Intelligent image recognition method for massage of acupoints by medicine stick based on image processing

Abstract

This invention belongs to the field of image recognition technology, specifically relating to an intelligent image recognition method for acupoint massage using a medicinal stick based on image processing. The method involves acquiring a basic color image before the medicinal stick contacts the acupoint to extract an initial two-dimensional coordinate set; simultaneously acquiring a high-frequency depth image sequence during the pressure application phase, and calculating the pixel displacement vector field between adjacent frames; inputting the pixel displacement vector field into a musculoskeletal biomechanics prior model to solve for a three-dimensional skin deformation tensor of the target region; and performing spatial inverse mapping compensation calculation on the initial two-dimensional coordinate set of acupoints physically obscured by the medicinal stick, outputting the three-dimensional dynamic coordinates of the acupoints under pressure. This method eliminates the interference of local skin stretching caused by physical pressure and instrument obstruction on coordinate positioning, maintaining the spatial topological consistency of acupoint coordinates during dynamic physical interaction.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to an intelligent image recognition method for acupoint massage using a medicinal stick based on image processing. Background Technology

[0002] Existing acupoint recognition schemes for massage sticks typically employ external visual acquisition devices to obtain two-dimensional color images of the human body surface. An image processing network then extracts geometric features such as skin texture, pigment distribution, and muscle contours. These extracted features are matched against a pre-stored acupoint anatomical location model to pinpoint the static coordinates of the acupoints within the two-dimensional image plane. Throughout the recognition process, the visual acquisition device maintains a fixed distance from the human body surface, and the image processing network performs feature extraction and coordinate regression calculations only on the skin surface in a state of no external force contact.

[0003] In the specific implementation of the aforementioned prior art, when the medicinal stick contacts the target skin area and applies pressure, the stick itself physically obstructs the viewing angle of the visual acquisition device. Simultaneously, the skin in the contact area undergoes nonlinear elastic deformation under external pressure. Existing image processing networks lose the surface texture features of the contact area due to the obstructed viewpoint, and the pixel arrangement of the two-dimensional color image cannot reflect the physical displacement state of the skin in the depth direction. Consequently, the two-dimensional coordinates of the acupoints, originally calibrated in a non-contact state, cannot adaptively update with skin deformation, resulting in spatial position drift and failing to output three-dimensional dynamic coordinates of the acupoints that match the actual physical pressure position. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent image recognition method for acupoints massaged with a medicinal stick based on image processing, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for intelligent image recognition of acupoints by a medicine stick massage based on image processing includes: acquiring a basic RGB image of the medicine stick before it contacts the target area, inputting the basic RGB image into a pre-trained human topology network, and extracting the initial two-dimensional coordinate set of acupoints in the target area; When the drug stick contacts the target area and is under pressure, a high-frequency depth image sequence of the target area is acquired simultaneously, and the pixel displacement vector field between adjacent frames in the high-frequency depth image sequence is calculated. The pixel displacement vector field is input into a pre-constructed musculoskeletal biomechanics prior model, and the pixel displacement vector field is solved into a three-dimensional skin deformation tensor of the target region. Based on the three-dimensional skin deformation tensor, spatial inverse mapping compensation calculation is performed on the initial two-dimensional coordinate set of acupoints that are physically obscured by the medicated stick, and the three-dimensional dynamic coordinates of the acupoints under the pressure state are output.

[0006] Preferably, the step of inputting the basic RGB image into a pre-trained human topology network and extracting the initial two-dimensional coordinate set of acupoints in the target region includes: performing edge detection and skin color segmentation processing on the basic RGB image to generate a binarized mask image of the target region; The binarized mask image is concatenated with the base RGB image through channels and then input into the human body topology network. The human body topology network includes a feature extraction backbone network and a graph convolutional network. The feature extraction backbone network outputs a multi-scale spatial feature map of the basic RGB image. The graph convolutional network uses key points of the human skeleton as graph nodes to map the multi-scale spatial feature map into graph node features and outputs the initial two-dimensional coordinate set of acupoints.

[0007] Preferably, the step of calculating the pixel displacement vector field between adjacent frames in the high-frequency depth image sequence includes: taking the first frame depth image corresponding to the initial moment when the stick touches the target area as a reference frame, and extracting the current frame depth image after the reference frame in the high-frequency depth image sequence; The reference frame and the current frame depth image are divided into image blocks with overlapping regions. Block matching search is performed on the image blocks to determine the depth direction displacement and horizontal direction displacement of each pixel in the current frame depth image relative to the reference frame. The depth displacement and the horizontal displacement are combined to generate the pixel displacement vector field.

[0008] Preferably, the step of inputting the pixel displacement vector field into a pre-constructed musculoskeletal biomechanics prior model and solving the pixel displacement vector field into a three-dimensional skin deformation tensor of the target region includes: the musculoskeletal biomechanics prior model stores skin elastic modulus parameters and subcutaneous bone normal vector parameters; The two-dimensional horizontal displacement in the pixel displacement vector field is combined with the skin elastic modulus parameter and converted into a three-dimensional spatial displacement through finite element mesh interpolation. The three-dimensional spatial displacement and the depth direction displacement are multiplied by a tensor, and then orthogonal projection correction is performed in combination with the subcutaneous bone normal vector parameters to generate the three-dimensional skin deformation tensor.

[0009] Preferably, the step of performing spatial inverse mapping compensation calculation on the initial acupoint two-dimensional coordinate set physically occluded by the medicinal stick based on the three-dimensional skin deformation tensor includes: extracting local image blocks corresponding to the initial acupoint two-dimensional coordinate set in the basic RGB image, and determining the pixel neighborhood set of the local image blocks; Obtain the current position coordinates of the stick in the target area, map the current position coordinates to the local image block, remove the pixels in the local image block that are within a preset radius of the current position coordinates, and generate a set of remaining pixels. The remaining pixel set is subjected to an affine transformation using the three-dimensional skin deformation tensor, and the geometric center coordinates of the transformed remaining pixel set are used as the compensated three-dimensional dynamic coordinates of the acupoint.

[0010] Preferably, after outputting the three-dimensional dynamic coordinates of the acupoints under the applied pressure state, the method further includes: constructing a sliding time window and extracting the three-dimensional dynamic coordinates of the acupoints corresponding to multiple historical moments, including the current moment, within the sliding time window; Using the timestamp of the current moment as a weight decay factor, a weighted Kalman filter is applied to the three-dimensional dynamic coordinates of the acupoints corresponding to the multiple historical moments to calculate the predicted coordinates of the current moment. Calculate the Euclidean distance between the predicted coordinates and the current three-dimensional dynamic coordinates of the acupoint. When the Euclidean distance is within a preset distance range, update the predicted coordinates to the final output locked coordinates.

