Facial aging grade evaluation method and system based on multi-dimensional data

Abstract

The present application relates to the field of medical data processing, in particular to a face aging grade evaluation method and system based on multi-dimensional data, comprising: obtaining a reference age span D1 of historical users in each dimension; obtaining a growth trend of a face data growth sequence based on D1, denoted as M1; obtaining a reference face data span D2 of each dimension for face data distribution of all similar users in each dimension between adjacent reference ages; obtaining an average difference M2 between face data of the current user and face data of the same-age history whose difference with the face data of the current user is greater than D2; and evaluating the face aging grade of the current user by combining M1 and M2 after obtaining the difference between the noise of the face data of all historical users when the face data changes with age under the reference age span and the noise of the face data distribution. The present application ensures the reliability of the face aging evaluation result.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, specifically to a method and system for assessing facial aging levels based on multi-dimensional data. Background Technology

[0002] Assessing facial aging typically requires data from multiple dimensions of the face, such as wrinkles, subcutaneous fat thickness, skin elasticity, stretch recovery rate, stratum corneum moisture content, sebum secretion, and pigmentation area. Machine learning models are then used to evaluate the level of facial aging based on this data. Theoretically, using multi-dimensional facial data can guarantee the objectivity and reliability of the assessment results. However, collecting multi-dimensional facial data requires specialized equipment. It's not feasible to use instruments frequently and continuously for the same user, resulting in a lack of long-term facial data that cannot describe the evolution of aging. Furthermore, the results of different dimensions of facial data are affected by individual differences. For example, different users have different basic physical conditions and recent lifestyle habits, and even the same user's lifestyle or physical condition varies at different times. This leads to different errors and interference in different dimensions of a single facial data test. Therefore, directly using all dimensions of facial data for an aging assessment may not yield objective results. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a method and system for assessing facial aging levels based on multi-dimensional data.

[0004] The facial aging level assessment method and system based on multi-dimensional data of the present invention adopts the following technical solution:

[0005] One embodiment of the present invention provides a method for assessing facial aging levels based on multi-dimensional data, the method comprising the following steps:

[0006] If the change in facial data for each dimension of all historical users with age across any age span is greater than the distribution range of facial data for each dimension across that age span, then that age span is recorded as the reference age span D1 for each dimension.

[0007] For each age group, the historical users whose facial data is most similar to the current user's are recorded as similar users;

[0008] Based on the difference between the current user's age and D1, several reference ages are set. The facial data of the current user in each dimension and the average facial data of similar users in several reference ages constitute a facial data growth sequence. The growth trend of the facial data growth sequence is denoted as M1. For the facial data distribution of all similar users between adjacent reference ages in each dimension, the distribution range of facial data with the same distribution in all adjacent reference ages is denoted as the reference facial data span D2 of each dimension.

[0009] For historical users of the same age whose facial data differs from the current user's facial data by more than D2, the average difference between the current user's facial data and the facial data of the historical users of the same age is denoted as M2.

[0010] The noise of all historical users' facial data as they change with age over a reference age span is used as the difference between the noise of the facial data growth sequence and the noise of each dimension. The combination validity is used to combine M1 and M2 to assess the current user's facial aging level.

[0011] Preferably, when the change in facial data for each dimension of all historical users with age over any age span is greater than the distribution range of facial data for each dimension over that age span, the age span is recorded as the reference age span for each dimension. The specific steps include the following:

[0012] The age difference between any two historical users is denoted as an age span; for any age span K, and for all ages of all historical users, any one of these ages is denoted as the initial age K0, and K+K0 is denoted as the ending age; under any dimension, the absolute value of the difference between the average facial data of all historical users under the ending age and the initial age is denoted as the change in facial data with age under age span K; the maximum difference in facial data of historical users within the age interval [K0, K0+K] is obtained and denoted as the distribution range of facial data of all historical users under age span K, and the difference between the change in facial data with age under age span K and the distribution range is denoted as the first difference of age span K;

[0013] After taking all the ages of all historical users as the initial age, the mean of all the first differences obtained for the age span K is used as the preferred index of the age span K. Among all age spans, the age span with the preferred index greater than 0 and the smallest value is selected and recorded as the reference age span for each dimension.

[0014] Preferably, the specific steps for setting several reference ages based on the difference between the current user's age and D1 are as follows:

[0015] The current user's age is denoted as K1. K1-D1, K1-2×D1, K1-3×D1, ..., K1-m×D1 are all used as reference ages. Here, m represents a preset integer. When there is a reference age that is greater than the minimum age of all historical users, the reference age is deleted.

[0016] Preferably, the facial data of the current user in each dimension and the average facial data of several users of similar ages constitute a facial data growth sequence, including the following specific steps:

[0017] For any given reference age, obtain the first mean of the facial data of all similar users in each dimension; sort all the first means of reference ages in ascending order of reference age to obtain the facial data growth sequence of the current user, and add the facial data of the current user in each dimension to the end of the facial data growth sequence.

