Human body acupoint automatic positioning method, system and device based on deep learning and storage medium

Abstract

The invention discloses a human body acupoint automatic positioning method, system and device based on deep learning and a storage medium. A body part area containing an acupoint is quickly positioned through a target detection model; utilizing a key point detection model to accurately detect anatomical key point coordinates in the region; innovatively, firstly, according to a preset traditional Chinese medicine acupoint positioning rule knowledge base, mathematical calculation is conducted on the key point coordinates, and initial acupoint coordinates are obtained; then a local image area with the initial coordinates as the center is intercepted, a lightweight self-adaptive fine tuning network is input, fine offset compensation is conducted on the initial coordinates according to local image features, and final high-precision acupoint coordinates are output. In addition, the invention further provides a double-model cooperative training process, weighted screening is carried out on training samples by utilizing a region detection model so as to improve the training quality of a key point detection model, and traditional Chinese medicine knowledge and a modern artificial intelligence technology are deeply fused so as to improve the training quality of the key point detection model. The method overcomes the defects of black box, high data dependence, poor generalization and the like of a pure data driven model, has the advantages of high positioning precision, high reliability, good interpretability and the like, and is suitable for multiple fields of traditional Chinese medicine clinic, teaching, health service and the like.

Description

Technical Field

[0001] This invention relates to the field of acupoint location technology, specifically to a method, system, device, and storage medium for automatic acupoint location based on deep learning. Background Technology

[0002] Meridians are the channels through which Qi and blood circulate in the human body, connecting the internal organs, communicating between the interior and exterior, and running through the upper and lower parts of the body. Acupoints are special locations on the body surface where the Qi of the internal organs and meridians is transported. When the internal organs become diseased, the Qi of the internal organs and meridians becomes imbalanced and cannot be transported to the body surface, resulting in pathological reactions at the acupoints. External treatments in Traditional Chinese Medicine, such as acupuncture, act on acupoints to stimulate the Qi of the meridians, harmonize the internal organs, and balance Yin and Yang. Currently, acupoint location is based on the existing standards "Anthropometry Methods for Acupoint Location" and "Acupoint Location Diagrams." Traditional methods of acupoint location include anatomical landmark location, "bone measurement" location, finger measurement, and simplified acupoint location, which rely on individual professional expertise and experience and lack objectivity, uniformity, and simplicity, thus hindering the promotion and application of Traditional Chinese Medicine.

[0003] With the advancement of technology and the continuous development of image recognition, computer vision, machine learning, and deep learning, the use of deep learning technology to achieve automatic acupoint location can promote the objectification, standardization, and intelligentization of acupuncture. However, due to the large number of acupoints in traditional Chinese medicine, deep learning requires a large amount of data for training, and the manual annotation workload is substantial, which is not conducive to optimization, improvement, and expansion. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device and storage medium for automatic location of human acupoints based on deep learning. Based on deep learning neural network technology, acupoints can be located automatically, quickly and accurately, which is conducive to the objectification, standardization, digitalization and intelligent development of traditional Chinese medicine.

[0005] To achieve the above objectives, the present invention provides a method for automatic localization of human acupoints based on deep learning, comprising the following steps:

[0006] S1, Data acquisition and preparation: Obtain a dataset of skin images of the target area and label the images in the dataset. The labeling includes the labeling of the bounding box of the acupoint area and the coordinate labeling of the anatomical key points related to the acupoint.

[0007] S2, Region detection model training: Based on a dataset labeled with the bounding boxes of the acupoint regions, a target detection neural network model is trained. The model is used to locate the region of interest containing the target acupoint from the input image.

[0008] S3, Keypoint Detection Model Training: Using a dataset labeled with the coordinates of the anatomical keypoints, a keypoint detection neural network model is trained. The model is used to accurately detect the coordinates of predefined anatomical keypoints from the region of interest.

[0009] S4, Acupoint localization: Input the image to be localized into the trained region detection model to obtain the region of interest; input the region of interest into the trained key point detection model to obtain the coordinates of the anatomical key points; based on the coordinates of the anatomical key points, calculate the final coordinates of the target acupoint using predefined mapping rules.

[0010] Furthermore, in step S2,

[0011] The trained region detection model is used to infer the training image set to generate predicted candidate regions.

[0012] Based on the sample weight values, a high-quality sample subset is selected from the training image set;

[0013] The key point detection model is trained using the high-quality sample subset and its labeled anatomical key point coordinates.

