WO2019223080A1 - Bmi预测方法、装置、计算机设备和存储介质 - Google Patents

Bmi预测方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2019223080A1
WO2019223080A1 PCT/CN2018/095386 CN2018095386W WO2019223080A1 WO 2019223080 A1 WO2019223080 A1 WO 2019223080A1 CN 2018095386 W CN2018095386 W CN 2018095386W WO 2019223080 A1 WO2019223080 A1 WO 2019223080A1
Authority
WO
WIPO (PCT)
Prior art keywords
bmi
model
training
sample data
interval determination
Prior art date
Application number
PCT/CN2018/095386
Other languages
English (en)
French (fr)
Inventor
王健宗
吴天博
马进
肖京
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019223080A1 publication Critical patent/WO2019223080A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/08—Insurance
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment

Definitions

  • the present application relates to the field of computers, and in particular, to a BMI prediction method, apparatus, computer equipment, and storage medium.
  • BMI Body Mass Index
  • BMI Body Index
  • Diabetes According to statistics of diabetic patients in China, 80% of diabetic patients are obese, and the longer the period of obesity, the higher the risk of diabetes.
  • Hypertension Obese friends are hypertrophic. They must increase blood volume and cardiac output to meet the needs of the body. In the long term, they will increase the burden on the heart and cause hypertension.
  • Fatty liver Due to the accumulation of body fat in the liver, fatty liver is formed, also known as fatty liver deformation.
  • fatty liver will further worsen, causing hepatitis and cirrhosis.
  • the health risks of obese people are still very large, so insurance companies will also pay attention to the BMI value of policyholders. It is inefficient to calculate the BMI every time you measure the height and weight of the insured person.
  • the main purpose of this application is to provide a BMI prediction method, device, computer equipment, and storage medium for quickly obtaining the BMI of the applicant.
  • a BMI prediction method is proposed in the present application, which includes: acquiring current face features of an insured person;
  • the current face features are input into a preset BMI interval decision model trained based on a neural network model for calculation; wherein, the BMI interval decision model is trained based on the face features and sample data composed of BMI categories associated with the face features Made
  • the present application also provides a BMI prediction device, including:
  • An obtaining unit configured to obtain the current facial features of the applicant
  • An input calculation unit is configured to input the current face feature into a preset BMI interval determination model trained based on a neural network model for calculation; wherein the BMI interval determination model is based on a face feature and a BMI associated with the face feature Training from sample data composed of categories;
  • An output unit is used to output the BMI result of the applicant.
  • the present application further provides a computer device including a memory and a processor.
  • the memory stores computer-readable instructions.
  • the processor implements the steps of any one of the foregoing methods when the computer-readable instructions are executed.
  • the present application also provides a computer non-volatile readable storage medium having computer-readable instructions stored thereon, characterized in that when the computer-readable instructions are executed by a processor, the steps of the method described in any one of the foregoing are implemented.
  • the BMI prediction method, device, computer equipment, and storage medium of the present application predicts the BMI of the policyholder through face features, which can quickly infer the type of the BMI of the policyholder, to prevent the policyholder from falsely reporting the BMI, and because it is calculated by a computer Therefore, no special BMI test equipment is needed, saving costs.
  • FIG. 1 is a schematic flowchart of a BMI prediction method according to an embodiment of the present application
  • FIG. 2 is a schematic flowchart of a BMI prediction method according to an embodiment of the present application.
  • FIG. 3 is a schematic block diagram of a structure of a BMI prediction apparatus according to an embodiment of the present application.
  • FIG. 4 is a schematic block diagram of a structure of an obtaining unit according to an embodiment of the present application.
  • FIG. 5 is a schematic block diagram of a structure of an input operation unit according to an embodiment of the present application.
  • FIG. 6 is a schematic block diagram of a training module according to an embodiment of the present application.
  • FIG. 7 is a schematic block diagram of a training module according to an embodiment of the present application.
  • FIG. 8 is a schematic block diagram of a structure of a BMI prediction apparatus according to an embodiment of the present application.
  • FIG. 9 is a schematic block diagram of a computer device according to an embodiment of the present application.
  • an embodiment of the present application provides a BMI prediction method, including steps:
  • the current face features of the insured person are generally extracted from the face picture taken by the insured person at the insured site, and the face picture is generally a front face of the face, and then the face picture Facial features, such as face width, length, ratio of length to width, proportion of facial features, degree of depression of eye contour, line features of contour of facial features, etc.
  • the neural network model includes multiple types, such as VGG19 model, VGG16 model, VGG-F model, ResNet50 model, ResNet152 model, DPN131 model, InceptionV3 model, Xception model, DenseNet model, and AlexNet model.
  • DPN Dual, Path, Network
  • the above DPN, ResNeXt, and DenseNet are existing network structures, and are not described in detail here.
  • the above-mentioned BMI interval determination model is a model trained on sample data.
  • facial features of a category to which a known BMI belongs are obtained, and then the facial features and their corresponding BMI categories are divided into a training set and a test set, and then the above DPN model is obtained by training the sample data of the training set.
  • a result training model, and then the sample data of the test set is input to the result training model for testing. If the test passes, the above result training model is used as the above BMI interval determination model.
  • the category to which the BMI belongs refers to the category obtained by classification according to the size of the BMI. The specific table is as follows:
  • the larger the category to which the BMI belongs the larger the corresponding BMI.
  • the output result is the corresponding category.
  • BMI is divided into six categories, with labels 0 to 5. In other embodiments, more classifications and the like may be set, which may be specifically set according to specific situations.
  • the above BMI result is the estimated BMI category of the insured person.
  • the staff or the corresponding recommendation system can recommend an insurance product suitable for the insured person according to the BMI category of the insured person, or determine whether the insured person can apply for insurance It has selected insurance products etc.
  • the above-mentioned step S1 of obtaining the current facial features of the policyholder includes:
  • S101 Acquire a current face picture of the insured person, and extract a face area picture in the current face picture;
  • the face area picture in the current face picture is extracted to facilitate the extraction of the face features of the face picture.
  • the method for extracting the face area picture in the current face picture can be operated using a face detection algorithm such as dlib, OpenCV, and dlib and OpenCV are conventional methods in the field of face recognition, and are not described herein again.
  • the extracted face area picture generally includes only the face area, in order to obtain more features, the detected face area picture is expanded.
  • the specific expansion method is as follows:
  • x, y, w, h represent the coordinate information of the face area obtained by face detection, where x, y represent the coordinates of the upper left corner, and w and h represent the width and height of the detected face.
  • the expanded face information is x_n, y_n, w_n, h_n, respectively:
  • x_n x- (w / 4)
  • h_n y_n + 3 * h / 2
  • the expanded face area picture can include not only the face area, but also the neck and head areas, and the features are more sufficient.
  • step S103 since there are more features on the expanded face region picture, such as neck features, the calculation results of the BMI interval determination model will be more accurate.
  • the method for obtaining the BMI interval determination model includes:
  • S21 Obtain a specified amount of sample data, and divide the sample data into a training set and a test set; wherein the sample data includes a face feature and a BMI corresponding to the face feature;
  • the sample data of the training set is input into a neural network model for training; wherein, during the training process, a stochastic gradient descent method is used, and parameters of each layer of the neural network model are updated by using a back conduction rule to obtain a result training model;
  • the specified amount is a set value of the sample amount of the sample data, and can be set according to specific requirements, such as the specified amount is 100,000 sample data.
  • a large number of known face features and BMI categories associated with the face features are used as sample data, and then a preset neural network model is trained.
  • the sample data is randomly divided into two sets, that is, the training set and the test set, and then the above neural network model is trained based on the training set sample data. After obtaining an input face feature, the corresponding BMI category is output. Results train the model. After training is completed to obtain the result training model, the result training model is verified through the sample data of the test set to determine whether the result training model is available.
  • the above stochastic gradient descent method is to randomly sample some training data to replace the entire training set. If the sample size is large (such as hundreds of thousands), then only tens of thousands or thousands of samples may be used to iterate to The optimal solution can improve the training speed.
  • the above-mentioned back conduction law (BP) is based on the gradient descent method.
  • the input-output relationship of the BP network is essentially a mapping relationship: the function completed by a BP neural network with n inputs and m outputs is from n-dimensional A continuous mapping of a finite field into a finite field in m-dimensional Euclidean space. This mapping is highly non-linear.
  • the information processing ability of BP network comes from the multiple recombination of simple nonlinear functions, so it has a strong function reproduction ability.
  • the method before step 22 above, inputting the sample data of the training set into the neural network model for training, the method includes:
  • step S22 in order to avoid overfitting during the process of training the model, usually we need to input a sufficient amount of sample data.
  • the sample data is limited, so the sample data volume needs to be made larger by special methods.
  • the following methods can be used to enhance the data: Rotation / reflection, randomly rotate the image by a certain angle; change the orientation of the image content; Flip transformation (flip), flip the image along the horizontal or vertical direction; zoom transformation (zoom), to enlarge or reduce the image according to a certain ratio; shift transformation (shift), the image on the image plane is translated in a certain way; using random Or manually specify the translation range and translation step, and pan in the horizontal or vertical direction; change the position of the image content, etc., that is, an original picture is processed by means such as reorientation and rotation to get more Sample data, which in turn solves the problem of low sample data and overfitting.
  • the above step S22 of inputting the sample data of the training set into a neural network model for training includes:
  • S2203 The training model is obtained by training the neural network model after initialization.
  • ImageNet is the name of a computer vision system identification project. It is the largest database of image recognition in the world. It is a computer scientist based in Stanford, USA, which simulates a human recognition system and contains 1.2 million pictures. DPN107 The model is a model trained using the ImageNet dataset, and the weight parameters of each layer have been trained.
  • the basis for training the BMI interval decision model in this application is also DPN107, then the parameters of the DPN107 model trained through the ImageNet data set can be initialized into the basic DPN107 model for training the BMI interval decision model, and then the BMI interval decision model is trained to reduce the training time.
  • the initial learning rate is set to 0.01.
  • the step S22 of inputting the sample data of the training set into a neural network model for training includes:
  • the sample data of the training set is input to an unfrozen layer of a neural network model for training, and the result training model is obtained.
  • the neural network is a DPN107 model trained on the ImageNet data set.
  • ImageNet is the name of a computer vision system recognition project and is the largest database of image recognition in the world. It is a computer scientist based in Stanford, USA, which simulates a human recognition system. It contains 1.2 million pictures.
  • the DPN107 model is a model trained using the ImageNet dataset. The weight parameters of each layer have been trained.
  • the BMI interval determination model includes a male BMI interval determination model and a female BMI interval determination model; and the current facial features are input into a preset BMI interval determination model trained based on a neural network model.
