Age prediction device and age prediction method

Abstract

The present invention aims to provide a more accurate age prediction device and age prediction method. [Solution] The age prediction device 100 of the present invention is an age prediction device that predicts age using a multimodal model with data relating to multiple faces of different properties, and comprises a learning model 200 which includes a face image feature extraction unit 201 that extracts features from face image data, and a face index feature extraction unit 202 that extracts features from face indices created by calculating the ratio of measurement data of each part constituting the face from the face image data, and the learning model 200 performs multimodal learning to fill in the difference with the actual age for the pairs of features extracted by the face image feature extraction unit 201 and the face index feature extraction unit 202.

Description

Technical Field

[0001] The present invention relates to an age prediction device and an age prediction method.

Background Art

[0002] In recent years, age prediction technology has been attempted to be utilized in various fields such as healthcare, security, marketing, etc., and accurately predicting age has become increasingly important. When predicting age, the face is particularly exposed and is a part that is easily affected by age, so the development of a system for predicting age from face images has been actively carried out. When performing age prediction using deep learning, architectures such as convolutional neural networks (CNNs) and Vision Transformers (ViTs) are used, and the accuracy is improving day by day. However, improving only the single-modal model that directly inputs such images has shown limitations in improving accuracy.

[0003] Patent Document 1 describes a specific method for predicting how the parts constituting the face change with aging. However, the age prediction method described in Patent Document 1 does not use face image data, nor does it use AI such as deep learning. Patent Document 2 discloses a method for predicting age by reading signs of the skin such as wrinkles, but Patent Document 2 also does not use face image data, nor does it use AI such as deep learning.

[0004] Furthermore, Non-Patent Document 1 discloses a method for predicting age by combining face image data with numbers and coordinates assigned to points that serve as marks for each part constituting the face, such as the inner and outer corners of the eyes, the contours, and the ends of the lips, and data on the distances and positions of these points. It is considered that the accuracy of age prediction is improved compared to the case of predicting age only from face image data.

Prior Art Documents

Patent Documents

[0005] [Patent Document 1] Japanese Patent Publication No. 2002-360544 [Patent Document 2] Special table 2023-531265 publication [Non-Patent Document 1] Zhang Beichen, "Age Prediction of Subjects in Camera Images Using Deep Learning," Visual Media Laboratory, Department of Information Science, Graduate School of Science and Engineering, Tokyo City University, April 22, 2022, Doctoral Dissertation [Overview of the project] [Problems that the invention aims to solve]

[0006] However, while Patent Documents 1, 2, and Non-Patent Document 1 describe a method that uses facial image data and measurement data of multiple parts that make up the face, there is room for improvement in methods to further enhance the accuracy of age prediction.

[0007] The present invention aims to provide a more accurate age prediction method and age prediction device by using facial image data and measurement data obtained using facial feature points derived from the facial image data as markers, calculating the ratio of multiple measurement data to create a numerical data called a facial index, and learning a multimodal model based on these two different types of data (image data and numerical data).

[0008] Furthermore, the present invention aims to provide an advice method and an advice device that provide advice on how to look younger based on the age prediction method and age prediction device of the present invention. [Means for solving the problem]

[0009] The age prediction device according to the present invention is an age prediction device configured by a multimodal model that accepts data relating to multiple faces with different properties as input, and comprises a learning model including a face image feature extraction unit that extracts features from face image data, and a face index feature extraction unit that extracts features from face indices created by calculating the ratio of measurement data of each part constituting the face from the face image data, and the multimodal model learns to bridge the gap with the actual age using the features extracted by the face image feature extraction unit and the face index feature extraction unit.

[0010] The age prediction device of the present invention, in addition to the above configuration, includes an image input unit that inputs a face image from an imaging device such as a camera or as image data, an age prediction unit that predicts the age using the learning model based on the face image data acquired from the image input unit, a real age input unit that inputs the actual age, and an advice presentation unit that provides advice on makeup or beauty. The learning model generates multiple age prediction data based on different facial indicators, and the advice presentation unit provides advice by referring to the facial indicators that were the basis for generating age prediction data that deviates significantly from the real age data input from the real age input unit. [Effects of the Invention]

[0011] According to the present invention, a specific method and apparatus can be provided that learns a multimodal model based on different types of data (image data and numerical data), namely facial images and facial indices, to predict age with greater accuracy. Furthermore, a method and apparatus can be provided that offers advice on how to look younger. [Brief explanation of the drawing]

