Anthropometry System

Abstract

An anthropometry system includes a first machine learning model, a second machine learning model, an image acquisition and transmission module, a result transmission module, a first training module, and a second training module. The image acquisition and transmission module acquires frontal and side-view standing images of person A, inputs these images into the first machine learning model to obtain a first processing result, and the result transmission module inputs the first processing result into the second machine learning model to obtain the second machine learning model's output of person A's anthropometry parameters. Through this system, the first machine learning model can generate image segmentation meshes, and the segmentation results, used as vectors, are more easily used for learning and processing by the second machine learning model, resulting in more accurate estimations of the final anthropometry parameters.

Description

Technical Field

[0001] This invention relates to a machine learning-based anthropometric system. Background Technology

[0002] With the widespread use of smartphones, tablets, and other smart terminals, technologies related to measuring human height and object dimensions based on images and videos captured by these devices are constantly developing. For example, (Application No.: CN 201910450077.1) proposes a machine learning-based method for predicting human 3D dimensions. This method proposes capturing frontal and side images of the human body and inputting them into a machine learning model to predict 3D dimensions. However, this machine learning-based method can only directly input frontal and side images and cannot obtain more accurate values ​​through machine learning. Summary of the Invention

[0003] Therefore, there is a need for a system that can better identify human body three-dimensional data parameters in order to solve the problem that the identification of human body measurement parameters is not accurate enough in the existing technology.

[0004] To achieve the above objectives, the inventors provide a human body measurement system, including a first machine learning model, a second machine learning model, an image acquisition and transmission module, a result transmission module, a first training module, and a second training module;

[0005] The image acquisition and transmission module is used to acquire front and side standing images of A, input the front and side standing images of A into a first machine learning model to obtain a first processing result, and the result transmission module is used to input the first processing result into a second machine learning model to obtain the output result of the second machine learning model on the human body parameters of A.

[0006] The first training module is used to train a first machine learning model and includes the following units: a first material acquisition unit, used to acquire first material, the first material including human body images and segmentation meshes of the human body; a first training input unit, used to control the human body images in the first material as input to the first machine learning model, and also includes a first training output unit, used to control the segmentation meshes of the human body in the first material as output to the first machine learning model; and a first parameter adjustment model, used to adjust the parameters of the first machine learning model according to the difference between the segmentation network output by the first training output unit and the segmentation network in the first material.

[0007] The second training module is used to train a second machine learning model and includes the following units: a second material acquisition unit, used to acquire second material, the second material including a second image, the second image being a frontal and side-view standing image of the same human body; the second material also includes the height and other body parameters of the human body in the second image; a first machine learning model invocation unit, used to input the second image into the first machine learning model for processing, and obtain the output result of the first machine learning model on the second image; a second training input unit, used to add the height information of the human body in the second image to the output result as the input information of the second machine learning model; a second training output unit, used to use the other body parameters of the human body in the second image as the output information of the second machine learning model; and a second parameter adjustment model, used to adjust the parameters of the second machine learning model according to the difference between the information output by the second training output unit and the other body parameters in the second material.

[0008] Specifically, the first material also includes several control points for each human body part in the segmented mesh, and the output of the first machine learning model includes several control points for each human body part in the segmented mesh.

[0009] Specifically, the first machine learning model is trained using a first material library, which includes more than 200 sets of first materials.

[0010] Specifically, the second machine learning model is trained using a second material library, which includes more than 50 sets of second materials.

[0011] Preferably, the first machine learning model is a deep neural network.

[0012] Preferably, the second machine learning model is LightGBM.

[0013] Furthermore, the second training input unit is also used to perform extended processing on the output result by adding specific numerical values ​​of the height information of the human body to which the second image belongs, specifically including:

[0014] The average value K is taken from the output of the first machine learning model on the second image;

[0015] Set up a Gaussian mixture model G with two sub-distributions.

[0016] G was trained using the height parameters of the individuals in the second dataset.

