BMI estimation method based on multi-dimensional feature extraction

Abstract

The invention discloses a BMI estimation method based on multi-dimensional feature extraction, and belongs to the technical field of BMI estimation. Comprising the following steps that three-dimensional human body measurement features are extracted based on human body posture estimation and a three-dimensional human body reconstruction method, two-dimensional human body measurement features are extracted based on human body skeleton point detection and a human body contour detection method, and depth features are extracted from an image through an Attention-Enhanced VGG model. And finally, fusing the three features, and mapping the fused features into a BMI value through a kernel ridge regression method.

Description

Technical Field

[0001] The present invention relates to the technical field of BMI estimation, and in particular to a BMI estimation method based on multi-dimensional feature extraction. Background Art

[0002] BMI is a commonly used indicator to measure the degree of human obesity, classifying individuals as underweight, normal, overweight or obese. BMI is calculated by dividing weight (kg) by height (m) squared, and individuals with a proper body shape are generally considered to be more attractive and healthier than those with abnormal obesity.

[0003] In recent years, the study of predicting BMI values ​​has received increasing attention. Traditionally, BMI values ​​are usually obtained by measuring an individual's height and weight. However, this method is difficult for bedridden patients because it requires the patient's cooperation and usually requires an upright standing posture. Therefore, in order to solve this problem, a visual-based BMI measurement method came into being.

[0004] Vision-based BMI measurement is a contactless measurement method that not only makes the measurement process more comfortable, but also effectively avoids the risk of infection caused by touch sensors. Compared with traditional contact measurement methods, this visual measurement method is more convenient and safer. In addition, the vision-based BMI estimation method also supports telemedicine applications and can perform measurements without face-to-face contact. This makes large-scale BMI measurement possible, and BMI values ​​can be estimated for multiple individuals at the same time, greatly improving efficiency and scope of application.

[0005] In some past studies, some scholars have proposed predicting BMI from facial images by analyzing the relationship between facial features and BMI values. However, compared with two-dimensional human body images, clear frontal facial images are usually more difficult to obtain due to privacy protection issues. In addition, BMI prediction based on facial images also faces certain accuracy challenges.

[0006] To overcome these challenges, some researchers use 3D human body data or RGB-D images to calculate BMI values. Although these methods can accurately estimate BMI values ​​under certain conditions, RGB-D image acquisition is very sensitive to sunlight exposure and measurement distance, which makes it difficult to obtain high-quality 3D human body data in outdoor environments.

[0007] With the continuous development of computer technology and image processing technology, some scholars have used 3D human body reconstruction methods to extract 3D anthropometric features such as waist circumference and hip circumference from 2D images, and then used regression methods to estimate BMI values. These 3D anthropometric features have been shown to be closely related to BMI, and can also be predicted using parametric models. However, learning from these 3D anthropometric features still cannot strongly represent human body shape information.

[0008] In recent years, many health science studies have shown that some two-dimensional anthropometric indicators are closely related to BMI. However, most two-dimensional anthropometric characteristics, such as waist-to-shoulder ratio, waist-to-hip ratio, hip-to-head ratio, waist-to-thigh ratio, etc., are only horizontal ratios of the human body and do not contain vertical information of the human body.

[0009] With the development of computing power and big data, deep learning, especially convolutional neural networks, has made significant progress in the field of image processing. Deep network models extract deep features and are applied to BMI estimation tasks. Some scholars have proposed a three-branch BMI estimation method that combines manually defined three-dimensional anthropometric features with features extracted by CNN models to improve the accuracy of BMI estimation. It can be seen that deep features can effectively represent image semantic features and improve the accuracy of BMI estimation.

[0010] To this end, those skilled in the art have proposed a BMI estimation method based on multi-dimensional feature extraction to solve the above problems. Summary of the invention

[0011] In view of the shortcomings of the prior art, the present invention provides a BMI estimation method based on multi-dimensional feature extraction, which starts from two-dimensional and three-dimensional perspectives and extracts deep features of images through deep learning methods, aiming to construct a BMI estimation method based on multi-dimensional feature extraction.

