi-head cross-attention, and weighted aggregation is performed according to the corresponding weights. Then, the aggregation result is residually connected with the query vector, and then processed by a feedforward network to finally obtain the first... Aggregated features corresponding to geometry descriptors The calculation formula is as follows:

[0050] ,

[0051] ,

[0052] in, This indicates a multi-headed cross-attention operation. Indicates a feedforward network. Indicates the first The type embedding of the geometric descriptor is performed; finally, the three types of aggregated features are concatenated and normalized along the feature sequence dimension to obtain the structural domain geometric features. The calculation formula is:

[0053]

[0054] in This indicates a concatenation operation; the geometric descriptor construction and encoding process is as follows: Figure 2 As shown;

[0055] 3) Visual Domain Feature Extraction and Enhancement: The normalized face image is input into the visual feature extraction network to obtain initial visual domain features. After processing by the visual aesthetic feature enhancement module, residual fusion is performed with the initial features to obtain the visual domain features. Specifically, the normalized face image is... Input the Swin Transformer visual domain feature extraction network and output the initial visual domain features. The calculation formula is: ,in This represents the pre-training parameters of the Swin Transformer visual feature extraction network; it represents the initial visual domain features. The input visual aesthetic feature enhancement module acquires local visual information through convolutional operations at different receptive fields and generates visual enhancement weights. Based on visual enhancement weights, the initial visual domain features are... Weighting is performed to obtain enhanced visual domain features. ,Will and The features are multiplied, and the fused features are flattened and normalized to obtain the visual domain features. The calculation formula is: , in, , Indicates the flattening operation. This represents an element-wise multiplication operation;

[0056] 4) Dual-domain semantic collaborative interaction: Utilizing bidirectional cross-attention interaction between visual domain features and structural domain geometric features, structurally perceptual visual features and visually perceptual structural features are obtained respectively; subsequently, through residual fusion and attention pooling, global visual domain features and global structural domain geometric features are obtained. Specifically: first, the visual domain features... Normalization and linear mapping are performed to obtain visual query features. Visual key features Visual value features The calculation formula is: ,

[0057] in, This represents a linear mapping operation applied to the geometric features of the structural domain. Normalization and linear mapping are performed to obtain structural query features. Structural bond features and structural value characteristics The calculation formula is: In the interaction direction of the visual query structure, with As a query, with and Using these as keys and values ​​respectively, we obtain structure-aware visual features. In the synergistic direction from structure to vision, with As a query, with and As keys and values, we obtain visual perceptual structural features. The structure-aware visual features are residually fused with the original visual domain features, and the visually-aware structural features are residually fused with the original structural domain geometric features. The resulting visual domain features and structural domain geometric features are then input into attention pooling layers. Normalized weights are first obtained through two layers of linear mapping and the Softmax function. and The calculation formula is as follows:

[0058]

[0059]

[0060] in, Let represent the hyperbolic tangent activation function; then, based on the attention weights... and The global visual domain features are obtained by weighted summation of the visual domain features and structural domain geometric features after interaction. and global structural domain geometric features The calculation formula is as follows: and The visual structure dual-domain semantic collaborative interaction process is as follows: Figure 3 As shown;

[0061] 5) Dual-domain adaptive fusion: Fusion weights are generated based on the global visual domain features and the global structural domain geometric features. Adaptive weighted fusion of the global visual domain features and the global structural domain geometric features is then performed based on these fusion weights to obtain dual-domain fused features. Specifically, the global visual domain features are... and global structural domain geometric features The concatenation is performed and input into the gating weight generation unit to obtain the fused weight. The calculation formula is as follows:

[0062] ,

[0063] in, Represents the activation function, based on Global visual domain features and global structural domain geometric features Adaptive weighted fusion is performed to obtain dual-domain fusion features. The calculation formula is: ;

[0064] 6) Expected Distribution and Residual Correction Score Prediction: Based on the dual-domain fusion features, predict the probability distribution and residual correction value of the face aesthetic score, and add the expected score corresponding to the probability distribution to the residual correction value to obtain the final face aesthetic prediction score. Specifically: First, the dual-domain fusion features... Input the rating distribution prediction layer and normalize it using the Softmax function to obtain the rating probability distribution. The calculation formula is: ,in, This represents the linear mapping operation of the score distribution prediction layer. This indicates the number of preset rating intervals. Indicates that the predicted score belongs to the first The probability of each rating interval, and satisfying Subsequently, the predicted probabilities are weighted and summed based on the center value of each scoring interval to obtain the expected score distribution. The calculation formula is: ,in Indicates the first The center value of each scoring interval; simultaneously, the residual correction branch is used to fuse features from the two domains. The predicted scoring bias is determined and its range is limited by the hyperbolic tangent function, then multiplied by the residual range control coefficient. The residual correction value is obtained. The calculation formula is: ,in, This represents a multilayer perceptron in the residual correction branch. The residual range control coefficient is used to obtain the final facial aesthetic score. .

