Age prediction method and system based on facial fine granularity

Abstract

The invention discloses an age prediction method and system based on face fine granularity, and relates to the technical field of computer vision image processing.The method includes the steps that a face image is obtained and subjected to standardization processing; dividing the face into a plurality of sub-regions according to the standardized face image; extracting multi-dimensional fineness features of each face sub-region, and quantifying fine structure features of the skin surface by adopting a local binary pattern of multi-scale direction perception; according to the fine structure features and contribution degrees of different face regions to age prediction, corresponding weights are distributed to the fineness features of all the regions, the weighted fineness features are input into a pre-trained age prediction model, and a predicted age value is output. According to the method, the condition that the aging degrees of different areas of the face are inconsistent can be better processed, and the age change information which is finer than that of a traditional method can be captured by analyzing the skin fineness index, so that the accuracy of age prediction is improved.

Description

[0001] A method and system for age prediction based on facial fine-grained features Technical Field This invention relates to the field of computer vision image processing technology, and in particular to a method and system for age prediction based on fine-grained facial features. Background Technology

[0002] Existing age prediction technologies primarily rely on overall facial features, such as wrinkle distribution and facial contour shape, for assessment. While these methods can predict age to some extent, they have the following limitations: Traditional methods refer to machine learning approaches based on hand-designed features, which primarily focus on obvious aging characteristics such as deep wrinkles and changes in facial contours. However, they have lower accuracy in identifying younger individuals or early signs of aging. Existing technologies often overlook skin texture, an important age indicator, yet changes in skin texture often appear earlier than changes in macroscopic features, providing more precise age information.

[0003] Furthermore, the robustness of the current method in handling different lighting conditions, shooting angles, and image quality needs to be improved, especially in everyday photos taken by mobile devices, where the prediction accuracy drops significantly. Summary of the Invention

[0004] In view of the above-mentioned prior art, the present invention provides a method and system for age prediction based on facial fine-grained features, which mainly solves the technical problems existing in the background art.

[0005] To achieve the above objectives, the technical solution of this invention is implemented as follows: In a first aspect, the present invention provides a facial fine-grained age prediction method, the method comprising the following steps: Acquire facial images and perform standardized processing on them, including face detection, key point localization, geometric correction, and illumination normalization; The face is divided into multiple sub-regions based on the standardized facial image, including the forehead region, the periorbital region, the cheek region, the nose region, and the jaw region. Multi-dimensional detail features are extracted for each facial sub-region, including texture roughness, pore density, skin smoothness, and micro-wrinkle density. The fine structural features of the skin surface are quantified using a multi-scale directional perception local binary mode. Based on the fine structural features and the contribution of different facial regions to age prediction, corresponding weights are assigned to the fine features of each region. The weighted fine features are then input into the pre-trained age prediction model, which outputs the predicted age value.

[0006] As a preferred embodiment of the present invention, a multi-task cascaded convolutional neural network is used to perform face detection and key point localization on the input image, as detailed below: Receive the input image after multi-scale scaling and generate preliminary face candidate regions and their bounding box coordinates; Based on the candidate regions, false detections are filtered out using a convolutional layer structure to refine the bounding box positions; based on the refined candidate regions, the final face bounding box and the coordinate values ​​of 68 facial key points are output.

[0007] As a preferred embodiment of the present invention, geometric correction processing is performed on the 68 facial key points, as follows: Multiple preset key points are selected as reference points for geometric correction, including the center point of the left eye, the center point of the right eye, the tip of the nose, the left corner of the mouth, and the right corner of the mouth. The mapping relationship between the coordinates of the reference point in the current image and the coordinates of the corresponding point in the standard frontal template is calculated, and based on the mapping relationship, the detected face image is transformed to the standard frontal pose through affine transformation.

[0008] As a preferred embodiment of the present invention, the step of transforming the detected face image to a standard frontal pose through affine transformation specifically includes: Calculate the transformation matrix from the current facial key points to the key points of the standard frontal template; The transformation matrix is ​​applied to the entire facial region, and the output is a frontal face image with a size normalized to 256×256 pixels.