  • the operation step S2 includes:
  • a BMI interval determination model for men and women is set to improve the accuracy of the judgment.
  • the training methods of the male BMI interval determination model and the female BMI interval determination model are the same as the training methods described above, except that the sample data used in training the male BMI interval determination model is male face features and their associated BMI, and trains women
  • the sample data used in the BMI interval determination model are women's facial features and their associated BMI.
  • the method for judging the gender of the insured person according to the current facial features can be judged by a preset gender recognition model, which is obtained through facial feature data and corresponding gender training. It is concluded that after inputting a face feature of an unknown gender, the gender corresponding to the face feature will be obtained.
  • the specific process is a conventional method, which is not described in detail here.
  • the above-mentioned BMI interval determination model includes BMI interval determination models of different age groups; and the input of the current facial features to a preset BMI interval determination model trained based on a neural network model is performed.
  • Step S2 includes:
  • a BMI interval determination model corresponding to the age group is called, and the current facial features are input into the BMI interval determination model of the corresponding age group for calculation.
  • the facial features of people in different age groups are different.
  • the age of a person can be judged by the skin state, gloss, contours and other characteristics of the face, and the facial features and BMI of people in different age groups
  • There are some differences in the relevance for example, the correlation between children's face features and BMI is different from young people, middle-aged people, and elderly people.
  • the above different age groups are divided into 7 stages, that is, one year old BMI interval determination model for children aged -6 years, BMI interval determination model for children aged 7-14 years, BMI interval determination model for teenagers 13 to 19 years old, BMI interval determination model for youths 20 to 39 years old, 40-59 years old
  • a model for determining the middle-aged BMI interval at the age of one year and a model for determining the BMI interval for the elderly at the age of 60 or more.
  • a corresponding BMI interval determination model may be set according to the actual customer population, so as to improve the accuracy of determining the human BMI index.
  • the BMI interval determination model for each age group is obtained through the facial features and BMI training of the corresponding age group.
  • the above-mentioned BMI interval determination model includes a height determination model, a weight determination model, and a BMI calculation model, and the above-mentioned input the current face features into a preset BMI interval determination model trained based on a neural network model
  • the operation of step S2 includes:
  • S221 Input the current facial features to the height judgment model and weight judgment model, and obtain the estimated height and estimated weight of the insured person;
  • the height determination model is obtained by training a known neural network model and a corresponding height on a specified neural network model.
  • the weight determination model is obtained by a known facial feature.
  • the corresponding weight are obtained by training the specified neural network model.
  • the above BMI calculation model is the calculation formula of BMI.
  • the specific BMI weight (kg) ⁇ height ⁇ 2 (m), and then obtain the category of BMI according to the classification table of the category to which the BMI belongs.
  • step S3 of outputting a BMI result of the policyholder the method includes:
  • the above insurance product list is an electronic form for the policyholder to view, which records a variety of insurance products preset by the insurance company, and the requirements of each insurance product for the BMI result, for example, the BMI result of the policyholder When it belongs to category 5, it is not applicable to purchase insurance products related to hypertension, diabetes, etc., then insurance products related to hypertension, diabetes, etc. can be hidden or cleared from the above insurance product list to prevent the policyholder from choosing to not Insurance products that can be insured, etc.
  • step S4 of hiding or clearing insurance products that are not suitable for an insured person from a preset insurance product list according to the BMI result the method includes:
  • the parameter information includes at least the age, gender, and job of the insured person.
  • a model for classifying policyholders based on parameter information is preset. Entering the age, gender, and job of the policyholder will output a customer type, and the direction for each type of customer to purchase insurance products is basically In the same way, the number of insurance products purchased by each type of customer is taken from the largest insurance product to the specified insurance product as the recommended insurance product for this type of customer. Once the customer type of the policyholder is determined, only Keep the recommended insurance products in the insurance product list, and hide or clear other ones to further improve the efficiency of the policyholders in purchasing insurance products.
  • the method before the step of obtaining the current facial features of the applicant, the method includes:
  • the ultrasonic wave is transmitted to the shooting area through the ultrasonic transmitting device, and then the reflected wave of the ultrasonic wave is received.
  • the distance between each reflection point and the transmission point is calculated according to the reflection time of the ultrasonic wave, and then the ultrasonic transmission area is drawn.
  • the shape and contour of each object in the object For example, if there is a basketball in the ultrasonic transmission area, the ultrasonic wave will be reflected by the basketball. Because the basketball is circular, the time it takes for the basketball to receive the ultrasonic wave varies with the distance.
  • the reflected waves also vary in time, which in turn draws the outline of the basketball.
  • the photo such as a poster
  • the photo such as a poster
  • the outline is a plane, and if it is a real person, the outline is a 3d outline. That is, the subject's current face picture is a real-life photo, in order to prevent others from using the photo deception system to obtain a lower BMI category, and insurance products that are insured by insurance companies are prohibited.
  • the BMI prediction method of the present application of this application predicts the BMI of the policyholder through face features, which can quickly infer the type of the BMI of the policyholder to prevent the policyholder from falsely reporting the BMI, and because it is calculated by a computer or the like, no special BMI test equipment saves costs.
  • an embodiment of the present application provides a BMI prediction device, including steps:
  • An obtaining unit 10 configured to obtain current face features of an applicant
  • the input operation unit 20 is configured to input a current face feature into a preset BMI interval determination model trained based on a neural network model for calculation; wherein, the BMI interval determination model is based on a face feature and a feature associated with the face feature. Trained from sample data consisting of BMI categories;
  • the output unit 30 is configured to output the BMI result of the applicant.
  • the current face features of the insured person are generally extracted from a face picture taken by the insured person at the insured site, and the face picture is generally a front face of the face, and then the front face picture is extracted.
  • Facial features such as face width, length, ratio of length to width, proportion of facial features, degree of depression of eye contour, line features of contour of facial features, etc.
  • the neural network model includes multiple types, such as a VGG19 model, a VGG16 model, a VGG-F model, a ResNet50 model, a ResNet152 model, a DPN131 model, an InceptionV3 model, a Xception model, a DenseNet model, and an AlexNet model.
  • DPN Downlink Network
  • DPN is a neural network structure that introduces the core content of DenseNet on the basis of ResNeXt, so that the model makes fuller use of features.
  • the above DPN, ResNeXt, and DenseNet are existing network structures, and are not described in detail here.
  • the above-mentioned BMI interval determination model is a model trained on sample data.
  • facial features of a category to which a known BMI belongs are obtained, and then the facial features and their corresponding BMI categories are divided into a training set and a test set, and then the above DPN model is obtained by training the sample data of the training set.
  • a result training model, and then the sample data of the test set is input to the result training model for testing. If the test passes, the above result training model is used as the above BMI interval determination model.
  • the category to which the BMI belongs refers to the category obtained by classification according to the size of the BMI. The specific table is as follows:
  • the larger the category to which the BMI belongs the larger the corresponding BMI.
  • the output result is the corresponding category.
  • BMI is divided into six categories, with labels 0 to 5. In other embodiments, more classifications and the like may be set, which may be specifically set according to specific situations.
  • the above BMI result is the BMI category of the insured person calculated.
  • the staff member or the corresponding recommendation system may recommend an insurance product suitable for the insured person according to the BMI category of the insured person, or determine whether the insured person can apply for insurance. It has selected insurance products etc.
  • the foregoing obtaining unit 10 includes:
  • An obtaining module 101 configured to obtain a current face picture of an insured person, and extract a face area picture in the current face picture;
  • An expansion module 102 configured to perform expansion processing on a face region picture
  • An extraction module 103 is configured to perform feature extraction on the expanded face area picture to obtain current face features.
  • a face area picture in a current face picture is extracted to facilitate extracting a face feature of the face picture.
  • the method for extracting the face area picture in the current face picture can be operated using face detection algorithms such as dlib, OpenCV, and dlib and OpenCV are conventional methods in the field of face recognition, and are not described in detail here.
  • the extracted face area picture generally includes only the face area, in order to obtain more features, the detected face area picture is expanded.
  • the specific expansion method is as follows:
  • x, y, w, h represent the coordinate information of the face area obtained by face detection, where x, y represent the coordinates of the upper left corner, and w and h represent the width and height of the detected face.
  • the expanded face information is x_n, y_n, w_n, h_n, respectively:
  • x_n x- (w / 4)
  • h_n y_n + 3 * h / 2
  • the expanded face area picture can include not only the face area, but also the neck and head areas, and the features are more sufficient.
  • the input operation unit 20 includes:
  • the acquisition classification module 21 is configured to acquire a specified amount of sample data and divide the sample data into a training set and a test set; wherein the sample data includes a face feature and a BMI corresponding to the face feature;
  • the training module 22 is used to input the training set sample data into the neural network model for training.
  • the random gradient descent method is used in the training process, and the parameters of each layer of the neural network model are updated using the back conduction rule to obtain the result training. model;
  • a testing module 23 configured to train a model by using sample data verification results of a test set
  • a labeling module 24 is configured to record the result training model as a BMI interval determination model if the result training model passes the verification.
  • the specified amount is a set value of the sample amount of the sample data, and can be set according to specific requirements, such as the specified amount of 100,000 Sample data, etc.
  • a large number of known face features and BMI categories associated with the face features are used as sample data, and then a preset neural network model is trained.
  • the sample data is randomly divided into two sets, that is, the training set and the test set, and then the above neural network model is trained based on the training set sample data. After obtaining an input face feature, the corresponding BMI category is output. Results train the model.
  • the result training model is verified by the sample data of the test set to determine whether the result training model is available.
  • the above stochastic gradient descent method is to randomly sample some training data to replace the entire training set. If the sample size is large (such as hundreds of thousands), then only tens of thousands or thousands of samples may be used to iterate to The optimal solution can improve the training speed.
  • the above-mentioned back conduction law (BP) is based on the gradient descent method.
  • the input-output relationship of the BP network is essentially a mapping relationship: the function completed by a BP neural network with n inputs and m outputs is from n-dimensional A continuous mapping of a finite field into a finite field in m-dimensional Euclidean space. This mapping is highly non-linear.
  • the information processing ability of BP network comes from the multiple recombination of simple nonlinear functions, so it has a strong function reproduction ability.