[0017] The height parameters of each person in the second data set are calculated using G, resulting in two sub-distribution probabilities P and Q for each person.

[0018] The height numerical dimension in the input information of the second machine learning model is expanded into two dimensions, assigned values ​​P*K and Q*K respectively, without retaining the original height numerical dimension.

[0019] Train a second machine learning model;

[0020] During the measurement phase, the first processing result is combined with A's height data, and two sub-probability values ​​are calculated by G and multiplied by K. These values ​​are then input into the second machine learning model to obtain the output results of the second machine learning model on A's human body parameters.

[0021] Furthermore, the second training input unit is also used to perform extended processing on the output result by adding specific numerical values ​​of the height information of the human body to which the second image belongs, specifically including:

[0022] Let the height of the human body be H. After calculating P and Q, compare the size of P and Q. If P is greater than Q, add the following values ​​to the input information of the second machine learning model: H*0.515, H*0.345, H*0.542; if P is not greater than Q, add the following values ​​to the input information of the second machine learning model: H*0.612, H*0.421, H*0.643.

[0023] At the same time, the height value dimension in the input information of the second machine learning model will no longer be expanded into two dimensions, but the original height value dimension will be retained.

[0024] Specifically, the second training input unit is further configured to perform extended processing on the output result by adding specific numerical values ​​of the height information of the human body to which the second image belongs, specifically including:

[0025] Add the following values ​​to the input information of the second machine learning model:

[0026] (0.515*P+0.612*Q) / (P+Q),

[0027] (0.345*P+0.6421*Q) / (P+Q),

[0028] (0.542*P+0.643*Q) / (P+Q),

[0029] At the same time, the height value dimension in the input information of the second machine learning model will no longer be expanded into two dimensions, and the original height value dimension will not be retained.

[0030] Furthermore, the first training output unit is also used to perform skin color detection on the segmented mesh of the human body, perform linear regression on the boundary line between the skin color region and the non-skin color region in the segmented mesh of the human body to obtain the boundary line, obtain the width of the skin color region at a preset distance on the skin color region side of the boundary line, obtain the width of the non-skin color region at a preset distance on the non-skin color region side of the boundary line, and if the width of the non-skin color region is greater than the width of the skin color region, then the following step is performed: generating a reference curve with a distance S from the edge of the non-skin color region inward from the edge of the non-skin color region, wherein...

[0031] S = width of non-skin-colored area - width of skin-colored area.

[0032] Using the above method, we can generate image segmentation grids using the first machine learning model. The segmentation results, as vectors, are more easily used for learning and processing by the second machine learning model, resulting in more accurate measurement and estimation results of human body parameters. Attached Figure Description

[0033] Figure 1 This is a flowchart of the human body measurement method according to a specific embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the human body parameter measuring device according to a specific embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the camera device for measuring human body parameters according to a specific embodiment of the present invention;

[0036] Figure 4 This is a block diagram of the human body measurement system according to a specific embodiment of the present invention. Detailed Implementation

[0037] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0038] Please see Figure 1A human body measurement method includes the following steps: S100: acquiring a first source material, the first source material including a human body image and a segmentation mesh of the human body; S100: setting a first machine learning model, the input of the first machine learning model being the human body image in the first source material, and the output including the segmentation mesh of the human body in the first source material, and training the first machine learning model; the human body image in the first source material is not limited to a frontal or side view of the human body, and can include human body images from any angle. Training with human body images from multiple angles enables the first machine learning model to achieve more detailed and intelligent segmentation results for human body images. The method may further include a step before acquiring the first source material: manually preprocessing the human body image in the first source material by drawing a segmentation mesh of the human body in the first source material. The segmentation mesh is a segmentation processing mesh drawn for segmenting human body part regions. Through the above steps, the technical effect of training the first machine learning model to learn the differentiation of human body parts can be achieved.