[0009] As a preferred embodiment of the present invention, the geometrically corrected image is subjected to illumination normalization processing, as follows: The geometrically corrected image is converted from the RGB color space to the LAB color space, thereby separating the luminance channel and the color channel; The separated luminance channel is divided into multiple local regions by limiting contrast adaptive histogram equalization. Histogram equalization is performed independently on each region, and artifacts at the region boundaries are eliminated by bilinear interpolation. The equalized luminance channel is then merged back with the original color channel and converted back to the RGB color space to obtain the image after illumination normalization.

[0010] As a preferred embodiment of the present invention, the face is divided into multiple sub-regions based on the standardized facial image, specifically including: For the forehead region, its upper boundary is defined by the key point set {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17}, and its lower boundary is defined by the key point set {18,19,20,21,22,23,24,25,26}. By connecting these key points, a closed polygon is formed, which extends upward to the hairline, forming the complete forehead region.

[0011] For the periorbital region, including the left eye region and the right eye region, the eyelid contour of the left eye region is defined by the key point set {37,38,39,40,41,42}, and extended outward by 5-10 pixels to form a rectangular region including the corner of the eye. The eyelid contour of the right eye region is defined by the key point set {43,44,45,46,47,48} using the same method. For the cheek area, the left cheek area is defined by the key point set {1,2,3,4,5,28,29,30}, and the right cheek area is defined by the key point set {13,14,15,16,17,32,33,34,35}. These key points are connected by a curve fitting algorithm to form a cheek area that includes the apple cheek and the area around the nasolabial folds. For the nasal region, the contours of the bridge and alar are defined by the key point set {28,29,30,31,32,33,34,35,36}, and the key point {31} is used as the tip of the nose. The complete nasal region, including the bridge, alar and tip of the nose, is constructed by the triangulation algorithm. For the mandibular region, the mandibular contour line is defined by the key point set {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17} and extends upward to the mandibular angle, forming an area that includes the chin and the edge of the mandible. For each facial sub-region, generate the corresponding binary mask: First, based on the boundary key points of each region, an initial mask is generated using a polygon filling algorithm. For complex regions such as the eye area and nose, morphological closing operations are used to fill any possible holes. Second, for overlapping areas of adjacent regions, they are divided according to preset priority rules, specifically including the eye area with the highest priority, followed by the nose area, and the cheek area with the lowest priority, to ensure that each pixel belongs to only one main area.

[0012] As a preferred embodiment of the present invention, the extraction of multi-dimensional detail features for each facial sub-region is as follows: For each central pixel in the image, a predetermined number of sampling points are selected within its circular neighborhood. The LBP feature and two orientation-sensitive features are calculated. The two orientation-sensitive features include: first, obtaining the gradient orientation consistency feature by calculating the gradient orientation difference between each sampling point and the center point; second, obtaining the orientation distribution histogram feature by dividing the 0-360 degree orientation range into multiple orientation intervals and counting the number of sampling points with consistent gradient orientations in each interval. Principal component analysis was used to reduce the dimensionality of the LBP features at each scale, while retaining a predetermined proportion of the variance contribution rate. We constructed three-scale features, including basic scale, intermediate scale and macro scale, to capture skin structure features at different levels and obtain inter-scale correlation features. The channel attention module automatically learns the weights of features at each scale and performs weighted fusion; the LBP features, two-direction sensitive features and inter-scale correlation features are concatenated in a predetermined order to form the final feature vector.

[0013] As a preferred embodiment of the present invention, the method of quantifying the fine structural features of the skin surface using a multi-scale directional sensing local binary mode specifically includes: First, a multi-scale variance map is calculated using sliding windows of three different scales. For each scale, a variance map of the corresponding scale is generated by sliding a window on the image and calculating the local variance value within the window. Second, the variance maps generated at the corresponding scales are subjected to variance map fusion and enhancement processing, including calculating the saliency map of each scale variance map and determining the fusion weight accordingly, fusing the variance maps of multiple scales; calculating the local autocorrelation matrix and performing eigenvalue decomposition to obtain the principal direction features and direction consistency measure. Third, statistical features, structural features, and multi-scale correlation features are extracted from the fused variance plot. The statistical features include the variance plot mean, standard deviation, skewness, and kurtosis. The structural features include the proportion of high-variance regions, the variance plot entropy, and the spatial autocorrelation length. The multi-scale correlation features include the inter-scale variance ratio and the inter-scale correlation.