  • the input operation unit 20 further includes:
  • the data enhancement module performs data enhancement on the sample data of the training set.
  • the sample data enhancement module in order to avoid overfitting during the process of training the model, we usually need to input a sufficient amount of sample data.
  • the sample data is limited, so the sample data volume needs to be made larger by special methods.
  • the following methods can be used to enhance the data: Rotation / reflection, randomly rotate the image by a certain angle; change the orientation of the image content; Flip transformation (flip), flip the image along the horizontal or vertical direction; zoom transformation (zoom), to enlarge or reduce the image according to a certain ratio; shift transformation (shift), the image on the image plane is translated in a certain way; using random Or manually specify the translation range and translation step, and pan in the horizontal or vertical direction; change the position of the image content, etc., that is, an original picture is processed by means such as reorientation and rotation to get more Sample data, which in turn solves the problem of low sample data and overfitting.
  • the training module 22 includes:
  • a calling submodule 2201 configured to call the weight parameters of each layer of the known neural network model that has been trained and corresponding to the neural network model;
  • An initialization sub-module 2202 configured to initialize the weight parameters of each layer to the weight parameters of each layer of the neural network model
  • the first training sub-module 2203 is configured to obtain a result training model by training the initialized neural network model.
  • the above-mentioned calling sub-module 2201, the initialization sub-module 2202, and the first training sub-module 2203 are sub-module devices used for transfer learning and completing training.
  • the weight parameters of each layer of the already trained model based on the same neural network are called and initialized to the initial weight parameters of the current neural network to be trained.
  • ImageNet is the name of a computer vision system identification project. It is the largest database for image recognition in the world. It is a computer scientist based in Stanford, USA, which simulates a human recognition system and contains 1.2 million pictures. DPN107
  • the model is a model trained using the ImageNet dataset, and the weight parameters of each layer have been trained.
  • the basis for training the BMI interval decision model in this application is also DPN107, then the parameters of the DPN107 model trained through the ImageNet data set can be initialized into the basic DPN107 model for training the BMI interval decision model, and then the BMI interval decision model is trained to reduce the training time.
  • the initial learning rate is set to 0.01.
  • the training module 22 includes:
  • a freezing sub-module 2211 for freezing the weight parameters of a specified layer in the neural network model
  • a second training sub-module 2212 is configured to input sample data of the training set into an unfrozen layer of the neural network model for training, and obtain a result training model.
  • the frozen sub-module 2211 and the second training sub-module 2212 are also sub-module devices for transfer learning and training models.
  • the neural network is the DPN107 model trained through the ImageNet dataset.
  • ImageNet is the name of a computer vision system identification project. At present, the largest database for image recognition in the world is created by computer scientists at Stanford, USA, which simulates human recognition systems. It contains 1.2 million pictures.
  • the DPN107 model is a model trained using the ImageNet dataset. The weight parameters of each layer have been The training is complete.
  • freezing the weight parameter of the specified layer in the DPN107 model means that, in accordance with the sequence between the layers, the volume base layer or / and the fully connected layer before the N ordering is used as the specified layer, where N is greater than 1 and less than the DPN107 model.
  • the BMI interval determination model includes a male BMI interval determination model and a female BMI interval determination model.
  • the input calculation unit 20 includes:
  • a gender recognition module configured to determine the gender of the policyholder based on current facial features
  • the first calling module is configured to call a BMI interval determination model corresponding to a gender according to a determination result, and input a current face feature into a BMI interval determination model corresponding to a gender for calculation.
  • a BMI interval determination model for men and women is set up to improve the accuracy of the judgment.
  • the training methods of the male BMI interval determination model and the female BMI interval determination model are the same as the training methods described above, except that the sample data used in training the male BMI interval determination model is male face features and their associated BMI, and trains women
  • the sample data used in the BMI interval determination model are women's facial features and their associated BMI.
  • the method for judging the gender of an insured person based on the current facial features can be judged by a preset gender recognition model, which is obtained by training on facial feature data and corresponding genders. After the facial features of the unknown gender, the gender corresponding to the facial features will be obtained.
  • the specific process is a conventional method, which will not be repeated here.
  • the above-mentioned BMI interval determination model includes BMI interval determination models of different age groups; the above-mentioned input operation unit 20 includes:
  • An age recognition module configured to determine the age range of the insured according to current facial features
  • the second calling module is configured to call a BMI interval determination model corresponding to the age group according to the determination result, and input the current facial features into the BMI interval determination model of the corresponding age group for calculation.
  • the facial features of people of different age groups are different.
  • the skin state, gloss, contour, and other characteristics of the human face can be used to judge the age of people, and people of different age groups
  • the correlation between facial features of children and BMI is different from young people, middle-aged people, and elderly people.
  • the above different age groups are divided into 7 stages , That is, the BMI interval determination model for children aged 1 to 6 years, the BMI interval determination model for children aged 7-14 years, the BMI interval determination model for teenagers 13 to 19 years old, the BMI interval determination model for youths 20 to 39 years old, A 40-59-year-old middle-aged BMI interval judgment model, and a 60-year-old or older BMI interval judgment model.
  • a corresponding BMI interval determination model may be set according to the actual customer population, so as to improve the accuracy of determining the human BMI index.
  • the BMI interval determination model for each age group is obtained through the facial features and BMI training of the corresponding age group.
  • the BMI interval determination model includes a height determination model, a weight determination model, and a BMI calculation model.
  • the input calculation unit 20 includes:
  • the input module is used to input the current facial features to the height judgment model and weight judgment model, respectively, to obtain the estimated height and estimated weight of the insured person;
  • a calculation module is used to input the estimated height and estimated weight into the BMI calculation model for calculation, and obtain the BMI result of the insured person.
  • the height determination model is obtained by training a known neural network model and a corresponding height through a specified neural network model.
  • the weight determination model is obtained by a known facial feature.
  • the corresponding weight are obtained by training the specified neural network model.
  • the foregoing BMI prediction apparatus further includes:
  • the first shielding unit 40 is configured to hide or clear insurance products that are not suitable for the applicant from a preset insurance product list according to the BMI result.
  • the insurance product list is an electronic form for the policyholder to view, which records a variety of insurance products preset by the insurance company, and the requirements of each insurance product for the BMI result, for example, the policyholder's
  • the BMI result belongs to category 5
  • the insurance products related to hypertension, diabetes, etc. can be hidden or cleared from the above insurance product list to prevent the policyholder from choosing To insurance products that cannot be insured.
  • the above BMI prediction device further includes:
  • An acquisition parameter unit 50 configured to acquire parameter information of an insured person, the parameter information including at least the age, gender, and work of the insured person;
  • a classification unit 60 configured to classify policyholders according to parameter information
  • the second shielding unit 70 is configured to retain the insurance products corresponding to the classification results in the insurance product list according to the classification results, and hide or clear the rest.
  • a model for classifying policyholders based on parameter information is preset, and inputting the age, gender, and job of the policyholder will output a customer type, and each The directions for a group of customers to purchase insurance products are basically the same.
  • the number of insurance products purchased by each type of customers is taken from the largest to the smallest insurance product as the recommended insurance product for this type of customers. After the type of customer of the insured, only the insurance products recommended in the insurance product list are kept, and the others are hidden or cleared, which further improves the efficiency of the insured purchasing insurance products.
  • the BMI prediction apparatus further includes:
  • the scanning unit 11 is configured to perform ultrasonic scanning on the shooting area and receive reflected waves of the ultrasonic waves;
  • a determining unit 12 configured to determine an outline of an object in a shooting area according to a reflected wave
  • the generating instruction unit 13 is configured to determine that the currently captured picture is a real-life picture if the contour meets a preset standard, and generate an instruction for acquiring a current face feature of the policyholder.
  • the ultrasonic wave is transmitted to the shooting area through the ultrasonic transmitting device, and then the reflected wave of the ultrasonic wave is received, and the distance between each reflection point and the emission point is calculated according to the reflection time of the ultrasonic wave. Then draw the outline of the shape of each object in the ultrasonic emission area. For example, if there is a basketball in the ultrasonic emission area, the ultrasonic wave will be reflected by the basketball. Because the basketball is circular, the time it takes for the basketball to receive the ultrasonic wave varies with the distance. The time of the reflected wave received by the ultrasonic receiver is also different, and the outline of the basketball is drawn.
  • the photo such as a poster
  • the photo such as a poster
  • the outline is a plane, and if it is a real person, the outline is a 3d outline. That is, the subject's current face picture is a real-life photo, in order to prevent others from using the photo deception system to obtain a lower BMI category, and insurance products that are insured by insurance companies are prohibited.
  • the BMI prediction device of the present application of the present application predicts the BMI of the insured person through facial features, which can quickly infer the type of BMI of the insured person, to prevent the insured person from falsely reporting the BMI, and because the calculation is performed by a virtual device in the computer, etc. So no special BMI test equipment is needed, saving costs.
  • an embodiment of the present invention further provides a computer device.
  • the computer device may be a server, and its internal structure may be as shown in FIG.
  • the computer device includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the computer design processor is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer-readable instructions, and a database.
  • the memory provides an environment for operating systems and computer-readable instructions in a non-volatile storage medium.
  • the database of the computer equipment is used to store data such as the BMI interval decision model.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions are executed by a processor to implement the processes of the embodiments of the methods described above.
  • FIG. 9 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied.
  • An embodiment of the present invention also provides a computer non-volatile readable storage medium having computer-readable instructions stored thereon.
  • the computer-readable instructions are executed by a processor, the processes of the embodiments of the foregoing methods are implemented.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Public Health (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Marketing (AREA)
  • Development Economics (AREA)
  • Biomedical Technology (AREA)
  • Databases & Information Systems (AREA)
  • Strategic Management (AREA)
  • Pathology (AREA)
  • Technology Law (AREA)
  • Economics (AREA)
  • General Business, Economics & Management (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种BMI预测方法、装置、计算机设备和存储介质，其中，方法包括：获取投保人的当前人脸特征（S1）；将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，所述BMI区间判定模型基于人脸特征，以及与所述人脸特征关联的BMI类别组成的样本数据训练而成（S2）；输出所述投保人的BMI结果（S3）。

