Human body meridian point automatic detection method and system based on AI machine vision

Abstract

The invention discloses a human body meridian point automatic detection method and system based on AI machine vision, and relates to the crossing field of artificial intelligence and medical health, the method comprises the following steps: hand face direction determination: determining a palm center face or a hand back face according to a continuous distribution curve extraction result in a hand image; hand face feature extraction: segmenting a finger region and a palm region by finger root cross grains, and extracting feature marks according to a hand face selection corresponding strategy; establishing a global reference, constructing a coordinate system by taking the midpoint of the wrist cross grain as a global original point, segmenting the refined region and forming a mapping library; local adaptation: establishing a local coordinate system and calculating an offset with a global origin; candidate points are screened, the optimal acupuncture points are screened and verified from the refined area in combination with multiple features, and the system correspondingly comprises the function modules; the method improves the automation degree, positioning precision and adaptability of acupoint detection, and meets the requirements of traditional Chinese medicine diagnosis and treatment.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology between artificial intelligence and medical health, and more specifically to a method and system for automatic detection of human meridian acupoints based on AI machine vision. Background Art

[0002] With the rapid development of artificial intelligence and machine vision technologies, automated detection of human meridian and acupoints based on AI machine vision has become an important area of ​​research in the modernization of Traditional Chinese Medicine (TCM). This type of technology enables precise location of meridian and acupoints, providing technical support for TCM diagnosis and treatment. However, existing technologies still have limitations in automation, positioning accuracy, and adaptability, limiting their widespread adoption in practical applications.

[0003] After searching, a method for preparing a customized facial mask device based on visual three-dimensional reconstruction was disclosed with publication number CN111639553B. This patent takes multiple facial pictures and combines them with three-dimensional reconstruction technology to mark facial acupuncture points to produce a customized facial mask device that fits the user's face. However, this technical solution mainly relies on manual marking of acupuncture points and lacks the ability to automatically identify meridian acupuncture points, resulting in low efficiency and accuracy limited by the accuracy of manual operation.

[0004] In addition, a mobile articulated pulse diagnosis instrument with publication number CN115778320B has been disclosed. This patent uses an image acquisition module and a robotic arm module, combined with visual positioning technology, to achieve automatic recognition and pressing of the Cun, Guan, and Chi pulse acupoints. However, this technical solution is mainly aimed at locating specific acupoints on the wrist, and its scope of application is limited. It is difficult to cover meridian acupoints in other parts of the human body. At the same time, this method relies on vascular imaging technology, and has weak detection capabilities for non-vascular-related acupoints, which cannot meet the needs of traditional Chinese medicine meridian theory for diversified acupoint detection.

[0005] The above issues indicate that existing acupoint location methods based on vision technology still have significant deficiencies in terms of automation, scope of application, and adaptability to diverse acupoints. Therefore, the present invention provides a method for automatic detection of human meridian acupoints based on AI machine vision. This method aims to achieve high-precision, automated detection of meridian acupoints throughout the human body through intelligent algorithms, improving detection efficiency and applicability, and meeting the demand for efficient and accurate acupoint detection technology in modern Traditional Chinese Medicine diagnosis and treatment. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of the present invention is to solve the shortcomings of the existing acupoint positioning methods based on vision technology in terms of automation level, scope of application and adaptability to diversified acupoints. Therefore, an automatic detection method of meridian acupoints on the human hand based on AI machine vision is provided. Through AI machine vision technology, high-precision and automated detection of meridian acupoints on the human hand can be achieved, thereby improving positioning efficiency and applicability and meeting the needs of modern Chinese medicine diagnosis and treatment.

[0007] To achieve the above object, the present invention provides the following technical solutions: The method for automatically detecting acupuncture points on the human hand meridians based on AI machine vision includes the following steps: A hand face direction determination step is performed to determine whether a continuous distribution curve can be extracted from the user's hand image captured by the visual camera, and output the palm face or the back of the hand according to the extraction result; a hand surface feature extraction step, dividing the hand image into a finger area and a palm area with the finger root crease as the boundary, and selecting a corresponding feature extraction strategy according to the surface of the hand image to extract characteristic landmarks in the hand image; A global reference establishment step, selecting the midpoint of the wrist crease as the global origin, defining the horizontal and vertical coordinate axes, segmenting the hand image according to the characteristic marks to obtain a refined region, extracting the boundary coordinates of the refined region in the global coordinate system and forming a mapping library; In the local adaptation step, a local coordinate system is established for each refined area with its geometric center as the local origin. The offset between the local origin and the global origin is calculated and a coordinate transformation is performed to enable the local positioning result to be directly mapped to the global space. In the candidate point screening step, the spatial relationship parameters between the refined area and the feature marks are extracted as the initial conditions for subsequent candidate point screening. The proportionate candidate points are screened out from the refined area in combination with the texture intersection features, grayscale gradient features and acupoint group correlation verification. The optimal candidate point is retained as the target point by performing concave-convex feature matching verification on the candidate points.

