A method and system for human acupoint recognition based on artificial intelligence image analysis

Abstract

A method and system for human acupoint recognition based on artificial intelligence image analysis, wherein the human acupoint recognition method is carried out through 6 steps. The present invention can accurately recognize acupoints and has the advantages of strong real-time performance and strong adaptability. The beneficial effects of the present invention are as follows: 1. By adopting the architecture of a multi-task CNN backbone network and a cascaded regressor, accurate positioning and real-time recognition of acupoints are realized; 2. The present invention also combines a position encoding mechanism of anatomical knowledge to improve the accuracy and robustness of recognition; 3. Through multi-scale feature extraction and iterative optimization strategies, it adapts to different body shapes and pose changes; 4. It has a real-time guidance function, improving the efficiency and safety of acupuncture operations; 5. It has good scalability and practicability, providing technical support for the standardization and modernization of acupuncture treatment.

Description

Technical Field

[0001] The present invention relates to the field of computer science and technology, and particularly relates to a method and system for human acupoint recognition based on artificial intelligence image analysis. Background Art

[0002] As a traditional treatment method passed down for thousands of years, traditional Chinese acupuncture has significant curative effects in treating pain, regulating endocrine, improving sleep, enhancing immunity, etc. However, the effect of acupuncture treatment depends to a large extent on the accuracy of acupoint location. Traditional acupoint location methods mainly rely on acupuncturists to complete it through palpation and visual inspection, combined with their years of clinical experience and professional training. This manual location method has the following problems:

[0003] 1. Time-consuming operation: Each acupoint needs to be carefully touched and measured, and the whole location process is time-consuming;

[0004] 2. Strong subjectivity: The location results of different physicians may vary;

[0005] 3. Prone to interference: Factors such as the physical state of the physician and environmental light can affect the location accuracy;

[0006] 4. Individual differences: Each patient has differences in body shape and bone structure, increasing the difficulty of accurate location.

[0007] In recent years, computer vision and deep learning technologies have made significant breakthroughs in the fields of human key point detection, pose estimation, etc. Models trained based on convolutional neural networks and large-scale datasets can accurately locate human bone joint points and facial feature points, and effectively learn the spatial relationships between them. However, applying these technologies to the field of acupoint recognition still faces many challenges:

[0008] 1. The corresponding relationship between acupoint positions and anatomical landmarks is complex, and a more refined feature mapping needs to be established;

[0009] 2. There are significant differences in the relative positions of acupoints among people with different body types, increasing the difficulty of model generalization;

[0010] 3. Some acupoints lack obvious visual features, and it is difficult to accurately locate them only relying on image information.

[0011] Therefore, in view of the deficiencies of the prior art, it is very necessary to provide a method and system for human acupoint recognition based on artificial intelligence image analysis to solve the deficiencies of the prior art. Summary of the Invention

[0012] The first object of the present invention is to avoid the deficiencies of the prior art and provide a method for identifying human acupoints based on artificial intelligence image analysis. The method for identifying human acupoints based on artificial intelligence image analysis can accurately identify acupoints and has the advantages of strong real-time performance and strong adaptability.

[0013] The above object of the present invention is achieved by the following technical measures:

[0014] Provide a method for identifying human acupoints based on artificial intelligence image analysis, which is carried out through the following steps:

[0015] S1. Collect the human body surface image to be processed;

[0016] S2. Preprocess the human body surface image obtained in S1;

[0017] S3. Extract the visual feature map and human body pose parameters in the preprocessed human body surface image obtained in S2 through the multi-task CNN backbone network in the trained acupoint recognition model; the trained acupoint recognition model is obtained by training the acupoint recognition model with an acupoint database, wherein the acupoint recognition model is composed of a multi-task CNN backbone network, a 3D human meridian acupoint template, and a cascade regressor;

[0018] S4. Generate the initial estimated position of the acupoint on the human body surface image according to the 3D human meridian acupoint template in the trained acupoint recognition model and the human body pose parameters obtained in S3 ;

[0019] S5. Process the initial estimated position obtained in S4 and the visual feature map obtained in S3 through the cascade regressor in the trained acupoint recognition model to obtain the acupoint coordinates in the human body surface image;

[0020] S6. Post-process the acupoint coordinates obtained in S5 to obtain the acupoint recognition result.

[0021] Preferably, the above acupoint database includes basic acupoint data.

[0022] Preferably, the acquisition process of the above basic acupoint data is as follows:

[0023] A1. Collect multiple human body surface images;

[0024] A2. Perform image preprocessing and label acupoint information on the human body surface images obtained in A1. The labeled acupoint information is the standard name of the acupoint, the anatomical location description of the acupoint, the meridian system to which the acupoint belongs, the main treatment function of the acupoint, and the indications of the acupoint, to obtain the above basic acupoint data.

[0025] Preferably, the above acupoint database further includes spatial positioning data.

[0026] Preferably, the process of obtaining the above spatial positioning data is as follows:

[0027] B1. Scan the three-dimensional data of multiple human bodies;

[0028] B2. Mark the three-dimensional coordinate information of acupoints, the positional relationship between acupoint anatomical landmarks on the human body, the relative distance between different acupoints, and the angular data between different acupoints in the three-dimensional data of the human body to obtain the above spatial positioning data.

[0029] Preferably, the above 3D human meridian acupoint template is constructed from the above spatial positioning data.

