Acupuncture point positioning method and system based on artificial intelligence

Abstract

The invention relates to the technical field of image sensing, general image recognition and feature positioning under computer vision, and discloses an acupuncture point positioning method and system based on artificial intelligence, and the method comprises the steps: obtaining a non-rigid target section image through an image collection module; the key point identification module determines topological anchor points at two ends of the section; the reference construction module establishes a structure reference vector; a deflection calculation module extracts surface texture feature points, obtains a high-frequency component and a low-frequency component through multi-scale frequency domain decomposition, and determines a spatial deflection parameter based on the gradient of the energy distribution ratio of the high-frequency component and the low-frequency component; the proportion compensation module corrects a target distribution proportion coefficient; according to the method, the pose is inversed by using the scale invariance between the frequency domain components, the decoupling of the anatomical positioning logic and the pixel space is realized, and the non-linear interference of the non-uniformity of the biological surface and the dynamic environment illumination on the positioning precision is eliminated.

Description

TECHNICAL FIELD

[0001] The application relates to an artificial intelligence-based acupuncture point positioning method and system, and belongs to the technical field of image sensing, general image recognition and feature positioning under computer vision. BACKGROUND

[0002] In the current image recognition field, a two-dimensional pixel coordinate regression model is used to determine a target point. However, when dealing with a non-rigid target with complex biological characteristics, the traditional Chinese medicine positioning follows the proportional bone measurement method, which is based on the relative proportion system of human skeletal joint weights. The positioning point is required to maintain a geometric mapping relationship between the image plane and the anatomical space. This is essentially a complex image space geometric transformation and feature point reconstruction process. During the two-dimensional imaging process, geometric projection information is lost. When the imaging device and the limb plane have a deflection angle, the pixel distance of the image plane is nonlinearly changed due to the perspective shortening effect, resulting in physical displacement deviation of the fixed proportional positioning logic in the pixel space. That is, during the projection mapping process of the image data of the visual sensor, the scalar scale factor is nonlinearly distorted due to the perspective transformation.

[0003] The conventional scheme introduces a three-dimensional depth sensor or constructs a high-dimensional human body grid model to supplement spatial information, which increases the hardware deployment cost and data processing load. There are limitations in the software control level. For example, the utility model patent with the authorization publication number CN201618114U discloses an acupuncture point positioning treatment device, which uses a conical head to stimulate the skin and cooperates with a plastic tube to guide the needle insertion. This scheme belongs to mechanical physical assistance, and the positioning logic depends on the experience of the operator. It cannot cope with individual differences and is difficult to solve the imaging distortion positioning deviation at the visual sensing level. The hardware improvement scheme exposes the lag of automatic algorithm and cannot adaptively optimize the imaging software control logic.

[0004] Therefore, how to construct a general image geometric transformation model with projection consistency compensation capability and general image feature point distribution vector, use the local feature components captured by the image sensing module to perceive the image space pose of the non-rigid target, perform high-precision image feature recognition and inversion, eliminate the nonlinear interference of the visual angle deviation and non-rigid deformation on the pixel positioning in the monocular image sensing environment, and realize high-precision image data pixel coordinate mapping and positioning output, become the technical problems to be solved by the application at the image recognition algorithm level. SUMMARY

[0005] To solve the problems in the background art, the technical scheme of the application is as follows: an artificial intelligence-based acupuncture point positioning system, comprising: an image acquisition module, a key point recognition module, a reference construction module, a deflection calculation module, a proportional compensation module and a position output module.

[0006] The image acquisition module is configured to acquire image data containing a non-rigid target segment;

[0007] The key point identification module is configured to identify a first topological anchor point and a second topological anchor point of the non-rigid target segment in the image data.

[0008] The reference construction module is configured to establish a structural reference vector with the first topological anchor point as a starting point and the second topological anchor point as an ending point.

[0009] The deflection calculation module is configured to extract surface texture feature points in a region covered by the structural reference vector, perform multi-scale frequency domain decomposition on a local image of the non-rigid target segment, extract a first feature component corresponding to a high-frequency response component and a second feature component corresponding to a low-frequency response component, respectively, calculate an energy distribution ratio between the first feature component and the second feature component, and determine a spatial deflection parameter of the non-rigid target segment relative to an imaging plane according to a gradient slope of the energy distribution ratio in the direction of the structural reference vector.

[0010] The scale compensation module is configured to perform projection consistency compensation on a preset target distribution scale factor based on the spatial deflection parameter, and determine a corrected scale factor.