[0008] Furthermore, when the hand image is the palm surface, the feature extraction strategy includes using an edge detection algorithm to extract the palm main line, the palm main line reflects three feature lines, and the palm area is divided into several macro areas according to the spatial distribution of the palm main line, and the boundary coordinates are output. The palm muscle texture analysis is then performed on the macro areas, and the texture direction is extracted through Hough transform. The texture direction includes longitudinal, transverse and radial directions. The preset palm acupoints are mapped one by one to each macro area to obtain the acupoint location information.

[0009] Furthermore, when the hand image is of the back of the hand, the feature extraction strategy includes using a contour detection algorithm to identify the metacarpal bone vertices and the dorsal wrist creases, analyzing the local gradient direction and intensity changes of the tendon through a texture tracking algorithm to extract the texture direction of the extensor tendon, and mapping the preset back of the hand acupoints one by one to each macro region to obtain the acupoint location information.

[0010] Furthermore, the global benchmark establishment step includes a refined region segmentation strategy, and the refined region segmentation strategy includes a palm face segmentation step and a back hand segmentation step. The palm surface segmentation step further divides the macro region into fine regions by combining the texture arrangement and the number of texture intersections per square millimeter; The step of segmenting the back of the hand divides the metacarpal bone apex and the dorsal wrist crease into fine regions by combining tendon directions and bony landmarks.

[0011] Furthermore, the global benchmark establishment step includes a benchmark establishment strategy, and the benchmark establishment strategy includes a coordinate axis construction step and a boundary coordinate determination step. In the coordinate axis construction step, the midpoint of the wrist transverse crease pointing toward the thumb is defined as the horizontal coordinate axis, and the midpoint of the wrist transverse crease pointing toward the fingertips is defined as the vertical coordinate axis; The boundary coordinate determination step is to identify the contour edges of each refined area through an edge detection algorithm, and obtain the coordinate values ​​of at least four vertices as boundary coordinates.

[0012] Furthermore, in the local adaptation step, the average of the horizontal coordinates of all boundary vertices in the refined area is used as the horizontal coordinate of the local origin, and the average of the vertical coordinates of all boundary vertices in the refined area is used as the vertical coordinate of the local origin. The offset includes a horizontal offset and a vertical offset.

[0013] Furthermore, the spatial relationship parameters include distance parameters and angle parameters, and the candidate point screening step also includes a multi-feature fusion screening strategy, which includes a texture cross-matching step and a grayscale gradient analysis step. The pattern cross-matching step identifies the intersection points, turning points and dense texture clusters of the palm print texture in the refined area, calculates the Euclidean distance between the candidate point and the area boundary and the angle with the feature line, and selects the top 5 candidate points that meet the preset ratio. The grayscale gradient analysis step extracts gradient changes around candidate points in the hand image and retains candidate points with annular gradient features.

[0014] Furthermore, the multi-feature fusion screening strategy also includes a step of verifying the correlation of acupoint groups; The acupoint group correlation verification step verifies the spatial topological relationship of the candidate points based on the distribution position correlation between the acupoints in the palm area, and eliminates the candidate points whose relative position deviation from the adjacent acupoints exceeds a preset threshold.

[0015] Furthermore, the multi-feature fusion screening strategy also includes a deep verification step; The depth verification step samples the depth values ​​of the screened candidate points, extracts their three-dimensional coordinates and analyzes the depth value distribution, and retains the optimal candidate points that meet the conditions through concave-convex feature recognition and feature matching verification.

[0016] The human meridian and acupoint automatic detection system based on AI machine vision includes: The hand direction determination module determines whether a continuous distribution curve can be extracted from the user's hand image captured by the visual camera, and outputs the palm surface or back of the hand based on the extraction result; A hand face feature extraction module, which divides the hand image into a finger region and a palm region based on the finger root creases, and selects a corresponding feature extraction strategy according to the face of the hand image to extract characteristic landmarks in the hand image; A global reference establishment module selects the midpoint of the wrist crease as the global origin, defines the horizontal and vertical coordinate axes, segments the hand image according to the characteristic landmarks to obtain a refined region, extracts the boundary coordinates of the refined region in the global coordinate system and forms a mapping library; The local adaptation module establishes a local coordinate system for each refined area with its geometric center as the local origin, calculates the offset between the local origin and the global origin, and forms a coordinate transformation so that the local positioning results can be directly mapped to the global space; The candidate point screening module extracts the spatial relationship parameters between the refined area and the feature marks as the initial conditions for subsequent candidate point screening. It also combines the texture intersection features, grayscale gradient features and acupoint group correlation verification to screen out proportional candidate points from the refined area. The optimal candidate point is retained as the target point by performing concave-convex feature matching verification on the candidate points.