[0039] Then, our anthropometry method further includes the following steps: acquiring second material, which includes a second image, specifically frontal and side-view standing images of the same human body; the second material also includes the height and other body parameters of the human body in the second image. These other body parameters can include all measurable body data indicators such as chest circumference, waist circumference, hip circumference, hand length, palm length, shoulder width, and foot length. S101: The second image is input into a first machine learning model for processing, yielding the output of the first machine learning model on the second image. Here, the output of the first machine learning model on the second image refers to the already trained first machine learning model, which, after processing the second image, can output segmentation meshes for different human body parts in the second image according to the aforementioned principles. S102: The height information of the human body in the second image is added to the output of the first machine learning model on the second image as input information for the second machine learning model, and the other body parameters of the human body in the second image are used as output information for the second machine learning model to train it. Thus, after training, the second machine learning model can learn and estimate the other body parameters of the human body in the second image by receiving the second image, its segmentation mesh information, and the height information. In a specific embodiment, during the application measurement phase, frontal and side-view standing images of person A are acquired. In step S103, these images are input into a first machine learning model to obtain a first processing result. This first processing result includes block segmentation results of person A's body parts in the frontal and side-view standing images. Then, in step S104, the first processing result is input separately into a second machine learning model, or the first processing result and person A's height are input together into the second machine learning model to obtain the output result of the second machine learning model on person A's body parameters. These body parameters may include parameters such as chest circumference, waist circumference, hip circumference, hand length, palm length, shoulder width, and foot length. Through the above steps, the present invention achieves the technical effect of "measuring" body parameters.

[0040] In some further embodiments, the first material also includes several control points for each human body part in the segmented mesh. These control points can also be completed during the material annotation step, that is, when the human body image in the first material is preprocessed manually, the control points in the human body image in the first material are also manually annotated. Here, annotating control points is mainly to increase the amount of information. The output of the first machine learning model includes several control points for each human body part in the segmented mesh. Training by adding control points as output can also make the training of the first machine learning model faster and more stable. The control points can be annotated as joint nodes of the human body in the image. By adding control point annotation information, more data can be provided to the first machine learning model. Choosing joint nodes as control points can also make the classification of the first machine learning model more scientific.

[0041] While training the first and second machine learning models using only a single set of first and second materials is feasible, the output results are not stable enough and lack practical significance. We need to provide a method where the first machine learning model is trained using a first material library, and the second machine learning model is trained using a second material library. The first materials in the first material library have all undergone manual annotation preprocessing and must include at least 20 sets of first materials. The second material library must include at least 10 sets of materials. In some specific embodiments, to achieve better training results—that is, to ensure convergence of the outputs of the first and second machine learning models or to achieve the required confidence level—it is preferable to use a first material library containing at least 200 sets of first materials to train the first machine learning model, and a second material library containing at least 50 sets of second materials to train the second machine learning model.

[0042] In some other specific embodiments, the first machine learning model is a deep neural network architecture. The second machine learning model is a LightGBM architecture. These settings enable more accurate measurement parameter results.

[0043] After performing the above steps, in order to further improve the fitting degree of the scheme to different heights, after executing the step "attach the height information of the human body to the second image to the output result", the height value is expanded, specifically including:

[0044] The average value K is taken from the output of the first machine learning model on the second image;

[0045] Set up a Gaussian mixture model G with several sub-distributions.

[0046] The model G is trained using the height parameters of each person in the second dataset. This approach can better adapt to the peak height distribution that may occur in a population. Specifically, our technical solution involves setting up a Gaussian mixture model G with two sub-distributions.

[0047] The model G is trained using the height parameters of individuals in the second dataset. The height parameters of each individual in the second dataset are used to calculate two sub-distribution probabilities P and Q for each person. This is to accommodate the possibility of two peaks in height values ​​within the population, caused by gender differences. The height value dimension in the input information of the second machine learning model is expanded into two dimensions, assigned values ​​P*K and Q*K respectively, without retaining the original height value dimension, and the second machine learning model is then trained.