[0011] Preferably, the graph convolutional network uses key points of the human skeleton as graph nodes to map the multi-scale spatial feature map into graph node features, including: performing global average pooling operation on the feature maps of different scales in the multi-scale spatial feature map to generate feature vectors of different scales. The feature vectors of different scales are concatenated to generate the full feature vector. The full feature vector is assigned to an adjacency matrix with key points of the human skeleton as graph nodes. Feature transfer calculation is performed on the adjacency matrix through multiple graph convolutional layers to update the graph node features. Fully connected classification calculation is performed using the updated graph node features to output the initial acupoint two-dimensional coordinate set.

[0012] Preferably, performing block matching search on the image block to determine the depth direction displacement and horizontal direction displacement of each pixel in the current frame depth image relative to the reference frame includes: extracting the three-dimensional coordinates of the center pixel and the three-dimensional coordinates of the neighboring pixels of each image block in the reference frame; In the current frame depth image, a search window is constructed using the three-dimensional coordinates of the center pixel. Within the search window, the three-dimensional spatial distance between the three-dimensional coordinates of the neighboring pixels and the three-dimensional coordinates of the corresponding candidate pixels is calculated. A nonlocal mean weighted calculation is performed on the three-dimensional spatial distance within the search window. The difference between the three-dimensional coordinates of the candidate pixel and the three-dimensional coordinates of the center pixel when the weighted calculation result is at its minimum is used as the displacement in the depth direction and the displacement in the horizontal direction.

[0013] Preferably, the two-dimensional horizontal displacement in the pixel displacement vector field, combined with the skin elastic modulus parameter, is converted into a three-dimensional spatial displacement through finite element mesh interpolation calculation, including: using the edge contour of the target region as the boundary condition, dividing the target region into regularly distributed hexahedral finite element meshes; The two-dimensional horizontal displacement is mapped onto the surface nodes of the hexahedral finite element mesh, serving as the initial boundary displacement constraint for the surface nodes. The stiffness matrix of the hexahedral finite element mesh is constructed based on the skin elastic modulus parameter. The linear equations of the stiffness matrix are solved in combination with the initial boundary displacement constraints. The displacement vectors of the internal nodes of the hexahedral finite element mesh are calculated, and the displacement vectors of the internal nodes are combined to form the three-dimensional spatial displacement.

[0014] Preferably, the remaining pixel set is subjected to an affine transformation using the three-dimensional skin deformation tensor, including: obtaining the base radius of the stick and the tilt angle of the stick relative to the target region; Based on the bottom radius and the tilt angle, extract the edge deformation tensor that is outside the geometric projection region of the stick from the three-dimensional skin deformation tensor; Using the edge pixels in the edge deformation tensor as control points, an offset linear regression equation is constructed. Input the coordinates of each pixel in the remaining pixel set into the offset linear regression equation, calculate the coordinate offset value corresponding to each pixel, and add the coordinate offset value to the coordinates of each pixel to complete the affine transformation.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention simultaneously acquires high-frequency depth image sequences during the skin-contacting and pressure-application phase of the medicated stick, calculates the pixel displacement vector field between adjacent frames, and introduces a musculoskeletal mechanics prior model to solve the pixel displacement vector field into a three-dimensional skin deformation tensor. This three-dimensional skin deformation tensor is then used to perform spatial inverse mapping compensation calculations on the initial two-dimensional coordinate set physically occluded by the medicated stick. This processing method couples the image displacement extraction process with the physical deformation process of the subcutaneous muscles and bones, eliminating the interference of local skin stretching caused by physical pressure and instrument occlusion on coordinate positioning. It maintains the spatial topological consistency of acupoint coordinates during dynamic physical interaction, realizing the output of three-dimensional dynamic coordinates of acupoints under pressure.

[0016] 2. In extracting the initial two-dimensional coordinate set, this invention uses a graph convolutional network to map multi-scale spatial features into graph node features with key points of the human skeleton as graph nodes, thus eliminating interference from background pixels in non-acupoint areas. When calculating the three-dimensional spatial displacement, it combines the skin elastic modulus to construct a hexahedral finite element mesh stiffness matrix and solves a system of linear equations, making the calculation results of the three-dimensional spatial displacement conform to the mechanical transmission characteristics of the subcutaneous physical structure. After outputting the dynamic coordinates, it combines multiple historical coordinates within the sliding time window to perform weighted Kalman filtering calculation, filtering out spatial position fluctuation errors during the acquisition of single-frame depth images and maintaining the geometric continuity of the three-dimensional dynamic coordinates of acupoints during continuous recognition. Attached Figure Description

[0017] Figure 1 This is a flowchart of the overall method for intelligent image recognition of acupoints using a medicated stick massage according to the present invention. Figure 2 This is a flowchart of the initial acupoint two-dimensional coordinate set extraction method of the present invention; Figure 3 This is a flowchart of the pixel displacement vector field calculation method of the present invention; Figure 4 This is a flowchart of the three-dimensional skin deformation tensor solution method of the present invention; Figure 5 This is a flowchart of the acupoint coordinate space inverse mapping compensation method of the present invention; Figure 6 This is a flowchart of the weighted Kalman filter locking method for acupoint coordinates according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please refer to Figure 1 This embodiment provides an intelligent image recognition method for acupoint massage using a medicinal stick based on image processing. The target area is the area on the human body surface to be massaged, including but not limited to the back, shoulders, neck, and limbs. The medicinal stick is a stick-shaped instrument used for acupoint pressing in traditional Chinese medicine massage scenarios. The state before the medicinal stick contacts the target area is that the pressing end of the stick has no physical contact with the skin of the target area, and the entire stick is within the field of view of the visual acquisition device. The visual acquisition device simultaneously acquires RGB image data and depth image data. During the acquisition process, the spatial position and posture of the visual acquisition device remain fixed, and the field of view completely covers the target area. Before the medicinal stick contacts the target area, the visual acquisition device acquires the basic RGB image of the target area. After acquisition, distortion correction processing is performed on the basic RGB image. The distortion correction processing is based on the intrinsic distortion coefficients pre-calibrated by the visual acquisition device. During the distortion correction process, the coordinates in the pixel coordinate system of the basic RGB image are first converted to normalized imaging plane coordinates. The distortion-corrected coordinates are calculated using a preset distortion model, and then the corrected coordinates are converted back to the pixel coordinate system to obtain the distortion-corrected basic RGB image. The pixel coordinate system of the distortion-free base RGB image is registered with the pixel coordinate system of the depth image. The registration process is completed based on the extrinsic parameters pre-calibrated by the vision acquisition device. After registration, each pixel in the base RGB image and the pixel at the same coordinate position in the depth image correspond to the same physical point in three-dimensional space, ensuring spatial consistency in the subsequent coordinate mapping process.