[0017] Beneficial effects of the present invention: 1. The full-process automated detection of hand meridian acupoints is realized through AI machine vision technology. No manual intervention is required from hand face direction judgment, feature extraction to candidate point screening, which completely solves the problems of low efficiency and poor consistency caused by traditional technology relying on manual marking. Specifically, the palm face and the back of the hand are accurately distinguished through the hand face direction judgment step, and targeted feature extraction strategies are adopted, such as extracting feature lines through edge detection on the palm face and identifying metacarpal bone vertices and tendon textures through contour detection on the back of the hand. Combined with the collaborative construction of the global benchmark and the local coordinate system, hierarchical segmentation from macroscopic areas to refined areas is realized, ensuring the unified spatial reference for acupoint positioning. At the same time, the candidate point screening link integrates texture intersection features, grayscale gradient features, acupoint group correlation verification and deep concave-convex feature verification, and eliminates interference points in multiple dimensions, so that the positioning error of the optimal acupoint is controlled within a subtle range, which is significantly better than the positioning accuracy of the existing technology and meets the strict requirements of traditional Chinese medicine diagnosis and treatment for acupoint accuracy. 2. Through a multi-level adaptation mechanism, it effectively copes with the influence of individual hand morphological differences (such as palm size, depth of lines, tendon direction variation, etc.) and complex imaging environments. In the global benchmark establishment step, a standardized coordinate system is constructed with the midpoint of the wrist crease as the origin, combined with the boundary coordinate mapping library of feature markers to provide a unified spatial reference for hand images of different individuals; the local adaptation step realizes the precise conversion of local and global coordinates through the calculation of geometric center offset, so that the positioning results of each refined area can be mapped to the global space, avoiding the errors caused by individual hand proportion differences. Positioning deviation. In addition, according to the different physiological characteristics of the palm surface and the back of the hand, feature extraction strategies are designed respectively (such as focusing on palm muscle texture analysis on the palm surface and focusing on tendon and bone landmark recognition on the back of the hand), and the regional segmentation granularity is dynamically adjusted through parameters such as texture intersection density and texture direction to ensure full coverage of various acupoints (such as Laogong point on the palm and Hegu point on the back of the hand). This adaptability not only expands the scope of applicable population of the technology, but also improves its robustness under different lighting and shooting angles, providing reliable technical support for scenarios such as clinical diagnosis and treatment of traditional Chinese medicine and family health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the overall flow chart of the present invention; Figure 2 This is a flow chart of hand face feature extraction in the present invention; Figure 3 This is a flowchart of multiple feature screening of candidate points in the present invention; Figure 4 This is a system module connection diagram of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom," "top," "inner," and "outer" refer to directions toward or away from the geometric center of a particular component, respectively.

[0020] The present invention provides a method and system for automatic detection of human meridian acupoints based on AI machine vision. The specific implementation method is described in detail in conjunction with the figures in the accompanying drawings. The following content will gradually expand from image acquisition and elaborate on the coordination relationship between each step.

[0021] The hand-face direction judgment step includes hand image acquisition and preprocessing and hand-face direction judgment. First, the user's hand image is obtained through a visual camera. The visual camera used can be a high-resolution industrial camera. Its installation position maintains a fixed distance from the user's hand to ensure image clarity. In actual operation, the user places the hand on a fixed platform. A light source is provided under the platform for uniform fill light, thereby reducing the impact of ambient light on image quality. The image is denoised by a Gaussian filter to eliminate the impact of ambient noise. At the same time, a grayscale normalization operation is used to reduce the impact of illumination changes on image quality. The core parameters of the Gaussian filter include kernel size and standard deviation. , the kernel size is usually set to 5×5 pixels, The value is dynamically adjusted according to the image noise level, ranging from 0.8 to 1.2. The purpose of grayscale normalization is to map the brightness range of the image to a fixed interval, such as 0 to 255, to improve the contrast of subsequent processing. At this point, the image quality is significantly optimized, providing a basis for further feature extraction.