Description

BMI预测方法、装置、计算机设备和存储介质
本申请要求于2018年5月25日提交中国专利局、申请号为2018105154574，申请名称为“BMI预测方法、装置、计算机设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及到计算机领域，特别是涉及到一种BMI预测方法、装置、计算机设备和存储介质。
背景技术
目前比较常用BMI(体重指数)来判断肥胖，BMI的全称是Body Mass Index，BMI与体脂重量相关性很好，早些年测评体脂不方便的时候，医学上还曾经用BMI诊断肥胖。如果BMI值在28以上就算是肥胖的。BMI最大的好处就是简单，知道身高体重就能算出来。有些疾病与肥胖密切相关，如：
糖尿病：在我国糖尿病患者统计，80％糖尿病患者都是肥胖者，而当肥胖的时间越长，患糖尿病的风险就越高。高血压：肥胖的朋友体形肥大，必须增加血容量和心输出量才能满足身体需求，长期下来会加重心脏负担引起高血压。脂肪肝：由于体内脂肪在肝脏的堆积，从而形成脂肪肝，又称作脂肪肝变形。严重的是，脂肪肝还会进一步恶化，引发肝炎、肝硬化疾病。综上，肥胖人群健康风险还是非常大的，所以保险公司也会关注投保人的BMI值。而每次测量投保人的身高、体重等计算BMI，效率低下。
技术问题
本申请的主要目的为提供一种快速得出投保人BMI的BMI预测方法、装置、计算机设备和存储介质。
技术解决方案
为了实现上述发明目的，本申请提出一种BMI预测方法，包括：获取投保人的当前人脸特征；
将当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，BMI区间判定模型基于人脸特征，以及与人脸特征关联的BMI类别组成的样本数据训练而成；
输出投保人的BMI结果。
本申请还提供一种BMI预测装置，包括：
获取单元，用于获取投保人的当前人脸特征；
输入运算单元，用于将当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，BMI区间判定模型基于人脸特征，以及与人脸特征关联的BMI类别组成的样本数据训练而成；
输出单元，用于输出投保人的BMI结果。
本申请还提供一种计算机设备，包括存储器和处理器，存储器存储有计算机可读指令处理器执行计算机可读指令时实现上述任一项方法的步骤。
本申请还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，其特征在于，计算机可读指令被处理器执行时实现上述任一项所述的方法的步骤。
有益效果
本申请的BMI预测方法、装置、计算机设备和存储介质，通过人脸特征预测投保人的BMI，可以快速地推断出投保人的BMI类别，以防止投保人虚报BMI，又因为通过计算机等进行测算，所以无需专门的BMI测试设备，节约成本。
附图说明
图1为本申请一实施例的BMI预测方法的流程示意图；
图2为本申请一实施例的BMI预测方法的流程示意图；
图3为本申请一实施例的BMI预测装置的结构示意框图；
图4为本申请一实施例的获取单元的结构示意框图；
图5为本申请一实施例的输入运算单元的结构示意框图；
图6为本申请一实施例的训练模块的结构示意框图；
图7为本申请一实施例的训练模块的结构示意框图；
图8为本申请一实施例的BMI预测装置的结构示意框图；
图9为本申请一实施例的计算机设备的结构示意框图。
本发明的最佳实施方式
参照图1，本申请实施例提供一种BMI预测方法，包括步骤：
S1、获取投保人的当前人脸特征；
S2、将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，所述BMI区间判定模型基于人脸特征，以及与所述人脸特征关联的BMI类别组成的样本数据训练而成；
S3、输出所述投保人的BMI结果。
如上述步骤S1所述，上述投保人的当前人脸特征一般是由投保人在投保现场拍摄的人脸图片中提取的，该人脸图片一般为人脸的正面照，然后提取该正面照上的人脸特征，人脸特征包括如人脸的宽度、长度、长度与宽度的比例、五官的比例、眼睛轮廓的凹陷程度、五官轮廓的线条特征等。
如上述步骤S2所述，上述神经网络模型包括多种，如VGG19模型、VGG16模型、VGG-F模型、ResNet50模型、ResNet152模型、DPN131模型、InceptionV3模型、Xception模型、DenseNet模型和AlexNet模型等，本申请使用DPN模型，DPN(Dual Path Network)是一种神经网络结构，是在ResNeXt的基础上引入了DenseNet的核心内容，使得模型对特征的利用更加充分。上述DPN、ResNeXt和DenseNet是现有的网络结构，在此不在赘述。上述BMI区间判定模型是一种通过样本数据训练而得的模型。在一个实施例中，获取已知BMI所属类别的人脸特征，然后将人脸特征及其对应的BMI类别分成训练集和测试集等，然后通过训练集的样本数据对上述DPN模型进行训练得到一个结果训练模型，然后将测 试集的样本数据输入到结果训练模型中进行测试，若测试通过，则将上述结果训练模型作为上述的BMI区间判定模型。上述的BMI所属类别是指根据BMI的大小进行分类后得到的类别，具体如下表：
BMI 类别
16～18.5 0
18.5～25 1
25～30 2
30～35 3
40 4
>40 5
也就是说，BMI所属类别越大，其对应的BMI越大。本申请中，将当前人脸特征输入到BMI区间判定模型后，其输出的结果即为对应类别。本申请中，将BMI分为六个类别，标签为0～5。在其它实施例中，还可以设置更多的分类等，可以根据具体的情况具体设定。
如上述步骤S3所述，上述BMI结果即为推算出的投保人的BMI类别，工作人员或者相应的推荐系统可以根据投保人的BMI类别推荐适合投保人的保险产品，或者确定投保人是否可以投保其已经选择的保险产品等。
在一个实施例中，上述获取投保人的当前人脸特征的步骤S1，包括：
S101、获取所述投保人的当前人脸图片，并提取出所述当前人脸图片中的人脸区域图片；
S102、对所述人脸区域图片进行扩充处理；
S103、对扩充后的所述人脸区域图片进行特征提取，得到所述当前人脸特征。
如上述步骤S101所述，将当前人脸图片中的人脸区域图片提取出来，方便提取人脸图片的人脸特征。提取出所述当前人脸图片中的人脸区域图片的方法可以使用诸如dlib，OpenCV等人脸检测算法进行操作，dlib和OpenCV是人脸识别领域的常规手段，在此不在赘述。
如上述步骤S102所述，由于上述提取出的人脸区域图片一般仅包含了人脸的区域，为了获得更多的特征，将检测到的人脸区域图片进行扩充，具体的扩充方法如下：
令x,y,w,h表示人脸检测得到的人脸区域坐标信息，其中x，y表示左上角的坐标，w和h表示检测到的人脸的宽和高。扩充后的人脸信息则为x_n,y_n,w_n,h_n,分别为：
x_n＝x-(w/4)
y_n＝y-(h/4)
w_n＝x_n+3*w/2
h_n＝y_n+3*h/2
扩充后的人脸区域图片可以不仅包含了人脸区域，还包含了颈部、头部的区域，特征更为充分。
如上述步骤S103所述，由于扩充后的人脸区域图片上的特征更多，比如脖子的特征等，所以BMI区间判定模型的运算结果会越加的准确。
在一个实施例中，上述BMI区间判定模型的获取方法，包括：
S21、获取指定量的样本数据，并将样本数据分成训练集和测试集；其中，所述样本数据包括人脸 特征，以及与所述人脸特征对应的BMI；
S22、将训练集的样本数据输入到神经网络模型中进行训练；其中，训练的过程中采用随机梯度下降法，利用反向传导法则更新所述神经网络模型各层的参数，得到结果训练模型；
S23、利用所述测试集的样本数据验证所述结果训练模型；
S24、如果验证通过，则将所述结果训练模型记为所述BMI区间判定模型。
如上述步骤S21-S24所述，上述指定量是对样本数据的样本量的设定值，可以根据具体的要求进行设定，如指定量为10万个样本数据等。在本实施例中，即为将大量的已知的人脸特征，以及与人脸特征关联的BMI类别作为样本数据，然后对预设的神经网络模型进行训练。在训练之前，将样本数据进行随机分成两组集合，即训练集和测试集，然后通过训练集的样本数据对上述神经网络模型进行训练，得到一个输入人脸特征后，输出对应的BMI类别的结果训练模型。训练完成得到结果训练模型后，通过测试集的样本数据验证所述结果训练模型，以判断结果训练模型是否可用。上述随机梯度下降法就是随机取样一些训练数据，替代整个训练集，如果样本量很大的情况(例如几十万)，那么可能只用其中几万条或者几千条的样本，就已经迭代到最优解了，可以提高训练速度。上述反向传导法则(BP)它建立在梯度下降法的基础上，BP网络的输入输出关系实质上是一种映射关系：一个n输入m输出的BP神经网络所完成的功能是从n维欧氏空间向m维欧氏空间中一有限域的连续映射，这一映射具有高度非线性。BP网络的信息处理能力来源于简单非线性函数的多次复合，因此具有很强的函数复现能力。
在一个实施例中，上述将训练集的样本数据输入到神经网络模型中进行训练的步骤22之前，包括：
S22’、对所述训练集的样本数据进行数据增强。
如上述步骤S22’所述，为了在训练模型的过程避免出现过拟合(Overfitting)，通常我们需要输入充足的样本数据量。而样本数据有限，所以需要通过特殊的手段将样本数据量变得更大，具体的可以使用如下手段进行数据增强：旋转反射变换(Rotation/reflection)，随机旋转图像一定角度；改变图像内容的朝向；翻转变换(flip)，沿着水平或者垂直方向翻转图像；缩放变换(zoom)，按照一定的比例放大或者缩小图像；平移变换(shift)，在图像平面上对图像以一定方式进行平移；采用随机或人为定义的方式指定平移范围和平移步长,沿水平或竖直方向进行平移；改变图像内容的位置等，即将一张原始的图片通过变向、旋转等手段进行处理，以得到更多的样本数据，进而解决样本数据量低，出现过拟合的问题。
在一个实施例中，上述将训练集的样本数据输入到神经网络模型中进行训练的步骤S22，包括：
S2201、调用对应所述神经网络模型的已经训练完成的已知神经网络模型的各层权重参数；
S2202、将各层的所述权重参数初始化为所述神经网络模型的各层权重参数；
S2203、通过初始化后的所述神经网络模型训练得到所述结果训练模型。
如上述步骤S2201至S2202所述，即为迁移学习的过程，将基于同一神经网络的，且已经训练好的模型的各层权重参数进行调用，初始化为当前待训练的神经网络的初始权重参数。在一具体实施例中， ImageNet是一个计算机视觉系统识别项目名称，是目前世界上图像识别最大的数据库，是美国斯坦福的计算机科学家，模拟人类的识别系统建立的，其中包含120万张图片，DPN107模型是利用ImageNet数据集训练完成的模型，其各层的权重参数已经训练完成。本申请训练BMI区间判定模型的基础同样是DPN107，那么可以将通过ImageNet数据集训练完成DPN107模型的参数初始化到训练BMI区间判定模型的基础DPN107模型中，然后进行BMI区间判定模型的训练，降低训练时间。本申请中将初始学习率设置为0.01。