[0065] In this example, the SCUT-FBP5500 face aesthetics scoring dataset is used to train the dual-domain collaborative scoring model. The dataset contains 5500 face images, with face aesthetics scores ranging from 1 to 5. A five-fold cross-validation method is used to partition the dataset, selecting four folds as the training set and the remaining fold as the test set each time. The face images are uniformly adjusted to 224×224 pixels and pixel normalized. Simultaneously, MediaPipe Face Mesh is used to extract 478 two-dimensional facial key points, resulting in the face images, facial key point sets, and corresponding face aesthetics scoring labels used for model training. The resulting face images to be scored and the facial key point region segmentation results are as follows: Figure 4 As shown; the face image to be scored and its facial key points are input into the trained dual-domain collaborative scoring model to obtain visual domain features and structural domain geometric features, and perform dual-domain semantic collaborative interaction. The structurally-aware visual feature response results are as follows: Figure 5 As shown; during model training, the rating regression loss, rating distribution supervision loss, and distribution consistency loss are calculated separately, and the three losses are then weighted and summed. Rating regression loss The calculation formula is as follows:

[0066]

[0067] in, This indicates the number of training samples in each batch. and They represent the first The true scores and final predicted scores for each sample; supervised loss based on score distribution. The calculation formula is as follows:

[0068]

[0069] in, Indicates the number of discrete rating levels. and They represent the first The sample at the th Predicted probability and target probability at each discrete rating level; distribution consistency loss The calculation formula is as follows:

[0070]

[0071] in, Indicates the first The expected score of the sample distribution is calculated using the following formula: Here, the representative value of the discrete rating level is used; finally, the joint loss is obtained by weighted summing of the rating regression loss, rating distribution supervision loss, and distribution consistency loss, and the calculation formula is as follows:

[0072] ,

[0073] in, and , represent the weight coefficients corresponding to the supervision loss and distribution consistency loss of the rating distribution, respectively.

[0074] In this example, the Adam optimizer is used to iteratively update the model parameters, and the learning rate is set to... The weight decay coefficient is set to During training, a cosine annealing scheduler is used to dynamically adjust the learning rate. Set to 50. Set the value to 0, the number of training rounds to 150, and after training, you will get the trained dual-domain collaborative scoring model.