[0012] [Figure 1] Block diagram showing the configuration of the age prediction device 100 in the learning phase of Embodiment 1. [Figure 2] Flowchart showing the operation of the age prediction device 100 during the learning phase of Embodiment 1 [Figure 3]Figure showing the measurement positions of measurement data for calculating facial indices [Figure 4] Figure explaining 24 types of facial indices and their definitions [Figure 5] Figure showing the configuration of learning using a deep learning model in an embodiment [Figure 6] Block diagram showing the configuration of the age prediction device 100 in the inference phase of Embodiment 1 [Figure 7] Flowchart showing the operation of the age prediction device 100 in the inference phase of Embodiment 1 of the present invention [Figure 8] Figure showing the configuration of inference using a deep learning model in an embodiment [Figure 9] Figure showing 68 landmarks on the face [Figure 10] Figure showing the evaluation results for each facial index when inference evaluation is performed using 24 types of facial indices [Figure 11] Block diagram showing the configuration when giving advice by the age prediction device 100 in Embodiment 2

Mode for Carrying Out the Invention

[0013] Hereinafter, embodiments of the present invention will be described. Note that the present invention is not limited to the following embodiments.

[0014] (Embodiment 1) The age prediction device in Embodiment 1 of the present invention will be described separately for the learning phase and the inference phase. [1. Learning Phase]

[0015] [1-1. Outline of Configuration] Figure 1 is a block diagram showing the configuration of the age prediction device 100 in the learning phase of this embodiment. The age prediction device 100 comprises a learning model 200 and a storage unit 300. The learning model 200 includes a face image feature extraction unit 201 that takes a face image learning dataset 401 as input and extracts image features for predicting age from face image data, and a face index feature extraction unit 202 that takes a face index learning dataset 402 as input and extracts face index features for predicting age from face index data. The learning model 200 learns a multimodal model using the features extracted by the face image feature extraction unit 201 and the face index feature extraction unit 202. The storage unit 300 stores the learned model.

[0016] The face image learning dataset 401 and the face index learning dataset 402 include real age data, and the face image feature extraction unit 201 and the face index feature extraction unit 202 perform supervised learning. Note that it is not necessary for all of the face image learning dataset 401 and the face index learning dataset 402 to include real age data; semi-supervised learning is also acceptable. Furthermore, while it is assumed that each of the face image learning dataset 401 and the face index learning dataset 402 includes real age data, this is not a limitation. The face image learning dataset 401, the face index learning dataset 402, and the real age data may be configured as a set. In this case, the face image learning dataset 401, the face index learning dataset 402, and the real age data are stored in a single database.

[0017] The learning model 200 is a multimodal model that performs deep learning and consists of a face image feature extraction unit 201 that handles face images, a face index feature extraction unit 202 that handles face indices, and an integration unit 203 that integrates these two. Details of the face index data will be described later, but face indices are calculated by using measurement data (indicated by symbols (b), (c), (d), etc.) for the parts of the face shown in Figure 3 to calculate the ratio of measurement data.

[0018] The age prediction device 100 consists of memory and a processor, and each block shown inside the age prediction device 100 in Figure 1 is implemented by a program in which the memory and processor work together. It may also be implemented as a program that runs on a local computer. It is not necessary to implement all of each block using cloud computing; some may be implemented on a local computer. In addition, although the face image learning dataset 401 and the face index learning dataset 402 are configured as data input from outside the age prediction device 100, they may also be stored inside the age prediction device 100.

[0019] [1-2. Overview of Operation] Figure 2 is a flowchart showing the operation of the age prediction device 100 during the learning phase of this embodiment.

[0020] A face image learning dataset 401 is prepared and fed into the face image feature extraction unit 201 (step S201). The face image feature extraction unit 201 extracts features from the face image (step S202). Furthermore, in parallel with or independently of the extraction of face image features, measurement data for facial parts is extracted from the face image data to prepare a face index learning dataset 402, which is then fed into the face index feature extraction unit 202 (step S211). The face index feature extraction unit 202 extracts features from the face index (step S212). The feature extraction results from steps S202 and S212 are reflected in the integration unit 203. As described above, the integration unit 203 combines the face image feature extraction unit 201 and the face index feature extraction unit 202. The learning model 200 uses the features extracted by the face image feature extraction unit 201 and the face index feature extraction unit 202 to learn how to bridge the gap with the actual age. The trained model of the learning model 200 is stored in the memory unit 300 (step S204).