在另一个实施例中，上述将训练集的样本数据输入到神经网络模型中进行训练的步骤S22，包括：
S2211、冻结所述神经网络模型中指定层的权重参数；
S2212、将所述训练集的样本数据输入到神经网络模型未冻结层中进行训练，得到所述结果训练模型。
如上述步骤S2211和S2212所述，同样为迁移学习的过程，比如神经网络为通过ImageNet数据集训练过的DPN107模型，ImageNet是一个计算机视觉系统识别项目名称，是目前世界上图像识别最大的数据库，是美国斯坦福的计算机科学家，模拟人类的识别系统建立的，其中包含120万张图片，DPN107模型是利用ImageNet数据集训练完成的模型，其各层的权重参数已经训练完成。本申请中，冻结所述DPN107模型中指定层的权重参数是指，按照各层之间的先后顺序，将排序前N的卷基层或/和全连接层作为指定层，其中N为大于1小于DPN107模型总层数的正整数，且N为预设值，比如，N＝100时，即为将DPN107模型前100层的权重参数，在训练上述结果训练模型时只需要重新训练后7层的权重即可，训练的更加快速。本申请将初始学习率设置为0.01。
在一个实施例中，上述BMI区间判定模型包括男性BMI区间判定模型、女性BMI区间判定模型；上述将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算的步骤S2，包括：
S201、根据所述当前人脸特征判断所述投保人的性别；
S202、根据判断结果调用对应性别的BMI区间判定模型，并将所述当前人脸特征输入到对应性别的BMI区间判定模型中进行运算。
如上述步骤S201和S202所述，因为男性和女性的人脸特征存在一定的差异，所以设置针对男性和女性的BMI区间判定模型，以提高判断的准确地性。上述男性BMI区间判定模型、女性BMI区间判定模型的训练方法与上述的训练方法相同，区别在于，训练男性BMI区间判定模型时使用的样本数据为男性的人脸特征及其关联的BMI，训练女性BMI区间判定模型时使用的样本数据为女性的人脸特征及其关联的BMI。本实施例中，根据所述当前人脸特征判断所述投保人的性别的方法，可以通过预设的性别识别模型进行判断，该性别识别模型是通过人脸特征数据及其对应的性别训练而得，当输入未知性别的人脸特征后，会得出该人脸特征对应的性别，具体过程为惯用手段，在此不在赘述。
在另一个实施例总，上述BMI区间判定模型包括不同年龄段的BMI区间判定模型；上述将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算的步骤S2，包括：
S211、根据所述当前人脸特征判断所述投保人的年龄段；
S212、根据判断结果调用对应年龄段的BMI区间判定模型，并将所述当前人脸特征输入到对应年龄段的BMI区间判定模型中进行运算。
如上述S211和S212所述，不同年龄段的人的人脸特征不同，通过人脸的皮肤状态、光泽、轮廓等等特征可以判断人的年龄，而不同年龄段的人的人脸特征与BMI的关联性存在一定的区别，比如，儿童的人脸特征与BMI的关联性，与青年人、中年人、老年人不同，本申请中，上述不同年龄段分为7个阶段，即1周岁-6周岁的幼儿BMI区间判定模型、7周岁-14周岁的儿童BMI区间判定模型、13周岁-19周岁的少年BMI区间判定模型、20周岁-39周岁的青年BMI区间判定模型、40周岁-59周岁的中年BMI区间判定模型、以及60周岁以上的老年BMI区间判定模型等。在其它实施例中，还可以根据实际的客户人群，设定对应的BMI区间判定模型，以提高对人体BMI指数的判断准确性。各年龄段的BMI区间判定模型是通过对应年龄段的人脸特征和BMI训练而得。
在一个实施例中，上述BMI区间判定模型包括身高判断模型、体重判断模型和BMI计算模型，上述上述将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算的步骤S2，包括：
S221、将所属当前人脸特征分别输入到身高判断模型和体重判断模型，得到投保人的估算身高和估算体重；
S222、将所述估算身高和估算体重输入到BMI计算模型进行计算，得出投保人的BMI结果。
如上述步骤S221和S222所述，上述身高判断模型是通过已知的人脸特征和对应的身高通过指定的神经网络模型训练而得，同样的，上述体重判断模型是通过已知的人脸特征和对应的体重通过指定的神经网络模型训练而得。上述BMI计算模型即为BMI的计算公式，具体的BMI＝体重(kg)÷身高^2(m)，然后根据上述BMI所属类别的分类表得到BMI的类别。
参照图2，在一个实施例中，上述输出所述投保人的BMI结果的步骤S3之后，包括：
S4、根据所述BMI结果，将不适于投保人的保险产品在预设的保险产品列表中隐藏或清除。
如上述步骤S4所述，上述保险产品列表是给投保人查看的电子表格，其中记载了保险公司预设的多种保险产品，以及各保险产品对BMI结果的要求，比如，投保人的BMI结果属于类别5时，其不适用购买与高血压、糖尿病等相关的保险产品，那么可以将与高血压、糖尿病等相关的保险产品从上述保险产品列表中隐藏或清除，以防止投保人选择到不可以投保的保险产品等。
参照图2，进一步地，上述根据所述BMI结果，将不适于投保人的保险产品在预设的保险产品列表中隐藏或清除的步骤S4之后，包括:
S5、获取投保人的参数信息，该参数信息至少包括投保人的年龄、性别和工作；
S6、根据所述参数信息对所述投保人进行分类；
S7、根据分类结果，将所述保险产品列表中与分类结果对应的保险产品保留，其余的隐藏或清除。
如上述步骤S5至S7所述，预设一个根据参数信息对投保人进行分类的模型，输入投保人的年龄、性别和工作即会输出一个客户类型，而每一类客户购买保险产品的方向基本相同，将每一类客户购买的保险产品的数量按照从大到小的排列顺序取指定名次之前的保险产品作为针对该类客户的推荐保险产品，当确定好投保人的客户类型后，则只保留保险产品列表中推荐的保险产品，其它的隐藏或清除，进一步地提高投保人购买保险产品的效率。
在一个实施例中，上述获取投保人的当前人脸特征的步骤之前，包括：
S11、向拍摄区域进行超声波扫描，并接收超声波的反射波；
S12、根据反射波判定拍摄区域的物体的轮廓；
S13、若轮廓符合预设的标准，则判定当前拍摄的图片是真人图片，生成获取投保人的当前人脸特征的指令。
如上述S11、S12和S13所述，通过超声波发射装置向拍摄区域发射超声波，然后接收超声波的反射波，根据超声波的反射时间计算各反射点与发射点之间的距离，进而描绘出超声波发射区域中各物体的形状轮廓，比如，超声波的发射区域有一个篮球，那么超声波会被篮球反射，因为篮球是圆形的，所以篮球接收到超声波的时间因为距离的不同而不同，超声波接收装置接收到的反射波的时间也各不相同，进而描绘出篮球的轮廓。本实施例中，如果当前拍摄的图片是一个海报等照片时，那么海报等照片必须较为平整的展开，此时其轮廓是一个平面，而如果是真人，其轮廓是一个3d轮廓。即，对拍摄物进上述当前人脸图片是一个真人照片，以防止他人利用照片欺骗系统而获取到一个较低的BMI类别，投保保险公司禁止投保的保险产品。
本申请的本申请的BMI预测方法，通过人脸特征预测投保人的BMI，可以快速地推断出投保人的BMI类别，以防止投保人虚报BMI，又因为通过计算机等进行测算，所以无需专门的BMI测试设备，节约成本。
参照图3，本申请实施例提供一种BMI预测装置，包括步骤：
获取单元10，用于获取投保人的当前人脸特征；
输入运算单元20，用于将当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，BMI区间判定模型基于人脸特征，以及与人脸特征关联的BMI类别组成的样本数据训练而成；
输出单元30，用于输出投保人的BMI结果。
在上述获取单元10中，上述投保人的当前人脸特征一般是由投保人在投保现场拍摄的人脸图片中 提取的，该人脸图片一般为人脸的正面照，然后提取该正面照上的人脸特征，人脸特征包括如人脸的宽度、长度、长度与宽度的比例、五官的比例、眼睛轮廓的凹陷程度、五官轮廓的线条特征等。
在上述输入运算单元20中，上述神经网络模型包括多种，如VGG19模型、VGG16模型、VGG-F模型、ResNet50模型、ResNet152模型、DPN131模型、InceptionV3模型、Xception模型、DenseNet模型和AlexNet模型等，本申请使用DPN模型，DPN(Dual Path Network)是一种神经网络结构，是在ResNeXt的基础上引入了DenseNet的核心内容，使得模型对特征的利用更加充分。上述DPN、ResNeXt和DenseNet是现有的网络结构，在此不在赘述。上述BMI区间判定模型是一种通过样本数据训练而得的模型。在一个实施例中，获取已知BMI所属类别的人脸特征，然后将人脸特征及其对应的BMI类别分成训练集和测试集等，然后通过训练集的样本数据对上述DPN模型进行训练得到一个结果训练模型，然后将测试集的样本数据输入到结果训练模型中进行测试，若测试通过，则将上述结果训练模型作为上述的BMI区间判定模型。上述的BMI所属类别是指根据BMI的大小进行分类后得到的类别，具体如下表：
BMI 类别
16～18.5 0
18.5～25 1
25～30 2
30～35 3
40 4
>40 5
也就是说，BMI所属类别越大，其对应的BMI越大。本申请中，将当前人脸特征输入到BMI区间判定模型后，其输出的结果即为对应类别。本申请中，将BMI分为六个类别，标签为0～5。在其它实施例中，还可以设置更多的分类等，可以根据具体的情况具体设定。
在上述输出单元30中，上述BMI结果即为推算出的投保人的BMI类别，工作人员或者相应的推荐系统可以根据投保人的BMI类别推荐适合投保人的保险产品，或者确定投保人是否可以投保其已经选择的保险产品等。
参照图4，在一个实施例中，上述获取单元10，包括：
获取模块101，用于获取投保人的当前人脸图片，并提取出当前人脸图片中的人脸区域图片；
扩充模块102，用于对人脸区域图片进行扩充处理；
提取模块103，用于对扩充后的人脸区域图片进行特征提取，得到当前人脸特征。
在上述获取模块101中，将当前人脸图片中的人脸区域图片提取出来，方便提取人脸图片的人脸特征。提取出当前人脸图片中的人脸区域图片的方法可以使用诸如dlib，OpenCV等人脸检测算法进行操作，dlib和OpenCV是人脸识别领域的常规手段，在此不在赘述。