[0020] The distortion-free base RGB image is input into a pre-trained human topology network to extract the initial two-dimensional coordinate set of acupoints in the target region. The human topology network is a pre-trained supervised neural network model. The training sample set includes RGB images of human body surfaces of different genders, ages, and body types. Each sample is labeled with corresponding skeletal key points and acupoint two-dimensional coordinates. During training, the labeled acupoint two-dimensional coordinates are used as the ground truth, and the error between the network output coordinates and the ground truth is used as the loss function to iteratively update the network parameters. The input to the human topology network is the base RGB image data, and the output is the two-dimensional coordinates of all acupoints to be identified within the target region in the pixel coordinate system of the base RGB image. The combination of all acupoint two-dimensional coordinates forms the initial acupoint two-dimensional coordinate set, with each acupoint corresponding to a unique two-dimensional coordinate value, and the unit of the coordinate value is pixels.

[0021] When the medicated stick contacts the target area and is in the pressure application phase, the visual acquisition device simultaneously acquires a high-frequency depth image sequence of the target area. The pressure application phase is the stage where the pressing end of the medicated stick makes physical contact with the skin of the target area, and the medicated stick continuously applies pressure to the skin. During this phase, there is no relative sliding between the pressing end of the medicated stick and the skin, or the displacement of relative sliding is less than a preset threshold. The acquisition frequency of the high-frequency depth image sequence is matched with the skin deformation rate during the pressure application phase to ensure that the skin deformation between adjacent frames is within a preset linear range, avoiding the failure of subsequent displacement calculations due to excessive deformation between adjacent frames. Each frame of the high-frequency depth image sequence stores the three-dimensional spatial depth value corresponding to each pixel. The physical meaning of the depth value is the straight-line distance between the physical point corresponding to the pixel and the optical center of the visual acquisition device. The pixel coordinate system of each frame of the depth image is registered with the pixel coordinate system of the basic RGB image.

[0022] Calculate the pixel displacement vector field between adjacent frames in a high-frequency depth image sequence. Adjacent frames are two depth images acquired at adjacent times within the high-frequency depth image sequence. The pixel displacement vector field contains the displacement vector of each pixel in the current frame relative to the pixel at the same coordinate position in the preceding frame. The displacement vector includes horizontal and depth components. The horizontal direction corresponds to the x and y axes of the pixel coordinate system of the depth image, and the depth direction corresponds to the optical axis of the visual acquisition device. During the calculation, the preceding frame of the two adjacent depth images is used as a reference. Matching calculations are performed on the pixels in both frames to obtain the displacement vector of each pixel. The displacement vectors of all pixels are combined to form the complete pixel displacement vector field.

[0023] The pixel displacement vector field is input into a pre-constructed musculoskeletal mechanics prior model, and the pixel displacement vector field is solved into a three-dimensional skin deformation tensor of the target region. The musculoskeletal mechanics prior model is a mechanical model built based on the anatomical characteristics of human skin, subcutaneous muscles, and bones. The model stores the skin elasticity parameters and skeletal anatomical position parameters corresponding to the target region. The input of the model is the pixel displacement vector field, and the output is the three-dimensional skin deformation tensor of the target region. The three-dimensional skin deformation tensor is a second-order tensor used to describe the deformation state of the skin surface of the target region in three-dimensional space. It includes the tensile, compressive, and shear deformation components of the skin. Each component of the tensor corresponds to the degree of deformation in one direction in three-dimensional space, which can completely reflect the spatial deformation state of the skin under pressure.

[0024] Based on the 3D skin deformation tensor, spatial inverse mapping compensation calculation is performed on the initial 2D coordinate set of acupoints physically occluded by the medicinal stick, outputting the 3D dynamic coordinates of the acupoints under pressure. The physically occluded area of ​​the medicinal stick is the projection area of ​​the medicinal stick in the pixel coordinate system of the basic RGB image. The pixels within this projection area correspond to skin texture features that are occluded by the medicinal stick, making it impossible to directly extract feature information from the RGB image. The spatial inverse mapping compensation calculation, based on the skin deformation state described by the 3D skin deformation tensor, maps the initial 2D coordinate set of acupoints in the non-contact state to the 3D spatial coordinate system under pressure, compensating for the coordinate offset caused by skin deformation and instrument occlusion, thus obtaining the dynamic coordinates of the acupoints in 3D space under pressure. The unit of the 3D dynamic coordinates of the acupoints is the International Standard Unit of Length (ISU).

[0025] In this embodiment, the parameters and physical meanings of each component of the pixel displacement vector field are shown in Table 1.

[0026] Table 1. Comparison of Pixel Displacement Vector Field Component Parameters and Physical Meanings

[0027] Table 1 is used to clarify the definition and physical meaning of each component in the pixel displacement vector field, to ensure that those skilled in the art can accurately understand the calculation logic of the displacement vector, avoid confusion of parameter meanings, and ensure full disclosure of the technical solution.

[0028] In this embodiment, an initial two-dimensional coordinate set of acupoints is extracted from the basic RGB image under non-contact conditions. During the pressure application phase, a high-frequency depth image sequence is simultaneously acquired to calculate the pixel displacement vector field. The three-dimensional skin deformation tensor is solved by combining the musculoskeletal mechanics prior model. Finally, the three-dimensional dynamic coordinates of acupoints under pressure are obtained through spatial inverse mapping compensation calculation. This fully realizes the dynamic recognition of acupoint coordinates under the scenarios of instrument occlusion and skin deformation during the massage with the medicated stick. It solves the problem of positioning failure caused by the drift of acupoint coordinates with skin deformation and the loss of features in the occluded area in the prior art.