[0018] Preferably, the specific steps for the facial data distribution of all similar users between adjacent reference ages in each dimension, where the distribution range of facial data with the same distribution among all adjacent reference ages is denoted as the reference facial data span D2 for each dimension, are as follows:

[0019] The facial data of all similar users between any two reference ages constitutes the facial data distribution of adjacent reference ages in each dimension. For all facial data distributions of all adjacent reference ages, the centers of all facial data distributions are aligned. After linear normalization of the facial data in all the center-aligned facial data distributions, mean-shift clustering is performed to obtain all categories. The category containing the most facial data is taken as the target category. The standard deviation of the facial data before normalization in the target category is denoted as the reference facial data span of each dimension, and is called D2.

[0020] Preferably, the difference between the noise of the facial data of all historical users as they change with age over a reference age span and the noise of the facial data growth sequence is denoted as the binding validity of each dimension, and the specific steps include the following:

[0021] For all historical users at all ages, obtain the average facial data of all historical users in each dimension for each age.

[0022] Starting from any age, several ages are sampled at equal intervals. The difference between the sampled ages is equal to the reference age span. The average facial data corresponding to the sampled ages constitute the first sequence.

[0023] Obtain the noise distribution F1 of the first sequence and the noise distribution F2 of the facial data growth sequence. Let (F1-F2) / F1 denote the binding effectiveness of each dimension.

[0024] Preferably, after obtaining the combination validity of each dimension, the specific steps include the following:

[0025] Obtain the mean density of all categories other than the target category. The ratio of this mean density to the density of the target category is recorded as the uncertainty coefficient of each dimension. Obtain the correction priority index of the reference age span for each dimension. The correction priority index is negatively correlated with the uncertainty coefficient and positively correlated with the combination effectiveness of each dimension. The reference age span when the correction priority index is at its maximum is recorded as the updated reference age span.

[0026] The combined validity of each dimension, as well as the values ​​of M1 and M2, are re-acquired using the updated reference age span. The combined validity of M1 and M2 is then used to assess the current user's facial aging level.

[0027] Preferably, the specific steps for obtaining the growth trend are as follows:

[0028] After linearly normalizing the facial data growth sequence, the least squares method is used to fit the element value at each position in the facial data growth sequence into a straight line, and the slope of the straight line is recorded as the growth trend.

[0029] Preferably, the specific steps for assessing the current user's facial aging level after combining M1 and M2 using the binding effectiveness are as follows:

[0030] Normalize the combination validity of all dimensions; obtain the combination coefficient, where the combination coefficient is equal to 1 when the normalized combination validity is greater than the preset valid threshold, and equal to 0 when the normalized combination validity is less than or equal to the preset valid threshold.

[0031] The average of the growth trend and the relative difference among peers is multiplied by a combination coefficient to denot the aging change feature of the current user in each dimension. The aging change feature of the current user is used to represent the result of combining the growth trend and the relative difference among peers. The facial data of the current user in each dimension and the aging change feature obtained in each dimension are used as the facial aging detection result for each dimension. The facial aging detection results of all dimensions are input into the machine learning model to obtain the facial aging level.

[0032] Another embodiment of the present invention provides a facial aging level assessment system based on multi-dimensional data. The system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the computer program, it performs all the steps of the above-described facial aging level assessment method based on multi-dimensional data.

[0033] The beneficial effects of the technical solution of the present invention are:

[0034] This invention analyzes the growth trend of current users' facial data in each dimension based on an appropriate age span (i.e., a reference age span) to describe the characteristics of facial aging changes with age. It ensures that the trend of facial data changes with age is more obvious and avoids the trend being masked by the differences and distinctions in the facial data of historical users, which helps to make a more reliable assessment of facial aging.

[0035] In addition, the reference facial data span D2 for each dimension is analyzed based on an appropriate age span (i.e., reference age span) to ensure that the average difference of the facial data of historical users of the same age can avoid the interference of noise distribution caused by individual differences of historical users.

[0036] Finally, by combining the growth trend with the average difference among historical users of the same age for aging assessment, this method can simultaneously assess aging from the perspectives of the current user's horizontal aging evolution over time and the vertical aging distribution among individuals of the same age. This mitigates the unreliability of assessment results caused by the lack of long-term monitoring results for users, the differences in physical condition of the same user at different times, and individual differences among different users. Specifically, this invention combines the noise of facial data from all historical users as it changes with age over a reference age span with the noise difference in the facial data distribution. This avoids the inaccuracy or invalidity of the growth trend and average difference among historical users of the same age analyzed based on the reference age span, further ensuring the reliability of aging assessment. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 The flowchart illustrates the steps of a facial aging level assessment method based on multi-dimensional data, as provided in an embodiment of the present invention. Detailed Implementation

[0039] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the facial aging level assessment method and system based on multi-dimensional data proposed by the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0040] 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.

[0041] The following description, in conjunction with the accompanying drawings, details the specific scheme of the facial aging level assessment method and system based on multi-dimensional data provided by this invention.

[0042] Example 1:

[0043] Please see Figure 1 The diagram illustrates a flowchart of a facial aging level assessment method based on multi-dimensional data provided in an embodiment of the present invention. The method includes the following steps:

[0044] Step S101: Collect facial data of the user in several dimensions.