[0011] The position output module is configured to perform coordinate mapping on the structural reference vector according to the corrected scale factor, and output pixel coordinates of the feature point to be positioned in the image data after superimposing a preset orthogonal offset.

[0012] Preferably, when determining the spatial deflection parameter, the deflection calculation module is specifically configured to calculate anisotropic components of the surface texture feature points in the region covered by the structural reference vector, determine a local covariance matrix constructed by the surface texture feature points, extract a feature vector principal axis direction distribution rule and an envelope ellipse deformation trend of the local covariance matrix, invert an axial rotation component of the non-rigid target segment relative to the imaging plane using the anisotropic components, and perform spatial dimension compensation on the corrected scale factor using the axial rotation component; extract a distribution residual between the image data and a projection invariance model, and when the distribution residual is outside a preset safety interval, stop the projection consistency compensation, and determine the pixel coordinates of the feature point to be positioned using a preset geometric distribution parameter.

[0013] Preferably, the scale compensation module is further configured to calculate a spatial distribution variance of the surface texture feature points in the region covered by the structural reference vector, generate a local strain compensation operator of the non-rigid target segment according to a dispersion feature of the spatial distribution variance, and perform nonlinear correction on the target distribution scale factor using the local strain compensation operator.

[0014] Preferably, the reference construction module is further configured to obtain a pixel length of the reference topological vector fully displayed in the image data when the first topological anchor point or the second topological anchor point is blocked; determine a logical length proportion relationship between the reference topological vector and the non-rigid target segment according to a preset object structure characteristic parameter; and reconstruct a structure reference vector by using the logical length proportion relationship, the pixel length of the reference topological vector, and a pixel position of the fully displayed first topological anchor point or second topological anchor point.

[0015] Preferably, the reference construction module is further configured to calculate a feature principal axis included angle between the first topological anchor point and the second topological anchor point, determine a flexion curvature of the non-rigid target segment, perform a tangent integral mapping on the structure reference vector based on the flexion curvature to generate a non-linear geometric trajectory matching a shape of the non-rigid target segment, and replace the structure reference vector with the non-linear geometric trajectory when determining the pixel coordinates in the image data.

[0016] Preferably, the method further comprises a dynamic reinforcement module configured to obtain a motion residual vector of a surface texture feature point between adjacent image frames, perform an anisotropic Gaussian filtering on the image data by using the motion residual vector to suppress feature diffusion of the non-rigid target segment in a motion direction, and update the correction scale factor according to the filtered image data.

[0017] Preferably, the deflection calculation module is configured to perform a differential frequency domain filtering on the local image to obtain the first feature response component and the second feature response component corresponding to different spatial frequencies when extracting the first feature component and the second feature component, and to count a dot array energy distribution of the first feature response component and the second feature response component.

[0018] Preferably, the deflection calculation module is further configured to count a deviation value of an energy distribution ratio of adjacent pixel points in a quadrature sampling window of the structure reference vector, identify an unstable feature region whose deviation value exceeds a preset deviation threshold, and eliminate a data component corresponding to the unstable feature region from the energy distribution ratio.

[0019] Preferably, the deflection calculation module is configured to calculate a texture anisotropy index when determining the axial rotation component, the texture anisotropy index satisfies the following relationship: wherein, is a first eigenvalue of a local covariance matrix, is a second eigenvalue of the local covariance matrix.

[0020] A needle and acupuncture point positioning method based on artificial intelligence, comprising the following steps:

[0021] Step S101, acquire image data containing a non-rigid target segment;

[0022] Step S102, identify the first and second topological anchor points of the non-rigid target segment in the image data;

[0023] Step S103, establish a structure reference vector with the first topological anchor point as the starting point and the second topological anchor point as the ending point;

[0024] Step S104, extract surface texture feature points within the coverage area of the structure reference vector, perform multi-scale frequency domain decomposition on the local image of the non-rigid target segment, extract the first feature component corresponding to the high-frequency response component and the second feature component corresponding to the low-frequency response component, respectively, calculate the energy distribution ratio between the first feature component and the second feature component, and determine the spatial deflection parameter of the non-rigid target segment relative to the imaging plane according to the gradient slope of the energy distribution ratio in the direction of the structure reference vector;

[0025] Step S105, perform projection consistency compensation on the preset target distribution proportion coefficient based on the spatial deflection parameter, and determine the correction scale factor;

[0026] Step S106, perform coordinate mapping on the structure reference vector according to the correction scale factor, and output the pixel coordinates of the to-be-positioned feature points in the image data after superimposing the preset orthogonal offset.