[0048] During the measurement phase, the first processing result, along with A's height data, is used to calculate two sub-probability values ​​multiplied by K using G. These sub-probability values ​​are then input into the second machine learning model to obtain the model's output on A's anthropometric parameters. This approach allows the second machine learning model to have a significant advantage in processing anthropometric parameters. When considering potential factors affecting group height, such as race and nationality, employing technical solutions with 4, 6, or 8 sub-distributions is feasible.

[0049] In other specific embodiments, after performing the step "attaching the height information of the human body to the second image to the output result", the height value is expanded, specifically including:

[0050] Let the height be H. After calculating P and Q, compare their values. If P is greater than Q, add the following values ​​to the input of the second machine learning model: H*0.515, H*0.345, H*0.542; if P is not greater than Q, add the following values ​​to the input of the second machine learning model: H*0.612, H*0.421, H*0.643. Assuming P and Q are the two peaks with relatively high and low values ​​on the height axis, the corresponding gender's three-dimensional array is added as two dimensions to the input information, thus making the output of the second machine learning model more stable after training. In this embodiment, the height value dimension in the input information of the second machine learning model is no longer expanded into two dimensions, and P*K and Q*K are no longer assigned separately, thus retaining the original height value dimension.

[0051] Specifically, after executing the step "attach the height information of the human body to the second image to the output result", the height value is expanded, including:

[0052] Add the following values ​​to the input information of the second machine learning model:

[0053] (0.515*P+0.612*Q) / (P+Q),

[0054] (0.345*P+0.6421*Q) / (P+Q),

[0055] (0.542*P+0.643*Q) / (P+Q),

[0056] This technical solution is derived by considering the relationship between the normal height distribution of a population and the three-dimensional distribution of different populations. In this embodiment, it is not necessary to expand the height numerical dimension in the input information of the second machine learning model into two dimensions, and no longer assign values ​​P*K and Q*K separately, but retain the original height numerical dimension. Through the above solution, the output results generated by the second machine learning model after training are also more stable.

[0057] In other embodiments, considering the presence of clothing on the human body in the captured image, to further improve recognition accuracy, our method further includes the steps of: performing skin color detection on the segmented mesh of the human body; performing linear regression on the boundary line between the skin-colored and non-skin-colored regions in the segmented mesh of the human body to obtain the boundary line; obtaining the width of the skin-colored region at a preset distance on the skin-colored region side of the boundary line; and obtaining the width of the non-skin-colored region at a preset distance on the non-skin-colored region side of the boundary line. Here, the length of the preset distance is chosen to be proportional to the length of the boundary line segment in the boundary line, for example:

[0058] Preset distance = Length of boundary segment * 0.2

[0059] If the width of the non-skin-colored region is greater than the width of the skin-colored region, it indicates that the interference from clothing needs to be excluded. The following steps are then taken: Generate a reference curve with a distance S from the edge of the non-skin-colored region, extending inwards from the edge of the non-skin-colored region.

[0060] S = width of non-skin-colored area - width of skin-colored area.

[0061] The generated reference curve can be used for display or as input into a second machine learning model to improve the second machine learning model's ability to recognize clothing.

[0062] This plan also includes an anthropometry system; please refer to [link / reference needed]. Figure 4 The system of the solution includes a first machine learning model 400, a second machine learning model 402, an image acquisition and sending module 404, a result sending module 406, a first training module 41, and a second training module 42.

[0063] The image acquisition and transmission module 404 is used to acquire front and side standing images of A, input the front and side standing images of A into a first machine learning model to obtain a first processing result, and the result transmission module is used to input the first processing result into a second machine learning model to obtain the output result of the second machine learning model on the human body parameters of A.

[0064] The first training module 41 is used to train a first machine learning model and includes the following units: a first material acquisition unit 410, used to acquire first material, the first material including human body images and segmentation meshes of the human body; a first training input unit 412, used to control the human body images in the first material as input to the first machine learning model, and also includes a first training output unit 414, used to control the segmentation meshes of the human body in the first material as output to the first machine learning model; and a first parameter adjustment model, used to adjust the parameters of the first machine learning model according to the difference between the segmentation network output by the first training output unit and the segmentation network in the first material.