[0045] The dimensions described in this embodiment include: average wrinkle depth, wrinkle density, average wrinkle length, subcutaneous fat thickness, and skin rebound speed.

[0046] The average wrinkle depth, wrinkle density, and average wrinkle length can be obtained using image processing techniques; subcutaneous fat thickness can be obtained using ultrasound imaging; and skin elasticity can be obtained by negative pressure adsorption of the skin and measuring the deformation recovery time. The detection or test results for each of the above dimensions are recorded as facial data for that dimension.

[0047] Step S102: When the change in facial data for each dimension of all historical users with age across any age span is greater than the distribution range of facial data for each dimension across the age span, the age span is recorded as the reference age span for each dimension.

[0048] Users who have already had their facial data collected and undergone aging assessments are recorded as historical users. All dimensions of the facial data of historical users are stored in the database, and the age of each user is also recorded in the database, with the age accurate to each month. In other embodiments, an age can be recorded every three months.

[0049] For any dimension, and all historical users' facial data for that dimension, the average facial data of all historical users at the same age is calculated to obtain the average facial data for each age. The age difference (i.e., the absolute value of the age difference) between any two historical users is obtained. For all historical users' age differences (only one is retained if the age differences are the same), each age difference is recorded as an age span.

[0050] If the change in facial data for each dimension for all historical users across any age span is greater than the distribution range of facial data for each dimension across that age span, then that age span is recorded as the reference age span for each dimension, denoted as D1.

[0051] The reference age span for each dimension represents the smallest age span at which historical users' facial data shows a clear pattern of change with age.

[0052] As an example, if the change in facial data for each dimension of all historical users across any age span is greater than the distribution range of facial data for each dimension within that age span, then that age span is recorded as the reference age span for each dimension. The steps include:

[0053] For any age span K, and for all ages of all historical users, let any one age be denoted as the initial age K0, and K+K0 as the ending age; under any dimension, the absolute value of the difference between the average facial data of the ending age and the average facial data of the initial age is denoted as the change in facial data with age under age span K; obtain the facial data of all historical users within the age interval [K0, K0+K], and the difference between the maximum and minimum values ​​of these facial data is denoted as the distribution range of facial data of all historical users under age span K, and the difference between the change in facial data with age under age span K and the distribution range of facial data of all historical users under age span K is denoted as the first difference of age span K.

[0054] The greater the first difference (≥0) or the larger the first difference, the more pronounced the trend of facial data changes with age. This trend will not be masked by the differences in facial data from historical users. The differences in facial data from historical users are due to individual differences among them. These individual differences include differences in basic physiological conditions, lifestyle habits, and physical condition before facial data testing. These differences cannot be used to describe differences in aging. For example, differences in facial data from some historical users are not due to differences in aging, but rather to differences in their lifestyle, basic skin conditions, or even the same historical user's facial data may differ between two tests within a short period due to changes in their physical condition.

[0055] In special cases, if there are no historical users for a certain age, a linear interpolation algorithm is used to interpolate the average facial data for that age by using the average facial data corresponding to the ages on both sides of that age.

[0056] Thus, for any given age of all historical users, the first difference of the age span K is obtained by using that age as the initial age. For all ages of all historical users, multiple first differences are obtained for each age span K by using these ages as initial ages. The mean of all first differences is used as the optimal index for the age span K. Among all age spans, the age span with the optimal index being greater than 0 and having the smallest value is selected and recorded as the reference age span for each dimension.

[0057] Step S103: For each age group, the historical users whose facial data is most similar to the current user are recorded as similar users.

[0058] The current user refers to a user whose facial data has been collected in all dimensions but whose aging assessment has not yet been performed; for each user (including historical users and current users), the vector formed by the facial data in all dimensions is recorded as the facial features of each user.

[0059] Identifying similar users based on the facial features of historical users and the current user, as an example, this process specifically includes:

[0060] Linear normalization is performed on facial data of the same dimension in the facial features of all historical users and current users to obtain normalized facial features; the cosine similarity between the normalized facial features of each historical user and the normalized facial features of the current user is recorded as the similarity between each historical user and the current user.

[0061] Among all historical users at each age, the N0 historical users with the highest similarity to the current user are obtained and recorded as similar users of the current user at each age; in this embodiment, N0 represents half of the total number of historical users at each age (rounded down).

[0062] Step S104: Set several reference ages based on the reference age span. The facial data of the current user in each dimension and the average facial data of similar users in several reference ages constitute a facial data growth sequence, and obtain the growth trend of the facial data growth sequence.

[0063] For any given dimension, perform the following processing:

[0064] The current user's age is denoted as K1. K1-D1, K1-2×D1, K1-3×D1, ..., K1-m×D1 are all used as reference ages; where m represents an integer, and in this embodiment, m is set to 10. Specifically, if there is a reference age that is greater than the minimum age of all historical users, that reference age is deleted.