[0027] Compared with the prior art, the beneficial effects of the present application are:

[0028] 1. In the present image processing system, projection consistency compensation is performed by establishing an anatomical reference vector and using local image feature density gradient, the perspective shortening effect of the image plane is converted into a linear mapping process within the image algorithm scale, the anatomical positioning logic is decoupled from the pixel coordinate space and the coordinate field is adaptively corrected, the deflection of the target object in the image space is perceived through local feature density, the distortion of the positioning scale caused by posture twisting or view angle change is avoided, and the image feature point calibration is stable under various imaging angles and image sensing environments.

[0029] 2, analyze the spatial distribution variance of the texture feature points in the coverage area of the anatomical reference vector, identify the gathered and sparse features caused by skin surface muscle contraction or external force pulling, construct a local strain compensation to nonlinearly correct the proportion coefficient of the bone degree, make the positioning scale stretch and shrink with the difference of the local tension state of the skin surface, eliminate the non-rigid displacement of the skin surface relative to the deep skeletal anchor point, solve the positioning drift caused by body position fine-tuning in clinical dynamic operation; use multi-scale decomposition to extract high-frequency texture components and low-frequency tissue components of the limb segment, establish a parallax perception model based on the energy distribution ratio between the frequency components, use the scale invariance of the energy ratio projection process to invert the posture of the limb depth dimension space, perceive the posture through the geometric correlation between scale information, get rid of the dependence on the absolute number of feature points, keep the positioning accuracy stable in uneven lighting or skin feature deficient environment, and solve the feature gradient misjudgment caused by environmental noise interference.

[0030] 3, extract the affine anisotropic features of the texture feature point cloud, analyze the distribution law and envelope ellipse deformation trend of the principal axis direction of the local feature operator, invert the axial rotation angle of the limb segment relative to the imaging plane, perceive the deep rolling state of the limb under monocular vision, compound the space transformation of the rotation component and perspective deflection parameter, synthesize nonlinear compensation for the acupoint coordinates, solve the acupoint overturning offset caused by limb rotation, and improve the three-dimensional mapping ability of the system in non-standard body position; introduce anatomical mirror constraint, use the statistical consistency of human limb proportion, use the reference limb vector and the logical length proportion relationship of the target segment to reconstruct the missing anatomical anchor point and close the reference vector when the joint key point is blocked, convert the anatomical structure redundancy into a visual space constraint solving path, and make the positioning function remain coherent when some key anatomical markers are missing. BRIEF DESCRIPTION OF DRAWINGS

[0031] Fig. 1 The flow chart of the acupuncture acupoint positioning method of the present application is consistent with the multi-scale frequency domain decomposition and projection compensation.

[0032] Fig. 2 The positioning deviation numerical value comparison curve graph of the test group and the control group of the present application changes with the space deflection parameter.

[0033] Fig. 3 The functional architecture diagram of the acupuncture acupoint positioning system of the present application integrates the deflection calculation and dynamic reinforcement modules. DETAILED DESCRIPTION

[0034] The following disclosure is used to explain and illustrate the present application, and is not used to limit the protection scope of the present application. The technical features in each embodiment and example can be combined with each other without conflict.

[0035] This invention provides an artificial intelligence-based acupuncture point location method and system, including an image acquisition module, a key point recognition module, a reference construction module, a deflection calculation module, a scaling compensation module, and a position output module. The image acquisition module, as an image sensing unit, is used to acquire image data containing non-rigid target segments. The recognition module is used to identify key feature points of the non-rigid target segments in the image data, i.e., the first topological anchor points, through an image recognition process. With the second topological anchor point This process enables nonlinear feature extraction and analysis from the original pixel point cloud to the anatomical topological space; the benchmark construction module is used to establish a topological anchor point. Starting from the second topological anchor point The structural reference vector at the endpoint The deflection calculation module is used to extract the structural reference vector. Surface texture feature points within the coverage area are used to perform multi-scale frequency domain decomposition on the local image of the non-rigid target segment, and the first feature components corresponding to the high-frequency response components are extracted respectively. and the second characteristic component corresponding to the low-frequency response component. Calculate the first eigencomponent With the second characteristic component The ratio of energy distribution between And based on the energy distribution ratio In structural reference vector The gradient slope in the direction determines the spatial deflection parameters of the non-rigid target segment relative to the imaging plane. The proportional compensation module is used based on spatial deflection parameters. For the preset target distribution ratio coefficient Perform projection consistency compensation to determine the correction scale factor; the position output module is used to determine the structural reference vector based on the correction scale factor. Coordinate mapping is performed on the top, and a preset orthogonal offset is superimposed. Then, the pixel coordinates of the feature point to be located in the image data are output.