[0014] Furthermore, the formula for calculating the sample weight value is: Sample weight = IoU * C, where IoU is the intersection-union ratio of the predicted region and the real region, and C is the confidence of the region detection model for the category of the predicted region.

[0015] Furthermore, the keypoint detection model is trained through iterative optimization, wherein:

[0016] Initially, an initial version of the region detection model is trained using the entire training image set;

[0017] The initial version of the region detection model is used to perform the screening process to obtain a first-generation high-quality sample subset and train a first-generation keypoint detection model.

[0018] Use the first-generation keypoint detection model with better performance to assist in annotation or generate negative samples that are harder to distinguish, in order to train the next generation of region detection models with better performance.

[0019] The above process is repeated iteratively until the performance of both the region detection model and the key point detection model converges.

[0020] Furthermore, the region detection model is based on the YOLO architecture, and the key point detection model is based on the RTMPose architecture.

[0021] Furthermore, the specific steps for calculating the final coordinates of the target acupoint based on the coordinates of the anatomical key points using predefined mapping rules include:

[0022] Based on the coordinates of the key anatomical points, the initial coordinates of the target acupoint are calculated using a predefined mathematical rule, which is determined according to the measurement methods in traditional Chinese medicine literature.

[0023] Extract a local image region centered at the initial coordinates;

[0024] Input the initial coordinates and / or the local image region into an adaptive fine-tuning network;

[0025] The adaptive fine-tuning network outputs a coordinate offset based on the input image features;

[0026] The initial coordinates are added to the coordinate offset to obtain the final coordinates.

[0027] Furthermore, the adaptive fine-tuning network is a lightweight multilayer perceptron or convolutional neural network. The adaptive fine-tuning network is trained using a mean squared error loss function, and its training objective is to minimize the difference between the coordinate offset predicted by the network and the actual manual fine-tuning offset.

[0028] On the other hand, the present invention also provides an acupoint positioning system, comprising:

[0029] The data acquisition and preparation module acquires a dataset of skin images of the target area and annotates the images in the dataset. The annotation includes the annotation of the bounding box of the acupoint area and the coordinate annotation of the anatomical key points related to the acupoint.

[0030] The region detection model training module trains a target detection neural network model based on a dataset labeled with the bounding boxes of the acupoint regions. The model is used to locate the region of interest containing the target acupoint from the input image.

[0031] The key point detection model training module uses a dataset labeled with the coordinates of the anatomical key points to train a key point detection neural network model. The model is used to accurately detect the coordinates of predefined anatomical key points from the region of interest.

[0032] The acupoint localization module inputs the image to be localized into the trained region detection model to obtain the region of interest; inputs the region of interest into the trained key point detection model to obtain the coordinates of the anatomical key points; and calculates the final coordinates of the target acupoint based on the coordinates of the anatomical key points through a predefined mapping rule.

[0033] On the other hand, the present invention also provides a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method described above.

[0034] On the other hand, a computer-readable storage medium is characterized in that the computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the steps of the method as described in any one of claims 1 to 7.

[0035] This invention has the following technological advancements and beneficial effects:

[0036] 1. This invention innovatively proposes a hybrid paradigm combining anatomical structure perception, traditional Chinese medicine rule calculation, and image feature fine-tuning. The model first locates clearly defined anatomical landmarks, then performs calculations based on traditional Chinese medicine theoretical rules, and finally uses a small network for adaptive fine-tuning to resolve uncertainties. This method deeply embeds human prior knowledge into the artificial intelligence process, greatly enhancing the credibility and interpretability of the results.

[0037] 2. This invention solves the problems of small sample learning and model generalization. By decomposing the task, it greatly reduces the requirements for data scale and annotation. The YOLO and REMPose models are easy to use for annotation, and the rule calculation directly encodes TCM knowledge without training. This invention can achieve excellent performance on a small amount of professional data and has good generalization ability.

[0038] 3. The present invention employs a dual-model collaborative training mechanism. Through a weight selection mechanism, it ensures that the key point detection model is trained using only high-quality regional samples, thereby improving the robustness of the model from the source. The fine-tuned network can compensate for the initial coordinates calculated based on rules according to the subtle features of the local texture and wrinkles of the image, thus adapting to the skin condition and body shape differences of different individuals. This is something that fixed rule or pure data models cannot achieve. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a method for automatic localization of human acupoints based on deep learning, according to an embodiment of the present invention.

[0041] Figure 2 This is a system architecture diagram of an automatic human acupoint positioning system based on deep learning, according to an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0043] 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 embodiments of the present invention, and not all embodiments. 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.