[0014] As a preferred embodiment of the present invention, the weighted granularity features are input into a pre-trained age prediction model to output a predicted age value, specifically including: A regional contribution benchmark weight allocation mechanism based on the physiological characteristics of facial skin aging was established, in which the periorbital region was assigned the first weight coefficient, the forehead region the second weight coefficient, the cheek region the third weight coefficient, the nose region the fourth weight coefficient, and the chin region the fifth weight coefficient. For each facial region, feature quality assessment metrics were calculated, including feature saliency, feature stability, and light adaptability metrics. Based on the aforementioned feature quality evaluation index, the final fusion weight of each region is calculated, and the detail feature vector of each facial region is weighted and summed with the corresponding final fusion weight to generate a global detail feature vector. The weighted and fused global granularity feature vector is input into the pre-trained age prediction model to obtain the actual age prediction value.

[0015] Secondly, the present invention also provides a facial fine-grained age prediction system, the system comprising: The image preprocessing module is used to standardize facial images, including face detection, key point localization, geometric correction, and illumination normalization. The facial region segmentation module is used to divide the standardized facial image into multiple sub-regions, including the forehead region, the periorbital region, the cheek region, the nose region, and the jaw region. The detail feature extraction module is used to extract multi-dimensional detail features for each facial sub-region, including texture roughness, pore density, skin smoothness, and micro-wrinkle density, and uses a multi-scale directional perception local binary mode to quantify the fine structural features of the skin surface. The age prediction module assigns weights to the detail features of each region based on the fine structural features and the contribution of different facial regions to age prediction. The weighted detail features are then input into the pre-trained age prediction model, and the predicted age value is output through the non-linear mapping relationship between detail features and age.

[0016] The beneficial effects of this invention are as follows: by adopting a regional analysis strategy, it can better handle the situation where the degree of aging is inconsistent in different areas of the face, and by analyzing skin texture indicators, it can capture more refined information on age changes, thereby improving the accuracy of age prediction, especially significantly enhancing the ability to identify young people and early aging characteristics.

[0017] Furthermore, the method provided by this invention has good robustness and can adapt to different lighting conditions, shooting angles and image qualities, and can maintain high prediction accuracy in everyday photos taken by mobile devices. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method steps provided by the present invention; Figure 2 This is a schematic diagram of facial region segmentation using this method; Figure 3 This is a schematic diagram illustrating the fine-grained feature extraction method of this approach. Figure 4 This is a comparative chart of the average absolute projection histograms in the experimental results of this method; Figure 5 This is a comparison chart of the prediction accuracy trends in the experimental results of this method; Figure 6 This is a schematic diagram of the system structure provided by the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0020] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0021] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0022] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0023] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0024] Please refer to the attached document. Figure 1 In a first aspect, the present invention provides a method for age prediction based on fine facial granularity, the method comprising the following steps: Step S1: Obtain a facial image and perform standardization processing on the facial image, including face detection, key point localization, geometric correction, and illumination normalization.

[0025] As a preferred embodiment of the present invention, a multi-task cascaded convolutional neural network is used to perform face detection and key point localization on the input image, as detailed below: Receive the input image after multi-scale scaling and generate preliminary face candidate regions and their bounding box coordinates; Based on the candidate regions, false detections are filtered out using a convolutional layer structure to refine the bounding box positions; based on the refined candidate regions, the final face bounding box and the coordinate values ​​of 68 facial key points are output. In this embodiment, the 68 facial key points are distributed according to the facial anatomy, including: 17 points on the jawline, 5 points on each eyebrow, 9 points on the nose, 6 points on each eye, and 20 points on the lip contour. These key points together define the topological structure of the face.

[0026] As a preferred embodiment of the present invention, geometric correction processing is performed on the 68 facial key points, as follows: Multiple preset key points are selected as reference points for geometric correction, including the center point of the left eye, the center point of the right eye, the tip of the nose, the left corner of the mouth, and the right corner of the mouth. The mapping relationship between the coordinates of the reference point in the current image and the coordinates of the corresponding point in the standard frontal template is calculated, and based on the mapping relationship, the detected face image is transformed to the standard frontal pose through affine transformation.