[0036] In the pose sensing procedure for non-rigid target sections, the deflection calculation module determines the spatial deflection parameters through the following steps. The processor performs multi-scale frequency domain decomposition on local images of limb segments to extract high-frequency response components that characterize dermal texture details. The second characteristic component corresponding to the low-frequency response component characterizing tissue morphology ; in the structural reference vector Within the orthogonal sampling window, calculate and The ratio of energy distribution between Energy distribution ratio satisfies the following relationship: wherein, is the first feature component corresponding energy response value, is the second feature component corresponding energy response value, the system performs a log-Gabor filter based multi-scale decomposition with a center frequency of cycles per pixel and a bandwidth of octaves, low frequency response component is the standard deviation is high frequency response component is the difference between the original image and the low frequency response component energy distribution ratio is calculated within a pixel window centered at the sampling point, the first feature component energy response value and the second feature component energy response value are summed and divided by the square of their amplitudes to obtain a dimensionless value reflecting the local surface microscopic detail density, which serves as the input parameter for spatial pose inversion at the bottom layer, the system compares the energy ratio change rate at the sampling point and the first topological anchor point neighborhood, and uses the scale invariance of the energy ratio in the projection process to map the change rate to the rotation vector of the non-rigid target segment in the depth dimension of the image plane, thereby determining the spatial deflection parameter ; to suppress the interference caused by ambient light, the deflection calculation module calculates the deviation value of the energy distribution ratio of adjacent pixel points within the orthogonal sampling window of the structure reference vector , identifies unstable feature regions with a deviation value exceeding a preset deviation threshold, and removes the data components corresponding to the unstable feature regions from the energy distribution ratio .

[0037] spatial deflection parameter The pre-calibration based on the determination of the geometric projection model, by shooting the standard skin texture sample with known tilt angle, recording the energy distribution ratio decay rate at different angles and fitting to generate a feature mapping curve, the system obtains the energy distribution ratio inside the coverage area of the structure reference vector gradient slope along the vector direction, input the real-time angle between the imaging plane and the limb plane output by the feature mapping curve, and adjust the target distribution proportion coefficient by calculating the spatial deflection parameter trigonometric function projection value Mapping weights, determining a correction scale factor for coordinate mapping, eliminating pixel stepping errors caused by imaging perspective changes; for the rotation of the limb axis, the deflection calculation module extracts the structure reference vector Anisotropy components of surface texture feature points in the coverage area, determine the local covariance matrix constructed by the surface texture feature points, extract the feature vector principal axis direction distribution rule and envelope ellipse deformation trend of the local covariance matrix; the deflection calculation module involves the calculation of texture anisotropy index The texture anisotropy index Complies with the following relationship: Wherein, is the first eigenvalue of the local covariance matrix, is the second eigenvalue of the local covariance matrix, the system uses the anisotropy component determined by the texture anisotropy index Calculate the axial rotation component of the non-rigid target section relative to the imaging plane And use the axial rotation component To perform spatial dimension compensation on the correction scale factor; at the same time, the position output module extracts the distribution residual between the image data and the projection invariance model When the distribution residual Is outside the preset safe interval, the projection consistency compensation is stopped, and the pixel coordinates of the to-be-positioned feature points are determined using the preset geometric distribution parameters, and the distribution residual And the deviation degree of the preset safe interval, according to the deviation degree, output the positioning confidence index of the corresponding to-be-positioned feature point; when processing the non-rigid deformation of the skin surface, the proportion compensation module calculates the spatial distribution variance of the surface texture feature points in the coverage area of the structure reference vector According to the spatial distribution variance The discrete degree characteristics generate the local strain compensation operator of the corresponding non-rigid target section, and use the local strain compensation operator to nonlinearly correct the target distribution proportion coefficient .