[0044] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0046] Figure 1 This is a flowchart of a method for automatic acupoint localization based on deep learning, as described in this invention. Figure 1 As shown, the present invention provides a method for automatic localization of human acupoints based on deep learning, comprising the following steps:

[0047] S1, Data Acquisition and Preparation.

[0048] In a specific embodiment, this method acquires human surface image data including target acupoints, including but not limited to specific areas of the face, hands, feet, and limbs. Using commonly used clinical postures, vertical perspectives, and field of vision for different body parts, the method maintains relative consistency in shooting distance, angle, and lighting conditions during acquisition, and annotates the acquired image data.

[0049] After collection, the data is cleaned to remove unclear or incomplete data, and then stored, categorized, and numbered.

[0050] Preferably, data annotation includes acupoint region boundary box annotation, anatomical key point annotation, and acupoint coordinate annotation. Specifically, the acupoint region boundary box annotation outlines the area where the acupoint is located and labels the object name; the anatomical key point annotation marks the anatomical key points and provides annotations within the object area; and the acupoint coordinate annotation marks each acupoint within the object area according to acupoint location standards and annotates the acupoint name. After annotation is completed, the object name and coordinate box, acupoint name and coordinates, anatomical feature points and coordinates, and annotation information are saved.

[0051] Preferably, in another embodiment, after the data is labeled, new images can be generated through color, spatial transformation, geometric transformation, etc. The human body part in the image is extracted and combined with images with different backgrounds to generate new images. The new images and the original labels are matched to expand the data, thereby forming more sample data.

[0052] S2, Region Detection Model Training.

[0053] In a specific embodiment, the region detection model used in this implementation is built on the YOLOv8 (You Only LookOnce version 8) architecture.

[0054] Preferably, the backbone network of the target detection neural network model adopts the CSPDarknet structure. Through cross-stage partial connections, the gradient vanishing problem in deep networks is greatly alleviated, ensuring that the network can still effectively train and update while deeply extracting subtle features in areas such as the "nose" and "hand".

[0055] The neck network employs an FPN+PAN structure because the scale of acupoint regions in the human body varies greatly (e.g., the entire back region vs. the ear region). PAN transmits strong semantic features from top to bottom, enhancing the feature representation ability of small regions; PAN, on the other hand, transmits precise localization information from bottom to top, ensuring the accuracy of bounding box regression for large regions. This structure ensures that the model can accurately locate the target region regardless of its size.

[0056] The detection head employs a decoupled detection head, improving the accuracy of the final bounding box localization. The model training loss function uses CIoU Loss as the bounding box regression loss. This loss function comprehensively considers overlapping area, center point distance, and aspect ratio, directly optimizing localization accuracy and providing high-quality region input for subsequent keypoint detection.

[0057] Preferably, in this embodiment, after the target detection neural network model locates the region of interest containing the target acupoint, it calculates a sample weight for each predicted candidate region. This weight comprehensively considers the accuracy of localization (IoU) and the confidence level of classification, and can effectively measure the quality of the predicted region. This weight is defined by the following formula.

[0058] Sample weight = IoU(predicted region, true region) × region detection confidence

[0059] Based on the above sample weights, a weight threshold is set, and training samples with weights greater than this threshold are retained to form a high-quality sample subset. This step automatically filters out samples with inaccurate region detection model localization (low IoU) and samples with ambiguous region detection model judgment (low confidence).

[0060] Taking a hand training image as an example, the region detection model infers from a hand training image and outputs a predicted region bounding box with an IoU of 0.85 and a confidence score of 0.92. Therefore, the sample weight is approximately 0.85 * 0.92 ≈ 0.78. Assuming a threshold T = 0.6, if 0.78 > 0.6, the sample is considered a high-quality sample and retained in a subset.

[0061] The keypoint detection model will be trained using this precisely cropped image of the hand region to learn the locations of key points such as knuckles and wrist bones. Because the bounding box scores are high and the localization is accurate, there is little irrelevant background in the image, and the keypoint features are significant, the model's learning efficiency is higher.

[0062] S3, Keypoint Detection Model Training.

[0063] S3, training of the key point detection model.