[0065] The second training module 42 is used to train a second machine learning model and includes the following units: a second material acquisition unit 420, used to acquire second material, the second material including a second image, the second image being a frontal and side-view standing image of the same human body; the second material also includes the height and other body parameters of the human body in the second image; a first machine learning model invocation unit 422, used to input the second image into the first machine learning model for processing, and obtain the output result of the first machine learning model on the second image; a second training input unit 424, used to add the height information of the human body in the second image to the output result as input information for the second machine learning model; a second training output unit 426, used to use the other body parameters of the human body in the second image as the output information of the second machine learning model; and a second parameter adjustment model, used to adjust the parameters of the second machine learning model according to the difference between the information output by the second training output unit and the other body parameters in the second material. Through the above system, in the formal application stage, as long as the frontal and side images of the human body are input into the system, the output of human body measurement parameters can be obtained. Therefore, the present invention achieves the technical effect of "measuring" human body parameters.

[0066] Specifically, the first material also includes several control points for each human body part in the segmented mesh, and the output of the first machine learning model includes several control points for each human body part in the segmented mesh.

[0067] Specifically, the first machine learning model is trained using a first material library, which includes more than 200 sets of first materials.

[0068] Specifically, the second machine learning model is trained using a second material library, which includes more than 50 sets of second materials.

[0069] Preferably, the first machine learning model is a deep neural network.

[0070] Preferably, the second machine learning model is LightGBM.

[0071] To further improve the fitting degree of the scheme to different heights, the second training input unit 424 of this anthropometric system is also used to extend the output result by adding specific numerical values ​​of the height information of the human body to which the second image belongs, specifically including:

[0072] The average value K is taken from the output of the first machine learning model on the second image;

[0073] Set up a Gaussian mixture model G with several sub-distributions.

[0074] The model G is trained using the height parameters of each person in the second dataset. This approach can better adapt to the peak height distribution that may occur in a population. Specifically, our technical solution involves setting up a Gaussian mixture model G with two sub-distributions.

[0075] The model G is trained using the height parameters of individuals in the second dataset. The height parameters of each individual in the second dataset are used to calculate two sub-distribution probabilities P and Q for each person. This is to accommodate the possibility of two peaks in height values ​​within the population, caused by gender differences. The height value dimension in the input information of the second machine learning model is expanded into two dimensions, assigned values ​​P*K and Q*K respectively, without retaining the original height value dimension, and the second machine learning model is then trained.

[0076] During the measurement phase, the first processing result, along with A's height data, is used to calculate two sub-probability values ​​multiplied by K using G. These sub-probability values ​​are then input into the second machine learning model to obtain the model's output on A's anthropometric parameters. This approach allows the second machine learning model to have a significant advantage in processing anthropometric parameters. When considering potential factors affecting group height, such as race and nationality, employing technical solutions with 4, 6, or 8 sub-distributions is feasible.

[0077] In other specific embodiments, the second training input unit 424 of this anthropometry system is further used to perform extended processing on the output result, specifically adding the height information of the human body to which the second image belongs, including:

[0078] Let the height be H. After calculating P and Q, compare their values. If P is greater than Q, add the following values ​​to the input of the second machine learning model: H*0.515, H*0.345, H*0.542; if P is not greater than Q, add the following values ​​to the input of the second machine learning model: H*0.612, H*0.421, H*0.643. Assuming P and Q are the two peaks with relatively high and low values ​​on the height axis, the corresponding gender's three-dimensional array is added as two dimensions to the input information, thus making the output of the second machine learning model more stable after training. In this embodiment, the height value dimension in the input information of the second machine learning model is no longer expanded into two dimensions, and P*K and Q*K are no longer assigned separately, thus retaining the original height value dimension.