在上述扩充模块102中，由于上述提取出的人脸区域图片一般仅包含了人脸的区域，为了获得更多的特征，将检测到的人脸区域图片进行扩充，具体的扩充方法如下：
令x,y,w,h表示人脸检测得到的人脸区域坐标信息，其中x，y表示左上角的坐标，w和h表示检测到的人脸的宽和高。扩充后的人脸信息则为x_n,y_n,w_n,h_n,分别为：
x_n＝x-(w/4)
y_n＝y-(h/4)
w_n＝x_n+3*w/2
h_n＝y_n+3*h/2
扩充后的人脸区域图片可以不仅包含了人脸区域，还包含了颈部、头部的区域，特征更为充分。
在上述提取模块103中，由于扩充后的人脸区域图片上的特征更多，比如脖子的特征等，所以BMI区间判定模型的运算结果会越加的准确。
参照图5，在一个实施例中，上述输入运算单元20，包括：
获取分类模块21，用于获取指定量的样本数据，并将样本数据分成训练集和测试集；其中，样本数据包括人脸特征，以及与人脸特征对应的BMI；
训练模块22，用于将训练集的样本数据输入到神经网络模型中进行训练；其中，训练的过程中采用随机梯度下降法，利用反向传导法则更新神经网络模型各层的参数，得到结果训练模型；
测试模块23，用于利用测试集的样本数据验证结果训练模型；
标记模块24，用于如果结果训练模型验证通过，则将结果训练模型记为BMI区间判定模型。
在上述获取分类模块21、训练模块22、测试模块23和标记模块24中，上述指定量是对样本数据的样本量的设定值，可以根据具体的要求进行设定，如指定量为10万个样本数据等。在本实施例中，即为将大量的已知的人脸特征，以及与人脸特征关联的BMI类别作为样本数据，然后对预设的神经网络模型进行训练。在训练之前，将样本数据进行随机分成两组集合，即训练集和测试集，然后通过训练集的样本数据对上述神经网络模型进行训练，得到一个输入人脸特征后，输出对应的BMI类别的结果训练模型。训练完成得到结果训练模型后，通过测试集的样本数据验证结果训练模型，以判断结果训练模型是否可用。上述随机梯度下降法就是随机取样一些训练数据，替代整个训练集，如果样本量很大的情况(例如几十万)，那么可能只用其中几万条或者几千条的样本，就已经迭代到最优解了，可以提高训练速度。上述反向传导法则(BP)它建立在梯度下降法的基础上，BP网络的输入输出关系实质上是一种映射关系：一个n输入m输出的BP神经网络所完成的功能是从n维欧氏空间向m维欧氏空间中一有限域的连续映射，这一映射具有高度非线性。BP网络的信息处理能力来源于简单非线性函数的多次复合，因此具有很强的函数复现能力。
在一个实施例中，上述输入运算单元20，还包括：
数据增强模块、对训练集的样本数据进行数据增强。
在上述数据增强模块中，为了在训练模型的过程避免出现过拟合(Overfitting)，通常我们需要输入充足的样本数据量。而样本数据有限，所以需要通过特殊的手段将样本数据量变得更大，具体的可以 使用如下手段进行数据增强：旋转反射变换(Rotation/reflection)，随机旋转图像一定角度；改变图像内容的朝向；翻转变换(flip)，沿着水平或者垂直方向翻转图像；缩放变换(zoom)，按照一定的比例放大或者缩小图像；平移变换(shift)，在图像平面上对图像以一定方式进行平移；采用随机或人为定义的方式指定平移范围和平移步长,沿水平或竖直方向进行平移；改变图像内容的位置等，即将一张原始的图片通过变向、旋转等手段进行处理，以得到更多的样本数据，进而解决样本数据量低，出现过拟合的问题。
参照图6，在一个实施例中，上述训练模块22，包括：
调用子模块2201，用于调用对应神经网络模型的已经训练完成的已知神经网络模型的各层权重参数；
初始化子模块2202，用于将各层的权重参数初始化为神经网络模型的各层权重参数；
第一训练子模块2203，用于通过初始化后的神经网络模型训练得到结果训练模型。
上述调用子模块2201、初始化子模块2202和第一训练子模块2203，即为用于迁移学习，并完成训练的子模块装置。将基于同一神经网络的，且已经训练好的模型的各层权重参数进行调用，初始化为当前待训练的神经网络的初始权重参数。在一具体实施例中，ImageNet是一个计算机视觉系统识别项目名称，是目前世界上图像识别最大的数据库，是美国斯坦福的计算机科学家，模拟人类的识别系统建立的，其中包含120万张图片，DPN107模型是利用ImageNet数据集训练完成的模型，其各层的权重参数已经训练完成。本申请训练BMI区间判定模型的基础同样是DPN107，那么可以将通过ImageNet数据集训练完成DPN107模型的参数初始化到训练BMI区间判定模型的基础DPN107模型中，然后进行BMI区间判定模型的训练，降低训练时间。本申请中将初始学习率设置为0.01。
参照图7，在另一个实施例中，上述训练模块22，包括：
冻结子模块2211，用于冻结神经网络模型中指定层的权重参数；
第二训练子模块2212，用于将训练集的样本数据输入到神经网络模型未冻结层中进行训练，得到结果训练模型。
上述冻结子模块2211和第二训练子模块2212，同样为迁移学习和训练模型的子模块装置，比如神经网络为通过ImageNet数据集训练过的DPN107模型，ImageNet是一个计算机视觉系统识别项目名称，是目前世界上图像识别最大的数据库，是美国斯坦福的计算机科学家，模拟人类的识别系统建立的，其中包含120万张图片，DPN107模型是利用ImageNet数据集训练完成的模型，其各层的权重参数已经训练完成。本申请中，冻结DPN107模型中指定层的权重参数是指，按照各层之间的先后顺序，将排序前N的卷基层或/和全连接层作为指定层，其中N为大于1小于DPN107模型总层数的正整数，且N为预设值，比如，N＝100时，即为将DPN107模型前100层的权重参数，在训练上述结果训练模型时只需要重新训练后7层的权重即可，训练的更加快速。本申请将初始学习率设置为0.01。
在一个实施例中，上述BMI区间判定模型包括男性BMI区间判定模型、女性BMI区间判定模型；上述输入运算单元20，包括：
性别识别模块，用于根据当前人脸特征判断投保人的性别；
第一调用模块，用于根据判断结果调用对应性别的BMI区间判定模型，并将当前人脸特征输入到对应性别的BMI区间判定模型中进行运算。
在上述性别识别模块和第一调用模块中，因为男性和女性的人脸特征存在一定的差异，所以设置针对男性和女性的BMI区间判定模型，以提高判断的准确地性。上述男性BMI区间判定模型、女性BMI区间判定模型的训练方法与上述的训练方法相同，区别在于，训练男性BMI区间判定模型时使用的样本数据为男性的人脸特征及其关联的BMI，训练女性BMI区间判定模型时使用的样本数据为女性的人脸特征及其关联的BMI。本实施例中，根据当前人脸特征判断投保人的性别的方法，可以通过预设的性别识别模型进行判断，该性别识别模型是通过人脸特征数据及其对应的性别训练而得，当输入未知性别的人脸特征后，会得出该人脸特征对应的性别，具体过程为惯用手段，在此不在赘述。
在另一个实施例总，上述BMI区间判定模型包括不同年龄段的BMI区间判定模型；上述输入运算单元20，包括：
年龄识别模块，用于根据当前人脸特征判断投保人的年龄段；
第二调用模块，用于根据判断结果调用对应年龄段的BMI区间判定模型，并将当前人脸特征输入到对应年龄段的BMI区间判定模型中进行运算。
在上述年龄识别模块和第二调用模块中，不同年龄段的人的人脸特征不同，通过人脸的皮肤状态、光泽、轮廓等等特征可以判断人的年龄，而不同年龄段的人的人脸特征与BMI的关联性存在一定的区别，比如，儿童的人脸特征与BMI的关联性，与青年人、中年人、老年人不同，本申请中，上述不同年龄段分为7个阶段，即1周岁-6周岁的幼儿BMI区间判定模型、7周岁-14周岁的儿童BMI区间判定模型、13周岁-19周岁的少年BMI区间判定模型、20周岁-39周岁的青年BMI区间判定模型、40周岁-59周岁的中年BMI区间判定模型、以及60周岁以上的老年BMI区间判定模型等。在其它实施例中，还可以根据实际的客户人群，设定对应的BMI区间判定模型，以提高对人体BMI指数的判断准确性。各年龄段的BMI区间判定模型是通过对应年龄段的人脸特征和BMI训练而得。
在一个实施例中，上述BMI区间判定模型包括身高判断模型、体重判断模型和BMI计算模型，上述输入运算单元20，包括：
输入模块，用于将所属当前人脸特征分别输入到身高判断模型和体重判断模型，得到投保人的估算身高和估算体重；
计算模块，用于将估算身高和估算体重输入到BMI计算模型进行计算，得出投保人的BMI结果。
在上述输入模块和计算模块中，上述身高判断模型是通过已知的人脸特征和对应的身高通过指定的神经网络模型训练而得，同样的，上述体重判断模型是通过已知的人脸特征和对应的体重通过指定的神经网络模型训练而得。上述BMI计算模型即为BMI的计算公式，具体的BMI＝体重(kg)÷身高^2(m)， 然后根据上述BMI所属类别的分类表得到BMI的类别。
参照图8，在一个实施例中，上述BMI预测装置还包括：
第一屏蔽单元40，用于根据BMI结果，将不适于投保人的保险产品在预设的保险产品列表中隐藏或清除。
在上述第一屏蔽单元40中，上述保险产品列表是给投保人查看的电子表格，其中记载了保险公司预设的多种保险产品，以及各保险产品对BMI结果的要求，比如，投保人的BMI结果属于类别5时，其不适用购买与高血压、糖尿病等相关的保险产品，那么可以将与高血压、糖尿病等相关的保险产品从上述保险产品列表中隐藏或清除，以防止投保人选择到不可以投保的保险产品等。
进一步地，上述BMI预测装置还包括:
获取参数单元50，用于获取投保人的参数信息，该参数信息至少包括投保人的年龄、性别和工作；
分类单元60，用于根据参数信息对投保人进行分类；
第二屏蔽单元70，用于根据分类结果，将保险产品列表中与分类结果对应的保险产品保留，其余的隐藏或清除。
在上述获取参数单元50、分类单元60和第二屏蔽单元70中，预设一个根据参数信息对投保人进行分类的模型，输入投保人的年龄、性别和工作即会输出一个客户类型，而每一类客户购买保险产品的方向基本相同，将每一类客户购买的保险产品的数量按照从大到小的排列顺序取指定名次之前的保险产品作为针对该类客户的推荐保险产品，当确定好投保人的客户类型后，则只保留保险产品列表中推荐的保险产品，其它的隐藏或清除，进一步地提高投保人购买保险产品的效率。
在一个实施例中，上述BMI预测装置还包括：
扫描单元11，用于向拍摄区域进行超声波扫描，并接收超声波的反射波；
判定单元12，用于根据反射波判定拍摄区域的物体的轮廓；
生成指令单元13，用于若轮廓符合预设的标准，则判定当前拍摄的图片是真人图片，生成获取投保人的当前人脸特征的指令。