[0038] When the first topological anchor point Or the second topological anchor point There is visual occlusion, the reference topological vector pixel length Completely displayed in the image data, according to the preset object structure feature parameter to determine the logical length proportion relationship between the reference topological vector and the non-rigid target section , using the logical length proportion relationship , the pixel length of the reference topological vector And the first topological anchor point Or the second topological anchor point ​the predicted length of the non-rigid target segment is calculated according to the following formula : wherein, is the reconstruction structure reference vector the pixel length required, based on the predicted length , the system reconstructs the structure reference vector ; if the non-rigid target segment is in a curved state, the reference construction module calculates the feature principal axis angle between the first topological anchor point and the second topological anchor point , to determine the flexion curvature of the non-rigid target segment ; based on the flexion curvature , the system performs a tangent integral mapping on the structure reference vector to generate a non-linear geometric trajectory matching the morphology of the non-rigid target segment; when determining the pixel coordinates in the image data, the position output module replaces the structure reference vector with the non-linear geometric trajectory ; the system further comprises a dynamic reinforcement module for obtaining the motion residual vector of the surface texture feature points between adjacent image frames, performing anisotropic Gaussian filtering on the image data using the motion residual vector to suppress the feature diffusion of the non-rigid target segment in the motion direction, and updating the correction scale factor according to the filtered image data; during the output of the pixel coordinates of the feature points to be positioned, the system processes the positioning results between consecutive frames through time domain smoothing filtering; all image data is processed by the processor to perform a local de-identification procedure before processing, only the feature point cloud coordinates and frequency energy distribution parameters used for calculation are retained, and the original biometric image is not stored, so as to ensure user data security.

[0039] Embodiment 1: In a home rehabilitation scene where the ambient light intensity is lower than 50 lux and the positioning object has skin relaxation and limb tremor, the contrast of the skin texture of the subject's forearm is reduced, and the anatomical landmark points are displaced irregularly with the soft tissue; the pixel compression caused by limb rotation and the feature clustering caused by skin wrinkles have similar morphological characteristics in two-dimensional space, and the traditional positioning method cannot distinguish between the above two physical deformations, resulting in a nonlinear displacement deviation of more than 5.0 mm of the positioning scale on the forearm image plane; for the above working conditions, the system collects forearm image data, determines the first topological anchor point and the second topological anchor point , and establishes the structure reference vector ; the processor performs multi-scale frequency domain decomposition on the local image of the forearm, calculates the energy distribution ratio between the first feature component representing the skin texture details and the second feature component representing the tissue morphology It conforms to the following formula: ,in, The first characteristic component The energy response value, The second characteristic component The energy response value, since the evolution trend of the frequency domain energy components during projection is only controlled by geometric perspective, is obtained by analyzing the energy distribution ratio. Along the structural reference vector The gradient slope is used to determine the spatial deflection parameters of the forearm relative to the imaging plane. This process utilizes the scale invariance of the frequency domain response to counteract the interference of low-light environments on texture extraction, providing anatomically stable input parameters for subsequent scaling compensation.

[0040] To handle feature flipping caused by limb axial rotation, the system uses the structural reference vector. Construct a local covariance matrix within the orthogonal region and extract the first eigenvalue. With the second eigenvalue Calculate the texture anisotropy index Texture anisotropy index It conforms to the following formula: ,in, The first eigenvalue of the local covariance matrix. The second eigenvalue of the local covariance matrix is ​​used by the system to utilize the texture anisotropy index. The determined directional distribution pattern determines the axial rotation component, and at the same time, the structural reference vector is extracted. Spatial distribution variance within the coverage area To identify the local strain state of the skin surface; by using spatial deflection parameters With respect to the variance of spatial distribution The system couples the determined strain data to a preset target distribution scaling factor. The system performs corrections to generate a correction scale factor; finally, it maps the target position along the corrected path on the image plane and superimposes an orthogonal offset. Output the pixel coordinates of the acupoint; under this condition, the positioning deviation of the target acupoint is maintained within 0.5mm, and the calculation time is less than 30ms on the embedded processor.

[0041] Example 2: The test platform is equipped with an image acquisition unit with a resolution of 1920x1080 pixels, with a pixel clock frequency set to 74.25MHz, meeting the real-time sampling requirement of 30fps, and the illumination is controlled in the range of 10lux to 300lux by adjusting the LED light source array; To simulate the photoelectric noise interference in the home environment, the image processor superimposes additive white Gaussian noise with a signal-to-noise ratio of 25dB to the brightness channel after obtaining the original pixel array; The test data source is selected from 50 groups of limb images of subjects with different skin tension characteristics, and the sampling window size is selected as 32x32 pixels. The parameter setting takes into account the balance between the completeness of the feature point statistical sample size and the local affine transformation calculation amount; In the test process of verifying the effectiveness of spatial pose perception, the system identifies the first topological anchor point of the subject's forearm and the second topological anchor point , establishes the structure reference vector , and calibrates the real anatomical position of the acupoint as the reference point; The test group uses the projection consistency compensation method, which includes calculating the energy distribution ratio and the texture anisotropy index , and correcting the target distribution proportion coefficient accordingly; Group A removes the deflection calculation module and uses a fixed proportion coefficient for linear mapping on the structure reference vector ; Group B removes the proportion compensation module and does not perform nonlinear correction for skin strain, and the measurement value is obtained by calculating the Euclidean distance between the pixel coordinates and the calibrated reference point.