[0029] In a preferred embodiment, reference Figure 2Edge detection and skin color segmentation are performed on the base RGB image to generate a binary mask of the target region. Edge detection uses the Canny edge detection operator to extract the surface contour edges of the target region from the grayscale version of the base RGB image. Grayscale processing uses a weighted average method to convert the pixel values ​​of the R, G, and B channels of the base RGB image to grayscale values, following a preset weight ratio based on the human eye's sensitivity to different color channels. Skin color segmentation is performed in the YCbCr color space. The base RGB image is first converted from the RGB color space to the YCbCr color space, following a preset color space transformation matrix. After conversion, preset threshold ranges are set for the pixel values ​​of the Cb and Cr channels, and pixels within these ranges are extracted as skin color pixels to generate the initial skin color mask. The body surface contour edges obtained from edge detection are logically ANDed with the initial skin color mask to remove non-skin color pixels outside the contour edges. At the same time, non-skin color hole areas within the contour edges are filled to generate a binary mask image of the target area. In the binary mask image, the pixel value of the pixel corresponding to the target area is 1, and the pixel value of the pixel corresponding to the background area is 0.

[0030] A binarized mask is concatenated with the base RGB image through multiple channels and then input into the human body topology network. During the concatenation process, the binarized mask serves as the fourth channel, concatenating dimensionally with the R, G, and B channels of the base RGB image. The resulting concatenated image has four channels, and its width and height are identical to the base RGB image. This concatenated image data is then input into the human body topology network, enabling the network to distinguish target and background regions based on the binarized mask during feature extraction. This suppresses the feature weights of the background region, enhances the feature extraction effect of the target region, and reduces the interference of background pixels on acupoint coordinate extraction.

[0031] The human body topology network comprises a feature extraction backbone network and a graph convolutional network. The feature extraction backbone network outputs multi-scale spatial feature maps of the base RGB image. The graph convolutional network uses key points of the human skeleton as graph nodes, mapping the multi-scale spatial feature maps to graph node features and outputting an initial two-dimensional coordinate set of acupoints. The feature extraction backbone network adopts a residual network structure, containing multiple sequentially connected convolutional and pooling modules. Each convolutional module contains a convolutional layer, a batch normalization layer, and an activation function layer. The pooling module uses a max pooling layer. The feature extraction backbone network outputs spatial feature maps of different scales after different convolutional modules. Different scale spatial feature maps correspond to different receptive fields: low-scale feature maps correspond to large receptive fields, containing global contour features of the target region, while high-scale feature maps correspond to small receptive fields, containing local texture features of the target region.

[0032] Global average pooling is performed on feature maps of different scales in a multi-scale spatial feature map to generate feature vectors of different scales. For each channel of the feature map at each scale, global average pooling calculates the average of all pixel feature values ​​within that channel, and combines these averages to form the feature vector for that scale. The formula for global average pooling is as follows:

[0033] in, Let be the global average pooling output value of the i-th channel in the feature map at the s-th scale. The height of the feature map at the s-th scale. Let be the width of the feature map at the s-th scale. Let be the pixel feature value at coordinate (x, y) of the i-th channel in the feature map of the s-th scale.

[0034] The feature vectors at different scales are concatenated to generate a full feature vector. The concatenation operation follows the order of feature map scale from largest to smallest, sequentially concatenating the feature vectors at each scale. The dimension of the concatenated full feature vector is the sum of the dimensions of the feature vectors at all scales. The full feature vector simultaneously contains both global contour features and local texture features of the target region, thus fully reflecting the spatial features of the target region.

[0035] The full feature vector is assigned to an adjacency matrix with human skeletal keypoints as graph nodes. Multiple graph convolutional layers perform feature transfer calculations on the adjacency matrix to update the graph node features. The adjacency matrix of human skeletal keypoints is constructed based on the anatomical connections of the human skeleton. An element value of 1 in the adjacency matrix indicates an anatomical connection between the corresponding two skeletal keypoints, while an element value of 0 indicates no anatomical connection between the corresponding two skeletal keypoints. The full feature vector is assigned to the graph nodes corresponding to the skeletal keypoints according to their indexes, serving as the initial node features for the graph convolutional network. Multiple graph convolutional layers sequentially perform feature transfer and update calculations on the node features. The calculation process of each graph convolutional layer follows the following formula:

[0036] in, For the first The node feature matrix output by the layer graph convolutional layer. For the first The node feature matrix input to the layer graph convolutional layer, Here, A represents the adjacency matrix of the human skeleton keypoints with self-loops added, and A is the original adjacency matrix of the human skeleton keypoints. It is the identity matrix. for The degree matrix is ​​a set of diagonal elements representing the degree of the corresponding node, and the off-diagonal elements are 0. For the first The trainable weight matrix of a layered graph convolutional layer It is a non-linear activation function.

[0037] The updated graph node features are used to perform fully connected classification computation, outputting an initial set of two-dimensional coordinates for acupoints. There is a pre-defined anatomical topological relationship between key points of the human skeleton and acupoints; each acupoint corresponds to at least one adjacent key point of the skeleton. Based on the updated graph node features and combined with the topological relationship between key points of the skeleton and acupoints, the fully connected layer performs regression computation to obtain the two-dimensional coordinates of each acupoint in the pixel coordinate system. The combination of the two-dimensional coordinates of all acupoints forms the initial set of two-dimensional coordinates for acupoints.

[0038] In this embodiment, the parameters and output dimensions of the multi-scale feature map of the human body topology network are shown in Table 2.

[0039] Table 2 Comparison of Multi-Scale Feature Map Parameters and Output Dimensions of Human Topological Network

[0040] Table 2 is used to clarify the parameter configuration of the multi-scale spatial feature map output by the feature extraction backbone network, so that those skilled in the art can clearly understand the extraction and cascading logic of multi-scale features, ensure that the network structure is fully disclosed, and enable those skilled in the art to configure the corresponding network structure according to the table and repeat the technical solution of this embodiment.

[0041] In this embodiment, a binary mask image is generated by edge detection and skin color segmentation. After being concatenated with the basic RGB image channels, it is input into a human topology network containing a feature extraction backbone network and a graph convolutional network. Through global average pooling, feature concatenation and multi-layer graph convolution calculation, an initial two-dimensional coordinate set of acupoints is output. This effectively suppresses feature interference from the background region, strengthens the correlation between the human skeletal topology and acupoints, and improves the accuracy of acupoint coordinate extraction in a non-contact state.

[0042] In a preferred embodiment, reference Figure 3 The first frame depth image corresponding to the initial moment when the medicated stick contacts the target area is used as the reference frame. The current frame depth image after the reference frame in the high-frequency depth image sequence is extracted. The initial moment when the medicated stick contacts the target area is the moment when the pressing end of the medicated stick first physically contacts the skin of the target area. The first frame depth image corresponding to this moment is the starting frame of the pressure application phase. In this frame image, the skin of the target area has not yet deformed due to the pressure. Therefore, this frame is used as the reference frame for the displacement calculation benchmark of all subsequent frames, ensuring that the displacement calculation of all frames is based on the undeformed skin state and avoiding the generation of cumulative errors.