[0027] As a preferred embodiment of the present invention, the step of transforming the detected face image to a standard frontal pose through affine transformation specifically includes: Calculate the transformation matrix from the current facial key points to the key points of the standard frontal template; The transformation matrix is ​​applied to the entire facial region, and the output is a frontal face image with a size normalized to 256×256 pixels, where the line connecting the eyes remains horizontal and the facial region occupies the main part of the image.

[0028] As a preferred embodiment of the present invention, the geometrically corrected image is subjected to illumination normalization processing, as follows: The geometrically corrected image is converted from the RGB color space to the LAB color space, thereby separating the luminance channel and the color channel; The separated luminance channel is divided into multiple local regions by limiting contrast adaptive histogram equalization. Histogram equalization is performed independently on each region, and artifacts at the region boundaries are eliminated by bilinear interpolation. In this embodiment, the specific parameters of the contrast-limited adaptive histogram equalization method include: the cropping limit is set to 2.0, and the local region grid size is set to 8×8 pixels.

[0029] The equalized luminance channel is then merged back with the original color channel and converted back to the RGB color space to obtain the image after illumination normalization.

[0030] Therefore, this invention provides a reliable data foundation for age prediction based on facial detail by locating facial key points, standardizing geometric pose through affine transformation, normalizing illumination through adaptive histogram equalization, and combining multi-index quality assessment. It also effectively eliminates the interference of non-physiological factors such as pose and illumination on age prediction, significantly improving the accuracy and robustness of the prediction model.

[0031] Step S2: Divide the face into multiple sub-regions based on the standardized facial image, including the forehead region, the periorbital region, the cheek region, the nose region, and the jaw region.

[0032] As a preferred embodiment of the present invention, please refer to the appendix. Figure 2 The face is divided into multiple sub-regions based on the standardized facial image. The black dots in the image represent partial example locations of 68 facial key points. In practical applications, the system automatically detects and locates all 68 key points, and then precisely divides each analysis region based on the coordinates of these points. The contribution of detail features in different regions to age prediction varies, with the periorbital and cheek regions having the highest weight. Details are as follows: For the forehead region, its upper boundary is defined by the key point set {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17}, and its lower boundary is defined by the key point set {18,19,20,21,22,23,24,25,26}. By connecting these key points, a closed polygon is formed, which extends upward to the hairline, forming the complete forehead region.

[0033] For the periorbital region, including the left eye region and the right eye region, the eyelid contour of the left eye region is defined by the key point set {37,38,39,40,41,42}, and extended outward by 5-10 pixels to form a rectangular region including the corner of the eye. The eyelid contour of the right eye region is defined by the key point set {43,44,45,46,47,48} using the same method. For the cheek area, the left cheek area is defined by the key point set {1,2,3,4,5,28,29,30}, and the right cheek area is defined by the key point set {13,14,15,16,17,32,33,34,35}. These key points are connected by a curve fitting algorithm to form a cheek area that includes the apple cheek and the area around the nasolabial folds. For the nasal region, the contours of the bridge and alar are defined by the key point set {28,29,30,31,32,33,34,35,36}, and the key point {31} is used as the tip of the nose. The complete nasal region, including the bridge, alar and tip of the nose, is constructed by the triangulation algorithm. For the mandibular region, the mandibular contour line is defined by the key point set {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17} and extends upward to the mandibular angle, forming an area that includes the chin and the edge of the mandible. In this embodiment, a corresponding binary mask is generated for each facial sub-region: First, based on the boundary key points of each region, an initial mask is generated using a polygon filling algorithm. For complex regions such as the eye area and nose, morphological closing operations are used to fill any possible holes. Second, for overlapping areas of adjacent regions, they are divided according to preset priority rules, specifically including the eye area with the highest priority, followed by the nose area, and the cheek area with the lowest priority, to ensure that each pixel belongs to only one main area.

[0034] Step S3: Extract multi-dimensional detail features for each facial sub-region, including texture roughness, pore density, skin smoothness, and micro-wrinkle density, and quantify the fine structural features of the skin surface using a multi-scale directional perception local binary mode.