[0022] Subsequently, the continuous distribution curve of the main line of the palm is extracted by the Canny operator. If the feature line is detected, it is determined to be the palm surface; if not detected, it can be determined to be the back of the hand. The feature lines include the first feature line, the second feature line and the third feature line. The first feature line is usually the uppermost end of the palm among the three feature lines, and the starting position of the first feature line starts near the intersection of the middle finger and the palm area, specifically in the edge area on the side where the root of the middle finger connects to the palm surface. Its direction and shape are curved as a whole. After starting from the starting point, it gradually extends away from the thumb and ends at the edge position of the palm away from the thumb. The first feature line is an important reference for dividing the macro area of ​​the upper palm; the second feature line is used as the third feature line. The lowest line among the characteristic lines starts from the area between the index finger and the thumb, more specifically, near the connection area between the base of the index finger and the base of the thumb, close to the above the transverse line at the junction of the palm area and the finger area. Its direction and shape extend in a curved posture toward the transverse line of the palm area, which is a key mark for distinguishing the macro area of ​​the lower part of the palm; the relative position of the second characteristic line is between the first and third characteristic lines. Its starting position starts from the edge of the third interphalangeal joint cavity of the index finger, and it may start between the index finger and the thumb. Its direction and shape extend to the bottom of the middle finger, ring finger or little finger. Common ending positions include the midline of the ring finger, below the ring finger and below the little finger. It is an important basis for dividing the macro area of ​​the middle part of the palm.

[0023] For some reason, the feature line was not identified. In order to improve the judgment accuracy, the texture tracking algorithm was switched to extract the extensor tendon texture (the texture tracking algorithm uses the improved SUSAN corner detection combined with the gradient descent method, the local gradient direction calculation window is set to 3×3 pixels, and the intensity change threshold is set to 20-30 grayscale values). The texture tracking algorithm identifies the tendon texture by analyzing the local gradient direction and intensity change, and combines the metacarpal bone landmarks for matching verification. The specific matching process includes calculating the spatial consistency of the texture direction and the bony landmark points. If the consistency is higher than the set threshold (for example, 90%), it is judged to be the back of the hand; this process ensures that the image can be accurately judged as belonging to the palm or back of the hand. During the image acquisition process, the camera shooting angle is calibrated according to the position of the reference point mapped in the hand image to ensure the geometric consistency of the image.

[0024] The hand face feature extraction step includes region segmentation and feature extraction, such as Figure 2As shown, first, during region segmentation, the YOLO or Faster R-CNN target detection algorithm is used to segment the finger and palm regions. The finger base ridges serve as the dividing line to separate the distal finger region from the proximal palm region. When outputting the segmentation results, the boundary coordinates of the finger and palm regions and their area ratios in the entire image are recorded. The finger base ridge position coordinates extracted by the target detection algorithm serve as the dividing line to ensure the accuracy of the region segmentation. In this step, the target detection algorithm model training dataset must contain a large number of clearly labeled hand images to improve segmentation accuracy. In addition, the area ratio of the finger and palm regions is calculated as the ratio of the number of pixels in each region to the total number of pixels in the entire image. This ratio is used in the subsequent feature extraction and benchmark anchoring steps.

[0025] Subsequently, in feature extraction, based on the face selection corresponding strategy of the hand image, the Hough transform is used to extract the spatial distribution of the palm main lines, such as the first feature line, the second feature line, and the third feature line, for the palm surface. The palm area is initially divided into macro regions, which are further divided into refined regions based on the number of texture intersections per square millimeter. The refined regions are then divided into multiple refined subregions based on the difference in texture density. For example, the thenar eminence is divided into the medial thenar subregion and the lateral thenar subregion. For the back of the hand, a contour detection algorithm is used to identify the metacarpal bone vertices and dorsal wrist transverse lines. The bony landmarks include the metacarpal bone vertices and dorsal wrist transverse lines. The metacarpal bone vertices identified by the contour detection algorithm are the distal vertices of the 1st to 5th metacarpal bones, which are the protrusions near the base of the fingers. They are the key positioning marks of the skeletal structure of the dorsal hand. The dorsal wrist transverse line is the line distributed laterally on the back of the wrist. Its connection position with the base of the metacarpal bone is the proximal boundary of the macro region. The macro region of the dorsal hand is initially divided into multiple macro functional areas based on the metacarpal bone vertices and dorsal wrist transverse lines. For example, the area between the 1st and 2nd metacarpal bone vertices and the dorsal wrist transverse line is a macro region.

[0026] Since the macro-region can only roughly frame the functional area where the acupoint is located, but the range is wide and may include non-acupoint areas or interference features, the macro-region is divided into refined areas based on the direction of the extensor tendon texture and bony landmarks, and then the refined area is divided into multiple refined sub-areas. Specifically, the division of the refined area needs to rely on three rules. First, spatial distance association, taking the metacarpal vertex as the starting point, calculate the vertical distance between the extensor tendon texture and the metacarpal vertex; second, angle association, the angle between the direction of the extensor tendon texture and the metacarpal axis needs to meet the preset range; third, texture density association, analyze the data volume of the tendon texture intersection to distinguish the gap area between tendons and the tendon area itself.