[0079] Specifically, the second training input unit 424 of this anthropometry system is also used to perform extended processing on the output result, specifically adding the height information of the human body to which the second image belongs, including:

[0080] Add the following values ​​to the input information of the second machine learning model:

[0081] (0.515*P+0.612*Q) / (P+Q),

[0082] (0.345*P+0.6421*Q) / (P+Q),

[0083] (0.542*P+0.643*Q) / (P+Q),

[0084] This technical solution is derived by considering the relationship between the normal height distribution of a population and the three-dimensional distribution of different populations. In this embodiment, it is not necessary to expand the height numerical dimension in the input information of the second machine learning model into two dimensions, and no longer assign values ​​P*K and Q*K separately, but retain the original height numerical dimension. Through the above solution, the output results generated by the second machine learning model after training are also more stable.

[0085] In other embodiments, considering the presence of clothing on the human body in the captured image, to further improve recognition accuracy, in a specific embodiment, the first training output unit 414 of the human body measurement system is further used to perform skin color detection on the segmented mesh of the human body, and to perform linear regression on the boundary line between the skin color region and the non-skin color region in the segmented mesh of the human body to obtain the boundary line. The width of the skin color region at a preset distance on the skin color region side of the boundary line is obtained, and the width of the non-skin color region at a preset distance on the non-skin color region side of the boundary line is also obtained. Here, the length of the preset distance is chosen to be proportional to the length of the boundary line segment in the boundary line, for example:

[0086] Preset distance = Length of boundary segment * 0.2

[0087] If the width of the non-skin-colored region is greater than the width of the skin-colored region, it indicates that the interference from clothing needs to be excluded. The following steps are then taken: Generate a reference curve with a distance S from the edge of the non-skin-colored region, extending inwards from the edge of the non-skin-colored region.

[0088] S = width of non-skin-colored area - width of skin-colored area.

[0089] The generated reference curve can be used for display or as input into a second machine learning model to improve the second machine learning model's ability to recognize clothing.

[0090] In other specific embodiments, please see here. Figure 2 We also provide a human body parameter measurement device. This can be a handheld device equipped with an intelligent system and a camera function, including a prompting module 200, an image detection module 202, and a processing unit 204. The prompting module prompts the user to take a frontal photo, acquires the photo taken by the user, and the image detection module detects whether the user's photo is a frontal image. If not, the prompting module prompts the user to take a new photo. The prompting module also prompts the user to take a side view photo, acquires the photo taken by the user, and the image detection module detects whether the user's photo is a side view. If not, the prompting module prompts the user to take a new photo.

[0091] The processing unit is used to input the first machine learning model to obtain a first processing result, and input the first processing result into a second machine learning model to obtain the output result of the second machine learning model on the human body parameters of person A. The first machine learning model is trained through the following steps: acquiring first material, which includes a human body image and a segmentation mesh of the human body; the human body image in the first material is the input of the first machine learning model, and the segmentation mesh of the human body in the first material is also used as the output of the first machine learning model; the second machine learning model is trained through the following steps: acquiring second material, which includes a second image, and the second image is a frontal and side standing image of the same human body; the second material also includes the height and other body parameters of the human body in the second image; inputting the second image into the first machine learning model for processing to obtain the output result of the first machine learning model on the second image; adding the height information of the human body in the second image to the output result as the input information of the second machine learning model, and using the other body parameters of the human body in the second image as the output information of the second machine learning model. Here, the first machine learning model can be integrated into this device or set up in the cloud, and the processing unit only needs to input it into the machine learning model for processing.

[0092] In our technical solution, the method for detecting whether a user's photo is a frontal image specifically includes:

[0093] Using Face++ face detection and facial landmark analysis (FLAGI) interfaces, the absolute value of the `yaw_angle` field under the `headpose` field of a frontal photo is checked to see if it is less than 20. For side-view photos, the method involves using Face++ face detection and FLAGI to check if the absolute value of the `yaw_angle` field under the `headpose` field is greater than 150. The choice of Face++ face detection interface is flexible, and the criteria for determining frontal and side-view faces can be adjusted based on confidence levels. For frontal faces, the criterion can be adjusted to whether the absolute value of the `yaw_angle` field is less than 30 or 15; for side-view faces, it can be adjusted to whether it is greater than 160 or 140, etc. Through this approach, we have implemented a service that allows users to interactively obtain photos and receive the image processing results, thus improving the user experience.