[0035] Please see the appendix Figure 3The model is simulated using CSS gradients and background patterns. In practical applications, high-resolution microscope images are used. The feature values ​​are standardized by 0-1 to facilitate comparison and fusion between different features. A 128-dimensional feature vector is extracted for each region, containing the above four main features and their sub-parameters. As age increases, skin smoothness decreases significantly, providing a reliable basis for age prediction.

[0036] As a preferred embodiment of the present invention, the extraction of multi-dimensional detail features for each facial sub-region is as follows: For each center pixel in the image, a predetermined number of sampling points are selected within its circular neighborhood, and LBP features and two orientation-sensitive features are calculated, wherein the two orientation-sensitive features include: One method is to obtain the gradient direction consistency feature by calculating the gradient direction difference between each sampling point and the center point; Secondly, the directional distribution histogram features are obtained by dividing the 0-360 degree directional range into multiple directional intervals and counting the number of sampling points with consistent gradient directions in each interval. In this embodiment, the 0°-360° directional range is divided into 8 directional intervals, and the number of sampling points with consistent gradient directions in each interval is counted to form a directional distribution histogram.

[0037] Principal component analysis was used to reduce the dimensionality of the LBP features at each scale, while retaining a predetermined proportion of the variance contribution rate. We constructed three-scale features, including basic scale, intermediate scale and macro scale, to capture skin structure features at different levels and obtain inter-scale correlation features. The channel attention module automatically learns the weights of features at each scale and performs weighted fusion; the LBP features, two orientation-sensitive features, and inter-scale correlation features are concatenated in a predetermined order to form the final feature vector. In this embodiment, the LBP feature adopts the improved rotation-invariant uniform LBP mode, the directional consistency feature adopts the distribution histogram form of multi-directional intervals, and the inter-scale correlation feature is obtained by calculating the mutual information metric between features at each scale.

[0038] Step S4: Based on the fine structural features and the contribution of different facial regions to age prediction, assign corresponding weights to the fine features of each region, input the weighted fine features into the pre-trained age prediction model, and output the predicted age value through the non-linear mapping relationship between fine features and age.

[0039] As a preferred embodiment of the present invention, the method of quantifying the fine structural features of the skin surface using a multi-scale directional sensing local binary mode specifically includes: First, a multi-scale variance map is calculated using sliding windows of three different scales. In this embodiment, the three different scales include a micro-scale window, a meso-scale window, and a macro-scale window. The micro-scale window, with a resolution of 3×3 pixels, is used to capture pore-level changes. The meso-scale window, with a resolution of 5×5 pixels, is used to capture skin texture changes. The macro-scale window, with a resolution of 7×7 pixels, is used to capture overall skin roughness. For each scale, a variance map for that scale is generated by sliding a window across the image and calculating the local variance within the window.

[0040] Second, the variance maps generated at the corresponding scales are subjected to variance map fusion and enhancement processing, including calculating the saliency map of each scale variance map and determining the fusion weight accordingly, fusing the variance maps of multiple scales; calculating the local autocorrelation matrix and performing eigenvalue decomposition to obtain the principal direction feature and direction consistency measure.

[0041] Third, statistical features, structural features, and multi-scale correlation features are extracted from the fused variance plot. The statistical features include the variance plot mean, standard deviation, skewness, and kurtosis. The structural features include the proportion of high-variance regions, the variance plot entropy, and the spatial autocorrelation length. The multi-scale correlation features include the inter-scale variance ratio and the inter-scale correlation.

[0042] In this embodiment, the extracted statistical features, structural features, and multi-scale correlation features are subjected to illumination normalization, which is achieved by subtracting the mean of the background region features and dividing by the standard deviation of the background region features. Simultaneously, the feature extraction parameters are adjusted according to the physiological characteristics of different facial regions, including using a smaller sliding window for the periorbital region, a medium window for the cheek region, and a larger window for the forehead region.