[0065] Among them, the difference between adjacent reference ages is equal to the reference time. As mentioned above, the facial data of historical users under adjacent reference ages have obvious trends. These trends are not masked by the differences and distinctions between the facial data of all historical users between adjacent reference ages. This helps to analyze the long-term trend of facial data based on different reference ages and is beneficial to describing the temporal evolution of facial aging.

[0066] Furthermore, for any reference age, all similar users have a high degree of similarity to the current user in all dimensions of facial data. Compared to other historical users besides similar historical users, the facial aging process of similar users has a certain degree of similarity to that of the current user.

[0067] For any given reference age, obtain the first mean of the facial data for all similar users in that dimension. Sort all the first means of reference ages in ascending order to obtain the facial data growth sequence for the current user. Add the current user's facial data for that dimension to the end of the facial data growth sequence. The facial data growth sequence describes how the current user's facial data changes with age.

[0068] The growth trend of the facial data growth sequence is obtained and denoted as M1. The growth trend describes the growth of the current user's facial data as they age. The process of obtaining the growth trend avoids the problem that the facial aging changes of historical users are masked by individual differences, and enables the assessment of aging changes for users who have not had long-term facial data collection.

[0069] As an example, methods for obtaining the growth trend of facial data growth sequences include:

[0070] The facial data growth sequence is linearly normalized, and the element value at each position in the facial data growth sequence is fitted into a straight line using the least squares method. The slope of this straight line is recorded as the growth trend.

[0071] Thus, this embodiment has obtained the growth trend of the current user's facial data in each dimension.

[0072] Step S105: Obtain the reference facial data span for each dimension for all similar users between adjacent reference ages, and obtain the relative differences of the same age for each dimension based on the reference facial data span.

[0073] For any set of adjacent reference ages, the facial data of historical users within these adjacent reference ages exhibits a clear trend. This trend is not masked by the differences in facial data among all historical users between adjacent reference ages. Specifically, these differences manifest as noise in the facial data. For all historical users within adjacent reference ages, due to individual differences (including lifestyle habits and physical conditions), their facial data does not show obvious characteristics of age-related evolution and change. Instead, it clearly shows the noise distribution within the facial data caused by these individual differences. For different adjacent reference ages, the noise distribution caused by individual differences among historical users is the same or similar. For example, the noise distribution caused by individual differences among historical users between K1-D1 and K1-2×D1 is the same or similar to the noise distribution caused by individual differences among historical users between K1-2×D1 and K1-3×D1. In other words, the noise distribution caused by individual differences is independent of age.

[0074] Based on this, the following processing is performed on any dimension:

[0075] For any two reference ages, all similar users' facial data constitute the facial data distribution for adjacent reference ages. For all facial data distributions for all adjacent reference ages, all facial data distributions are centered, and the facial data in all centered facial data distributions are linearly normalized and then subjected to mean-shift clustering to obtain all categories.

[0076] In this embodiment, the bandwidth in mean-drift clustering is set to 0.15. In other embodiments, mean-drift clustering can be replaced with a clustering algorithm based on Gaussian mixture model.

[0077] In this embodiment, the category containing the most facial data is taken as the target category. The facial data in the target category represents facial data with the same distribution pattern in different facial data distributions. These facial data describe the noise distribution caused by individual differences among historical users. The standard deviation of the facial data before normalization in the target category is denoted as the reference facial data span for each dimension, and is called D2.

[0078] The reference facial data span is expressed as follows: only when the difference in facial data between users is greater than D2 can it be determined that there is a clear difference in aging between users. If the difference in facial data between users is less than or equal to D2, then even if the facial data between users is different, it is impossible to determine whether there is a clear difference in aging between users.

[0079] As an example, methods for center-aligning all facial data distributions include:

[0080] For any given facial data distribution, obtain the mean of all facial data in that distribution. Then, calculate the difference between the mean and all facial data in that distribution to obtain a decentralized facial data distribution. All decentralized facial data distributions are the center-aligned facial data distributions.

[0081] The purpose of center alignment is to eliminate the influence of age in the distribution of all facial data.

[0082] Furthermore, for any given dimension, the following processing is performed:

[0083] Obtain historical users of the same age as the current user. In this embodiment, historical users whose age is greater than K1-D1 / 2 and less than K1+D1 / 2 are recorded as historical users of the same age as the current user. Among all historical users of the same age, if the absolute value of the difference between the facial data of any historical user of the same age and the facial data of the current user is greater than D2, then that historical user of the same age is recorded as a valid user.

[0084] The difference between the current user's facial data and the facial data of valid users is calculated. The ratio of the mean of all differences to the current user's facial data is recorded as the relative difference of the current user's facial data in that dimension, denoted as M2.

[0085] The aforementioned valid users and the current user are in the same age group. The differences in facial data between the valid users and the current user are not affected by individual differences and do not affect the judgment of the differences in facial aging between the valid users and the current user. This ensures that the obtained relative differences among users of the same age can reliably reflect the aging situation of the current user relative to other users among users of the same age.

[0086] Step S106: Obtain the binding validity of each dimension based on the noise of the facial data growth sequence, and use the binding validity to combine M1 and M2 to evaluate the current user's facial aging level.