[0012] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0013] A BMI estimation method based on multi-dimensional feature extraction specifically comprises the following steps:

[0014] Step 1, 3D human body measurement feature extraction: extracting 3D human body measurement features based on human body posture estimation and 3D human body reconstruction methods;

[0015] Step 2, 2D human body measurement feature extraction: extracting 2D human body measurement features based on human skeleton point detection and human body contour detection methods;

[0016] Step 3, deep feature extraction: propose a deep network model integrating attention mechanism to extract deep features;

[0017] Step 4, model training and selection: Use SVR, KRR, DTR and GRR regression methods to train the model on the training set; input the test set images into the four established regression methods for BMI estimation, and verify the performance of the model through evaluation indicators, and select the regression method with the best performance; among them, SVR is support vector regression; KRR is kernel ridge regression; DTR is decision tree regression; GPR is Gaussian process regression.

[0018] As a further preferred solution of the BMI estimation method based on multi-dimensional feature extraction of the present invention, the step 1 specifically comprises the following steps:

[0019] Step 1.1, preprocess the data using the OpenPose method to estimate human posture, thereby obtaining human posture information;

[0020] Step 1.2, using the human body 3D model SMPL-X model for human body 3D reconstruction, thereby obtaining body shape, body offset and posture parameters;

[0021] Step 1.3, use the three-dimensional body regression Shapy method to calculate the three-dimensional anthropometric features of waist circumference, hip circumference and the ratio of waist circumference to hip circumference.

[0022] As a further preferred solution of the BMI estimation method based on multi-dimensional feature extraction of the present invention, step 2 specifically comprises the following steps:

[0023] Step 2.1, obtaining skeleton point information through a human skeleton point detection method;

[0024] Step 2.2, perform human body contour detection;

[0025] Step 2.3, calculate the two-dimensional anthropometric characteristics waist to thigh width ratio, waist to hip ratio, waist to head width ratio, hip width to head width ratio, body area between waist and hips, Euclidean distance from nose to knee to waist width ratio, waist width to shoulder width ratio, and Euclidean distance from nose to knee to hip ratio.

[0026] As a further preferred solution of the BMI estimation method based on multi-dimensional feature extraction of the present invention, step 3 specifically comprises the following steps:

[0027] A deep network model integrating attention mechanism is proposed to extract deep features. This deep learning model is based on the deep convolutional neural network model VGG model, and adds a lightweight visual Transformer MobileViT attention mechanism, which can help the network focus more on important areas in the image and improve the network's ability to extract key features.

[0028] As a further preferred embodiment of the BMI estimation method based on multi-dimensional feature extraction of the present invention, in step 1.3, the specific steps of the three-dimensional anthropometric feature calculation method are as follows:

[0029] Step 1.31, calculate the waist circumference WC, the calculation formula is as follows:

[0030]

[0031] where i and j are two points in the convex hull, β is the SMPL-X shape parameter; (x i ,y i , z i ) is the coordinate of the waist landmark point, t i Index for the body triangle;

[0032] Step 1.32, calculate the hip circumference HC; given the coordinates of the hip landmark point (h i , m i , n i ) and the body triangle index t i ; According to formula (1), the hip circumference can be calculated;

[0033] Step 1.33, calculate the waist circumference to hip circumference ratio WCHCR; based on the waist and hip circumferences, it can be calculated using the following formula:

[0034]

[0035] As a further preferred embodiment of the BMI estimation method based on multi-dimensional feature extraction of the present invention, in step 2.3, the specific steps of the two-dimensional anthropometric feature calculation method are as follows:

[0036] Step 2.31, calculate the waist-to-shoulder ratio WSR. Under similar BMI conditions, some people have longer shoulders or shorter waists. WSR is used to represent this difference. Based on the width of the waist and shoulders, the calculation formula is as follows:

[0037]

[0038] Among them, D w is waist width, D s is the shoulder width; given the ordinate of the waist, the waist width calculation formula is as follows:

[0039]

[0040] Among them, st is the constraint condition, l is the linear function, and b is the intercept; x1 and x2 have numerical ranges under the constraint condition, which means they are two variables; x ew is the horizontal coordinate of the waist center point, which can be approximated to the horizontal coordinate x of the center shouldercs , and its calculation formula is as follows:

[0041]

[0042] Among them, x rs is the horizontal coordinate of the right shoulder, x ls is the horizontal coordinate of the left shoulder; y in formula (4) w is the ordinate of the waist, and its calculation formula is as follows:

[0043]

[0044] Among them, y cs is the ordinate of the center shoulder; h is the ordinate of the hip, and its calculation formula is as follows:

[0045]

[0046] Among them, y lh is the ordinate of the left hip, y rh is the ordinate of the right hip; cs is the ordinate of the center shoulder, and its calculation formula is as follows:

[0047]

[0048] Among them, y rs is the ordinate of the right shoulder, y ls is the vertical coordinate of the left shoulder; D s The calculation formula is as follows:

[0049]

[0050] Step 2.32, calculate the ratio of waist circumference to thigh width WTR; take the distance from the left thigh boundary to the left thigh center plus the distance from the right thigh boundary to the right thigh center as the thigh width, and refer to formula (5) to get the thigh width D t ; Based on the width of the waist and thigh, WTR can be calculated by the following formula:

[0051]

[0052] Step 2.33, calculate the waist-to-hip ratio WHpR; given the coordinates of the left and right hip joint skeletons, the hip circumference D hp It can be calculated in the same way as formula (4); WHpR can be calculated by the following formula:

[0053]

[0054] Step 2.34, calculate the ratio of waist circumference to head width WHdR; given the coordinates of the left and right ear skeleton joints (xle ,y le ) and (x re ,y re ), head width D hd It can be approximated by the Euclidean distance between two points, calculated as follows:

[0055]

[0056] Step 2.35, calculate the ratio of hip width to head width HpHdR. Based on the width of the hips and head, HpHdR can be calculated using the following formula:

[0057]

[0058] Step 2.36, calculate the body area Area between waist and hips; since the area between waist and hips can be used as an indicator of thin people and obese people, when the scale information is unknown, the characteristic area can be used to represent the information; Area can be calculated by the following formula:

[0059]

[0060]

[0061] Among them, pixcel num The number of pixels in the area, pix x，y is the number of pixels under the (x, y) coordinate point;

[0062] Step 2.37, the Euclidean distance from nose to knee and the waist width ratio HWR; HWR can be calculated by the following formula:

[0063]

[0064] Among them, D nk is the Euclidean distance from the nose to the knee, calculated as follows:

[0065]

[0066] The knee coordinate point is (x k ,y k ), which is calculated as follows:

[0067]

[0068] y k =y lk =y rk (twenty one)

[0069] The coordinate point of the left knee is (x lk y lk), the coordinate point of the right knee is (x rk ,y rk );

[0070] Step 2.38, calculate the Euclidean distance from nose to knee and hip ratio H2H; the two-dimensional human body measurement features of waist-thigh width ratio, waist-hip ratio, waist-head width ratio, hip width-head width ratio, body area between waist and hips, and waist width-shoulder width ratio only contain the horizontal feature ratio of the human body, and do not contain the longitudinal information of the human body; although the two-dimensional human body measurement features of the Euclidean distance from nose to knee and waist width ratio contain the longitudinal information of the human body, for people with similar waist width but large difference in hip circumference, the two-dimensional human body measurement features of the Euclidean distance from nose to knee and waist width ratio cannot be used to distinguish well; H2H is proposed to represent the aspect ratio; its calculation formula is as follows:

[0071]

[0072] Compared with the prior art, the technical advantages of the present invention are:

[0073] 1) A BMI estimation framework based on multi-dimensional feature extraction is proposed, which can estimate the BIM value from a single two-dimensional image.

[0074] 2) Three-dimensional anthropometric features of waist circumference, hip circumference and the ratio of waist circumference to hip circumference, two-dimensional anthropometric features of the Euclidean distance from nose to knee and hip circumference ratio, and depth features are proposed.

[0075] 3) Comprehensive comparisons with state-of-the-art methods are performed and extensive ablation studies are conducted to verify the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a data processing flow chart of Embodiment 1 of the present invention;

[0077] Figure 2 The extracted depth feature map of embodiment 1 of the present invention;

[0078] Figure 3 This is a performance comparison chart of different deep learning models in Example 1 of the present invention;

[0079] Figure 4 This is a performance comparison chart of different attention modules in Example 1 of the present invention;

[0080] Figure 5 This is a performance comparison chart of different numbers of deep feature extractions in Example 1 of the present invention. DETAILED DESCRIPTION

[0081] The following will be combined with the drawings of the present invention specification to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] A BMI estimation method based on multi-dimensional feature extraction, comprising:

[0083] Step 1, 3D human body measurement feature extraction: extracting 3D human body measurement features based on human body posture estimation and 3D human body reconstruction method;

[0084] Step 2, 2D human body measurement feature extraction: extracting 2D human body measurement features based on human skeleton point detection and human body contour detection methods;

[0085] Step 3, deep feature extraction: propose an Attention-EnhancedVGG (deep network integrating attention mechanism) model to extract deep features;

[0086] Step 4, model training and selection: Use SVR, KRR, DTR and GRR regression methods to train the model on the training set; input the test set images into the four established regression methods for BMI estimation, and verify the performance of the model through evaluation indicators, and select the regression method with the best performance; among them, SVR is support vector regression; KRR is kernel ridge regression; DTR is decision tree regression; GPR is Gaussian process regression.

[0087] Step (1) Specific steps:

[0088] Firstly, the OpenPose method is used to preprocess the data to obtain the human posture information. The SMPL-X model is used for 3D reconstruction of the human body to obtain the body shape, body offset and posture parameters. The 3D body regression Shapy method is used to calculate the 3D human body measurement features such as waist circumference, hip circumference and the ratio of waist circumference to hip circumference. The OpenPose method is used to preprocess the data to obtain the human posture information.

[0089] Step (1) Specific steps for calculating three-dimensional human body measurement features:

[0090] 1. Calculate the waist circumference WC. The calculation formula is as follows:

[0091]

[0092] where i and j are two points in the convex hull and β is the SMPL-X shape parameter. (x i ,yi , z i ) is the coordinate of the waist landmark point, t i Index to the body triangle.

[0093] 2. Calculate the hip circumference HC. Given the coordinates of the hip landmark point (h i ,h i ,h i ) and the body triangle index t i According to formula (2), the hip circumference can be calculated.

[0094] 3. Calculate the ratio of waist circumference to hip circumference WCHCR. Based on the circumference of the waist and hip, it can be calculated using the following formula:

[0095]

[0096] Step (2) Feature extraction specific steps:

[0097] The skeleton point information is obtained by the human skeleton point detection method, and then the human body contour is detected, and finally the two-dimensional human body measurement characteristics of waist circumference to thigh width ratio, waist circumference to hip circumference ratio, waist circumference to head width ratio, hip width to head width ratio, body area between waist and hips, Euclidean distance from nose to knee to waist width ratio, waist width to shoulder width ratio and Euclidean distance from nose to knee to hip circumference ratio are calculated;

[0098] Step (2) Specific steps for calculating two-dimensional human body measurement features:

[0099] 1. Calculate the waist-to-shoulder ratio (WSR). With similar BMIs, some people have longer shoulders or shorter waists, so this article uses WSR to represent this difference. Based on the width of the waist and shoulders, the calculation formula is as follows:

[0100]

[0101] Among them, D w is waist width, D s is the shoulder width. Given the ordinate of the waist, the waist width calculation formula is as follows:

[0102]

[0103] Among them, st is the constraint condition, l is the linear function, and b is the intercept; x1 and x2 have numerical ranges under the constraint condition, which means they are two variables; x cw is the horizontal coordinate of the waist center point, which can be approximated to the horizontal coordinate x of the center shoulder cs , and its calculation formula is as follows:

[0104]

[0105] Among them, x rs is the horizontal coordinate of the right shoulder, x ls is the horizontal coordinate of the left shoulder. w is the ordinate of the waist, and its calculation formula is as follows:

[0106]

[0107] Among them, y cs is the ordinate of the center shoulder. h is the ordinate of the hip, and its calculation formula is as follows:

[0108]

[0109] Among them, y lh is the ordinate of the left hip, y rh is the ordinate of the right hip. cs is the ordinate of the center shoulder, and its calculation formula is as follows:

[0110]

[0111] Among them, y rs is the ordinate of the right shoulder, y ls is the ordinate of the left shoulder. s The calculation formula is as follows:

[0112]

[0113] 2. Calculate the ratio of waist circumference to thigh width WTR. In this paper, the distance from the left thigh boundary to the left thigh center plus the distance from the right thigh boundary to the right thigh center is taken as the thigh width. Referring to formula (5), the thigh width D can be obtained. t Based on the waist and thigh width, WTR can be calculated using the following formula:

[0114]