[0043] As a preferred embodiment of the present invention, the weighted granularity features are input into a pre-trained age prediction model to output a predicted age value, specifically including: Establish a regional contribution benchmark weight allocation mechanism based on the physiological characteristics of facial skin aging. The periorbital region is assigned a first weight coefficient, which can be set to 0.3; the forehead region is assigned a second weight coefficient, which can be set to 0.15; the cheek region is assigned a third weight coefficient, which can be set to 0.3; the nose region is assigned a fourth weight coefficient, which can be set to 0.13; and the chin region is assigned a fifth weight coefficient, which can be set to 0.12. For each facial region, feature quality assessment metrics were calculated, including feature saliency, feature stability, and light adaptability metrics. In this embodiment, the feature saliency index is determined by calculating the relative value of the Euclidean distance between the feature vector of the region and the baseline feature vector of the corresponding age group; the feature stability index is evaluated by calculating the coefficient of variation of the feature values ​​of the local region; and the light adaptability index is determined by evaluating the color difference index before and after light normalization.

[0044] Based on the aforementioned feature quality evaluation index, the final fusion weight of each region is calculated, and the detail feature vector of each facial region is weighted and summed with the corresponding final fusion weight to generate a global detail feature vector. In this embodiment, the fusion process can be represented as a linear combination of the feature vectors of each region and their corresponding weights, wherein the feature vector of each region is a 128-dimensional feature representation, and the global granularity feature vector is a fusion feature representation of the same dimension.

[0045] The weighted and fused global granularity feature vector is input into the pre-trained age prediction model to obtain the actual age prediction value.

[0046] In this embodiment, the age prediction model expands the input feature dimension from 128 to 256 dimensions through a fully connected layer, and uses the ReLU activation function and batch normalization. It adopts a three-layer fully connected network structure, with each layer having dimensions of 256, 128, and 64 dimensions respectively, and uses the ELU activation function and Dropout function for regularization. The 64-dimensional features are mapped to a 1-dimensional output through an age regression layer, and the output value is constrained to between 0 and 1 using the Sigmoid activation function. Finally, the actual age prediction value is obtained through age inverse normalization.

[0047] For example, please refer to the appendix. Figure 4 and attached Figure 5 As shown in Table 1, this experiment was conducted on a dataset containing 10,000 facial images of people of different ages. The results show that the mean absolute error (MAE) of the method presented in this invention is 2.8 years, a 33% improvement compared to traditional methods based on overall facial features (MAE = 4.2 years). Particularly in the 18-30 age group, the prediction error decreased from 5.1 years for the traditional method to 3.2 years, demonstrating better recognition capabilities for younger populations.

[0048] Table 1. Prediction results of this method and traditional methods at different age groups.

[0049] Moreover, this method can be applied to a variety of fields. For example, firstly, in pharmacological and aging research, it can serve as an objective endpoint in clinical trials to evaluate the effectiveness of anti-aging drugs or therapies, helping scientists to understand the trajectory of human aging and individual differences more precisely.

[0050] Secondly, in the field of consumer beauty and skincare, it can be combined with smartphone apps or professional equipment to analyze the user's skin texture and accurately recommend skincare ingredients most suitable for their skin age and condition.

[0051] Third, it can be applied to the quantitative evaluation of beauty effects, providing medical aesthetic institutions with objective tools to quantitatively assess the degree of skin improvement before and after phototherapy, injectable treatments, and skincare product use. This helps consumers demonstrate the effectiveness of their skincare investments and enhances brand loyalty.

[0052] Secondly, this invention also provides a facial fine-grained age prediction system, please refer to the appendix. Figure 6 The system includes: The image preprocessing module is used to standardize facial images, including face detection, key point localization, geometric correction, and illumination normalization. The facial region segmentation module is used to divide the standardized facial image into multiple sub-regions, including the forehead region, the periorbital region, the cheek region, the nose region, and the jaw region. The detail feature extraction module is used to extract multi-dimensional detail features for each facial sub-region, including texture roughness, pore density, skin smoothness, and micro-wrinkle density, and uses a multi-scale directional perception local binary mode to quantify the fine structural features of the skin surface. The age prediction module assigns weights to the detail features of each region based on the fine structural features and the contribution of different facial regions to age prediction. The weighted detail features are then input into the pre-trained age prediction model, and the predicted age value is output through the non-linear mapping relationship between detail features and age.

[0053] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described face-based fine-grained age prediction method.

[0054] In this embodiment, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0055] Fourthly, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores a plurality of instructions; the processor loads instructions from the memory to execute steps in any of the facial fine-grained age prediction methods provided in embodiments of this application.