[0094] Others such as Figure 3 In the illustrated embodiment, we will introduce a camera device for measuring human body parameters. Its features include two orthogonal cameras 300, a voice prompt module 302, and a processing unit 304. The intersection area of ​​the two cameras is the user's photo-taking area. The processing unit 304 detects whether a human body exists in the user's photo-taking area. The voice prompt module prompts the user to face one of the cameras. The two cameras respectively acquire images of the user's human body. The processing unit analyzes the acquired images, setting the image from the camera that identifies the face and has a high confidence level as a frontal image, and the other image as a side image. The processing unit also performs the following steps: detecting whether the user's photo is a frontal image; if not, prompting the user to retake the photo; prompting the user to take a side image; detecting whether the user's photo is a side image; if not, prompting the user to retake the photo. The main difference between this camera device and the aforementioned smart device is that it requires two sets of orthogonal cameras and can intelligently identify the presence of a human body and automatically recognize the camera facing the human body as a frontal or side image, facilitating subsequent processing.

[0095] Subsequent processing also includes: the processing unit inputs the first machine learning model to obtain a first processing result, and inputs the first processing result into a second machine learning model to obtain the output result of the second machine learning model on the human body parameters of person A. The first machine learning model is trained through the following steps: acquiring first material, which includes a human body image and a segmentation mesh of the human body; the human body image in the first material is the input of the first machine learning model, and the segmentation mesh of the human body in the first material is also used as the output of the first machine learning model; the second machine learning model is trained through the following steps: acquiring second material, which includes a second image, and the second image belongs to the same human body as a frontal and side standing image; the second material also includes the height and other body parameters of the human body in the second image; inputting the second image into the first machine learning model for processing to obtain the output result of the first machine learning model on the second image; adding the height information of the human body in the second image to the output result as the input information of the second machine learning model, and using the other body parameters of the human body in the second image as the output information of the second machine learning model. Through the above scheme, we can also realize the service of user interaction from obtaining photos to receiving image processing results, thus improving the user experience.

[0096] In a specific embodiment, the method for detecting whether a user's photo is a frontal image includes:

[0097] Using Face++ face detection and facial landmark processing APIs, we check if the absolute value of the yaw_angle field under the headpose field of a frontal photo is less than 20.

[0098] Methods for detecting whether a user's photo is a side profile image include:

[0099] Using Face++ face detection and facial landmark processing, we can check whether the absolute value of the yaw_angle field under the headpose field of a side view photo is greater than 150.

[0100] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of patent protection of the present invention.