[0115] 3. Calculate the waist-to-hip ratio WHpR. Given the coordinates of the left and right hip joint skeletons, hip circumference D hp It can be calculated in the same way as formula (4). Therefore, WHpR can be calculated by the following formula:

[0116]

[0117] 4. Calculate the ratio of waist circumference to head width WHdR. Given the coordinates of the left and right ear skeleton joints (x le ,y le ) and (x re ,y re ), head width D hdIt can be approximated by the Euclidean distance between two points, calculated as follows:

[0118]

[0119] 5. Calculate the ratio of hip width to head width HpHdR. Based on the width of the hips and head, HpHdR can be calculated using the following formula:

[0120]

[0121] 6. Calculate the body area between waist and hips. Since the area between waist and hips can be used as an indicator of thin and obese people, the characteristic area can be used to represent this information when the scale information is unknown. Area can be calculated using the following formula:

[0122]

[0123] Among them, pixcel num The number of pixels in the area, pix x，y is the number of pixels at the (x, y) coordinate point.

[0124] 7. The Euclidean distance from nose to knee and the waist width ratio HWR. HWR can be calculated by the following formula:

[0125]

[0126] Among them, D nk is the Euclidean distance from the nose to the knee, calculated as follows:

[0127]

[0128] The knee coordinate point is (x k ,y k ), which is calculated as follows:

[0129]

[0130] y k =y lk =y rk (twenty one)

[0131] The coordinate point of the left knee is (x lk y lk ), the coordinate point of the right knee is (x rk ,y rk ).

[0132] 8. Calculate the Euclidean distance from nose to knee and hip circumference ratio H2H. The two-dimensional anthropometric features of waist-to-thigh width ratio, waist-to-hip ratio, waist-to-head width ratio, hip width-to-head width ratio, body area between waist and hips, and waist width-to-shoulder width ratio only contain the horizontal feature ratio of the human body, and do not contain the longitudinal information of the human body. Although the two-dimensional anthropometric features of the Euclidean distance from nose to knee and waist width ratio contain the longitudinal information of the human body, for people with similar waist widths but large differences in hip circumference, the two-dimensional anthropometric features of the Euclidean distance from nose to knee and waist width ratio cannot be used to distinguish them. Therefore, H2H is proposed to represent the aspect ratio. The calculation formula is as follows:

[0133]

[0134] Embodiment 1:

[0135] 1. Data source: The 2D image-to-BMI public dataset is used to evaluate the proposed method. The dataset has 4189 images, including 1477 males and 2712 females. The dataset is randomly divided into training set and test set in a ratio of 7:3. The training set contains images of 1032 males and 1903 females, and the training set is randomly divided into training set and validation set in a ratio of 4:1. There is no overlap between the training set and the validation set. The test set contains images of 445 males and 809 females.

[0136] 2. 3D human body measurement feature extraction: Extract 3D human body measurement features based on human body posture estimation and 3D human body reconstruction methods.

[0137] 3. Two-dimensional anthropometric feature extraction: Extract two-dimensional anthropometric features based on human skeleton point detection and human contour detection methods.

[0138] 4. Deep feature extraction: Propose an Attention-EnhancedVGG model to extract deep features, such as Figure 2 shown.

[0139] 5. Model training and selection: Four regression methods, SVR, KRR, DTR and GRR, were used to train the model on the training set. Next, the test set images were input into the four established models for BMI estimation, and the performance of the model was verified by the evaluation index. The model with the best performance was selected to determine the final BMI evaluation method. The comparison results of the four regression methods are shown in Table 1. As can be seen from Table 1, in terms of the mean absolute error index, the KRR regression method performed best in the three cases of men, women and the whole, followed by the SVR regression method. KRR outperformed the DTR and GRR regression methods in both the mean absolute error and the mean absolute percentage error indicators. Therefore, KRR was selected as the regression method.

[0140] Table 1

[0141]

[0142] 4. Performance comparison of different deep learning models: The performance of deep learning models is compared under the same data set. The results are as follows: Figure 3 As shown. Figure 3 It can be seen that compared with other deep network models, the Attention-EnhancedVGG model proposed in this paper shows the best performance in both MAE and MAPE indicators. This result shows that by introducing the attention mechanism, the model in this paper can more effectively capture the intrinsic characteristics of the data, thereby achieving higher accuracy.