[0030] Preferably, the above human body posture parameters are obtained by the following steps:

[0031] S3.1. Calculate the rotation matrix of Euler angles , and the rotation matrix is obtained by combining three single-axis rotation matrices in a specific order. The rotation matrix is obtained from Equations (1) to (4):

[0032] ... Equation (1);

[0033] ... Equation (2);

[0034] ... Equation (3);

[0035] ... Equation (4);

[0036] where , and are three Euler angles respectively;

[0037] S3.2. The above human body posture parameters are composed of Euler angle rotation parameters and a translation vector, which is represented by Equation (5), and ;

[0038] ... Equation (5);

[0039] where is the rotation parameter, and the first three vectors in the rotation parameter are the rotation matrix , is the translation vector, is the human body posture parameter.

[0040] Preferably, the above initial estimated position is obtained by the following steps:

[0041] S4.1. Obtain the internal parameter matrix K for describing the internal parameters of the camera and the external parameter matrix E for describing the position and orientation of the camera relative to the world coordinate system.

[0042] The internal parameter matrix K is calculated by Equation (6), and the external parameter matrix E is calculated by Equation (7).

[0043] …… Equation (6);

[0044] Wherein, and are the focal lengths, is the optical center coordinate;

[0045] …… Equation (7);

[0046] S4.2. Obtain the total projection matrix P according to Equation (8), so as to map the three-dimensional world coordinates to the coordinates of the human body surface image to be processed.

[0047] …… Equation (8);

[0048] S4.3. According to the coordinates of the 3D acupoint points X on the 3D human meridian acupoint template projected onto the human body surface image to be processed, obtain the 2D image acupoints corresponding to the initial estimated positions and , .

[0049] Preferably, the above S4.3 includes the following steps:

[0050] S4.3.1. Represent the 3D acupoint points on the 3D human meridian acupoint template as homogeneous coordinates and , where the homogeneous coordinates are represented by Equation (9);

[0051] …… Equation (9);

[0052] Wherein, x, y, and z are the coordinate points of the homogeneous coordinates respectively, T is the matrix transpose;

[0053] S4.3.2. Transform the 3D acupoint points on the 3D human meridian acupoint template from the world coordinate system to the camera coordinate system through the external parameter matrix E, which is specifically represented by Equation (10);

[0054] …… Equation (10);

[0055] Wherein, is a three-dimensional point in the camera coordinate system.

[0056] S4.3.3. Project the 3D acupoint points on the 3D human meridian acupoint template onto the two-dimensional image plane using the internal parameter matrix K and perform normalization, which is specifically represented by Equation (11);

[0057] …… Equation (11);

[0058] where is the homogeneous image coordinate, and , x h 、 y h and z h are the coordinate points of the homogeneous image coordinate respectively;

[0059] S4.3.4. Obtain the initial estimated position by normalizing the homogeneous image coordinate , which is specifically represented by Equation (12), and the initial estimated position is represented by the complete projection function , where the projection function is obtained from Equation (13):

[0060] …… Equation (12);

[0061] …… Equation (13).

[0062] Preferably, the above cascade regressor is provided with a multi-step cascade regressor, and each step of the regressor is provided with a feature extraction module, a GAT processing module of the graph attention network, and a decoder.

[0063] Preferably, the processing process of each step of the regressor includes the following steps:

[0064] S5.1. The feature extraction module in the current step of the regressor extracts the local features of the window from the feature map F, resamples it to 7×7×256, and then obtains a 512-dimensional visual feature vector through two layers of convolution; at the same time, calculate the relative position vector between each feature point in the feature map F, encode it into a position feature through a multi-layer perceptron, and add the relative position vector and the position feature to obtain a fused feature ; when the current step of the regressor is the first-step regressor, the feature map F is the visual feature map; when the current step of the regressor is a regressor other than the first-step regressor, the feature map F is the fused feature embedded in the previous step of the regressor Visual feature map;

[0065] S5.2. Input the fusion feature and the estimated position into the GAT processing module in the current-step regressor, and perform multi-step iteration through the decoder MLP network and activation function of the current-step regressor, and finally output the acupoint position correction vector , and obtain the updated acupoint coordinates, where the updated acupoint coordinates are , where is the estimated position; when the current-step regressor is the first-step regressor and the GAT processing module is the first-layer graph attention network, the estimated position is the initial estimated position ; when the current-step regressor is a regressor other than the first-step regressor, the estimated position is the updated acupoint coordinates of the previous graph attention network; when the current-step regressor is the first-step regressor but the GAT processing module is not the first-layer graph attention network, the estimated position is the updated acupoint coordinates of the previous graph attention network.

[0066] Preferably, the above S6 is specifically: define the last updated acupoint coordinates in the cascade regressor as the final predicted acupoint coordinates, and perform validity judgment and smoothing processing on the final predicted acupoint coordinates to obtain the final acupoint coordinates; then superimpose and display the acupoint points corresponding to the final acupoint coordinates on the human body surface image in S1, and output the corresponding acupoint names and positioning information at the same time.

[0067] Preferably, the training method of the above acupoint recognition model is trained through a two-stage strategy, specifically, only the multi-task CNN backbone network is trained in the first stage; in the second stage, the multi-task CNN backbone network is frozen, and the cascade regressor is trained at the same time, and finally the trained acupoint recognition model is obtained.