[0042] Table 1: Positioning deviation comparison data table of different spatial deflection parameters of each group

[0043]

[0044] Spatial deflection parameters are the angles between the limb plane and the image plane, in degrees; Energy distribution ratio is the energy ratio of the first and second characteristic components, and texture anisotropy index is the ratio of the first and second eigenvalues of the local covariance matrix, see Table 1. When the spatial deflection parameter increases within the range of 0° to 60°, the positioning deviation of Group A increases nonlinearly with the perspective shortening effect, while the test group uses the correction scale factor constructed by the energy distribution ratio and the texture anisotropy index to compensate for the pixel compression caused by geometric projection, and to stabilize the deviation within 0.5mm; When the spatial deflection parameter The positioning deviation appears an inflection point and increases to 1.65 mm when reaching 75°; by comparing the test group with the control group B with partial missing features, simply relying on pose inversion and ignoring skin local strain correction will cause the positioning result to deviate with limb muscle contraction; the test group introduces the spatial distribution variance The local strain compensation operator is generated to realize decoupling and cooperation of deep anatomical coordinates and surface skin deformation, all image data are subjected to local de-identification procedures by the processor before processing, only feature point cloud coordinates and frequency energy distribution parameters used for calculation are reserved, and the original biometric image is not stored.

[0045] Embodiment 3: This embodiment combines Figs. 1 to 3 , a kind of artificial intelligence-based acupuncture point positioning method and system are described, as shown in Fig. 1 , image acquisition module is used to obtain non-rigid target section image, its data flow is to key point identification module to determine the topological anchor point at both ends of section, reference construction module establishes structure reference vector, deflection calculation module determines spatial deflection parameter using multi-scale frequency domain decomposition, scale compensation module performs projection consistency compensation and determines correction scale factor, position output module outputs feature point pixel coordinates after superimposing orthogonal offset, and finally obtains the pixel coordinates of the feature point to be positioned.

[0046] As shown in Fig. 2 , the coordinate system shows the numerical relationship between the positioning deviation mm of the longitudinal axis and the spatial deflection parameter of the horizontal axis, the figure contains three broken lines respectively representing the change trend of the deviation of the test group, the deviation of the control group A and the deviation of the control group B, as the spatial deflection parameter increases from 0 to 75, the deviation of the test group maintains at a low level, while the deviation of the control group A shows an upward trend, and the deviation of the control group B shows a moderate upward trend; as shown in Fig. 3 , the system architecture associates multiple functional modules to the core processing logic, the left side is configured with an image acquisition module for acquiring a non-rigid target section, a reference construction module for establishing a structure reference vector, a scale compensation module for performing projection consistency compensation and determining a correction scale factor, and a dynamic reinforcement module for suppressing motion feature diffusion, the right side is configured with a key point identification module for identifying first and second topological anchor points, a deflection calculation module for performing multi-scale frequency domain decomposition and extracting high and low frequency components to calculate energy distribution ratio and determine spatial deflection parameter, and a position output module for performing coordinate mapping and superimposing orthogonal offset to output pixel coordinates, each module works cooperatively to finally realize high-precision acupuncture point positioning output.

[0047] Embodiment 4: In the initialization calibration scene of acupuncture point positioning algorithm, the individual differences of physiological characteristics of the skin of the subject cause the energy distribution ratio To address the computational bias, the system executes a frequency domain decomposition procedure based on image adaptive guided filtering, and utilizes the processor to perform continuous Gaussian smoothing on the original pixel array to construct the first-scale Gaussian kernel standard deviation. Low-pass component set to 1.2 And based on the original image and low-pass components The difference extraction characterizes the first feature component representing the details of the skin texture. Preset target distribution ratio coefficient It is necessary to map from anatomical bone size to image pixel space, based on the structural reference vector. Execute adaptive pixel stepping, target distribution scaling factor Based on the proportional position of the feature point to be located on the anatomical axis, the projection step distance of the feature point on the image plane is determined by the following formula. : ,in, The feature point to be located relative to the first topological anchor point The pixel displacement, in pixels; Structural reference vector The pixel modulus in the current image.

[0048] To address the deviation caused by limb axial rotation, the system utilizes the texture anisotropy index. With axial rotation component The mapping relationship between them performs spatial compensation, in orthogonal offset Introducing a rotation factor into the computational logic This allows the orthogonal displacement components to adjust their projection direction on the image plane in real time as the limb tilts. The pixel coordinates of the output feature points to be located have a coincidence of no less than 98% with the anatomical center point. Furthermore, the image data undergoes local de-identification processing before entering the algorithm core, extracting only abstract feature parameters without storing the original biological feature images.