[0056] Fifthly, embodiments of this application also provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps in any of the facial fine-grained age prediction methods provided in embodiments of this application.

[0057] In this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be accomplished by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0058] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the facial fine-grained age prediction methods provided in embodiments of this application.

[0059] It should be noted that, through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0060] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for face-based fine-grained age prediction, characterized in that, The method comprises the following steps: Obtaining a face image, and performing standardization processing on the face image, including face detection, key point positioning, geometric correction and illumination normalization; According to the face image after standardization processing, the face is divided into multiple sub-regions, including forehead region, eye region, cheek region, nose region and chin region; For each face sub-region, multi-dimensional delicacy features are extracted, including texture roughness, pore density, skin smoothness and micro-wrinkle density, and a multi-scale direction perception local binary pattern is used to quantify the fine structure features of the skin surface; According to the fine structure features and the contribution of different face regions to age prediction, corresponding weights are assigned to the delicacy features of each region, and the weighted delicacy features are input into a pre-trained age prediction model to output a predicted age value.

2. The method of claim 1, wherein, A multi-task cascaded convolutional neural network is used for face detection and key point positioning on the input image, specifically as follows: Receiving a multi-scale scaled input image, generating a preliminary face candidate region and its bounding box coordinates; According to the candidate region, false detection is filtered out through a convolutional layer structure to refine the bounding box position; according to the refined candidate region, the final face bounding box and the coordinate values of 68 face key points are output.

3. The method of claim 2, wherein, Geometric correction is performed on the 68 face key points, specifically as follows: Selecting multiple preset key points as reference points for geometric correction, including left eye center point, right eye center point, nose tip point, left corner point and right corner point; Calculating the mapping relationship between the coordinates of the reference points in the current image and the coordinates of the corresponding points in the standard front template, and based on the mapping relationship, the detected face image is transformed to a standard front posture through affine transformation.

4. The method of claim 3, wherein, The detected face image is transformed to a standard front posture through affine transformation, specifically including: Calculating the transformation matrix from the current face key points to the standard front template key points; Applying the transformation matrix to the entire face region to output a front face image with a size of 256x256 pixels.

5. The method of claim 4, wherein the method is based on facial fine-grained age prediction. The image after geometric correction is subjected to illumination normalization processing, specifically as follows: Convert the image after geometric correction from RGB color space to LAB color space to separate the brightness channel from the color channel; For the separated brightness channel, the contrast limited adaptive histogram equalization method is used to divide the brightness channel into multiple local regions, and histogram equalization is independently performed on each region, and the artifacts at the region boundaries are eliminated through bilinear interpolation; The equalized brightness channel and the original color channel are recombined and converted back to RGB color space to obtain the illumination normalized image.

6. The method of claim 5, wherein the method is based on facial fine-grained age prediction. According to the face image after standardization processing, the face is divided into multiple sub-regions, specifically including: For the forehead region, the upper boundary is defined by the key point set {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17}, and the lower boundary is defined by the key point set {18, 19, 20, 21, 22, 23, 24, 25, 26}, and by connecting these key points to form a closed polygon and extending upwards to the hairline position to form a complete forehead region; For the eye region, the left eye region is defined by the key point set {37, 38, 39, 40, 41, 42} to define the eyelid contour, and is extended outward by 5-10 pixels to form a rectangular region containing the eye corner, and the right eye region is defined by the key point set {43, 44, 45, 46, 47, 48} to define the eyelid contour in the same way; For the cheek region, the left cheek region is defined by the key point set {1, 2, 3, 4, 5, 28, 29, 30} to define its left boundary, and the right cheek region is defined by the key point set {13, 14, 15, 16, 17, 32, 33, 34, 35} to define its right boundary, and the key points are connected by a curve fitting algorithm to form a cheek region containing the apple muscle and the periphery of the nasolabial fold; For the nose region, the key point set {28, 29, 30, 31, 32, 33, 34, 35, 36} is used to define the contour of the bridge and wings of the nose, and the key point {31} is used as the tip of the nose, and a complete nose region containing the bridge, wings and tip of the nose is constructed by a triangulation algorithm; For the chin region, the key point set {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17} is used to define the contour line of the chin, and is extended upward to the position of the gonial angle to form a region containing the chin and the edge of the chin; For each facial sub-region, a corresponding binary mask is generated: First, according to the boundary key points of each region, an initial mask is generated by a polygon filling algorithm, and for complex regions such as the eye region and the nose region, a morphological closing operation is used to fill possible holes; Second, for the overlapping parts of adjacent regions, a preset priority rule is used for division, including the eye region with the highest priority, followed by the nose region, and the cheek region with the lowest priority, in order to ensure that each pixel point belongs to only one main region.