[0068] Preferably, the above acupoint database further includes clinical application data; the clinical application data is the acupuncture depth of the acupoint, the angle of the acupoint, the contraindication information of the acupoint, the acupoint combination plan, and the precautions for the use of the acupoint corresponding to special populations.

[0069] The second object of the present invention is to provide a human acupoint recognition system to avoid the deficiencies of the prior art. The human acupoint recognition system can accurately identify acupoints and has the advantages of strong real-time performance and strong adaptability.

[0070] The above object of the present invention is achieved by the following technical measures:

[0071] Provide a human acupoint recognition system, which is carried out by using the above human acupoint recognition method based on artificial intelligence image analysis.

[0072] The human acupoint recognition system of the present invention is provided with:

[0073] Multi - angle image acquisition module - acquires images of the human body surface;

[0074] Acupoint database storage module - stores the acupoint database;

[0075] Pre - processing and annotation module - pre - processes the human body surface image and simultaneously annotates the acupoint information of the human body surface image in the acupoint database;

[0076] Big data model learning module - trains the acupoint recognition model according to the acupoint database; simultaneously, the trained acupoint recognition model is used to recognize the to - be - processed human body surface image and obtain the acupoint recognition result;

[0077] Interaction interface module - displays the acupoint recognition result and conducts human - machine interaction.

[0078] A method and system for human acupoint recognition based on artificial intelligence image analysis according to the present invention, wherein the human acupoint recognition method is carried out through the following steps: S1. Acquire the to - be - processed human body surface image; S2. Pre - process the human body surface image obtained in S1; S3. Extract the visual feature map and human body pose parameters in the pre - processed human body surface image obtained in S2 through the multi - task CNN backbone network in the trained acupoint recognition model; the trained acupoint recognition model is obtained by training the acupoint recognition model with the acupoint database, wherein the acupoint recognition model is composed of a multi - task CNN backbone network, a 3D human meridian acupoint template, and a cascade regressor; S4. Generate the initial estimated positions of acupoints on the human body surface image according to the 3D human meridian acupoint template in the trained acupoint recognition model and the human body pose parameters obtained in S3 ; S5. Process the initial estimated positions obtained in S4 and the visual feature map obtained in S3 through the cascade regressor in the trained acupoint recognition model to obtain the acupoint coordinates in the human body surface image; S6. Post - process the acupoint coordinates obtained in S5 to obtain the acupoint recognition result. The present invention can accurately recognize acupoints and has the advantages of strong real - time performance and strong adaptability. The beneficial effects of the present invention are as follows: 1. Adopting the architecture of a multi - task CNN backbone network and a cascade regressor realizes the accurate positioning and real - time recognition of acupoints; 2. The present invention also combines the position encoding mechanism of anatomical knowledge to improve the accuracy and robustness of recognition; 3. Through multi - scale feature extraction and iterative optimization strategies, it adapts to different body types and pose changes; 4. It has a real - time guidance function, improving the efficiency and safety of acupuncture operations; 5. It has good scalability and practicability, providing technical support for the standardization and modernization of acupuncture treatment. Brief Description of the Drawings

[0079] The present invention will be further described with reference to the accompanying drawings, but the content in the drawings does not constitute any limitation to the present invention.

[0080] Figure 1 It is a flowchart of a method for identifying human acupoints based on artificial intelligence image analysis.

[0081] Figure 2 It is a flowchart of the identification method of the trained acupoint recognition model. Specific embodiments

[0082] The technical solution of the present invention will be further described in conjunction with the following embodiments.

[0083] Embodiment 1

[0084] A method for identifying human acupoints based on artificial intelligence image analysis, as Figure 1 shown, is carried out through the following steps:

[0085] S1. Collect the human body surface image to be processed;

[0086] S2. Preprocess the human body surface image obtained in S1;

[0087] S3. Extract the visual feature map and human body pose parameters in the preprocessed human body surface image through the multi-task CNN backbone network in the trained acupoint recognition model; the trained acupoint recognition model is obtained by training the acupoint recognition model with an acupoint database, where the acupoint recognition model is composed of a multi-task CNN backbone network, a 3D human meridian acupoint template, and a cascade regressor;

[0088] S4. Generate the initial estimated positions of the acupoints on the human body surface image according to the 3D human meridian acupoint template in the trained acupoint recognition model and the human body pose parameters obtained in S3 ;

[0089] S5. Process the initial estimated positions obtained in S4 and the visual feature map obtained in S3 through the cascade regressor in the trained acupoint recognition model to obtain the acupoint coordinates in the human body surface image;

[0090] S6. Post-process the acupoint coordinates obtained in S5 to obtain the acupoint recognition result.

[0091] It should be noted that the multi-task CNN backbone network of the present invention takes a 256×256 human body surface image as input, downsamples it to 64×64 through an initial encoder, and then processes it through 4 cascaded hourglass modules. Each hourglass module contains an encoder-decoder structure where the feature map is reduced to 8×8 at the bottleneck layer, and has two branch tasks of 3D pose parameter estimation and feature point heat map. Finally, pose parameters and a 64×64 visual feature map are output. Multi-task learning is a machine learning method that allows the model to learn multiple related tasks simultaneously during training, thereby improving the generalization ability and robustness of the model. Specifically, while the network of the present invention is learning human pose parameter estimation, it is also learning the acupoint recognition task. In this way, the global features extracted from the pose estimation task can be used to assist the acupoint recognition task, thereby improving the overall recognition accuracy.