[0027] Directly match the specific location of the acupoint to reduce interference in the subsequent screening of candidate points. In addition, the hand morphology of different individuals is different. When dividing the refined sub-regions, the size is not fixed, but dynamically adjusted based on local features. The back of the hand is divided according to the "relative angle between the tendon and the bony landmark", which can adapt to the hand features of different individuals and avoid positioning deviations caused by morphological differences. In addition, the macro region can only provide a rough boundary in the global coordinate system, and the boundary coordinates of the refined sub-region (at least 4 vertices) are stored in the mapping library. The millimeter-level coordinate conversion can be achieved by the offset between the local origin and the global origin, ensuring that the positioning result of each acupoint can be mapped to the global space and retain local fine features, ultimately achieving hierarchical accuracy assurance from "macro positioning to local refinement to global unification"; the output results of the hand feature extraction step include not only the texture distribution information, but also the boundary coordinates and geometric center of each sub-region. These data are passed to the global benchmark step and the local benchmark step; The distribution of the second and third characteristic lines and other main palm lines mentioned above, together with the positional relationship between the metacarpal bone vertices and the extensor tendon texture, form the basis for feature extraction. The key to this step lies in the parameter settings of the Hough transform and contour detection algorithm, such as the Hough transform threshold and the minimum area limit for contour detection. These parameters directly affect the granularity of feature extraction.

[0028] In the global benchmark establishment step, the midpoint of the wrist transverse line is selected as the global origin, the X-axis and Y-axis directions are defined, and the units are unified in millimeters. The boundary coordinates of all refined areas in the global coordinate system are extracted and stored in the mapping library. The definition of the global coordinate system enables the positioning of all areas to have a unified reference, which facilitates subsequent global positioning operations. At the same time, the global benchmark step is also responsible for converting the coordinates of the sub-area data output by the feature extraction step to adapt it to the requirements of the global coordinate system.

[0029] In the local adaptation step, a local coordinate system is established for each refined area with its geometric center as the local origin. The local origin coordinate calculation formula is: Let the coordinates of the vertex on the boundary of the refined area be 、 … , then the horizontal coordinate of the local origin is , vertical coordinate , offset , , calculate the offset between the local origin and the global origin and form the coordinate transformation, the global coordinate ;in, and They represent the offset between the local origin and the global origin in the X-axis and Y-axis directions respectively. At the same time, the spatial relationship parameters between the refined area and the feature mark, including distance parameters and angle parameters, are extracted and stored in the benchmark database so that the local positioning results can be directly mapped to the global space. For example, "the vertical distance between the upper edge of the medial sub-area of ​​the thenar eminence and the third feature line is 3-5mm" and "the angle between the central axis of the central area of ​​the palm and the second feature line is 30°±5°". These parameters provide an accurate positioning basis for subsequent candidate point screening.

[0030] like Figure 3 As shown in the figure, in the candidate point screening step, the texture intersection feature, grayscale gradient feature and acupoint group correlation verification are combined to screen candidate points that meet the anatomical proportions from the refined area. For example, the texture intersection dense area between the first and second metacarpal bones is identified as the candidate point in the area where the Hegu point is located. The top five candidate points are screened by calculating the Euclidean distance between the candidate point and the area boundary and the angle with the feature line. The gradient changes around the candidate points are analyzed, and the Sobel operator is used to extract the gradient information. The points that meet the annular gradient characteristics are retained. The specific quantitative indicators of the annular gradient can be the grayscale value difference range, the radius range of the annular distribution, and the angular span of the gradient change. The anatomical correlation of the acupoint group is used to verify the spatial topological relationship of the candidate points, and the points that do not meet the relative position deviation threshold are eliminated.

[0031] In addition, the candidate point screening step also includes depth information assistance, extracting the 3D coordinates of the candidate point, collecting the depth value distribution of a 5×5 pixel area centered on the candidate point, and calculating the standard deviation of the depth value in the area. Identify flatness differences, extract depth mutation points, and determine whether they are concave or convex edges. For concave acupuncture points, set the conditions that the depth value of the center of the area is less than the average depth value of the surrounding area and the depth mutation points are distributed in a ring-shaped manner; for convex acupuncture points, set the conditions that the depth value of the center of the area is greater than the average depth value of the surrounding area and the depth mutation points are distributed along the direction of the bony landmarks, and finally retain the optimal candidate point; among them, the standard deviation σ of the depth value is calculated as the square root of the variance of the depth values ​​in the area.