[0021] Furthermore, the training dataset should be prepared to include as many facial image data as possible. To improve the accuracy of age prediction, it is preferable for the facial image dataset to include diverse facial image data from various races, ages, etc. However, when using it for the purpose of predicting the age of a specific target group, it is desirable to prepare a facial image dataset that includes many facial image data with attributes similar to those of the specific target group. For age prediction training based on facial images and facial indices, either supervised learning or semi-supervised learning is acceptable.

[0022] [1-3. Specific Examples of Learning Phases] The following provides a detailed explanation of specific examples of the learning phases. [1-3-1. Dataset] As datasets corresponding to the facial image learning dataset 401 and the facial index learning dataset 402, we used the CFD (Chicago Face Database), a dataset on human faces created by The University of Chicago Center for Decision Research (2014). This dataset includes data on various races and ages, and in addition to facial images, a variety of measurement data about the face, such as eye width and the distance between the eyes and lips, is available. In this embodiment, the age prediction accuracy is improved by combining facial images and facial indexes.

[0023] The dataset was divided into training, validation, and test sets. The dataset of 1209 images was split into training, validation, and test sets in an 8:1:1 ratio. In the training phase, the model was tuned using the training and validation data (images), and in the inference phase, which will be explained later, evaluation was performed using the evaluation images. The evaluation images were those not used in training. Since this often results in bias, random numbers were used for the splitting. In addition, fixed random numbers were used to avoid the evaluation images being different each time, and shuffling was performed to ensure consistent evaluation. The facial features are explained below.

[0024] [1-3-2. Facial Indicators] A facial index is calculated by using the measurement data (indicated by symbols (b), (c), (d), etc.) for the facial parts shown in Figure 3, and determining the ratio of the measurement data. The reason for not using the measurement data directly as the facial index is that the size of facial parts differs depending on the size of the face, and also to ensure stable learning and to accommodate different image sizes. The facial index used in this study was uniquely defined using two or more measurement data sets. As shown in Figure 4(a), a total of 24 types of facial indexes were defined.

[0025] Figure 4(b) shows the symbols used in Figure 3 (indicated by symbols (b), (c), (d), etc.), the notation for the measurement data shown in Figure 4(a), and the definition of that measurement data.

[0026] [1-3-3. Learning Models] A deep learning model was constructed using Google Colaboratory as the execution environment and Keras and TensorFlow as the deep learning libraries. The deep learning model is a multimodal model that performs age prediction using face images and facial indices. The deep learning model corresponds to training model 200. The neural network that handles face images corresponds to the face image feature extraction unit 201, and the neural network that handles facial indices corresponds to the facial indice feature extraction unit 202. The integration unit 203 integrates the face image feature extraction unit 201 and the facial indice feature extraction unit 202. The learning model 200 performs overall learning on two types of paired data extracted by the face image feature extraction unit 201 and the face index feature extraction unit 202, in order to bridge the gap between the extracted data and the actual age.

[0027] In this example, a neural network for handling facial images, EfficientNetB7 (CNN architecture), pre-trained on ImageNet (a large image database designed for use in research on object recognition software), was fine-tuned, and the output layer was customized as a regression head to perform regression prediction using the image classification model EfficientNetB7. Figure 5 shows the configuration diagram for training using this deep learning model.

[0028] Using the dataset and execution environment described above, a learning model for age prediction was constructed and trained. The trained model is stored in the memory unit 300 and used as the trained model during inference.

[0029] The inference phase is described in detail below. [2. Inference Phase]

[0030] [2-1. Overview of the Structure] Figure 6 is a block diagram showing the configuration of the age prediction device 100 during the inference phase of this embodiment. The age prediction device 100 receives a face image 430 from the face image input unit 440. The face image 430 may be an image taken on the spot with a camera or the like, or it may be image data such as a photograph taken in the past. The face index data creation unit 450 generates face index data using measurement data that shows the feature quantities of facial parts measured based on the image data acquired from the face image input unit 440.