[0143] 5. Performance comparison of different attention modules: The performance of different attention modules is compared under the same data set. The results are as follows Figure 4 As shown. Figure 4 It can be seen that compared with other attention modules, the MobileViTAttention module in this paper performs best in MAE and MAPE indicators, showing the best performance. This result shows that the MobileViTAttention module is more effective in capturing data features, thereby achieving lower MAE and MAPE values, further improving the prediction accuracy of the model.

[0144] 6. Performance comparison of different numbers of deep feature extraction: The performance of different numbers of deep feature extraction is compared under the same data set. The results are as follows: Figure 5 As shown. Figure 5 It can be seen that as the number of deep features increases, the overall trend of MAE and MAPE is that they decrease from 5 to 10, increase from 10 to 20, decrease from 20 to 25, and then increase after 25. At the same time, the total number of network parameters is also increasing. Considering this trade-off, this paper sets the number of deep features to 10 as the final setting of this paper.

[0145] 7. Validation of multi-dimensional feature extraction: Under the same data set, the validity verification experiment was carried out on multi-dimensional feature extraction such as two-dimensional anthropometric features 2DFs, three-dimensional anthropometric features 3DFs and depth features DFs, and the experimental results are shown in Table 2. As can be seen from Table 2, the MAE and MAPE values ​​of the multi-dimensional feature extraction proposed in the present invention are much higher than those of other dimensional feature extractions. This shows that multi-dimensional feature extraction can effectively improve the prediction accuracy, provide more comprehensive information, and ensure better model performance.

[0146] Table 2

[0147]

[0148] The above detailed description is a specific description of one feasible embodiment of the present invention. The embodiment is not intended to limit the patent scope of the present invention. Any equivalent implementation or modification that does not deviate from the present invention should be included in the scope of the technical solution of the present invention.

Claims

1. A BMI estimation method based on multi-dimensional feature extraction, characterized in that: The specific steps include: Step 1, 3D human body measurement feature extraction: extracting 3D human body measurement features based on human body posture estimation and 3D human body reconstruction methods; Step 2, 2D human body measurement feature extraction: extracting 2D human body measurement features based on human skeleton point detection and human body contour detection methods; Step 3, deep feature extraction: propose a deep network model integrating attention mechanism to extract deep features; Step 4, model training and selection: Use SVR, KRR, DTR and GRR regression methods to train the model on the training set; input the test set images into the four established regression methods for BMI estimation, and verify the performance of the model through evaluation indicators, and select the regression method with the best performance; among them, SVR is support vector regression; KRR is kernel ridge regression; DTR is decision tree regression; GPR is Gaussian process regression.

2. The BMI estimation method based on multi-dimensional feature extraction according to claim 1, characterized in that: The step 1 specifically comprises the following steps: Step 1.1, preprocess the data using the OpenPose method to estimate human posture, thereby obtaining human posture information; Step 1.2, using the human body 3D model SMPL-X model for human body 3D reconstruction, thereby obtaining body shape, body offset and posture parameters; Step 1.3, use the three-dimensional body regression Shapy method to calculate the three-dimensional anthropometric features of waist circumference, hip circumference and the ratio of waist circumference to hip circumference.

3. The BMI estimation method based on multi-dimensional feature extraction according to claim 1, characterized in that: The step 2 specifically comprises the following steps: Step 2.1, obtaining skeleton point information through a human skeleton point detection method; Step 2.2, perform human body contour detection; Step 2.3, calculate the two-dimensional anthropometric characteristics waist to thigh width ratio, waist to hip ratio, waist to head width ratio, hip width to head width ratio, body area between waist and hips, Euclidean distance from nose to knee to waist width ratio, waist width to shoulder width ratio, and Euclidean distance from nose to knee to hip ratio.

4. The BMI estimation method based on multi-dimensional feature extraction according to claim 1, characterized in that: The step 3 specifically comprises the following steps: A deep network model integrating attention mechanism is proposed to extract deep features. This deep learning model is based on the deep convolutional neural network model VGG model, and adds a lightweight visual Transformer MobileViT attention mechanism, which can help the network focus more on important areas in the image and improve the network's ability to extract key features.