[0032] In multi-dimensional verification, after obtaining the optimal candidate point, the benchmark anchoring and candidate point screening are repeated, and the two results are compared. If the fluctuation unit is located in the same hand image, the candidate point is determined to be accurate, otherwise it is re-analyzed. The key to this step lies in the determination method of the fluctuation unit and the comparison rules of the two results. For example, the steep peak and trough values ​​of the fluctuation unit are calculated through trigonometric functions, and the relative position relationship between the acupuncture point and the adjacent acupuncture points is obtained by combining the external system library index. This process ensures the stability and reliability of the candidate points.

[0033] In the acupoint association feedback step, the steep peak and trough values ​​of the fluctuation unit are calculated, and the relative position relationship between the acupoint and the adjacent acupoints is obtained through trigonometric calculation or external system library index. The relative position relationship is compared with the threshold, and monitoring instructions and danger instructions are output. The key to this step lies in the calculation method of the relative position relationship and the setting rules of the threshold. For example, the relative position relationship of adjacent acupoints must meet a specific anatomical ratio, otherwise a danger instruction is output to prompt re-analysis.

[0034] Through the above steps, the present invention realizes multi-level analysis of hand images, gradually refining from the global to the local, and combining multiple features such as palm print texture, extensor tendon texture and bony landmarks to construct a high-precision acupoint positioning system; through multi-dimensional verification of candidate points, including the introduction of depth information and verification of anatomical relevance, the reliability of the positioning results is improved; at the same time, through automated algorithms, comprehensive coverage of diverse acupoints is achieved, solving the problems of low efficiency and insufficient consistency in traditional manual positioning, and meeting the needs of modern Chinese medicine diagnosis and treatment for efficient and accurate acupoint detection.

[0035] Based on the above steps, the present invention provides an embodiment including: Hand orientation determination: 1. Image acquisition and preprocessing: A 16-megapixel visual camera (resolution 4608×3456) is used to capture an image of the user's right palm (palm facing up, fingers naturally extended). Noise is removed using a Gaussian filter (kernel size 5×5, σ=0.9), and grayscale is normalized to the 0-255 range to eliminate lighting interference. 2. Direction Judgment Logic: The Canny edge detection algorithm (high threshold 70, low threshold 35) is used to extract continuous distribution curves in the image. The three main palm lines (the third characteristic line, the second characteristic line, and the first characteristic line) are successfully identified and determined to be the palm surface.

[0036] Hand feature extraction: 1. Region segmentation: Using the finger root crease as the boundary (identified by the Faster R-CNN algorithm, boundary coordinate y = 1100 pixels), the hand image is segmented into the finger area (distal end) and the palm area (proximal end). The palm area accounts for approximately 60% of the area. 2. Palm feature extraction strategy: ① Palm main line extraction: The third, second, and third characteristic lines were extracted using Hough transform (line detection threshold ρ = 1, θ = 1°). Based on the spatial distribution of the three main lines, the palm area was macroscopically segmented into the "thenar area," "palm central area," and "hypothenar area." The boundary coordinates of the "palm central area" where the Laogong acupoint is located were output as follows: left (1200, 900), right (1800, 1000), top (1100, 800), and bottom (1900, 1100).

[0037] ② Texture trend analysis: The palm muscle texture of the "central palm area" is analyzed. The texture trend (mainly longitudinal and radial) is extracted through Hough transform. The number of texture intersections per square millimeter is about 8-12, which provides a basis for subsequent refined segmentation.

[0038] Global datum establishment: 1. Coordinate system definition: Select the midpoint of the wrist crease O (1500, 2000) as the global origin (0, 0), define the horizontal coordinate axis (X axis: pointing to the thumb, horizontally to the right) and the vertical coordinate axis (Y axis: pointing to the fingertips, vertically upward), and convert the unit to millimeters (1 pixel = 0.04 mm); 2. Refined Region Segmentation: Based on the texture orientation (mainly vertical) and the density of texture intersections per square millimeter (10 / mm²), the "central palm area" is segmented into candidate Laogong acupoint regions. The boundary coordinates (in the global coordinate system) are: upper left (1300, 920), upper right (1700, 980), lower left (1350, 1050), and lower right (1650, 1100). These boundary coordinates are stored in the mapping library.

[0039] Local adaptation: 1. Establish a local coordinate system and calculate the geometric center of the refined area (local origin P): Horizontal axis: (1300+1700+1350+1650) / 4 =1500; Vertical coordinate: (920+980+1050+1100) / 4 =1012.5; The global coordinates of the local origin P are (1500, 1012.5); 2. Offset calculation: The offset between the local origin P and the global origin O is Pixel (0mm), Pixel (-39.5mm), the mapping between local positioning results and global space is achieved through the coordinate conversion formula (global coordinate = local coordinate + offset).