[0092] The 3D human meridian acupoint template is a general template. Through S4 processing, the pose parameters predicted by the multi-task CNN backbone network are projected onto the 3D human meridian acupoint template to obtain the initial estimated position. This projection of the 3D human meridian acupoint template utilizes the geometric information of the 3D model to improve the accuracy of the initial positioning.

[0093] It should be noted that the 3D human meridian acupoint template of the present invention can be constructed from the spatial positioning data in the acupoint database. The 3D human meridian acupoint template can also be sourced from public anatomical databases and medically annotated image datasets. The 3D human meridian acupoint template is a statistical average of the real human anatomical structure distribution, ensuring the accuracy and reliability of the model. In this embodiment, the construction of the 3D human meridian acupoint template from the spatial positioning data in the acupoint database is taken as an example for illustration.

[0094] Then the acupoint database of this embodiment includes basic acupoint data, spatial positioning data, and clinical application data. Among them, the clinical application data are the acupuncture depth of the acupoints, the angle of the acupoints, the contraindication information of the acupoints, the acupoint combination scheme, and the usage precautions for the corresponding acupoints of special populations (such as pregnant women, the elderly). The role of the clinical application data of the present invention is to output the clinical application data corresponding to the acupoints together when the acupoint recognition result is output, thereby enriching the acupoint information.

[0095] Among them, the process of obtaining the basic acupoint data is as follows:

[0096] A1. Collect multiple human body surface images;

[0097] A2. Perform image preprocessing on the human body surface image obtained in A1 and label acupoint information, where the labeled acupoint information includes the standard acupoint names (such as Chinese names, pinyin names, English names), anatomical location descriptions of acupoints, the meridian systems to which acupoints belong, the main treatment functions of acupoints, and the indications of acupoints, so as to obtain basic acupoint data.

[0098] It should be noted that the preprocessing of the human body surface image obtained in A1 and the human body surface image obtained in S1 of the present invention includes adjusting brightness, adjusting contrast, correcting color balance, removing noise, unifying image size resolution, aligning poses, correcting poses, performing light compensation, and eliminating unqualified images. Among them, adjusting brightness and contrast can improve image clarity, performing color balance correction can ensure accurate skin color restoration, using methods such as Gaussian filtering to remove noise, and at the same time unifying image size and resolution, performing pose alignment and correction, and performing light compensation to eliminate the influence of uneven illumination, and eliminating unqualified images such as blurred, overexposed, and underexposed images, and removing duplicate or redundant data.

[0099] When collecting the human body surface images in A1 and S1, a high-resolution camera array (≥12 million pixels) arranged in a surrounding manner can be used, which can capture the subtle features and anatomical landmarks on the human body surface in all directions. The bracket of the camera array adopts an electric lifting design, which can flexibly adjust the height and shooting angle according to the body type of the subject to ensure complete coverage of all parts of the body. The camera of the present invention is also equipped with an intelligent exposure control and an LED fill light array, which can provide uniform illumination in different light environments and highlight the body surface features. The high-precision autofocus system cooperates with the anti-shake bracket to effectively eliminate shooting jitter and ensure image clarity. This multi-angle, high-resolution image acquisition scheme provides a reliable data basis for subsequent acupoint recognition.

[0100] In the labeling of acupoint information in A2, acupuncture experts with rich clinical experience divide the human body into main anatomical regions such as the head, trunk, and limbs, and at the same time label the main anatomical landmark points of the human body (such as bone joints, muscle trends, etc.), record the relative distance and angular relationship between acupoints and surrounding anatomical landmark points, and label the recommended insertion depth and angle of acupoints. At the same time, when labeling acupoint information, quality control can also be carried out. Specifically, a multi-person cross-validation method is adopted for repeated labeling by different experts to ensure the accuracy of labeling, a standardized review process is established to ensure labeling quality, and the labeling data is continuously optimized and updated according to clinical feedback.

[0101] Among them, the process of obtaining spatial positioning data is as follows:

[0102] B1. Scan the three-dimensional data of multiple human bodies;

[0103] B2. Mark the three - dimensional coordinate information of acupoints, the positional relationship of anatomical landmark points of acupoints on the human body, the relative distance between different acupoints, and the angular data between different acupoints on the three - dimensional data of the human body to obtain spatial positioning data.

[0104] It should be noted that the three - dimensional data of the human body can be obtained through structured light, 3D scanning devices, CT scanners, etc., and the marking method of the three - dimensional data of the human body is a conventional technology in this field and will not be elaborated here one by one.

[0105] The human body posture parameters of the present invention are obtained by the following steps:

[0106] S3.1. Calculate the rotation matrix of Euler angles , and the rotation matrix is obtained by combining three single - axis rotation matrices in a specific order. The rotation matrix is obtained from equations (1) to (4):

[0107] ... Equation (1);

[0108] ... Equation (2);

[0109] ... Equation (3);

[0110] ... Equation (4);

[0111] where , and are three Euler angles respectively;

[0112] S3.2. The human body posture parameters are composed of Euler angle rotation parameters and a translation vector, which is represented by equation (5), and ;

[0113] ... Equation (5);

[0114] where is the rotation parameter, and the first three vectors in the rotation parameter are the rotation matrix , is the translation vector, is the human body posture parameter.