5. The BMI estimation method based on multi-dimensional feature extraction according to claim 2, characterized in that: In step 1.3, the specific steps of the three-dimensional human body measurement feature calculation method are as follows: Step 1.31, calculate the waist circumference WC, the calculation formula is as follows: where i and j are two points in the convex hull, β is the SMPL-X shape parameter; (x i ,y i , z i ) is the coordinate of the waist landmark point, t i Index for the body triangle; Step 1.32, calculate the hip circumference HC; given the coordinates of the hip landmark point (h i , m i , n i ) and the body triangle index t i ; According to formula (1), the hip circumference can be calculated; Step 1.33, calculate the waist circumference to hip circumference ratio WCHCR; based on the waist and hip circumferences, it can be calculated using the following formula:

6. The BMI estimation method based on multi-dimensional feature extraction according to claim 3, characterized in that: In step 2.3, the specific steps of the two-dimensional human body measurement feature calculation method are as follows: Step 2.31, calculate the waist-to-shoulder ratio WSR. Under similar BMI conditions, some people have longer shoulders or shorter waists. WSR is used to represent this difference. Based on the width of the waist and shoulders, the calculation formula is as follows: Among them, D w is waist width, D s is the shoulder width; given the ordinate of the waist, the waist width calculation formula is as follows: Among them, s. t. are constraints, l is a linear function, b is the intercept; x1 and x2 have numerical ranges under the constraints, which means they are two variables; x cw is the horizontal coordinate of the waist center point, which can be approximated to the horizontal coordinate x of the center shoulder cs , and its calculation formula is as follows: Among them, x rs is the horizontal coordinate of the right shoulder, x ls is the horizontal coordinate of the left shoulder; y in formula (4) w is the ordinate of the waist, and its calculation formula is as follows: Among them, y cs is the ordinate of the center shoulder; h is the ordinate of the hip, and its calculation formula is as follows: Among them, y lh is the ordinate of the left hip, y rh is the ordinate of the right hip; cs is the ordinate of the center shoulder, and its calculation formula is as follows: Among them, y rs is the ordinate of the right shoulder, y ls is the vertical coordinate of the left shoulder; D s The calculation formula is as follows: Step 2.32, calculate the ratio of waist circumference to thigh width WTR; take the distance from the left thigh boundary to the left thigh center plus the distance from the right thigh boundary to the right thigh center as the thigh width, and refer to formula (5) to get the thigh width D c ; Based on the width of the waist and thigh, WTR can be calculated by the following formula: Step 2.33, calculate the waist-to-hip ratio WHpR; given the coordinates of the left and right hip joint skeletons, the hip circumference D hp It can be calculated in the same way as formula (4); WHpR can be calculated by the following formula: Step 2.34, calculate the ratio of waist circumference to head width WHdR; given the coordinates of the left and right ear skeleton joints (x le ,y le ) and (x re ,y re ), head width D hd It can be approximated by the Euclidean distance between two points, calculated as follows: Step 2.35, calculate the ratio of hip width to head width HpHdR; based on the width of the hips and head, HpHdR can be calculated using the following formula: Step 2.36, calculate the body area Area between waist and hips; since the area between waist and hips can be used as an indicator of thin people and obese people, when the scale information is unknown, the characteristic area can be used to represent the information; Area can be calculated by the following formula: Among them, pixcel num The number of pixels in the area, pix x，y is the number of pixels under the (x, y) coordinate point; Step 2.37, the Euclidean distance from the nose to the knee and the waist-to-width ratio HWR; HWR can be calculated by the following formula: Among them, D nk is the Euclidean distance from the nose to the knee, calculated as follows: The knee coordinate point is (x k ,y k ), which is calculated as follows: and k =and lk =and rk (21) The coordinate point of the left knee is (x lk ,y lk ), the coordinate point of the right knee is (x rk ,y rk ); Step 2.38, calculate the Euclidean distance from nose to knee and hip ratio H2H; the two-dimensional human body measurement features of waist-thigh width ratio, waist-hip ratio, waist-head width ratio, hip width-head width ratio, body area between waist and hips, and waist width-shoulder width ratio only contain the horizontal feature ratio of the human body, and do not contain the longitudinal information of the human body; although the two-dimensional human body measurement features of the Euclidean distance from nose to knee and waist width ratio contain the longitudinal information of the human body, for people with similar waist width but large difference in hip circumference, the two-dimensional human body measurement features of the Euclidean distance from nose to knee and waist width ratio cannot be used to distinguish well; H2H is proposed to represent the aspect ratio; its calculation formula is as follows:

Images