[0031] The inference unit 230 uses the trained model stored in the memory unit 300 during the training phase. The face image feature extraction unit 201 receives image data from the face image input unit 440 and extracts features from the face image data. The face index feature extraction unit 202 receives face index data from the face index data creation unit 450 and extracts features from the face index data. The integration unit 203 integrates the face image feature extraction unit 201 and the face index feature extraction unit 202. The inference unit 230 performs age inference on the two pairs of data extracted by the face image feature extraction unit 201 and the face index feature extraction unit 202 as a whole and predicts the age. The predicted age, which is the result of the age prediction, is output to the predicted age output unit 460. The predicted age output unit 460 is a display screen such as a display, but it may also be a printing device such as a printer. Different execution environments may be used during training and inference.

[0032] The age prediction device 100 consists of memory and a processor. In Figure 6, each block shown inside the age prediction device 100 is implemented by a program in which the memory and processor work together. It may also be implemented as a program running on a local computer. It is not necessary to implement all of the blocks using cloud computing; some may be implemented on a local computer.

[0033] [2-2. Overview of Operation] Figure 7 is a flowchart showing the operation of the age prediction device 100 during the inference phase of this embodiment. The face image input unit 440 acquires a face image 430 (step S701). The acquired face image 430 is input to the face image feature extraction unit 201, and features are extracted based on the face image (step S702). Furthermore, in parallel with or independently of the feature extraction based on the face image, the acquired face image 430 is also input to the face index data creation unit 450 to generate face index data (step S703). The generated face index data is input to the face index feature extraction unit 202, and features are extracted based on the face index (step S704). The inference unit 230 predicts the age using the features extracted by the face image feature extraction unit 201 and the face index feature extraction unit 202. The predicted age is output to the predicted age output unit 460 (step S706).

[0034] [2-3. Specific Examples of the Inference Phase] The following provides a detailed explanation of specific examples of the inference phase. Figure 8 shows the configuration of the inference phase. For the neural network that handles facial images, fine tuning was performed using EfficientNetB7 (CNN architecture) pre-trained on ImageNet (a large image database designed for use in research on object recognition software), and the output layer was customized as a regression head to perform regression prediction with the image classification model EfficientNetB7.

[0035] [2-3-1. Facial Indicators] As shown in the example of the learning phase, it is possible to actually measure facial features from a face image and calculate facial indices using that measurement data. However, in this case, it was difficult to actually measure and obtain facial indices during inference, so we used the machine learning library Dlib to calculate facial indices from face images and input them into the deep learning model. Dlib can measure 68 landmarks as shown in Figure 9 (https: / / www.researchgate.net / figure / dentification-of-facial-landmarks-using-Dlib-a-Facial-landmarks-b-The-position-and_fig2_343699139). In Figure 9, "●" indicates a landmark, and a total of 68 landmarks are assigned. Figure 9(a) is an image showing landmarks assigned to a face image, and Figure 9(b) is an image showing only the landmarks remaining from the image in Figure 9(a).

[0036] [2-3-2. Evaluation of Inference Performance] We evaluated how much the accuracy improved compared to simply predicting age from facial images alone. The Mean Absolute Error (MAE) shown in Equation 1 below was used for evaluation. N is the number of images, y ip is the first predicted value, y ic This indicates the first correct answer. Note that only images in which facial features could be measured were used in this study.

[0037]

number

[0038] [2-3-3. Evaluation Results of Inference Performance] Figure 10 shows the evaluation results for each facial index. In Figure 10, index number 0 represents the case where age prediction was performed using only the face image without using a facial index, and the facial index column is labeled "ImageOnly" to indicate that a facial index was not used.

[0039] The inference results for all (24 types) using face images and facial indices showed higher accuracy in age prediction compared to the results of inference using only face images (MAE of 4.73 for "indicator number 0").

[0040] "Index number 20," PupilToLipL(o) / UpperFaceLength2(m), had the lowest MAE at 3.08, allowing for the most accurate age prediction. Next came "Index number 15," BottomLiptoChin(q) / MidCheakToChinL(r) at 3.19, and "Index number 13," BottomLipToChin(q) / MidCheakToChinR(r) at 3.21. PupilToLipL(o) is the distance from the pupil of the left eye to the upper lip, and UpperFaceLength2(m) is the distance from the eyelid (upper part of the eye) to the upper lip. This ratio of the two has the greatest impact on age prediction accuracy. Two possible reasons for this are that, relative to the height of the pupil, the distance between the pupil and the eyelid changes with age (CFD is only used for adult face images), and the position of the upper lip changes. It is thought that as we age, our muscles weaken, causing our eyelids to droop (the distance between the eyelids and pupils to decrease), and our upper lip to droop as well.