[0115] The initial estimated position of the present invention is obtained by the following steps:

[0116] S4.1. Obtain the intrinsic parameter matrix K for describing the internal parameters of the camera and the extrinsic parameter matrix E for describing the position and orientation of the camera relative to the world coordinate system. The intrinsic parameter matrix K is calculated by Equation (6), and the extrinsic parameter matrix E is calculated by Equation (7). The intrinsic parameter matrix K is used to describe the internal parameters of the camera, specifically including the focal length, the optical center, and the pixel ratio. The intrinsic parameter matrix K is usually represented as a 3×3 matrix:

[0117] …… Equation (6);

[0118] where, and are the focal lengths, is the optical center coordinate;

[0119] …… Equation (7);

[0120] S4.2. Obtain the total projection matrix P according to Equation (8), so as to map the three-dimensional world coordinates to the coordinates of the human body surface image to be processed; among them, the dimension of the total projection matrix P is 3×4;

[0121] …… Equation (8);

[0122] S4.3. Obtain the initial estimated positions of the 2D image acupoints corresponding to the 3D acupoints X on the 3D human meridian acupoint template projected onto the coordinates of the human body surface image to be processed , and , .

[0123] Among them, S4.3 includes the following steps:

[0124] S4.3.1. Represent the 3D acupoints on the 3D human meridian acupoint template as homogeneous coordinates , and , where the homogeneous coordinate is represented by Equation (9);

[0125] …… Equation (9);

[0126] Among them, x, y, and z are the coordinate points of the homogeneous coordinate respectively, T is the matrix transpose;

[0127] S4.3.2. Transform the 3D acupoints on the 3D human meridian acupoint template from the world coordinate system to the camera coordinate system through the extrinsic parameter matrix E, which is specifically represented by Equation (10);

[0128] …… Equation (10);

[0129] Among them, is a three-dimensional point in the camera coordinate system;

[0130] S4.3.3. Project the 3D acupoint points on the 3D human meridian acupoint template onto the two-dimensional image plane using the internal parameter matrix K and perform normalization, which is specifically represented by Equation (11);

[0131] …… Equation (11);

[0132] Among them, is the homogeneous image coordinate, and , x h , y h and z h are the coordinate points of the homogeneous image coordinate respectively;

[0133] S4.3.4. Obtain the initial estimated position by normalizing the homogeneous image coordinate , which is specifically represented by Equation (12), and the initial estimated position is represented by the complete projection function , where the projection function is obtained from Equation (13):

[0134] …… Equation (12);

[0135] …… Equation (13).

[0136] The cascaded regressor is set with a multi-step cascaded regressor, and each step of the regressor is set with a feature extraction module, a GAT processing module of the graph attention network, and a decoder.

[0137] It should be noted that the cascaded regressor adopts a multi-step iterative optimization structure. Usually, three steps are used by the cascaded regressor to achieve the ideal effect. In this embodiment, a three-step cascaded regressor is taken as an example. The GAT processing module has a total of 4 layers, and each layer includes operations such as query / key / value vector generation, attention weight calculation, feature aggregation, and residual connection. The multi-layer structure of the GAT processing module allows the model to first observe globally and then gradually focus on the local information most relevant to the target. The initial layer pays more attention to the global structural information, such as the unoccluded area, and the subsequent layers focus on local precise adjustment. This strategy helps to capture more accurate features at different scales and can capture feature information at different levels. The role of the decoder is to calculate the relative position of each acupoint relative to other anatomical landmark points for each acupoint, generate a position vector, which includes its relative position and geometric features, and use the embedding layer to convert the position vector into a high-dimensional feature representation so as to combine with other features and integrate the spatial relationship information of the acupoints into the feature representation. Feature aggregation combines the local visual features of the acupoints (such as color, texture) and their relative positions on the human body to form a complete feature representation. This can help the model understand the relative position relationship between acupoints, and by extracting features at different scales, gradually improve the expression ability of features from rough to fine.

[0138] Among them, as Figure 2 , the processing process of each step of the regressor includes the following steps:

[0139] S5.1. The feature extraction module in the current step regressor extracts the local features of the window from the feature map F, resamples them to 7×7×256, and then obtains a 512-dimensional visual feature vector through two layers of convolution; at the same time, calculates the relative position vector between each feature point in the feature map F, encodes it into a position feature through a multi-layer perceptron, and adds the relative position vector and the position feature to obtain a fused feature ; when the current step regressor is the first step regressor, the feature map F is the visual feature map; when the current step regressor is a regressor other than the first step regressor, the feature map F is the visual feature map embedded with the fused feature in the previous step regressor;

[0140] S5.2. Input the fused feature obtained in S5.1 and the estimated position into the GAT processing module in the current step regressor, and perform multi-step iteration through the decoder MLP network and activation function of the current step regressor, and finally output the acupoint position correction vector , and obtain the updated acupoint coordinates, where the updated acupoint coordinates are , where To estimate the position; when the current step regressor is the first step regressor and the GAT processing module is the first layer of graph attention network, the estimated position is the initial estimated position ; when the current step regressor is a regressor other than the first step regressor, the estimated position is the updated acupoint coordinates of the previous graph attention network; when the current step regressor is the first step regressor but the GAT processing module is not the first layer of graph attention network, the estimated position is the updated acupoint coordinates of the previous graph attention network.