[0087] In this embodiment, multiple dimensions of facial data are used simultaneously when assessing the current user's facial aging level, which theoretically ensures the objectivity and reliability of the assessment results. However, collecting facial data from multiple dimensions requires specialized equipment. For the same user, it is not feasible to use the instrument frequently for aging detection over a long period of time, resulting in a lack of long-term facial data detection results. Furthermore, the detection results of different dimensions of facial data are affected differently by individual differences. For example, different users have different basic physical conditions and recent lifestyle habits, and the same user's lifestyle habits or physical condition are also uncertain at different times. This leads to different errors and interference in the results of a single facial data detection for different dimensions. Therefore, directly using the facial data detection results of all dimensions of the current user for aging assessment may not yield objective evaluation results.

[0088] This embodiment analyzes and processes each dimension, obtaining the growth trend of facial data for each dimension and the relative difference among peers for each dimension. The growth trend of facial data can be used to estimate the long-term changes in facial data of the current user; the relative difference among peers for each dimension describes the difference in facial data between the current user and users of the same age. This embodiment combines the growth trend of facial data for each dimension and the relative difference among peers for each dimension to conduct aging assessment. It can simultaneously assess aging from the current user's horizontal time-based aging evolution and the vertical individual aging distribution among peers, mitigating the problems of unreliable assessment results caused by the lack of long-term monitoring results for users, differences in physical condition at different times for the same user, and individual differences between different users.

[0089] For any given dimension, when combining the growth trend of facial data in that dimension with the relative differences among peers in each dimension, the following processing is performed:

[0090] The noise of facial data of all historical users as they change with age within a reference age span is obtained, and the difference between this noise and that of similar users as they change with age within the reference age span is recorded as the binding validity of each dimension. The binding validity of all dimensions is normalized. In this embodiment, softmax is used for normalization. The normalized binding validity is used to combine the growth trend and the relative difference among peers, and then used for facial aging assessment.

[0091] The greater the combined validity of each dimension, the less noise interference there is in the facial data changes of similar users based on the reference age span compared to the facial data changes of all historical users based on the reference age span. This indicates that the growth trend and relative differences among peers calculated based on the reference age span under each dimension are more reliable, and the more attention should be paid to the combined results of the growth trend and relative differences among peers under that dimension. Conversely, the smaller the combined validity of each dimension, the greater the noise interference in the facial data changes of similar users based on the reference age span. The growth trend and relative differences among peers obtained based on the above reference age span and facial data changes of similar users may still not avoid the following problems: users lack long-term facial data detection results, and the detection results of different dimensions are affected differently by individual differences. The single facial data detection results of a user have different errors and interferences in different dimensions, leading to unreliable evaluation results.

[0092] As an example, the difference between the error of facial data of all historical users as a function of age across a reference age span and the error of similar users as a function of age across the reference age span is denoted as the combination validity of each dimension. The methods include:

[0093] For all historical users of all ages, each age corresponds to an average facial data point;

[0094] Starting from any age, sample several ages at equal intervals. For example, starting from age k0, sample the ages k0, k0+D1, k0+2×D1, ..., and so on, sampling several ages at equal intervals from all ages of all historical users.

[0095] The first sequence is obtained by collecting the average facial data corresponding to the sampled age.

[0096] Thus, starting with any age, we obtain a first sequence; starting with each age, we obtain multiple first sequences. It should be noted that if a first sequence is a subsequence of another first sequence, then that first sequence is deleted.

[0097] Obtain the noise distribution for each first sequence, and denote the mean of the noise distributions of all first sequences as F1;

[0098] Obtain the noise distribution F2 of the facial data growth sequence (obtained by step S104), and denote (F1-F2) / F1 as the binding validity of each dimension.

[0099] This concludes the example.

[0100] Example 2:

[0101] This embodiment provides a method for updating the reference age span for each dimension, specifically including:

[0102] In step S105 of Example 1, multiple categories are obtained under each dimension. In addition to the target category, other categories can also represent the noise distribution caused by individual differences. That is, when there are multiple categories, it means that the noise distribution caused by individual differences is not unique and is not deterministic.

[0103] Based on this, the mean density of all categories other than the target category is obtained. The ratio of this mean density to the density of the target category is denoted as the uncertainty coefficient for each dimension, where the density of a category refers to the number of facial data within that category. A larger uncertainty coefficient indicates the existence of other noise distributions caused by individual differences, in addition to the target category. This indicates strong uncertainty in the noise distribution caused by individual differences, making it impossible to accurately determine the noise distribution caused by individual differences. This is because the reference age span in Example 1 has an error.

[0104] Furthermore, the ratio of the normalized combined validity to the uncertainty coefficient for each dimension is denoted as the correction priority index for the reference age span.

[0105] For all age spans where the preferred index is greater than 0 in step S102 of Embodiment 1, each age span is used as the reference age span for each dimension. Then, steps S103 to S105 of Embodiment 1 are executed to obtain the growth trend and relative differences among peers for each dimension, as well as the normalized combination validity of each dimension. Then, the corrected priority index for each age span is obtained according to the above steps of this embodiment.