[0041] [3. Effects] As described above, according to this embodiment, age prediction is performed using both face image data and facial index data, rather than using only face image data, and it was confirmed that the accuracy of age prediction is significantly improved.

[0042] In particular, in the inference experiment in the example, when PupilToLipL(o) / UpperFaceLength2(m), "index number 20," was used as the facial index, the MAE was 3.08, which was the most accurate age prediction.

[0043] [4. Others] In the above embodiment, the facial feature extraction unit 202 extracted features based on one of the 24 types of facial features and performed learning and inference. However, it is also possible to learn and perform inference using a combination of multiple facial features. Furthermore, although the ratio of two measurement data shown in Figure 4 was used as the facial feature, it is also possible to adopt the ratio of multiple measurement data, such as the ratio of three measurement data, as the facial feature.

[0044] In the above explanation, the age prediction device for the learning phase and the age prediction device for the inference phase may be configured in separate and different execution environments.

[0045] (Embodiment 2) Generally, when undergoing makeup, including permanent makeup, or cosmetic surgery, there is a need to know in advance what effects can be expected, and to predict the impact on aging, such as how much younger one will look depending on the type of makeup or cosmetic surgery performed.

[0046] Embodiment 2 describes an age prediction device, as described in Embodiment 1, that can provide advice on how to look younger, and an age prediction device that can predict or suggest the effects of makeup and cosmetic surgery on age, such as how much younger one will look.

[0047] [Overview of the structure] For the 24 facial indicators described in Embodiment 1, there is a difference in predicted age compared to the actual age. Among these, the facial features of the facial indicators that predict an age higher than the actual age are likely to be altered to appear younger through makeup, including permanent makeup, cosmetic surgery, etc. Specifically, when comparing the predicted age values ​​using the 24 facial indicators, it becomes possible to suggest that the parts of the face that are calculated to be older represent areas with potential for looking younger.

[0048] The following will be explained in detail using the drawings. Figure 11 is a block diagram showing the configuration when the age prediction device 100 provides advice in Embodiment 2. In Figure 11, the memory unit 300 stores 24 types of facial indices and 24 types of trained models that have been trained on facial images in advance. These 24 types of trained models are used sequentially during inference, and the predicted age values ​​for each of the 24 types of facial indices are sequentially output from the predicted age output unit 460 and stored in the advice generation unit 480. The advice generation unit 480 receives the actual age of the person whose age prediction was performed from the actual age input unit 470. The advice generation unit 480 compares the actual age with the predicted age values ​​for each of the 24 types of facial indices and generates advice based on one or more facial indices where the difference between the actual age and the predicted age value is large. The generated advice is output from the advice output unit 490. The advice output unit 490 is a display screen such as a display, but it may also be a printing device such as a printer.

[0049] An example of specific advice is shown below. This is an example of advice for the case where the difference between actual age and predicted age is largest when using "Indicator Number 20" PupilToLipL(o) / UpperFaceLength2(m), with an actual age of 35 and a predicted age of 39. PupilToLipL(o) is the distance from the pupil of the left eye to the upper lip, and UpperFaceLength2(m) is the distance from the eyelid (upper part of the eye) to the upper lip. A large ratio suggests that the eyelid position has dropped (the distance between the eyelid and pupil has decreased) and the upper lip position has dropped more than the average for the actual age, so appropriate makeup or cosmetic surgery advice should be given according to the cause. The advice generation unit 480 can pre-store individual advice for cases where the age prediction for 24 types of facial indicators deviates significantly from the actual age, and output the necessary advice. Furthermore, the advice does not necessarily need to be stored in the advice generation unit 480. The advice generation unit 480 may also cooperate with other devices connected to a network such as the Internet to receive advice data for the corresponding facial indicators and output it from the advice output unit 490.

[0050] Furthermore, although we will omit the detailed explanation here, by using pre-makeup / pre-plastic surgery facial images and post-makeup / pre-plastic surgery simulated facial images, it becomes possible to see how the age prediction results from this age prediction device change before and after makeup / plastic surgery, thereby alleviating anxiety about makeup and plastic surgery. In this example, we have described an example of use that suggests the product is likely to produce a youthful appearance. However, this is just one example, and the present invention can also be used when you want to appear more dignified or look your age.