[0141] It should be noted that the activation function of the present invention is a conventional activation function in the art, such as Sigmoid function, ReLU function, Leaky ReLU function, Softmax function, etc., and these activation functions can all be used as the activation function of the present invention.

[0142] Specifically, S6 is: defining the last updated acupoint coordinates in the cascade regressor as the final predicted acupoint coordinates, and performing validity judgment and smoothing processing on the final predicted acupoint coordinates to obtain the final acupoint coordinates; then superimposing and displaying the acupoint points corresponding to the final acupoint coordinates on the human body surface image of S1, and outputting the corresponding acupoint names and positioning information at the same time.

[0143] It should be noted that the validity judgment is to delete the wrong acupoint coordinates, such as acupoints that are relatively close to each other, overlapping acupoints, etc.; and the coordinate smoothing processing is a conventional operation in the art and is not the focus of the present invention, so it will not be elaborated here.

[0144] The training method of the acupoint recognition model is trained through a two-stage strategy. Specifically, in the first stage, only the multi-task CNN backbone network is trained; in the second stage, the multi-task CNN backbone network is frozen, and the cascade regressor is trained at the same time, and finally the trained acupoint recognition model is obtained.

[0145] It should be noted that the end condition of the first stage training of the present invention is to reach the number of training times, and the number of training times can be 500 rounds or other times. After the first stage training, the second stage training is carried out, and the end condition of the second stage training is to reach the number of training times or convergence. The specific training end condition can be determined according to the actual situation. The training method of the acupoint recognition model of the present invention uses the Adaptive Moment Estimation (Adam) optimizer, the initial learning rate is set to 0.001, and the cosine annealing learning rate scheduling strategy is adopted, and it decays once every 100 rounds. The and parameters are set to 0.9 and 0.999 respectively, and the weight decay coefficient is 0.0001.

[0146] Beneficial effects of the human acupoint recognition method based on artificial intelligence image analysis: 1. By adopting the architecture of a multi-task CNN backbone network and a cascaded regressor, accurate positioning of acupoints is achieved. At the same time, the image acquisition and processing speed of the present invention reaches more than 30 frames per second, and the acupoint recognition response time is controlled within 100 ms, with no lag in real-time display and prompt, so that real-time recognition can be achieved; 2. The present invention also combines the position encoding mechanism of anatomical knowledge to improve the accuracy and robustness of recognition; 3. Through multi-scale feature extraction and iterative optimization strategies, it adapts to acupoint recognition and pose changes of different body types, ages, and genders; 4. It has a real-time guidance function, improving the efficiency and safety of acupuncture operations; 5. It has good scalability and practicability, providing technical support for the standardization and modernization of acupuncture treatment.

[0147] Example 2

[0148] Application of the human acupoint recognition method based on artificial intelligence image analysis in Example 1. Among them, the human body surface images of S1 and A1 are both collected in real time from patients through a camera or loaded from a saved image file. When collecting the human body surface image, a high-definition camera with a resolution not lower than 1080p is used, equipped with a multi-angle adjustable bracket and an optional infrared camera. The acquisition parameters are set to a frame rate of 30 - 60 fps, and the functions of autofocus, automatic exposure, and white balance are enabled. During the acquisition process, the target area is photographed in a 360-degree loop, and multiple images are taken at key angles such as 0°, 45°, 90°, 135°, 180°, etc., maintaining an optimal shooting distance of 20 - 50 cm. The present invention detects the image clarity and lighting conditions in real time, automatically filters out blurred or unqualified images, and ensures the acquisition quality.

[0149] The human body surface image of this embodiment needs to be preprocessed standardly. The preprocessing specifically includes adjusting the image size to a standard input size of 256×256 and converting the RGB color space. At the same time, noise is removed through Gaussian filtering, contrast is enhanced through histogram equalization, salt-and-pepper noise is eliminated through median filtering, and image enhancement and lighting correction processes such as adaptive gamma correction, local contrast enhancement, and shadow removal are performed. For input images of different sizes, bilinear interpolation is used for resampling, maintaining the aspect ratio of the image, and zero-padding is performed at the edges to meet the input requirements of the network. Other processes are the same as in Example 1.

[0150] This embodiment has accurate recognition: training based on deep learning and graph attention network; position encoding mechanism combined with anatomical knowledge; multi-scale feature extraction and iterative optimization strategy

[0151] Strong real-time performance: The image acquisition and processing speed reaches more than 30 frames per second; the acupoint recognition response time is controlled within 100 ms; there is no lag in real-time display and prompt

[0152] Strong adaptability: Supports acupoint recognition for different body sizes, ages, and genders; adapts to various lighting and shooting angle conditions; is robust to occlusion and pose changes; supports personalized parameter adjustment.

[0153] Intelligent interaction: Intuitive graphical operation interface; supports multiple interaction methods; allows switching between professional and simple modes; has remote guidance and teaching functions.

[0154] Example 3

[0155] Application of the human acupoint recognition method based on artificial intelligence image analysis in Example 2. This example is used to recognize acupoints on the human hand. The operator first uses a camera to take multiple images of the human body surface of the hand, and the system processes the images and uses a big data model to recognize acupoints such as "Hegu" and "Neiguan". When performing acupuncture, the camera will monitor the hand position in real time. When the tip of the needle approaches a certain acupoint, the system will prompt the name and function of the acupoint for the acupuncturist's reference.