[0043] The depth images of the reference frame and the current frame are divided into image blocks with overlapping regions. Block matching search is performed on these image blocks to determine the depth and horizontal displacement of each pixel in the current frame relative to the reference frame. During the image block division process, the size of each image block is a preset fixed value, and the size of the overlapping region between adjacent image blocks is also a preset value. The overlapping region is set to avoid discontinuities in the displacement calculation at the edges of the image blocks, ensuring the spatial continuity of the displacement vector field. The block matching search uses each image block in the reference frame as a reference, searching for the candidate image block in the current frame depth image with the highest feature similarity to the reference image block. The coordinate difference between the candidate image block and the reference image block is the displacement of the center pixel of the image block.

[0044] The 3D coordinates of the center pixel and neighboring pixels of each image block in the reference frame are extracted. The 3D coordinates of each pixel in the reference frame are calculated based on its 2D coordinates and corresponding depth value, combined with the intrinsic parameters of the visual acquisition device. The calculation process converts the 2D coordinates in the pixel coordinate system to 3D coordinates in the camera coordinate system, ensuring spatial consistency. Neighboring pixels are all pixels within a preset size neighborhood of the center pixel. The number of neighboring pixels matches the size of the image block. The 3D coordinates of the neighboring pixels are used to characterize the local spatial structure of the image block, improving the accuracy of block matching.

[0045] In the current frame depth image, a search window is constructed using the 3D coordinates of the center pixel. Within the search window, the 3D spatial distance between the 3D coordinates of neighboring pixels and the corresponding candidate pixels is calculated. The size of the search window matches the maximum expected displacement between adjacent frames, ensuring that the matching block corresponding to the reference image block is within the search window and avoiding matching failures. The search process employs a multi-step search strategy: a coarse search is performed first to determine candidate regions, followed by a fine search within the candidate regions to determine the optimal matching block. For each candidate pixel within the search window, the 3D coordinates of its corresponding neighboring pixels are calculated, along with the 3D spatial distance between them and the neighboring pixels in the reference frame. This 3D spatial distance is the Euclidean distance between the two 3D coordinate points and is used to characterize the similarity between the two neighborhood structures.

[0046] Non-local mean weighted calculation is performed on the 3D spatial distance within the search window. The difference between the 3D coordinates of the candidate pixel corresponding to the minimum weighted calculation result and the 3D coordinates of the center pixel is used as the depth and horizontal displacement. Non-local mean weighted calculation is used to enhance the consistency of pixel displacement within the neighborhood and suppress noise interference from single pixels. The formula for weighted calculation is as follows:

[0047] in, For the pixels in the reference frame The weighting coefficient between the candidate pixel q in the current frame and the candidate pixel q. This is a normalization constant used to ensure that the sum of the weighting coefficients of all candidate pixels is 1. Let p be the set of three-dimensional coordinates of the neighboring pixels. Let q be the set of three-dimensional coordinates of the neighboring pixels. is the L2 norm, and h is the smoothing coefficient used to control the decay rate of the weighting coefficients.

[0048] The candidate pixel corresponding to the minimum value of the weighted calculation result is the pixel with the highest matching degree to the center pixel of the reference frame. In the difference of their three-dimensional coordinates, the components in the x and y directions are the horizontal displacement, and the component in the z direction is the depth displacement. The depth displacement and the horizontal displacement are combined to generate a pixel displacement vector field. Each pixel in the pixel displacement vector field corresponds to a three-dimensional displacement vector, and the displacement vectors of all pixels are combined to form a complete pixel displacement vector field.

[0049] refer to Figure 4 The pixel displacement vector field is input into a pre-constructed musculoskeletal mechanics prior model, and the pixel displacement vector field is solved into a three-dimensional skin deformation tensor for the target region. The musculoskeletal mechanics prior model stores skin elastic modulus parameters and subcutaneous bone normal vector parameters. The skin elastic modulus parameters are preset based on the anatomical characteristics of human skin and correspond to the location of the target region. The skin elastic modulus parameters differ in different body surface regions. The subcutaneous bone normal vector parameters are the normal vectors of the subcutaneous bone surface corresponding to the target region. They are preset based on the human skeletal anatomy model, and the direction of the normal vector is from the inside of the bone to the skin surface.

[0050] The two-dimensional horizontal displacement in the pixel displacement vector field, combined with the skin elastic modulus parameter, is converted into a three-dimensional spatial displacement through finite element mesh interpolation. The edge contour of the target region is used as the boundary condition, and a regularly distributed hexahedral finite element mesh is generated within the target region. The edge contour of the target region is the outer contour of the target region in the binarized mask image. This contour is mapped to three-dimensional space and used as the boundary condition for the finite element mesh. Each element of the hexahedral finite element mesh is a regular hexahedral structure. The mesh nodes include surface nodes and internal nodes. Surface nodes correspond to points on the skin surface of the target region, and internal nodes correspond to points within the subcutaneous tissue. The mesh density matches the deformation gradient of the target region; the larger the deformation gradient, the higher the mesh density, ensuring the accuracy of deformation calculation.

[0051] The two-dimensional horizontal displacement is mapped onto the surface nodes of a hexahedral finite element mesh, serving as the initial boundary displacement constraint for the surface nodes. The mapping process is based on the registration relationship between the pixel coordinate system and the three-dimensional spatial coordinate system. Each surface node corresponds to a pixel in the depth image, and the horizontal displacement corresponding to that pixel is the horizontal displacement constraint of that surface node. The constraint condition is that the horizontal displacement of the surface node is consistent with the corresponding displacement in the pixel displacement vector field.

[0052] The stiffness matrix of a hexahedral finite element mesh is constructed based on the skin's elastic modulus parameter. The linear equations of the stiffness matrix are then solved using initial boundary displacement constraints to calculate the displacement vectors of the nodes within the hexahedral finite element mesh. The stiffness matrix is ​​constructed based on the skin's elastic modulus and Poisson's ratio parameters. The element stiffness matrix of each hexahedral element is calculated using the element's geometric parameters, elastic modulus, and Poisson's ratio. The element stiffness matrices of all elements are assembled according to their node numbers to obtain the overall stiffness matrix. The solution to the linear equations follows the formula:

[0053] Where K is the global stiffness matrix of the hexahedral finite element mesh, u is the displacement vector of all nodes in the mesh, including the known displacements of surface nodes and the unknown displacements of internal nodes, and F is the node load vector, which is generated by the boundary displacement constraints of the surface nodes.