[0106] The largest age span of the priority indicator will be used as the reference age span for updating each dimension.

[0107] By utilizing the updated reference age span for each dimension, steps S103 to S106 of Example 1 are executed to achieve an assessment of the current user's facial aging.

[0108] By further updating the reference age span in the above process, the accurate noise distribution caused by individual differences can be extracted, and the growth trend of each dimension and the relative differences among peers can be combined with the greatest possible effectiveness, further ensuring the reliability of the assessment of facial aging.

[0109] In other embodiments, the ratio of the correction priority index obtained in Embodiment 2 to lgG+1 can be recorded as the correction priority index for the reference age span in other embodiments. Then, in these embodiments, facial data is evaluated based on this correction priority index according to the method in Embodiment 2. Here, G represents the value of the reference age span, and lg represents the logarithm to base 10. A smaller G indicates a larger correction priority index, ensuring that the updated reference age span is as small as possible, thus facilitating a more accurate estimation of changes in the current user's facial data. The purpose of using lgG+1 as the denominator in the ratio calculation is twofold: to avoid a denominator of 0 and to avoid an excessively large denominator.

[0110] Example 3:

[0111] Step S106 of Embodiment 1 includes: obtaining the noise distribution of each first sequence.

[0112] As an optional example, the method for obtaining the noise distribution of each first sequence includes:

[0113] The first sequence is subjected to Gaussian filtering using a Gaussian filter kernel of length 5 to obtain the filtered sequence. The absolute value of the difference between the mean of all elements in the first sequence and the mean of all elements in the filtered sequence is denoted as the noise distribution of the first sequence.

[0114] As a preferred example, the method for obtaining the noise distribution of each first sequence includes:

[0115] All elements in the first sequence and all elements in the filtered sequence are linearly normalized. Then, the difference between each element in the normalized first sequence and each element in the normalized filtered sequence is calculated, and the absolute value is taken to obtain the difference sequence.

[0116] The standard deviation of all elements in the difference sequence is denoted as the noise distribution of the first sequence.

[0117] Step S106 of Example 1 includes: combining the growth trend with relative differences among peers.

[0118] As an example, combining growth trends with relative differences among peers includes the following methods:

[0119] The binding coefficient is obtained, where the binding coefficient equals 1 when the normalized binding validity is greater than a preset effective threshold, and equals 0 when the normalized binding validity is less than or equal to the preset effective threshold. In this embodiment, the preset effective threshold is set to 0.45.

[0120] The aging change characteristic of the current user is calculated by averaging the growth trend and the relative differences among peers, then multiplying the result by a combination coefficient. This aging change characteristic represents the combined result of the growth trend and the relative differences among peers.

[0121] In this example, when the binding validity is low, all aging change features are marked as 0 to indicate that the growth trend and relative differences among peers are not relevant (i.e., growth trend and relative differences among peers are not considered when conducting aging assessments).

[0122] It should be noted that in this embodiment, both the growth trend and the relative difference among peers are dimensionless, and the average of the two values ​​is directly calculated.

[0123] As another example, combining growth trends with relative differences among peers includes the following methods:

[0124] A three-dimensional vector, composed of growth trend, relative differences among peers, and the normalized combined validity, is used as the aging change feature of the current user. The aging change feature of the current user is used to represent the result of combining growth trend and relative differences among peers.

[0125] In this example, it is possible to distinguish the aging changes of different users in a refined manner based on the characteristics of aging changes.

[0126] In other examples, growth trends and relative differences among peers are combined, including the following methods:

[0127] The aging change feature of the current user is denoted by the average of the growth trend and the relative difference among peers, multiplied by the combined validity after normalization. This aging change feature represents the result of combining the growth trend and the relative difference among peers. This example avoids misjudging the current user as severely aged when the growth trend and the relative difference among peers are large, but the combined validity of each dimension is small.

[0128] Step S106 of Example 1 includes: combining the growth trend and age-relative differences using the normalized binding validity and applying it to facial aging assessment. As an example, the specific process includes:

[0129] The facial data of the current user in each dimension, as well as the aging change features obtained in each dimension, are used as the facial aging detection results for each dimension.

[0130] The facial aging detection results from all dimensions are input into a machine learning model to obtain the facial aging level.

[0131] The machine learning model in this embodiment is a fully connected neural network, and the training method for this fully connected neural network is as follows:

[0132] Half of the historical users are randomly selected as sample users (this part of the data will no longer be considered as historical users). For any sample user, this sample user is treated as the current user, and the facial aging detection results for this sample user in each dimension are obtained according to the above embodiment. The facial aging detection results for all sample users in all dimensions are recorded as sample data. The sample data of all sample users constitute the dataset. Each sample data is manually assigned an aging level label. In this embodiment, there are a total of 10 aging levels, represented by 0, 0.1, 0.2, ..., 0.9, where the smaller the value, the less serious the aging condition. The aging level of 0 indicates no aging phenomenon.