[0156] Example 4

[0157] Application of the human acupoint recognition method based on artificial intelligence image analysis in Example 2. This example is used for recognition of multiple parts of the whole body. The system can take images of different parts such as the head, torso, and limbs respectively. By combining the human body surface images of multiple parts, the present invention can recognize important acupoints such as "Baihui" and "Zusanli", and provide complete acupoint prompts through real-time monitoring of the human body surface image and the acupuncture site during the acupuncture process.

[0158] Example 5

[0159] A human acupoint recognition system uses the human acupoint recognition method based on artificial intelligence image analysis as in Example 1 or 2 for human acupoint recognition.

[0160] Among them, the human acupoint recognition system of the present invention is provided with:

[0161] Multi-angle image acquisition module - acquires human body surface images;

[0162] Acupoint database storage module - stores the acupoint database;

[0163] Preprocessing and annotation module - preprocesses the human body surface images and simultaneously annotates the labeled acupoint information of the human body surface images in the acupoint database;

[0164] Big data model learning module - trains the acupoint recognition model according to the acupoint database; simultaneously uses the trained acupoint recognition model to recognize the to-be-processed human body surface images and obtains the acupoint recognition results;

[0165] The interactive interface module - displays the acupoint recognition results and conducts human-computer interaction.

[0166] The beneficial effects of this human acupoint recognition system are as follows: 1. By adopting the architecture of a multi-task CNN backbone network and a cascaded regressor, accurate positioning of acupoints is achieved. At the same time, the image acquisition and processing speed of the present invention reaches more than 30 frames per second, and the acupoint recognition response time is controlled within 100 ms, with no lag in real-time display and prompt, so that real-time recognition can be achieved; 2. The present invention also combines the position encoding mechanism of anatomical knowledge to improve the accuracy and robustness of recognition; 3. Through multi-scale feature extraction and iterative optimization strategies, it adapts to acupoint recognition and pose changes of different body types, ages, and genders; 4. It has a real-time guidance function, improving the efficiency and safety of acupuncture operations; 5. It has good scalability and practicability, providing technical support for the standardization and modernization of acupuncture treatment.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for identifying human acupoints based on artificial intelligence image analysis, characterized in that, It is carried out through the following steps: S1. Collect the human body surface image to be processed; S2. Preprocess the human body surface image obtained in S1; S3. Extract the visual feature map and human body pose parameters in the preprocessed human body surface image through the multi-task CNN backbone network in the trained acupoint recognition model; the trained acupoint recognition model is obtained by training the acupoint recognition model with the acupoint database, where the acupoint recognition model is composed of a multi-task CNN backbone network, a 3D human meridian acupoint template, and a cascade regressor; S4. Generate the initial estimated position x0 of the acupoints on the human body surface image according to the 3D human meridian acupoint template in the trained acupoint recognition model and the human body pose parameters obtained in S3; S5. Process the initial estimated position x0 obtained in S4 and the visual feature map obtained in S3 through the cascade regressor in the trained acupoint recognition model to obtain the acupoint coordinates in the human body surface image; S6. Post-process the acupoint coordinates obtained in S5 to obtain the acupoint recognition result.

2. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 1, wherein: The acupoint database includes basic acupoint data; The acquisition process of the basic acupoint data is as follows: A1. Collect multiple human body surface images; A2. Perform image preprocessing and label acupoint information on the human body surface images obtained in A1. The labeled acupoint information is the standard name of the acupoint, the anatomical location description of the acupoint, the meridian system to which the acupoint belongs, the main treatment functions of the acupoint, and the indications of the acupoint, to obtain the basic acupoint data.

3. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 2, wherein: The acupoint database also includes spatial positioning data; The acquisition process of the spatial positioning data is as follows: B1. Scan multiple human body three-dimensional data; B2. Label the three-dimensional coordinate information of the acupoints, the positional relationship of the anatomical landmark points of the human body of the acupoints, the relative distance between different acupoints, and the angular data between different acupoints on the human body three-dimensional data to obtain the spatial positioning data; The 3D human meridian acupoint template is constructed from the spatial positioning data.

4. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 2 or 3, characterized in that, The human body pose parameters are obtained through the following steps: S3.

1. Calculate the rotation matrix R of the Euler angles. The rotation matrix R is obtained by combining three single-axis rotation matrices in a specific order. The rotation matrix R is obtained from equations (1) to (4): R = R z (θ z )·R y (θ y )·R x (θ x )..... Equation (1); where, θ y , θ x and θ z are three Euler angles, respectively; S3.

2. The human body pose parameters consist of Euler angle rotation parameters and a translation vector, are represented by Equation (5), and euler = p[:,0:3], m = p[:,3:6]...... Equation (5); Among them, euler is the rotation parameter, and the first three vectors in the rotation parameter euler are the rotation matrix R, m is the translation vector, and p is the human body pose parameter.

5. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 4, wherein The initial estimated position x0 is obtained through the following steps: S4.

1. Obtain the internal parameter matrix K for describing the internal parameters of the camera and the external parameter matrix E for describing the position and orientation of the camera relative to the world coordinate system. Among them, the internal parameter matrix K is calculated by equation (6), and the external parameter matrix E is calculated by equation (7), where f x and f y are the focal lengths, and (c x , c y ) are the optical center coordinates; E = [R|m]...... Equation (7); S4.