在上述扫描单元11、判定单元12和生成指令单元13中，通过超声波发射装置向拍摄区域发射超声波，然后接收超声波的反射波，根据超声波的反射时间计算各反射点与发射点之间的距离，进而描绘出超声波发射区域中各物体的形状轮廓，比如，超声波的发射区域有一个篮球，那么超声波会被篮球反射，因为篮球是圆形的，所以篮球接收到超声波的时间因为距离的不同而不同，超声波接收装置接收到的反射波的时间也各不相同，进而描绘出篮球的轮廓。本实施例中，如果当前拍摄的图片是一个海报等照片时，那么海报等照片必须较为平整的展开，此时其轮廓是一个平面，而如果是真人，其轮廓是一个3d轮廓。即，对拍摄物进上述当前人脸图片是一个真人照片，以防止他人利用照片欺骗系统而获取到一个较低的BMI类别，投保保险公司禁止投保的保险产品。
本申请的本申请的BMI预测装置，通过人脸特征预测投保人的BMI，可以快速地推断出投保人的BMI类别，以防止投保人虚报BMI，又因为通过计算机中的虚拟装置等进行测算，所以无需专门的BMI测试设备，节约成本。
参照图9，本发明实施例中还提供一种计算机设备，该计算机设备可以是服务器，其内部结构可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储BMI区间判定模型等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现如上述各方法的实施例的流程。
本领域技术人员可以理解，图9中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定。
本发明一实施例还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，计算机可读指令被处理器执行时实现如上述各方法的实施例的流程。
以上所述仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种BMI预测方法，其特征在于，包括：
    获取投保人的当前人脸特征；
    将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，所述BMI区间判定模型基于人脸特征，以及与所述人脸特征关联的BMI类别组成的样本数据训练而成；
    输出所述投保人的BMI结果。
  2. 根据权利要求1所述的BMI预测方法，其特征在于，所述获取投保人的当前人脸特征的步骤，包括：
    获取所述投保人的当前人脸图片，并提取出所述当前人脸图片中的人脸区域图片；
    对所述人脸区域图片进行扩充处理；
    对扩充后的所述人脸区域图片进行特征提取，得到所述当前人脸特征。
  3. 根据权利要求1所述的BMI预测方法，其特征在于，所述BMI区间判定模型的获取方法，包括：
    获取指定量的样本数据，并将样本数据分成训练集和测试集；其中，所述样本数据包括人脸特征，以及与所述人脸特征对应的BMI；
    将训练集的样本数据输入到神经网络模型中进行训练；其中，训练的过程中采用随机梯度下降法，利用反向传导法则更新所述神经网络模型各层的参数，得到结果训练模型；
    利用所述测试集的样本数据验证所述结果训练模型；
    如果验证通过，则将所述结果训练模型记为所述BMI区间判定模型。
  4. 根据权利要求3所述的BMI预测方法，其特征在于，所述将训练集的样本数据输入到神经网络模型中进行训练的步骤之前，包括：
    对所述训练集的样本数据进行数据增强。
  5. 根据权利要求3所述的BMI预测方法，其特征在于，所述将训练集的样本数据输入到神经网络模型中进行训练的步骤，包括：
    调用对应所述神经网络模型的已经训练完成的已知神经网络模型的各层权重参数；
    将各层的所述权重参数初始化为所述神经网络模型的各层权重参数；
    通过初始化后的所述神经网络模型训练得到所述结果训练模型。
  6. 根据权利要求1所述的BMI预测方法，其特征在于，所述BMI区间判定模型包括男性BMI区间判定模型、女性BMI区间判定模型；所述将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算的步骤，包括：
    根据所述当前人脸特征判断所述投保人的性别；
    根据判断结果调用对应性别的BMI区间判定模型，并将所述当前人脸特征输入到对应性别的BMI区间判定模型中进行运算。
  7. 根据权利要求1所述的BMI预测方法，其特征在于，所述BMI区间判定模型包括不同年龄段的BMI区间判定模型；所述将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算的步骤，包括：
    根据所述当前人脸特征判断所述投保人的年龄段；
    根据判断结果调用对应年龄段的BMI区间判定模型，并将所述当前人脸特征输入到对应年龄段的BMI区间判定模型中进行运算。
  8. 一种BMI预测装置，其特征在于，包括：
    获取单元，用于获取投保人的当前人脸特征；
    输入运算单元，用于将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，所述BMI区间判定模型基于人脸特征，以及与所述人脸特征关联的BMI类别组成的样本数据训练而成；
    输出单元，用于输出所述投保人的BMI结果。
  9. 根据权利要求8所述的BMI预测装置，其特征在于，所述获取单元，包括：
    获取模块，用于获取所述投保人的当前人脸图片，并提取出所述当前人脸图片中的人脸区域图片；
    扩充模块，用于对所述人脸区域图片进行扩充处理；
    提取模块，用于对扩充后的所述人脸区域图片进行特征提取，得到所述当前人脸特征。
  10. 根据权利要求8所述的BMI预测装置，其特征在于，所述输入运算单元，包括：
    获取分类模块，用于获取指定量的样本数据，并将样本数据分成训练集和测试集；其中，所述样本数据包括人脸特征，以及与所述人脸特征对应的BMI；
    训练模块，用于将训练集的样本数据输入到神经网络模型中进行训练；其中，训练的过程中采用随机梯度下降法，利用反向传导法则更新所述神经网络模型各层的参数，得到结果训练模型；
    测试模块，用于利用所述测试集的样本数据验证所述结果训练模型；
    标记模块，用于如果结果训练模型验证通过，则将所述结果训练模型记为所述BMI区间判定模型。
  11. 根据权利要求10所述的BMI预测装置，其特征在于，所述输入运算单元，还包括：
    数据增强模块、对所述训练集的样本数据进行数据增强。
  12. 根据权利要求10所述的BMI预测装置，其特征在于，所述训练模块，包括：
    调用子模块，用于调用对应所述神经网络模型的已经训练完成的已知神经网络模型的各层权重参数；
    初始化子模块，用于将各层的所述权重参数初始化为所述神经网络模型的各层权重参数；
    第一训练子模块，用于通过初始化后的所述神经网络模型训练得到所述结果训练模型。
  13. 根据权利要求8所述的BMI预测装置，其特征在于，所述BMI区间判定模型包括男性BMI区间判定模型、女性BMI区间判定模型；所述上述输入运算单元，包括：
    性别识别模块，用于根据所述当前人脸特征判断所述投保人的性别；
    第一调用模块，用于根据判断结果调用对应性别的BMI区间判定模型，并将所述当前人脸特征输入到对应性别的BMI区间判定模型中进行运算。
  14. 根据权利要求8所述的BMI预测装置，其特征在于，所述BMI区间判定模型包括不同年龄段的BMI区间判定模型；所述输入运算单元20，包括：
    年龄识别模块，用于根据所述当前人脸特征判断所述投保人的年龄段；
    第二调用模块，用于根据判断结果调用对应年龄段的BMI区间判定模型，并将所述当前人脸特征输入到对应年龄段的BMI区间判定模型中进行运算。
  15. 一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现BMI预测方法，该BMI预测方法包括：
    获取投保人的当前人脸特征；
    将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，所述BMI区间判定模型基于人脸特征，以及与所述人脸特征关联的BMI类别组成的样本数据训练而成；
    输出所述投保人的BMI结果。
  16. 根据权利要求15所述的计算机设备，其特征在于，所述获取投保人的当前人脸特征的步骤，包括：
    获取所述投保人的当前人脸图片，并提取出所述当前人脸图片中的人脸区域图片；
    对所述人脸区域图片进行扩充处理；
    对扩充后的所述人脸区域图片进行特征提取，得到所述当前人脸特征。
  17. 根据权利要求15所述的计算机设备，其特征在于，所述BMI区间判定模型的获取方法，包括：
    获取指定量的样本数据，并将样本数据分成训练集和测试集；其中，所述样本数据包括人脸特征，以及与所述人脸特征对应的BMI；
    将训练集的样本数据输入到神经网络模型中进行训练；其中，训练的过程中采用随机梯度下降法，利用反向传导法则更新所述神经网络模型各层的参数，得到结果训练模型；
    利用所述测试集的样本数据验证所述结果训练模型；
    如果验证通过，则将所述结果训练模型记为所述BMI区间判定模型。
  18. 根据权利要求17所述的计算机设备，其特征在于，所述将训练集的样本数据输入到神经网络模型中进行训练的步骤之前，包括：
    对所述训练集的样本数据进行数据增强。
  19. 一种计算机非易失性可读存储介质，其上存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现BMI预测方法，该BMI预测方法包括：
    获取投保人的当前人脸特征；
    将所述当前人脸特征输入到预设的基于神经网络模型训练完成的BMI区间判定模型中进行运算；其中，所述BMI区间判定模型基于人脸特征，以及与所述人脸特征关联的BMI类别组成的样本数据训练而成；
    输出所述投保人的BMI结果。
  20. 根据权利要求19所述的计算机非易失性可读存储介质，其特征在于，所述获取投保人的当前人脸特征的步骤，包括：
    获取所述投保人的当前人脸图片，并提取出所述当前人脸图片中的人脸区域图片；
    对所述人脸区域图片进行扩充处理；
    对扩充后的所述人脸区域图片进行特征提取，得到所述当前人脸特征。
PCT/CN2018/095386 2018-05-25 2018-07-12 Bmi预测方法、装置、计算机设备和存储介质 WO2019223080A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810515457.4A CN108875590A (zh) 2018-05-25 2018-05-25 Bmi预测方法、装置、计算机设备和存储介质
CN201810515457.4 2018-05-25