[0054] In the process of solving the linear equation system, the known displacements of the surface nodes are substituted into the equation system, and the unknown displacements of the internal nodes are obtained by Gaussian elimination. The displacement vectors of all nodes are combined to form a three-dimensional spatial displacement. The three-dimensional spatial displacement contains the displacement values ​​of all nodes in the target area in three-dimensional space, which can fully reflect the spatial deformation state of the skin and subcutaneous tissue.

[0055] The 3D spatial displacement and the depth direction displacement are multiplied by a tensor, and then orthogonal projection correction is performed using the subcutaneous bone normal vector parameters to generate a 3D skin deformation tensor. First, the deformation gradient tensor is calculated based on the 3D spatial displacement, using the following formula:

[0056] in, For the deformation gradient tensor, It is the identity matrix. Let be the gradient matrix of the displacement in three-dimensional space. The elements of the gradient matrix are the partial derivatives of the displacement components with respect to the spatial coordinates.

[0057] The deformation gradient tensor is orthogonally projected and corrected by combining the subcutaneous bone normal vector parameters. Rigid displacement components along the bone normal vector direction are removed, while the elastic deformation components of the skin are retained. The correction formula is as follows:

[0058] in, Here, n represents the corrected 3D skin deformation tensor, and n is the unit normal vector corresponding to the subcutaneous bone normal vector parameter. It represents the orthogonal projection component of the deformation gradient tensor onto the bone normal vector direction.

[0059] In this embodiment, the parameters and calculation logic of the hexahedral finite element mesh are shown in Table 3.

[0060] Table 3 Comparison of Hexahedral Finite Element Mesh Parameters and Calculation Logic

[0061] Table 3 clarifies the parameter configuration and calculation logic of the finite element mesh elements, enabling those skilled in the art to clearly understand the finite element interpolation calculation process of three-dimensional spatial displacement, ensuring the full disclosure of the mechanical calculation process, and enabling those skilled in the art to select the corresponding element type based on the table to complete the construction and solution of the finite element mesh.

[0062] In this embodiment, the horizontal and depth displacements of the pixel displacement vector field are obtained by block matching search with the initial pressure frame as the reference frame and by non-local mean weighted calculation. The stiffness matrix of a hexahedral finite element mesh is constructed based on the skin elastic modulus parameters of the musculoskeletal mechanics prior model. The three-dimensional spatial displacement is obtained by solving the linear equation system. After orthogonal projection correction with the subcutaneous bone normal vector, a three-dimensional skin deformation tensor is generated. The calculation result of the deformation tensor fits the mechanical transmission characteristics of human subcutaneous tissue and accurately reflects the real deformation state of the skin under pressure.

[0063] In a preferred embodiment, reference Figure 5 In the base RGB image, local image patches corresponding to the initial acupoint 2D coordinate set are extracted, and the pixel neighborhood set of the local image patches is determined. Each local image patch corresponding to the initial acupoint 2D coordinate set is a rectangular image patch of a preset size centered on the 2D coordinates of each acupoint in the initial acupoint 2D coordinate set. Each acupoint corresponds to an independent local image patch, and the size of the local image patch is larger than the projection size of the medicinal stick in the image, ensuring that the local image patch includes the skin area around the acupoint that is not obscured by the medicinal stick. The pixel neighborhood set is the set of coordinates of all pixels within the local image patch, including the center pixel of the acupoint and its surrounding neighboring pixels. The neighboring pixels are used to characterize the deformation trend of the skin around the acupoint.

[0064] The current position coordinates of the medicine stick in the target area are obtained and mapped to a local image patch. Pixels within a preset radius of the current position coordinates in the local image patch are removed, generating a set of remaining pixels. The current position coordinates of the medicine stick in the target area are the coordinates of the center of the pressing end face of the medicine stick in a three-dimensional spatial coordinate system. These coordinates are calculated based on the point cloud data of the medicine stick in the depth image sequence. The mapping process uses the intrinsic and extrinsic parameters of the vision acquisition device to convert the coordinates in the three-dimensional spatial coordinate system into two-dimensional coordinates in the pixel coordinate system of the local image patch. The preset radius matches the radius of the pressing end face of the medicine stick. The area corresponding to the pixels within the preset radius is the physically occluded area of ​​the medicine stick. After removing pixels in this area, the remaining pixels are all in the unoccluded area. The coordinates of all remaining pixels are combined to form the set of remaining pixels.

[0065] An affine transformation is performed on the remaining pixel set using the 3D skin deformation tensor, and the geometric center coordinates of the transformed remaining pixel set are used as the compensated 3D dynamic coordinates of the acupoint. The base radius of the medicinal stick and its tilt angle relative to the target area are obtained. The base radius of the medicinal stick is the radius of the pressing end of the stick, which is a preset fixed value; the tilt angle of the medicinal stick relative to the target area is the angle between the central axis of the medicinal stick and the normal vector of the skin surface of the target area, which is calculated based on the 3D point cloud data of the medicinal stick in the depth image sequence.

[0066] Based on the base radius and tilt angle, edge deformation sub-tensors located outside the geometric projection area of ​​the stick are extracted from the 3D skin deformation tensor. The geometric projection area of ​​the stick is the area projected onto the skin surface. The size and shape of this area are determined by the base radius and tilt angle; the larger the tilt angle, the greater the eccentricity of the ellipse in the projection area. The edge deformation sub-tensors are the deformation tensors corresponding to the areas outside the geometric projection area and whose distance from the edge of the projection area is less than a preset threshold. This part of the tensor reflects the skin deformation state at the edge of the stick pressing area, is not obscured by the stick, and can accurately reflect the deformation trend around the pressing area.

[0067] Using edge pixels in the edge deformation subtensor as control points, an offset linear regression equation is constructed. The least squares method is used to fit the linear regression equation. During the fitting process, the initial coordinates of the control points in the edge deformation subtensor are used as independent variables, and the displacements corresponding to the control points are used as dependent variables. The regression coefficients are solved, and the sum of squared residuals is less than a preset threshold to ensure that the regression equation accurately reflects the distribution trend of deformation. The formula for the offset linear regression equation is as follows:

[0068]

[0069]

[0070] in, , , These represent the coordinate offsets of the pixel along the x, y, and z axes, respectively. , Let be the initial two-dimensional coordinates of the pixel in the pixel coordinate system of the local image patch, and let a, b, c, d, e, f, g, h, i be the regression coefficients of the linear regression equation.