Claims

1. A dual-domain collaborative based face aesthetic score method, characterized in that, Includes the following steps: 1) Face Image and Facial Keypoint Acquisition: The face image to be scored is preprocessed to obtain a normalized face image, and a facial keypoint detection model is used to extract the set of facial keypoints. Specifically, let the face image to be scored be... ,in and Let the height and width of the face image to be scored be represented respectively. Then, the face image to be scored... Convert to an RGB three-channel image, then use the face image to be scored. Scaling the size to Next, the scaled-down face image is subjected to pixel normalization. Pixel normalization is performed using a channel-based normalization method to obtain a normalized face image. The calculation formula is: ,in, Indicates pixel position, Indicates the color channel index. and They represent the first The mean and standard deviation of each color channel were calculated; the normalized face image was calculated using the MediaPipe Face Mesh facial landmark detection model. The facial key points are obtained by identifying key facial points in the eyebrows, eyes, nose, lips, facial contours, and iris area. ,in, Indicates the first Key facial features and They represent the first The x and y coordinates of each facial key point in the image coordinate system This represents the total number of facial key points, with a value of 478. 2) Geometric descriptor construction and encoding: based on facial key point set Three types of geometric descriptors—distance descriptors, symmetry descriptors, and scale descriptors—are constructed, and the geometric features of the structural domain are obtained by encoding the distance descriptors, symmetry descriptors, and scale descriptors respectively. Specifically: 2-1) Select several groups of key points from the set of facial key points, calculate the Euclidean distance between each group of key points, and normalize it using the facial scale factor to obtain the distance descriptor. Specifically: Let the first... The and the first The two-dimensional coordinates of the facial key points are as follows: and The normalized distance between the two key points The calculation formula is: ,in Represents the face scale factor; 2-2) Select several sets of corresponding key points from the symmetrical regions of the face, calculate the mirror error in the horizontal direction and the positional error in the vertical direction for each set of key points, and normalize them using the facial scale factor to obtain the symmetry descriptor. The calculation formula is: and ,in Indicates the position of the midline of the face; 2-3) Select several sets of distances that can reflect the relative scale of the facial region, calculate the ratio between the two distances in each set, and obtain the scale descriptor. , No. proportional features corresponding to group distance The calculation formula is: ,in, and These represent the facial distances used as the numerator and denominator when calculating this proportional characteristic. This represents a constant used to prevent the denominator from being zero; 2-4) Obtaining the distance descriptor Symmetric descriptors and scale descriptor Then, the three types of geometric descriptors are encoded respectively, specifically as follows: Let Indicates the first Geometry-like descriptors, where For distance, symmetry, and scale descriptors respectively, first, each type of geometric descriptor is normalized. Then, each normalized descriptor is transformed through a feature mapping operation. We obtain a feature vector and then embed the corresponding descriptor identifier into it to get the descriptor feature sequence. The calculation formula is as follows: ， in, Indicates the number of this type of descriptor. This indicates a normalization operation. Indicates the first Feature mapping operations for class descriptors Then, set for each type of descriptor. Learnable query vectors These query vectors are automatically updated during model training and are used to derive features from descriptor sequences. Select and aggregate relevant geometric information to As a query feature As key and value features, the importance of each descriptor to different query vectors is calculated through multi-head cross-attention, and weighted aggregation is performed according to the corresponding weights. Then, the aggregation result is residually connected with the query vector, and then processed by a feedforward network to finally obtain the first... Aggregated features corresponding to geometry descriptors The calculation formula is as follows: , , in, This indicates a multi-headed cross-attention operation. Indicates a feedforward network. Indicates the first The type embedding of the geometric descriptor is performed; finally, the three types of aggregated features are concatenated and normalized along the feature sequence dimension to obtain the structural domain geometric features. The calculation formula is: in Indicates a splicing operation; 3) Visual Domain Feature Extraction and Enhancement: The normalized face image is input into the visual feature extraction network to obtain initial visual domain features. After processing by the visual aesthetic feature enhancement module, residual fusion is performed with the initial features to obtain the visual domain features. Specifically, the normalized face image is... Input the Swin Transformer visual domain feature extraction network and output the initial visual domain features. The calculation formula is: ,in This represents the pre-training parameters of the Swin Transformer visual feature extraction network; it represents the initial visual domain features. The input visual aesthetic feature enhancement module acquires local visual information through convolutional operations at different receptive fields and generates visual enhancement weights. Based on visual enhancement weights, the initial visual domain features are... Weighting is performed to obtain enhanced visual domain features. ,Will and The features are multiplied, and the fused features are flattened and normalized to obtain the visual domain features. The calculation formula is: , in, , Indicates the flattening operation. This represents an element-wise multiplication operation; 4) Dual-domain semantic collaborative interaction: Utilizing bidirectional cross-attention interaction between visual domain features and structural domain geometric features, structurally perceptual visual features and visually perceptual structural features are obtained respectively; subsequently, through residual fusion and attention pooling, global visual domain features and global structural domain geometric features are obtained. Specifically: first, the visual domain features... Normalization and linear mapping are performed to obtain visual query features. Visual key features Visual value features The calculation formula is: , in, This represents a linear mapping operation applied to the geometric features of the structural domain. Normalization and linear mapping are performed to obtain structural query features. Structural bond features and structural value characteristics The calculation formula is: In the interaction direction of the visual query structure, with As a query, with and Using these as keys and values ​​respectively, we obtain structure-aware visual features. In the synergistic direction from structure to vision, with As a query, with and As keys and values, we obtain visual perceptual structural features. The structure-aware visual features are residually fused with the original visual domain features, and the visually-aware structural features are residually fused with the original structural domain geometric features. The resulting visual domain features and structural domain geometric features are then input into attention pooling layers. Normalized weights are first obtained through two layers of linear mapping and the Softmax function. and The calculation formula is as follows: in, Let represent the hyperbolic tangent activation function; then, based on the attention weights... and The global visual domain features are obtained by weighted summation of the visual domain features and structural domain geometric features after interaction. and global structural domain geometric features The calculation formula is as follows: and ,in, and They represent the first Visual domain features and structural domain geometric features; 5) Dual-domain adaptive fusion: Fusion weights are generated based on the global visual domain features and the global structural domain geometric features. Adaptive weighted fusion of the global visual domain features and the global structural domain geometric features is then performed based on these fusion weights to obtain dual-domain fused features. Specifically, the global visual domain features are... and global structural domain geometric features The concatenation is performed and input into the gating weight generation unit to obtain the fused weight. The calculation formula is as follows: ， in, Represents the activation function, based on Global visual domain features and global structural domain geometric features Adaptive weighted fusion is performed to obtain dual-domain fusion features. The calculation formula is: ; 6) Expected Distribution and Residual Correction Score Prediction: Based on the dual-domain fusion features, predict the probability distribution and residual correction value of the face aesthetic score, and add the expected score corresponding to the probability distribution to the residual correction value to obtain the final face aesthetic prediction score. Specifically: First, the dual-domain fusion features... Input the rating distribution prediction layer and normalize it using the Softmax function to obtain the rating probability distribution. The calculation formula is: ,in, This represents the linear mapping operation of the score distribution prediction layer. This indicates the number of preset rating intervals. Indicates that the predicted score belongs to the first The probability of each rating interval, and satisfying Subsequently, the predicted probabilities are weighted and summed based on the center value of each scoring interval to obtain the expected score distribution. The calculation formula is: ,in Indicates the first The center value of each scoring interval; simultaneously, the residual correction branch is used to fuse features from the two domains. The predicted scoring bias is determined and its range is limited by the hyperbolic tangent function, then multiplied by the residual range control coefficient. The residual correction value is obtained. The calculation formula is: ,in, This represents a multilayer perceptron in the residual correction branch. The residual range control coefficient is used to obtain the final facial aesthetic score. .

Images