[0133] A fully connected neural network is trained using a dataset. The loss function during training is the cross-entropy loss function, and the neuron parameters are updated using the stochastic gradient descent algorithm. The fully connected neural network used in this embodiment has 4 intermediate layers, and each intermediate layer has 5 neurons.

[0134] The fully connected network and training method are well-known technologies, and will not be described in detail in this embodiment.

[0135] It should be noted that this embodiment is applicable to scenarios with a sufficient number of historical users, such as when the average number of historical users at each age is greater than 10, which can further obtain more reliable aging assessment results.

[0136] Specifically, when the average number of historical users for each age group is less than 10, Implementation Example 1 is no longer executed. Instead, the facial data of the current user for each dimension is directly used as the facial aging detection result for each dimension. Then, the facial aging detection results for all dimensions are input into the machine learning model according to the method of this embodiment to obtain the facial aging level.

[0137] Example 4:

[0138] As an example, the average wrinkle depth, wrinkle density, and average wrinkle length can be obtained using image processing techniques, including the following steps:

[0139] Facial images of users (including historical and current users) are captured from the front using a depth camera, including color and depth images of the face. The forehead and periorbital regions of the face are obtained using a semantic segmentation network (such as the DeepLabV3 network).

[0140] The Canny edge detection algorithm is used to detect edges in the forehead and peri-eye regions. Each edge is considered a wrinkle, and the ratio of the total number of edge pixels in each region to the total number of pixels in that region is called the wrinkle density. The average length of all edges in each region is calculated, and the ratio of this average length to the width of the facial image is called the average wrinkle length in that region. The width of the facial image is used as the denominator in the ratio calculation to normalize the average wrinkle length.

[0141] The mean grayscale value of all edge pixels in each region on the depth map is obtained. The ratio of this mean to 225 is recorded as the average wrinkle depth of each region. The purpose of using 255 as the denominator in the ratio is to normalize the average wrinkle depth.

[0142] In this embodiment, the wrinkle density in the forehead area and the wrinkle density in the periorbital area are weighted and summed to obtain facial data in the dimension of wrinkle density. The weights of the forehead area and the periorbital area are 0.2 and 0.8, respectively, during the weighted summation.

[0143] In this embodiment, the average length of wrinkles in the forehead region and the average length of wrinkles in the periorbital region are weighted and summed to obtain facial data in the dimension of average wrinkle length. The weights of the forehead region and the periorbital region are 0.3 and 0.7, respectively, during the weighted summation.

[0144] In this embodiment, the average wrinkle depth in the forehead region and the average wrinkle depth in the periorbital region are weighted and summed to obtain facial data in the dimension of average wrinkle depth. The weights of the forehead region and the periorbital region are 0.3 and 0.7, respectively, during the weighted summation.

[0145] As an example, subcutaneous fat thickness can be obtained using ultrasound imaging, including the following methods:

[0146] Ultrasound images are collected by closely placing an ultrasound device on the facial skin. The examiner reads the thickness of the subcutaneous fat on the face based on the ultrasound images and records it as L (in millimeters). L / L0 is recorded as the relative thickness of the subcutaneous fat, where L0 represents the maximum subcutaneous fat and is used for normalization. In this embodiment, L0 = 50 millimeters.

[0147] As an example, skin rebound speed can be obtained by negative pressure adsorption of the skin and measuring the deformation recovery time, including the following methods:

[0148] A depth camera is used to capture the rebound video of the user's face after pressure. The mean gray value P1 of the pressed area in the first frame depth image is obtained. Heat sources are detected to select a frame depth image when the facial deformation recovers. The mean gray value P2 of the pressed area in this depth image is obtained. The ratio of the absolute value of the difference between P1 and P2 to 255 is recorded as the facial data of the skin rebound speed dimension.

[0149] Example 5:

[0150] This embodiment provides a facial aging level assessment system based on multi-dimensional data. The system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the computer program, it performs all the steps of the above embodiment.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing facial aging levels based on multi-dimensional data, characterized in that, The method includes the following steps: If the change in facial data for each dimension of all historical users with age across any age span is greater than the distribution range of facial data for each dimension across that age span, then that age span is recorded as the reference age span D1 for each dimension. For each age group, the historical users whose facial data is most similar to the current user's are recorded as similar users; Based on the difference between the current user's age and D1, several reference ages are set. The facial data of the current user in each dimension and the average facial data of similar users in several reference ages constitute a facial data growth sequence. The growth trend of the facial data growth sequence is denoted as M1. For the facial data distribution of all similar users between adjacent reference ages in each dimension, the distribution range of facial data with the same distribution pattern in the facial data distribution of all adjacent reference ages is denoted as the reference facial data span D2 of each dimension. For historical users of the same age whose facial data differs from the current user's by more than D2, the average difference between the current user's facial data and the facial data of the historical users of the same age is denoted as the relative difference of the current user's facial data in that dimension, and is represented as M2. The noise of all historical users' facial data as they change with age over a reference age span is used as the difference between the noise of the facial data growth sequence and the noise of each dimension. The combination validity is used to combine M1 and M2 to assess the current user's facial aging level.

2. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, When the change in facial data for each dimension of all historical users with age across any age span exceeds the distribution range of facial data for each dimension across that age span, the age span is recorded as the reference age span for each dimension. The specific steps involved are as follows: The age difference between any two historical users is denoted as an age span; for any age span K, and for all ages of all historical users, any one of the ages is denoted as the initial age K0, and K+K0 is denoted as the termination age; under any dimension, the absolute value of the difference between the termination age and the average facial data of all historical users under the initial age is denoted as the change in facial data with age under the age span K. The maximum difference in facial data of historical users within the age range [K0, K0+K] is recorded as the distribution range of facial data of all historical users under the age span K. The difference between the change in facial data with age under the age span K and the distribution range is recorded as the first difference of the age span K. After taking all the ages of all historical users as the initial age, the mean of all the first differences obtained for the age span K is used as the preferred index of the age span K. Among all age spans, the age span with the preferred index greater than 0 and the smallest value is selected and recorded as the reference age span for each dimension.

3. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, The specific steps for setting several reference ages based on the difference between the current user's age and D1 are as follows: The current user's age is denoted as K1, and K1-D1, K1-2×D1, K1-3×D1, ..., K1-m×D1 are all used as reference ages; Where m represents a preset integer. When there is a reference age that is greater than the minimum age of all historical users, the reference age is deleted.

4. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, The facial data of the current user in each dimension and the average facial data of several similar users at reference ages constitute a facial data growth sequence, which includes the following specific steps: For any given reference age, obtain the first mean of the facial data of all similar users in each dimension; sort all the first means of reference ages in ascending order of reference age to obtain the facial data growth sequence of the current user, and add the facial data of the current user in each dimension to the end of the facial data growth sequence.

5. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, The specific steps for the facial data distribution of all similar users between adjacent reference ages in each dimension, where the distribution range of facial data with the same distribution among all adjacent reference ages is denoted as the reference facial data span D2 for each dimension, are as follows: The facial data of all similar users between any two reference ages constitutes the facial data distribution of adjacent reference ages in each dimension. For all facial data distributions of all adjacent reference ages, the centers of all facial data distributions are aligned. After linear normalization of the facial data in all the center-aligned facial data distributions, mean-shift clustering is performed to obtain all categories. The category containing the most facial data is taken as the target category. The standard deviation of the facial data before normalization in the target category is denoted as the reference facial data span of each dimension, and is called D2.

6. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, The difference between the noise of all historical users' facial data as they change with age over a reference age span and the noise of the facial data growth sequence is denoted as the combination validity of each dimension. The specific steps involved are as follows: For all historical users at all ages, obtain the average facial data of all historical users in each dimension for each age. Starting from any age, several ages are sampled at equal intervals. The difference between the sampled ages is equal to the reference age span. The average facial data corresponding to the sampled ages constitute the first sequence. Obtain the noise distribution F1 of the first sequence and the noise distribution F2 of the facial data growth sequence. Let (F1-F2) / F1 denote the binding effectiveness of each dimension.

7. The facial aging level assessment method based on multi-dimensional data according to claim 5, characterized in that, After obtaining the validity of the combination of each dimension, the specific steps are as follows: Obtain the mean density of all categories other than the target category. The ratio of this mean density to the density of the target category is recorded as the uncertainty coefficient of each dimension. Obtain the correction priority index of the reference age span for each dimension. The correction priority index is negatively correlated with the uncertainty coefficient and positively correlated with the combination effectiveness of each dimension. The reference age span at which the correction priority index is maximized is recorded as the updated reference age span; The combined validity of each dimension, as well as the values ​​of M1 and M2, are re-acquired using the updated reference age span. The combined validity of M1 and M2 is then used to assess the current user's facial aging level.

8. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, The specific steps for obtaining the growth trend are as follows: After linearly normalizing the facial data growth sequence, the least squares method is used to fit the element value at each position in the facial data growth sequence into a straight line, and the slope of the straight line is recorded as the growth trend.

9. The facial aging level assessment method based on multi-dimensional data according to claim 1, characterized in that, The specific steps involved in assessing the current user's facial aging level after combining M1 and M2 using the aforementioned binding effectiveness are as follows: Normalize the combination validity of all dimensions; obtain the combination coefficient, where the combination coefficient is equal to 1 when the normalized combination validity is greater than the preset valid threshold, and equal to 0 when the normalized combination validity is less than or equal to the preset valid threshold. The average of the growth trend and the relative difference among peers is multiplied by a combination coefficient to denot the aging change feature of the current user in each dimension. The aging change feature of the current user is used to represent the result of combining the growth trend and the relative difference among peers. The facial data of the current user in each dimension and the aging change feature obtained in each dimension are used as the facial aging detection result for each dimension. The facial aging detection results of all dimensions are input into the machine learning model to obtain the facial aging level.

10. A facial aging level assessment system based on multi-dimensional data, which includes: The memory, the processor, and the computer program stored in the memory and executable on the processor are characterized in that, when the processor runs the computer program, it performs all the steps of the facial aging level assessment method based on multi-dimensional data according to any one of claims 1 to 9.

Images