[0071] The coordinates of each pixel in the remaining pixel set are input into the offset linear regression equation to calculate the coordinate offset value for each pixel. This offset value is then added to the coordinates of each pixel to complete the affine transformation. After the affine transformation, the coordinates of each pixel in the remaining pixel set are converted into three-dimensional spatial coordinates under pressure. These converted coordinates reflect the positional shift caused by skin deformation. The geometric center coordinates of the remaining pixel set are the average of the three-dimensional coordinates of all transformed pixels; this coordinate is the compensated three-dimensional dynamic coordinate of the acupoint.

[0072] refer to Figure 6 A sliding time window is constructed to extract the three-dimensional dynamic coordinates of acupoints corresponding to multiple historical moments, including the current moment, within the sliding time window. The length of the sliding time window is a preset fixed value, and the number of historical moments included in the window is matched with the window length and the acquisition frequency of the high-frequency depth image sequence. The sliding time window slides synchronously with the current moment, always retaining the three-dimensional dynamic coordinates of acupoints within a preset time period before the current moment, ensuring that the filtering process can utilize continuous historical data and maintain the continuity of coordinates.

[0073] Using the current timestamp as a weight decay factor, a weighted Kalman filter is applied to the 3D dynamic coordinates of acupoints corresponding to multiple historical times to calculate the predicted coordinates for the current time. The weight decay factor adjusts the weights of data from different historical times; historical data closer to the current time has a higher weight and a greater impact on the prediction result. The formula for calculating the weight decay factor is as follows:

[0074] in, Let be the weight decay factor at time t. The timestamp of the current moment. t is the preset attenuation coefficient, and t is the timestamp of a historical moment.

[0075] The weighted Kalman filter is based on a uniform motion model. The state vector contains the position and velocity components of the acupoint's three-dimensional coordinates. The state transition matrix is ​​also based on the uniform motion model. Assuming that the acupoint's motion velocity remains constant within the sliding time window, the state prediction process of the filter follows the following formula:

[0076] in, For the first The prior state prediction vector of the frame corresponds to the predicted coordinates at the current time. Here is the state transition matrix. For the first The posterior state estimation vector of the frame. To estimate the covariance matrix a priori, Let Q be the covariance matrix for posterior estimation, and let Q be the process noise covariance matrix. The value of the process noise covariance matrix is ​​inversely proportional to the weight decay factor. The closer the historical data is to the current time, the smaller the value of the corresponding process noise covariance matrix.

[0077] The state update process of filtering follows the formula below:

[0078] in, Let H be the Kalman gain for the k-th frame, H be the observation matrix that maps the state vector to observation coordinates, and R be the observation noise covariance matrix. Let be the observation vector of the k-th frame, corresponding to the three-dimensional dynamic coordinates of the acupoint calculated at the current moment. Let be the posterior state estimation vector for the k-th frame. It is an identity matrix.

[0079] The Euclidean distance between the predicted coordinates and the current 3D dynamic coordinates of the acupoint is calculated. If the Euclidean distance is within a preset distance range, the predicted coordinates are updated to the final locked coordinates. The formula for calculating the Euclidean distance is as follows:

[0080] Where d is the Euclidean distance between the predicted coordinates and the current three-dimensional dynamic coordinates of the acupoint. It is an L2 norm.

[0081] The lower and upper limits of the preset distance interval are based on the measurement accuracy of the acupoint size and depth image. When the Euclidean distance is within the preset distance interval, it means that the continuity between the three-dimensional dynamic coordinates of the acupoint calculated at the current moment and the historical coordinates meets the requirements, and the predicted coordinates output by the weighted Kalman filter are updated to the final output locked coordinates. When the Euclidean distance exceeds the upper limit of the preset distance interval, it means that there is a jump in the coordinates at the current moment, and the posterior state estimation vector of the historical moment is used as the locked coordinates at the current moment to avoid the positioning error caused by the coordinate jump.

[0082] In this embodiment, the parameters and state update logic of the weighted Kalman filter are shown in Table 4.

[0083] Table 4. Comparison of Weighted Kalman Filter Parameters and State Update Logic

[0084] Table 4 clarifies the definition and calculation logic of each parameter in the weighted Kalman filtering process, enabling those skilled in the art to clearly understand the filtering and locking process of acupoint coordinates, ensuring the continuity and stability of dynamic coordinate output, and guaranteeing the full disclosure of the technical solution.

[0085] In this embodiment, by extracting local image blocks corresponding to acupoints, removing pixels in the area obscured by the medicinal stick to generate a set of remaining pixels, and extracting edge deformation tensors based on the size and tilt angle of the medicinal stick to construct a linear regression equation, performing affine transformation on the remaining pixel set to obtain the compensated three-dimensional dynamic coordinates of the acupoints, and combining sliding time windows and weighted Kalman filtering to output the final locked coordinates, the coordinate offset caused by the obscuration of the medicinal stick and skin deformation is effectively compensated, the acquisition noise of single-frame images is filtered out, and the geometric continuity of acupoint coordinates is maintained during continuous massage.

Claims

1. A method for intelligent image recognition of acupoints using a medicinal stick based on image processing, characterized in that, include: Acquire a basic RGB image of the target area before the medicine stick contacts it, input the basic RGB image into a pre-trained human topology network, and extract the initial two-dimensional coordinate set of acupoints in the target area; When the drug stick contacts the target area and is under pressure, a high-frequency depth image sequence of the target area is acquired simultaneously, and the pixel displacement vector field between adjacent frames in the high-frequency depth image sequence is calculated. The pixel displacement vector field is input into a pre-constructed musculoskeletal biomechanics prior model, and the pixel displacement vector field is solved into a three-dimensional skin deformation tensor of the target region. Based on the three-dimensional skin deformation tensor, spatial inverse mapping compensation calculation is performed on the initial two-dimensional coordinate set of acupoints that are physically obscured by the medicated stick, and the three-dimensional dynamic coordinates of the acupoints under the pressure state are output.

2. The intelligent image recognition method for acupoint massage using a medicinal stick based on image processing according to claim 1, characterized in that, The step of inputting the basic RGB image into a pre-trained human topology network and extracting the initial two-dimensional coordinate set of acupoints in the target region includes: performing edge detection and skin color segmentation processing on the basic RGB image to generate a binarized mask image of the target region; The binarized mask image is concatenated with the base RGB image through channels and then input into the human body topology network. The human body topology network includes a feature extraction backbone network and a graph convolutional network. The feature extraction backbone network outputs a multi-scale spatial feature map of the basic RGB image. The graph convolutional network uses key points of the human skeleton as graph nodes to map the multi-scale spatial feature map into graph node features and outputs the initial two-dimensional coordinate set of acupoints.