[0049] Example 5: In the field deployment scenario for heterogeneous camera hardware, the quantization noise component of the central region is extracted by acquiring a reference calibration image with a preset grayscale gradient. This component is then set as a soft threshold bias term in the multi-scale frequency domain decomposition process. Simultaneously, the surface texture distribution benchmark value of a specific region of the subject's limb is obtained while the subject's limb is in a relaxed state. This procedure transforms the differences in pixel response from different cameras and the initial dermal features of the subject into known boundary conditions for the algorithm input, thereby adjusting the energy distribution ratio under different hardware environments. When the limbs are at the same angle of deflection Maintain a consistent characteristic response.

[0050] and the working condition that the relative displacement between the surface skin and the deep anatomical axis is caused by the limb contraction, the system performs a controlled traction experiment to calibrate the skin strain law, by measuring the corresponding texture anisotropy component change under the preset anatomical reduction gradient, and according to the real-time measured spatial distribution variance and the reference value determined deviation ratio, the local strain compensation operator is determined according to the following formula : , wherein is the local strain compensation operator, is a proportional constant representing the skin strain characteristics, the system dynamically adjusts the step weight of the preset target distribution proportion coefficient using the local strain compensation operator , so that the correction scale factor expands and contracts with the soft tissue deformation, and the final determined offset of the acupoint pixel coordinates in the muscle continuous contraction process is not higher than 0.3mm.

[0051] Embodiment 6: In the application scene of mobile terminals equipped with different sensor noise floors and lens radial distortion distribution characteristics, in order to establish the calculation reference of the energy distribution ratio , the system performs a pre-calibration procedure based on sensor feature normalization, uses the image acquisition module to obtain a gray reference image containing a preset contrast distribution, calculates the pixel response non-uniformity deviation in the current imaging environment, and adds the deviation as a compensation amount to the energy response calculation logic of the first feature component , by adjusting the normal angle of the subject's forearm relative to the imaging plane, the system performs offline data sampling in the 0-60 degree interval with a step of 5 degrees, records the attenuation rate curve of the energy distribution ratio under different inclination angles, and uses the least squares method to fit to obtain the feature mapping coefficient characterizing the geometric projection relationship.

[0052] In view of the edge projection distortion challenge caused by the subject's limb deviating from the lens central axis, the system performs a coordinate field calibration procedure based on a radial distortion model, identifies the pixel distribution area of the structural reference vector on the image plane, obtains the local magnification component of the lens corresponding to the area, and the system performs nonlinear weighting on the preset target distribution proportion coefficient according to the local magnification component to determine the corrected displacement distance of the feature point to be positioned in the edge distortion field , and the corrected displacement distance meets the following relationship: , wherein is the structural reference vector pixel module length, and the system utilizes a correction displacement distance Compensate for scaling errors caused by optical system spherical aberration, and combine texture anisotropy index to axial rotation components Perform resampling check, automatically adjust orthogonal offset when the limb is in the image edge area mapping weight.

[0053] It is apparent for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An acupuncture point location system based on artificial intelligence, characterized in that, It includes an image acquisition module, a key point recognition module, a baseline construction module, a deflection calculation module, a scaling compensation module, and a position output module. The image acquisition module is used to acquire image data containing non-rigid target sections; The key point recognition module is used to identify the first and second topological anchor points of non-rigid target segments in image data; The baseline construction module is used to establish a structural baseline vector starting from the first topological anchor point and ending at the second topological anchor point; The deflection calculation module is used to extract surface texture feature points within the area covered by the structural reference vector, perform multi-scale frequency domain decomposition on the local image of the non-rigid target segment, extract the first feature component corresponding to the high frequency response component and the second feature component corresponding to the low frequency response component, calculate the energy distribution ratio between the first feature component and the second feature component, and determine the spatial deflection parameters of the non-rigid target segment relative to the imaging plane based on the gradient slope of the energy distribution ratio in the direction of the structural reference vector. The scaling compensation module is used to perform projection consistency compensation on the preset target distribution scaling coefficient based on the spatial deflection parameters, and to determine the correction scale factor; The position output module is used to perform coordinate mapping on the structural reference vector according to the correction scale factor, and after superimposing a preset orthogonal offset, output the pixel coordinates of the feature point to be located in the image data.

2. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, When determining the spatial deflection parameters, the deflection calculation module is specifically used to calculate the anisotropic components of surface texture feature points within the area covered by the structural reference vector, determine the local covariance matrix constructed from the surface texture feature points, and extract the distribution law of the principal axis direction of the eigenvector of the local covariance matrix and the deformation trend of the envelope ellipse. The axial rotation component of the non-rigid target segment relative to the imaging plane is inverted using the anisotropic component, and the spatial dimension compensation is performed on the correction scale factor using the axial rotation component; the distribution residual between the image data and the projection invariance model is extracted, and the projection consistency compensation is stopped when the distribution residual is outside the preset safety range, and the pixel coordinates of the feature point to be located are determined using preset geometric distribution parameters.

3. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, The scaling compensation module is also used to calculate the spatial distribution variance of surface texture feature points within the area covered by the structural reference vector; generate a local strain compensation operator for the corresponding non-rigid target section based on the dispersion characteristics of the spatial distribution variance; and use the local strain compensation operator to perform nonlinear correction on the target distribution scaling coefficient.

4. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, The baseline construction module is also used to obtain the pixel length of the reference topology vector that is fully displayed in the image data when the first or second topology anchor point is occluded; determine the logical length ratio between the reference topology vector and the non-rigid target segment based on preset object structure feature parameters; and reconstruct the structural baseline vector using the logical length ratio, the pixel length of the reference topology vector, and the pixel position of the fully displayed first or second topology anchor point.

5. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, The benchmark construction module is also used to calculate the angle between the characteristic principal axes of the first and second topological anchor points to determine the buckling curvature of the non-rigid target segment; and to perform tangent integral mapping on the structural benchmark vector based on the buckling curvature to generate a nonlinear geometric trajectory that matches the morphology of the non-rigid target segment. When determining pixel coordinates in image data, the position output module replaces the structural reference vector with a non-linear geometric trajectory.

6. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, It also includes a dynamic hardening module, which is used to obtain the motion residual vector of surface texture feature points between adjacent image frames; and to perform anisotropic Gaussian filtering on the image data using the motion residual vector to suppress feature diffusion of non-rigid target segments in the motion direction. The scale factor is updated and corrected based on the filtered image data.

7. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, When the deflection calculation module extracts the first feature component and the second feature component, it is specifically used to perform differential frequency domain filtering on the local image to obtain the first feature response component and the second feature response component corresponding to different spatial frequencies; and to statistically analyze the lattice energy distribution of the first feature response component and the second feature response component.

8. The acupuncture point positioning system based on artificial intelligence according to claim 1, characterized in that, The deflection calculation module is also used to calculate the deviation of the energy distribution ratio of adjacent pixels within the orthogonal sampling window of the structural reference vector; and to identify unstable feature regions where the deviation exceeds a preset deviation threshold. Remove the data components corresponding to unstable feature regions from the energy distribution ratio.

9. The acupuncture point positioning system based on artificial intelligence according to claim 2, characterized in that, When determining the axial rotation component, the deflection calculation module involves the texture anisotropy index. The calculation of texture anisotropy index It meets the following relationship: ,in, The first eigenvalue of the local covariance matrix. As the second eigenvalue of the local covariance matrix, the position output module is also used to calculate the deviation of the distribution residual from the preset safety interval while outputting the pixel coordinates; and output the positioning reliability index of the corresponding feature point to be located based on the deviation.

10. An artificial intelligence-based acupuncture point location method, used to implement the artificial intelligence-based acupuncture point location system described in claim 1, characterized in that, Includes the following steps: Step S101: Acquire image data containing non-rigid target segments; Step S102: Identify the first and second topological anchor points of the non-rigid target segment in the image data; Step S103: Establish a structural reference vector with the first topological anchor point as the starting point and the second topological anchor point as the ending point; Step S104: Extract surface texture feature points within the area covered by the structural reference vector, perform multi-scale frequency domain decomposition on the local image of the non-rigid target segment, extract the first feature component corresponding to the high frequency response component and the second feature component corresponding to the low frequency response component, calculate the energy distribution ratio between the first feature component and the second feature component, and determine the spatial deflection parameter of the non-rigid target segment relative to the imaging plane based on the gradient slope of the energy distribution ratio in the direction of the structural reference vector. Step S105: Perform projection consistency compensation on the preset target distribution scale coefficient based on the spatial deflection parameter to determine the correction scale factor; Step S106: Perform coordinate mapping on the structural reference vector according to the correction scale factor, and output the pixel coordinates of the feature point to be located in the image data after superimposing a preset orthogonal offset.

Images