7. The method of claim 6, wherein the method is based on facial fine-grained age prediction. The multi-dimensional delicacy features of each facial sub-region are extracted as follows: For each center pixel point in the image, a predetermined number of sampling points are selected in its circular neighborhood, and LBP features and two direction-sensitive features are calculated, wherein the two direction-sensitive features include: one is to obtain gradient direction consistency features by calculating the gradient direction difference between each sampling point and the center point; the other is to obtain direction distribution histogram features by dividing the 0-360 degree direction range into multiple direction intervals and counting the number of sampling points with consistent gradient direction in each interval; Principal component analysis method is used to reduce the dimension of LBP features at each scale, and a predetermined proportion of variance contribution rate is retained; Three-scale features are constructed, including basic scale, intermediate scale and macro scale, which are respectively used to capture different levels of skin structure features to obtain inter-scale correlation features; The channel attention module is used to automatically learn the weight of each scale feature and perform weighted fusion; then the LBP features, the two direction-sensitive features and the inter-scale correlation features are spliced in a predetermined order to form the final feature vector.

8. The method of claim 7, wherein, The multi-scale direction perception local binary pattern is used to quantify the fine structure features of the skin surface, which specifically includes: First, a multi-scale variance map is calculated by using three different size sliding windows, for each scale, a variance map of the corresponding scale is generated by sliding the window over the image and calculating the local variance value within the window; Second, the generated variance maps of the corresponding scales are fused and enhanced, including calculating the saliency map of each scale variance map and determining the fusion weight according to the saliency map, fusing the variance maps of multiple scales, calculating the local autocorrelation matrix and performing eigenvalue decomposition to obtain the principal direction feature and the direction consistency measure; Third, statistical features, structural features and multi-scale correlation features are extracted from the fused variance map, wherein the statistical features include variance map mean, standard deviation, skewness and kurtosis; the structural features include high variance area ratio, variance map entropy value and spatial autocorrelation length; the multi-scale correlation features include scale-to-scale variance ratio and scale-to-scale correlation.

9. The method of claim 8, wherein, The weighted fine texture features are input into a pre-trained age prediction model to output a predicted age value, specifically including: A regional contribution degree benchmark weight distribution mechanism based on facial skin aging physiology is established, wherein the eye area is assigned a first weight coefficient, the forehead area is assigned a second weight coefficient, the cheek area is assigned a third weight coefficient, the nose area is assigned a fourth weight coefficient, and the chin area is assigned a fifth weight coefficient; For each facial region, a feature quality evaluation index is calculated, including a feature saliency index, a feature stability index and a light adaptability index; Based on the feature quality evaluation index, the final fusion weight of each region is calculated, the fine texture feature vector of each facial region is weighted and summed with the corresponding final fusion weight to generate a global fine texture feature vector; The weighted and fused global fine texture feature vector is input into a pre-trained age prediction model to obtain an actual age prediction value.

10. A face fine-grained age prediction system based on, characterized by, The system comprises: An image preprocessing module for standardizing the facial image, including face detection, key point positioning, geometric correction and illumination normalization; A facial region segmentation module for dividing the standardized facial image into multiple sub-regions, including forehead region, eye region, cheek region, nose region and chin region; A fine texture feature extraction module for extracting multi-dimensional fine texture features from each facial sub-region, including texture roughness, pore density, skin smoothness and micro-wrinkle density, and using multi-scale direction perception to quantify the fine structure features of the skin surface; An age prediction module for assigning appropriate weights to the fine texture features of each region according to the contribution of the fine texture features to age prediction, inputting the weighted fine texture features into a pre-trained age prediction model, and outputting a predicted age value through a nonlinear mapping relationship between fine texture features and age.

Images