3. The intelligent image recognition method for acupoint massage with a medicinal stick based on image processing according to claim 1, characterized in that, The step of calculating the pixel displacement vector field between adjacent frames in the high-frequency depth image sequence includes: taking the first frame depth image corresponding to the initial moment when the stick touches the target area as a reference frame, and extracting the current frame depth image after the reference frame in the high-frequency depth image sequence; The reference frame and the current frame depth image are divided into image blocks with overlapping regions. Block matching search is performed on the image blocks to determine the depth direction displacement and horizontal direction displacement of each pixel in the current frame depth image relative to the reference frame. The depth displacement and the horizontal displacement are combined to generate the pixel displacement vector field.

4. The intelligent image recognition method for acupoint massage with a medicinal stick based on image processing according to claim 1, characterized in that, The step of inputting the pixel displacement vector field into a pre-constructed musculoskeletal biomechanics prior model and solving the pixel displacement vector field into a three-dimensional skin deformation tensor of the target region includes: the musculoskeletal biomechanics prior model stores skin elastic modulus parameters and subcutaneous bone normal vector parameters; The two-dimensional horizontal displacement in the pixel displacement vector field is combined with the skin elastic modulus parameter and converted into a three-dimensional spatial displacement through finite element mesh interpolation. The three-dimensional spatial displacement and the depth direction displacement are multiplied by a tensor, and then orthogonal projection correction is performed in combination with the subcutaneous bone normal vector parameters to generate the three-dimensional skin deformation tensor.

5. The intelligent image recognition method for acupoint massage using a medicinal stick based on image processing according to claim 1, characterized in that, The step of performing spatial inverse mapping compensation calculation on the initial acupoint two-dimensional coordinate set physically obscured by the medicinal stick based on the three-dimensional skin deformation tensor includes: extracting local image blocks corresponding to the initial acupoint two-dimensional coordinate set in the basic RGB image, and determining the pixel neighborhood set of the local image blocks; Obtain the current position coordinates of the stick in the target area, map the current position coordinates to the local image block, remove the pixels in the local image block that are within a preset radius of the current position coordinates, and generate a set of remaining pixels. The remaining pixel set is subjected to an affine transformation using the three-dimensional skin deformation tensor, and the geometric center coordinates of the transformed remaining pixel set are used as the compensated three-dimensional dynamic coordinates of the acupoint.

6. The intelligent image recognition method for acupoint massage with a medicinal stick based on image processing according to claim 1, characterized in that, After outputting the three-dimensional dynamic coordinates of the acupoints under the applied pressure state, the method further includes: constructing a sliding time window and extracting the three-dimensional dynamic coordinates of the acupoints corresponding to multiple historical moments, including the current moment, within the sliding time window; Using the timestamp of the current moment as a weight decay factor, a weighted Kalman filter is applied to the three-dimensional dynamic coordinates of the acupoints corresponding to the multiple historical moments to calculate the predicted coordinates of the current moment. Calculate the Euclidean distance between the predicted coordinates and the current three-dimensional dynamic coordinates of the acupoint. When the Euclidean distance is within a preset distance range, update the predicted coordinates to the final output locked coordinates.

7. The intelligent image recognition method for acupoint massage using a medicinal stick based on image processing according to claim 2, characterized in that, The graph convolutional network uses key points of the human skeleton as graph nodes and maps the multi-scale spatial feature map into graph node features, including: performing global average pooling operation on the feature maps of different scales in the multi-scale spatial feature map to generate feature vectors of different scales. The feature vectors of different scales are concatenated to generate the full feature vector. The full feature vector is assigned to an adjacency matrix with key points of the human skeleton as graph nodes. Feature transfer calculation is performed on the adjacency matrix through multiple graph convolutional layers to update the graph node features. Fully connected classification calculation is performed using the updated graph node features to output the initial acupoint two-dimensional coordinate set.

8. The intelligent image recognition method for acupoint massage with a medicinal stick based on image processing according to claim 3, characterized in that, Perform block matching search on the image block to determine the depth direction displacement and horizontal direction displacement of each pixel in the current frame depth image relative to the reference frame, including: extracting the three-dimensional coordinates of the center pixel and the three-dimensional coordinates of the neighboring pixels of each image block in the reference frame; In the current frame depth image, a search window is constructed using the three-dimensional coordinates of the center pixel. Within the search window, the three-dimensional spatial distance between the three-dimensional coordinates of the neighboring pixels and the three-dimensional coordinates of the corresponding candidate pixels is calculated. A nonlocal mean weighted calculation is performed on the three-dimensional spatial distance within the search window. The difference between the three-dimensional coordinates of the candidate pixel and the three-dimensional coordinates of the center pixel when the weighted calculation result is at its minimum is used as the displacement in the depth direction and the displacement in the horizontal direction.

9. The intelligent image recognition method for acupoint massage with a medicinal stick based on image processing according to claim 4, characterized in that, The two-dimensional horizontal displacement in the pixel displacement vector field, combined with the skin elastic modulus parameter, is converted into a three-dimensional spatial displacement through finite element mesh interpolation. This includes: using the edge contour of the target region as boundary conditions, dividing the target region into regularly distributed hexahedral finite element meshes. The two-dimensional horizontal displacement is mapped onto the surface nodes of the hexahedral finite element mesh, serving as the initial boundary displacement constraint for the surface nodes. The stiffness matrix of the hexahedral finite element mesh is constructed based on the skin elastic modulus parameter. The linear equations of the stiffness matrix are solved in combination with the initial boundary displacement constraints. The displacement vectors of the internal nodes of the hexahedral finite element mesh are calculated, and the displacement vectors of the internal nodes are combined to form the three-dimensional spatial displacement.

10. The intelligent image recognition method for acupoint massage with a medicinal stick based on image processing according to claim 5, characterized in that, Performing an affine transformation on the remaining pixel set using the three-dimensional skin deformation tensor includes: obtaining the base radius of the stick and the tilt angle of the stick relative to the target region; Based on the bottom radius and the tilt angle, extract the edge deformation tensor that is outside the geometric projection region of the stick from the three-dimensional skin deformation tensor; Using the edge pixels in the edge deformation tensor as control points, an offset linear regression equation is constructed. Input the coordinates of each pixel in the remaining pixel set into the offset linear regression equation, calculate the coordinate offset value corresponding to each pixel, and add the coordinate offset value to the coordinates of each pixel to complete the affine transformation.

Images