2. Obtain the total projection matrix P according to equation (8), so as to map the three-dimensional world coordinates to the coordinates of the human body surface image to be processed; P = K·[R|m]...... Equation (8); S4.

3. Obtain the 2D image acupoints corresponding to the initial estimated position x0 according to the coordinates of the 3D acupoint X on the 3D human meridian acupoint template projected onto the human surface image to be processed, and 6. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 5, characterized in that, The S4.3 includes the following steps: S4.3.

1. Represent the 3D acupoint on the 3D human meridian acupoint template as homogeneous coordinates X q , and where the homogeneous coordinates X q is represented by Equation (9); X q = [x, y, z, 1] T ...... Equation (9); where x, y, and z are the coordinate points of the homogeneous coordinates X q respectively, and T is the matrix transpose; S4.3.

2. Transform the 3D acupoint points on the 3D human meridian acupoint template from the world coordinate system to the camera coordinate system through the external parameter matrix E, which is specifically represented by Equation (10); X c = [R|m]·X q ...... Equation (10); where X c is a three-dimensional point in the camera coordinate system; S4.3.

3. Project the 3D acupoint points on the 3D human meridian acupoint template onto the two-dimensional image plane using the internal parameter matrix K and perform normalization, which is specifically represented by Equation (11); X h = K·X c ...... Equation (11); Among them, X h is the homogeneous image coordinate, and x h y h and z h are the coordinate points of the homogeneous image coordinate X h respectively; S4.3.

4. Obtain the initial estimated position \(x_0\) by normalizing the homogeneous image coordinate \(x\), which is specifically represented by Equation (12), and the initial estimated position \(x_0\) is represented by the complete projection function \(\pi(\cdot)\), where the projection function \(\pi(\cdot)\) is obtained from Equation (13): h ​ 7. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 2 or 3, characterized in that: The cascaded regressor is provided with a multi-step cascaded regressor, and each step of the regressor is provided with a feature extraction module, a GAT processing module with a multi-layer graph attention network, and a decoder; The processing process of each step of the regressor includes the following steps: S5.

1. The feature extraction module in the current step regressor extracts the local features of a w t ×w t window from the feature map F, resamples it to 7×7×256, and then obtains a 512-dimensional visual feature vector through two layers of convolution; meanwhile, calculates the relative position vector q t between each feature point in the feature map F, encodes it into a position feature r t through a multi-layer perceptron, and adds the relative position vector and the position feature to obtain a fused feature f t ; when the current step regressor is the first step regressor, the feature map F is the visual feature map; when the current step regressor is a regressor other than the first step regressor, the feature map F is the visual feature map embedded with the fused feature f t from the previous step regressor. S5.

2. Input the fused feature f t and the estimated position into the GAT processing module in the current-step regressor, and perform multi-step iteration through the decoder MLP network and activation function of the current-step regressor, and finally output the acupoint position correction vector Δx t to obtain the updated acupoint coordinates, where the updated acupoint coordinates are x t-1 +Δx t , where x t-1 is the estimated position; when the current-step regressor is the first-step regressor and the GAT processing module is the first-layer graph attention network, the estimated position is the initial estimated position x0; when the current-step regressor is a regressor other than the first-step regressor, the estimated position is the updated acupoint coordinates of the previous graph attention network; when the current-step regressor is the first-step regressor but the GAT processing module is not the first-layer graph attention network, the estimated position is the updated acupoint coordinates of the previous graph attention network.

8. The method for identifying human acupoints based on artificial intelligence image analysis according to claim 2 or 3, characterized in that, Specifically, S6 is: Define the acupoint coordinates updated last time in the cascaded regressor as the final predicted acupoint coordinates, and perform validity judgment and smoothing processing on the final predicted acupoint coordinates to obtain the final acupoint coordinates; then superimpose and display the acupoint points corresponding to the final acupoint coordinates on the human body surface image in S1, and output the corresponding acupoint names and positioning information at the same time; The training method of the acupoint recognition model is trained through a two-stage strategy. Specifically, in the first stage, only the multi-task CNN backbone network is trained; in the second stage, the multi-task CNN backbone network is frozen, and at the same time, the cascaded regressor is trained to finally obtain the trained acupoint recognition model; The acupoint database also includes clinical application data; the clinical application data is the acupuncture depth of the acupoint, the angle of the acupoint, the contraindication information of the acupoint, the acupoint combination plan, and the precautions for the use of acupoints corresponding to special populations.

9. A human acupoint recognition system, characterized in that: Use the human acupoint recognition method based on artificial intelligence image analysis according to any one of claims 1 to 8 for human acupoint recognition.

10. The human acupoint recognition system according to claim 9, wherein, There is provided: Multi-angle image acquisition module - to acquire the human body surface image; Acupoint database storage module - to store the acupoint database; Preprocessing and annotation module - to preprocess the human body surface image and at the same time annotate the annotated acupoint information of the human body surface image in the acupoint database; Big data model learning module - to train the acupoint recognition model according to the acupoint database; at the same time, identify the human body surface image to be processed through the trained acupoint recognition model and obtain the acupoint recognition result; Interaction interface module - to display the acupoint recognition result and perform human-computer interaction.

Images