Claims

1. A human body measurement system, characterized in that, It includes a first machine learning model, a second machine learning model, an image acquisition and transmission module, a result transmission module, a first training module, and a second training module; The image acquisition and transmission module is used to acquire front and side standing images of A, input the front and side standing images of A into a first machine learning model to obtain a first processing result, and the result transmission module is used to input the first processing result into a second machine learning model to obtain the output result of the second machine learning model on the human body parameters of A. The first training module is used to train a first machine learning model and includes the following units: a first material acquisition unit, used to acquire first material, the first material including human body images and segmentation meshes of the human body; a first training input unit, used to control the human body images in the first material as input to the first machine learning model, and also includes a first training output unit, used to control the segmentation meshes of the human body in the first material as output to the first machine learning model; and a first parameter adjustment model, used to adjust the parameters of the first machine learning model according to the difference between the segmentation network output by the first training output unit and the segmentation network in the first material. The second training module is used to train a second machine learning model and includes the following units: a second material acquisition unit, used to acquire second material, the second material including a second image, wherein the second image is a frontal and side-view standing image of the same human body; the second material also includes the height and other body parameters of the human body in the second image; a first machine learning model invocation unit, used to input the second image into the first machine learning model for processing, and obtain the output result of the first machine learning model on the second image; a second training input unit, used to add the height information of the human body in the second image to the output result as the input information of the second machine learning model; a second training output unit, used to use the other body parameters of the human body in the second image as the output information of the second machine learning model; a second parameter adjustment model, used to adjust the parameters of the second machine learning model according to the difference between the information output by the second training output unit and the other body parameters in the second material; the second training input unit is also used to perform extended processing on the output result by adding the specific value of the height information of the human body in the second image, specifically including: The average value K is taken from the output of the first machine learning model on the second image; Set up a Gaussian mixture model G with two sub-distributions. G was trained using the height parameters of the individuals in the second dataset. The height parameters of each person in the second data set are calculated using G, resulting in two sub-distribution probabilities P and Q for each person. The height numerical dimension in the input information of the second machine learning model is expanded into two dimensions, assigned values ​​P*K and Q*K respectively, without retaining the original height numerical dimension. Train a second machine learning model; During the measurement phase, the first processing result is combined with A's height data, and two sub-probability values ​​are calculated by G and multiplied by K. These values ​​are then input into the second machine learning model to obtain the output results of the second machine learning model on A's human body parameters.

2. The human body measurement system according to claim 1, characterized in that, The first material also includes several control points for each human body part in the segmented mesh, and the output of the first machine learning model includes several control points for each human body part in the segmented mesh.

3. The human body measurement system according to claim 1, characterized in that, The first machine learning model is trained using a first material library, which includes more than 200 sets of first materials.

4. The human body measurement system according to claim 1, characterized in that, The second machine learning model is trained using a second dataset, which includes more than 50 sets of second data.

5. The human body measurement system according to claim 1, characterized in that, The first machine learning model is a deep neural network.

6. The human body measurement system according to claim 1, characterized in that, The second machine learning model is LightGBM.

7. The human body measurement system according to claim 1, characterized in that, The second training input unit is also used to extend the output result by adding specific numerical values ​​of the height information of the human body to which the second image belongs, specifically including: Let the height of the human body be H. After calculating P and Q, compare the size of P and Q. If P is greater than Q, add the following values ​​to the input information of the second machine learning model: H*0.515, H*0.345, H*0.542; if P is not greater than Q, add the following values ​​to the input information of the second machine learning model: H*0.612, H*0.421, H*0.

643. At the same time, the height value dimension in the input information of the second machine learning model will no longer be expanded into two dimensions, but the original height value dimension will be retained.

8. The human body measurement system according to claim 1, characterized in that, The second training input unit is also used to extend the output result by adding specific numerical values ​​of the height information of the human body to which the second image belongs, specifically including: Add the following values ​​to the input information of the second machine learning model: (0.515*P+0.612*Q) / (P+Q), (0.345*P+0.6421*Q) / (P+Q), (0.542*P+0.643*Q) / (P+Q), At the same time, the height value dimension in the input information of the second machine learning model will no longer be expanded into two dimensions, and the original height value dimension will not be retained.

9. The human body measurement system according to claim 1, characterized in that, The first training output unit is further configured to: perform skin color detection on the segmented mesh of the human body; perform linear regression on the boundary line between the skin-colored and non-skin-colored regions in the segmented mesh of the human body to obtain the boundary line; obtain the width of the skin-colored region at a preset distance on the skin-colored region side of the boundary line; obtain the width of the non-skin-colored region at a preset distance on the non-skin-colored region side of the boundary line; if the width of the non-skin-colored region is greater than the width of the skin-colored region, then proceed to the step: generating a reference curve with a distance S from the edge of the non-skin-colored region inward from the edge of the non-skin-colored region, wherein... S = width of non-skin-colored area - width of skin-colored area.

Images