Publications (1)

Publication Number Publication Date
WO2019223080A1 true WO2019223080A1 (zh) 2019-11-28

Family

ID=64333794

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095386 WO2019223080A1 (zh) 2018-05-25 2018-07-12 Bmi预测方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN108875590A (zh)
WO (1) WO2019223080A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111401534A (zh) * 2020-04-29 2020-07-10 北京晶派科技有限公司 一种蛋白质性能预测方法、装置和计算设备
CN111461898A (zh) * 2020-02-28 2020-07-28 上海商汤智能科技有限公司 一种获取核保结果的方法及相关装置
CN113592024A (zh) * 2021-08-12 2021-11-02 燕山大学 冷轧铜带表面缺陷识别模型训练方法及识别方法、系统

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109086447B (zh) * 2018-08-20 2020-11-20 珠海格力电器股份有限公司 热泵设备的能耗检测方法、装置、计算机设备和存储介质
CN109637664A (zh) * 2018-11-20 2019-04-16 平安科技（深圳）有限公司 一种bmi评测方法、装置及计算机可读存储介质
CN109754068A (zh) * 2018-12-04 2019-05-14 中科恒运股份有限公司 基于深度学习预训练模型的迁移学习方法及终端设备
CN109726824A (zh) * 2018-12-05 2019-05-07 中科恒运股份有限公司 训练模型的迁移学习方法及终端设备
CN111292146B (zh) * 2018-12-07 2023-06-13 泰康保险集团股份有限公司 保险推荐方法及装置、计算机存储介质、电子设备
CN109829567A (zh) * 2019-01-02 2019-05-31 平安科技（深圳）有限公司 基于机器学习的学位调控方法、装置和计算机设备
CN111513673B (zh) * 2019-02-01 2023-10-31 百度在线网络技术（北京）有限公司 基于图像的生长状态监测方法、装置、设备及存储介质
CN110119988A (zh) * 2019-04-19 2019-08-13 深圳壹账通智能科技有限公司 基于决策树模型的产品交易方法、装置和计算机设备
CN110472509B (zh) * 2019-07-15 2024-04-26 中国平安人寿保险股份有限公司 基于人脸图像的胖瘦识别方法及装置、电子设备
CN110443620A (zh) * 2019-08-06 2019-11-12 中国工商银行股份有限公司 刷脸支付方法和装置
CN111012353A (zh) * 2019-12-06 2020-04-17 西南交通大学 一种基于人脸关键点识别的身高检测方法
CN111242193A (zh) * 2020-01-06 2020-06-05 深圳壹账通智能科技有限公司 基于迁移学习的电子商务欺诈检测方法、装置及存储介质
CN113436735A (zh) * 2020-03-23 2021-09-24 北京好啦科技有限公司 基于人脸结构度量的体重指数预测方法、设备和存储介质
CN111523501B (zh) * 2020-04-27 2023-09-15 阳光保险集团股份有限公司 体重指数预测方法及装置
CN113425254B (zh) * 2021-05-10 2022-10-18 复旦大学 基于混合数据输入的体脂率预测模型的青年男性体脂率预测方法
CN113591704B (zh) * 2021-07-30 2023-08-08 四川大学 体重指数估计模型训练方法、装置和终端设备
CN114496263B (zh) * 2022-04-13 2022-07-12 杭州研极微电子有限公司 用于体重指数估计的神经网络模型建立方法及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104699931A (zh) * 2013-12-09 2015-06-10 广州华久信息科技有限公司 一种基于人脸的神经网络血压预测方法及手机
US20170249502A1 (en) * 2013-05-10 2017-08-31 Tencent Technology (Shenzhen) Company Limited Systems and methods for facial property identification
CN107590460A (zh) * 2017-09-12 2018-01-16 北京达佳互联信息技术有限公司 人脸分类方法、装置及智能终端
CN107784482A (zh) * 2017-09-30 2018-03-09 平安科技（深圳）有限公司 招聘方法、电子装置及可读存储介质

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106339576A (zh) * 2016-07-20 2017-01-18 美的集团股份有限公司 一种健康管理方法和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170249502A1 (en) * 2013-05-10 2017-08-31 Tencent Technology (Shenzhen) Company Limited Systems and methods for facial property identification
CN104699931A (zh) * 2013-12-09 2015-06-10 广州华久信息科技有限公司 一种基于人脸的神经网络血压预测方法及手机
CN107590460A (zh) * 2017-09-12 2018-01-16 北京达佳互联信息技术有限公司 人脸分类方法、装置及智能终端
CN107784482A (zh) * 2017-09-30 2018-03-09 平安科技（深圳）有限公司 招聘方法、电子装置及可读存储介质

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461898A (zh) * 2020-02-28 2020-07-28 上海商汤智能科技有限公司 一种获取核保结果的方法及相关装置
CN111401534A (zh) * 2020-04-29 2020-07-10 北京晶派科技有限公司 一种蛋白质性能预测方法、装置和计算设备
CN111401534B (zh) * 2020-04-29 2023-12-05 北京晶泰科技有限公司 一种蛋白质性能预测方法、装置和计算设备
CN113592024A (zh) * 2021-08-12 2021-11-02 燕山大学 冷轧铜带表面缺陷识别模型训练方法及识别方法、系统
CN113592024B (zh) * 2021-08-12 2024-05-28 燕山大学 冷轧铜带表面缺陷识别模型训练方法及识别方法、系统

Also Published As

Publication number Publication date
CN108875590A (zh) 2018-11-23

Similar Documents

Publication Publication Date Title
WO2019223080A1 (zh) Bmi预测方法、装置、计算机设备和存储介质
US10962404B2 (en) Systems and methods for weight measurement from user photos using deep learning networks
CA3097712C (en) Systems and methods for full body measurements extraction
WO2021114892A1 (zh) 基于环境语义理解的人体行为识别方法、装置、设备及存储介质
WO2020215984A1 (zh) 基于深度学习的医学图像检测方法及相关设备
Yan et al. Modeling annotator expertise: Learning when everybody knows a bit of something
WO2019200749A1 (zh) 识别人脸的方法、装置、计算机设备和存储介质
CN115050064A (zh) 人脸活体检测方法、装置、设备及介质
CN110321968B (zh) 一种超声图像分类装置
WO2021068781A1 (zh) 一种疲劳状态识别方法、装置和设备
CN114359974B (zh) 一种人体姿态的检测方法、设备及存储介质
CN112750531A (zh) 一种中医自动化望诊系统、方法、设备和介质
CN116311539B (zh) 基于毫米波的睡眠动作捕捉方法、装置、设备及存储介质
WO2021217937A1 (zh) 姿态识别模型的训练方法及设备、姿态识别方法及其设备
Jin et al. Attention guided deep features for accurate body mass index estimation
CN110414562A (zh) X光片的分类方法、装置、终端及存储介质
CN114758382B (zh) 基于自适应补丁学习的面部au检测模型建立方法及应用
Xue et al. Early pregnancy fetal facial ultrasound standard plane‐assisted recognition algorithm
WO2024000233A1 (zh) 面部表情识别方法、装置、设备及可读存储介质
Rajasenbagam et al. Semantic content-based image retrieval system using deep learning model for lung cancer CT images
CN112417961B (zh) 一种基于场景先验知识的海面目标检测方法
CN118411744A (zh) 基于深度学习的面部视频抑郁检测系统、设备及介质
CN117253035B (zh) 一种基于极坐标下注意力的单目标医学图像分割方法
CN111723688A (zh) 人体动作识别结果的评价方法、装置和电子设备
CN118396930A (zh) 眼底图像分析方法、系统、终端及计算机可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18919729

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18919729

Country of ref document: EP

Kind code of ref document: A1