[0040] Candidate point screening: 1. Initial conditions: Extract the spatial relationship parameters between the refined area and the second feature line (3-5mm vertical distance from the second feature line, located below the second feature line); 2. Multi-feature fusion screening: ①, Texture intersection feature: Identify 6 intersection points of palmar muscle texture within the region and calculate the Euclidean distance to the region boundary (retain 4 candidate points within the range of 2-4mm); ② Grayscale gradient features: Use the Sobel operator to extract the grayscale gradient of 3×3 pixels around the candidate point, and retain the three candidate points with ring gradient features (the central grayscale value is 10-15 lower than the surrounding grayscale value); ③. Acupoint group correlation verification: Based on the spatial topological relationship between Laogong point and the adjacent Daling point and Zhongchong point (the distance from Daling point is 8-10 mm, and the angle of 60° with Zhongchong point is 60°), one deviation point was eliminated.

[0041] 3. Depth Verification: Depth information of candidate points is collected using a 3D depth camera. Laogong acupoint is a sunken acupoint. Candidate points with a central depth value 0.2-0.4mm lower than the surrounding average depth value and a circular distribution of depth mutation points are selected. Finally, one optimal candidate point is retained.

[0042] 4. Result output: The global coordinates of the optimal candidate point (1520, 1030) are converted to physical coordinates (0.8mm, -38.8mm), with a positioning error of ±0.2mm, which meets the Traditional Chinese Medicine positioning standard for the Laogong acupoint: "in the middle of the palm transverse line, between the second and third metacarpal bones."

[0043] Global coordinate system: The origin is the midpoint O of the wrist crease, with the X-axis pointing rightward (toward the thumb) and the Y-axis pointing upward (toward the fingertip). Feature markers: wrist crease, finger root crease (dividing the finger area / palm area), first feature line, second feature line, and third feature line. Refinement area: The rectangular area between the second and third metacarpal bones in the central area of ​​the palm (the candidate area for the Laogong acupoint). Candidate points and target points: The four initial candidate points and the final Laogong acupoint are marked, and the annular grayscale gradient range is marked.

[0044] In addition, if Figure 4 As shown in the figure, an automatic detection system for human meridian acupoints based on AI machine vision has been designed, including a hand and face direction judgment module, a hand and face feature extraction module, a global benchmark establishment module, a local adaptation module, and a candidate point screening module. Data interaction is achieved between the modules through hardware interfaces and software protocols to ensure the stability and accuracy of the system operation. In actual application scenarios, the user only needs to place the hand in the specified position, and the system can automatically complete the entire process from image acquisition to acupoint confirmation, ensuring fully automated operation from image acquisition to acupoint confirmation. The user can achieve efficient and accurate acupoint detection without manual intervention. Through multi-level processing and multi-feature fusion, the system significantly improves the automation level and positioning accuracy of acupoint detection, meeting the needs of traditional Chinese medicine diagnosis and treatment for efficient and accurate acupoint detection.

[0045] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that improvements and modifications that do not depart from the principles of the present invention are within the scope of protection of the present invention.

Claims

1. An automatic detection method for acupuncture points on the human hand based on AI machine vision, characterized by: The steps include: A hand face direction determination step is performed to determine whether a continuous distribution curve can be extracted from the user's hand image captured by the visual camera, and output the palm face or the back of the hand according to the extraction result; a hand surface feature extraction step, dividing the hand image into a finger area and a palm area with the finger root crease as the boundary, and selecting a corresponding feature extraction strategy according to the surface of the hand image to extract characteristic landmarks in the hand image; A global reference establishment step, selecting the midpoint of the wrist crease as the global origin, defining the horizontal and vertical coordinate axes, segmenting the hand image according to the characteristic marks to obtain a refined region, extracting the boundary coordinates of the refined region in the global coordinate system and forming a mapping library; In the local adaptation step, a local coordinate system is established for each refined area with its geometric center as the local origin. The offset between the local origin and the global origin is calculated and a coordinate transformation is performed to enable the local positioning result to be directly mapped to the global space. In the candidate point screening step, the spatial relationship parameters between the refined area and the feature marks are extracted as the initial conditions for subsequent candidate point screening. The proportionate candidate points are screened out from the refined area in combination with the texture intersection features, grayscale gradient features and acupoint group correlation verification. The optimal candidate point is retained as the target point by performing concave-convex feature matching verification on the candidate points.

2. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 1, characterized in that: When the hand image is of the palm surface, the feature extraction strategy includes using an edge detection algorithm to extract the palm main line, which reflects three feature lines. The palm area is divided into several macro areas according to the spatial distribution of the palm main line, and the boundary coordinates are output. The macro areas are then subjected to palm muscle texture analysis, and the texture direction is extracted through Hough transform. The texture direction includes longitudinal, transverse and radial directions. The preset palm acupoints are mapped one by one to each macro area to obtain acupoint location information.

3. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 2, characterized in that: When the hand image is of the back of the hand, the feature extraction strategy includes using a contour detection algorithm to identify the metacarpal bone vertices and the dorsal wrist creases, analyzing the local gradient direction and intensity changes of the tendon through a texture tracking algorithm to extract the texture direction of the extensor tendon, and mapping the preset dorsal hand acupoints one by one to each macro region to obtain acupoint location information.

4. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 3 is characterized by: The global benchmark establishment step includes a refined region segmentation strategy, and the refined region segmentation strategy includes a palm face segmentation step and a back hand segmentation step. The palm surface segmentation step further divides the macro region into fine regions by combining the texture arrangement and the number of texture intersections per square millimeter; The step of segmenting the back of the hand divides the metacarpal bone apex and the dorsal wrist crease into fine regions by combining tendon directions and bony landmarks.

5. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 4, characterized in that: The global benchmark establishment step includes a benchmark establishment strategy, which includes a coordinate axis construction step and a boundary coordinate determination step. In the coordinate axis construction step, the midpoint of the wrist transverse crease pointing toward the thumb is defined as the horizontal coordinate axis, and the midpoint of the wrist transverse crease pointing toward the fingertips is defined as the vertical coordinate axis; The boundary coordinate determination step is to identify the contour edges of each refined area through an edge detection algorithm, and obtain the coordinate values ​​of at least four vertices as boundary coordinates.

6. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 5, characterized in that: In the local adaptation step, the average of the horizontal coordinates of all boundary vertices in the refined area is used as the horizontal coordinate of the local origin, and the average of the vertical coordinates of all boundary vertices in the refined area is used as the vertical coordinate of the local origin. The offset includes a horizontal offset and a vertical offset.

7. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to any one of claims 1 to 6, characterized in that: The spatial relationship parameters include distance parameters and angle parameters. The candidate point screening step also includes a multi-feature fusion screening strategy, which includes a texture cross-matching step and a grayscale gradient analysis step. The pattern cross-matching step identifies the intersection points, turning points and dense texture clusters of the palm print texture in the refined area, calculates the Euclidean distance between the candidate point and the area boundary and the angle with the feature line, and selects the top 5 candidate points that meet the preset ratio. The grayscale gradient analysis step extracts gradient changes around candidate points in the hand image and retains candidate points with annular gradient features.

8. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 7, characterized in that: The multi-feature fusion screening strategy also includes an acupoint group correlation verification step; The acupoint group correlation verification step verifies the spatial topological relationship of the candidate points based on the distribution position correlation between the acupoints in the palm area, and eliminates the candidate points whose relative position deviation from the adjacent acupoints exceeds a preset threshold.

9. The method for automatically detecting acupuncture points on the human hand based on AI machine vision according to claim 8, characterized in that: The multi-feature fusion screening strategy also includes a depth verification step; The depth verification step samples the depth values ​​of the screened candidate points, extracts their three-dimensional coordinates and analyzes the depth value distribution, and retains the optimal candidate points that meet the conditions through concave-convex feature recognition and feature matching verification.

10. The automatic detection system for human hand meridian acupoints based on AI machine vision is characterized by: include: The hand direction determination module determines whether a continuous distribution curve can be extracted from the user's hand image captured by the visual camera, and outputs the palm surface or back of the hand based on the extraction result; A hand face feature extraction module, which divides the hand image into a finger region and a palm region based on the finger root creases, and selects a corresponding feature extraction strategy according to the face of the hand image to extract characteristic landmarks in the hand image; A global reference establishment module selects the midpoint of the wrist crease as the global origin, defines the horizontal and vertical coordinate axes, segments the hand image according to the characteristic landmarks to obtain a refined region, extracts the boundary coordinates of the refined region in the global coordinate system and forms a mapping library; The local adaptation module establishes a local coordinate system for each refined area with its geometric center as the local origin, calculates the offset between the local origin and the global origin, and forms a coordinate transformation so that the local positioning results can be directly mapped to the global space; The candidate point screening module extracts the spatial relationship parameters between the refined area and the feature marks as the initial conditions for subsequent candidate point screening. It also combines the texture intersection features, grayscale gradient features and acupoint group correlation verification to screen out proportional candidate points from the refined area. The optimal candidate point is retained as the target point by performing concave-convex feature matching verification on the candidate points.

Images