[0051] [effect] The age prediction device according to this embodiment has the important effect of objectively predicting the impact of makeup and cosmetic surgery on perceived age for people who desire to have a younger-looking face, thereby alleviating their anxiety about makeup and cosmetic surgery.

[0052] (Other embodiments) Passports and driver's licenses must be renewed at prescribed legal intervals, but if a considerable amount of time has passed since the last renewal, changes in facial photographs may make it difficult to identify the person, and age-related changes may affect identity verification. By using the age prediction device of the present invention, it is possible to raise doubts about the verification process when there is a significant discrepancy between the age predicted or estimated from the person's facial image and the age derived from objectively attached documents (such as the person's age from the date of birth stated on their passport or driver's license). This effect is achieved by the high age prediction accuracy of the age prediction device of the present invention.

[0053] Furthermore, the age prediction device of the present invention can also be used as "age authentication" for the sale of alcohol and tobacco, R-rated movie theaters, and other age-restricted matters, to warn customers when there is a clear discrepancy in age. This allows the person handling the situation to raise objective doubts without being rude to the customer, resulting in beneficial effects.

[0054] Although embodiments and examples of the present invention have been described in detail above, the present invention is not limited to these embodiments and examples, and various improvements and modifications are possible within the scope of the present invention. [Explanation of Symbols]

[0055] 100-year-old age prediction device 200 Learning Models 201 Facial Image Feature Extraction Unit 202 Facial Feature Extraction Unit 203 Integration Department 230 Reasoning Department 300 Memory Unit (stores trained models) 401 Face Image Training Dataset 402 Facial Indicator Training Dataset 430 facial images 440 Face Image Input Section 450 Facial Indicator Data Creation Department 460 Predicted Age Output Unit

Claims

1. An age prediction device that predicts age using a multimodal model with data on multiple faces of different properties, comprising a learning model that includes a face image feature extraction unit that extracts features from face image data, and a face index feature extraction unit that extracts features from face indices created by calculating the ratio of measurement data of each part constituting the face from the face image data, wherein the learning model learns a multimodal model to fill in the difference with the actual age for pairs of features extracted by the face image feature extraction unit and the face index feature extraction unit.

2. The age prediction device according to claim 1, wherein the calculation of the facial index uses the ratio of the distance from the pupil of the left eye to the upper lip and the distance from the eyelid to the upper lip as measurement data.

3. The age prediction device according to claim 1, comprising: an image input unit that inputs a face image from an imaging device such as a camera or as image data; and an age prediction unit that predicts the age of the face image data acquired from the image input unit using the learning model.

4. The age prediction device according to claim 3, comprising: an age input unit for inputting actual age; and an advice presentation unit for providing advice on makeup or beauty, wherein the learning model generates multiple age prediction data based on different facial indicators; and the advice presentation unit provides advice based on the facial indicators that caused the generation of age prediction data that deviates significantly from the actual age data input from the age input unit.

5. An age prediction method that predicts age by multimodal learning using data on multiple faces with different properties, comprising: a face image feature extraction step of extracting features from face image data; a face index feature extraction step of extracting features from face indices created by calculating the ratio of measurement data of each part constituting the face from face image data; and a learning step of learning a multimodal model to fill the difference with the actual age for pairs of features extracted in the face image feature extraction step and the face index feature extraction step.

6. An age prediction program that causes a computer to predict age by multimodal learning using data on multiple faces with different properties, the program comprising: a face image feature extraction step of extracting features from face image data; a face index feature extraction step of extracting features from face indices created by calculating the ratio of measurement data of each part constituting the face from the face image data; and a learning step of learning a multimodal model to bridge the gap with the actual age using pairs of features extracted in the face image feature extraction step and the face index feature extraction step.

7. An age prediction device comprising: an image input unit that inputs a face image from an imaging device such as a camera or as image data; and an inference unit that predicts age using a multimodal model based on the face image data acquired from the image input unit, wherein the multimodal model is a multimodal model that has been trained to fill in the difference between actual age and pairs of features extracted from a facial index created by calculating the ratio of feature quantities extracted from face image data and measurement data of each part constituting the face from the face image data.

Images