[0064] Specifically, in one embodiment, the RTMPose architecture is used as the keypoint detection model. Its backbone network extracts multi-scale features, maintaining high-resolution feature representations to construct parallel multi-resolution sub-networks for feature processing. The feature fusion mechanism enhances feature representation through repeated multi-scale fusion operations, uses upsampling and downsampling operations to align feature maps of different resolutions, and employs element-wise addition for feature fusion. The coordinate prediction head uses the SimCC coordinate classification mechanism for keypoint localization, decomposing the coordinate prediction task into two independent one-dimensional classification tasks, and using feature map compression techniques to reduce computational complexity.

[0065] Specifically, the training of the keypoint detection model is achieved through an iterative optimization process in conjunction with the region detection model, the steps of which include:

[0066] Initial model training: The initial version of the region detection model M_det0 is trained using the entire training image set; the initial version of the keypoint detection model M_kpt0 is trained using the entire labeled data based on the RTMPose architecture.

[0067] First-generation high-quality sample selection: M_det0 is used to infer the training image set to generate predicted candidate regions; the overlap ratio (IoU) between each predicted region and the ground truth region is calculated. Combined with the region confidence score output by the region detection model, sample weight values ​​are generated: Sample weight T_i = IoU(predicted region, ground truth region) × region detection confidence score. Based on the sample weight values, samples with weight values ​​greater than the threshold T1 are selected to form the first-generation high-quality sample subset S1.

[0068] Training of the first-generation keypoint detection model: The first-generation RTMPose keypoint detection model M_kpt1 was trained using a high-quality sample subset S1; a coordinate classification mechanism was used for model optimization; the performance of M_kpt1 was evaluated on the validation set, and the keypoint detection accuracy index was recorded.

[0069] Next-generation region detection model training: Use M_kpt1 to infer the training set and generate keypoint prediction results; identify difficult samples with high keypoint prediction confidence but low region detection quality; add these difficult samples as key training samples to the training set of the region detection model; use the enhanced training set to train the next-generation region detection model M_det1.

[0070] Iterative optimization process: Repeat the above steps for multiple rounds of iteration. In each round of iteration, adjust the sample weight threshold T_i and gradually improve the screening criteria. In each round of iteration, use the updated region detection model to screen a higher quality sample subset.

[0071] A new generation of RTMPose keypoint detection models is trained using a higher-quality subset of samples. When the improvement in keypoint detection accuracy is less than a preset threshold Δ in two consecutive iterations, the performance of both the region detection model and the keypoint detection model reaches a stable state on the validation set. The final trained RTMPose keypoint detection model, M_kpt_final, is output.

[0072] S4, Acupoint localization: Input the image to be localized into the trained region detection model to obtain the region of interest; input the region of interest into the trained key point detection model to obtain the coordinates of the anatomical key points; based on the coordinates of the anatomical key points, calculate the final coordinates of the target acupoint through a predefined mapping rule.

[0073] Specifically, the steps for calculating the final coordinates of the target acupoint based on the coordinates of the anatomical key points using predefined mapping rules include:

[0074] Based on the coordinates of key anatomical points, the initial coordinates of the target acupoints are calculated using a predefined mathematical rule. This predefined mathematical rule is determined according to measurement methods in traditional Chinese medicine literature. Specifically, this invention establishes mathematical mapping rules for each target acupoint based on classic TCM literature and national standards. The rule types include: midpoint calculation rules (e.g., Yintang acupoint = midpoint (left eyebrow, right eyebrow); proportional segmentation rules (e.g., Shousanli acupoint = Yangxi acupoint + (Quchi acupoint - Yangxi acupoint) × 0.4); and offset rules (e.g., Yingxiang acupoint = midpoint of the outer edge of the nasal ala + fixed offset vector).

[0075] Extract a local image region centered at the initial coordinates.

[0076] The initial coordinates and / or the local image region are input into an adaptive fine-tuning network.

[0077] Specifically, the adaptive fine-tuning network is a lightweight multilayer perceptron or convolutional neural network. The adaptive fine-tuning network is trained using a mean squared error loss function, and its training objective is to minimize the difference between the coordinate offset predicted by the network and the actual manual fine-tuning offset.

[0078] In one embodiment, the adaptive network adopts a lightweight dual-branch structure, which extracts local image features F_image by a 4-layer convolutional network, processes coordinate information F_coord by a 2-layer fully connected network, and then concatenates F_image and F_coord through a 3-layer fully connected network for feature fusion. The output layer includes 2 neurons that output the offsets (Δx, Δy) in the x and y directions, respectively.

[0079] The adaptive fine-tuning network outputs a coordinate offset based on the input image features;

[0080] The initial coordinates are added to the coordinate offset to obtain the final coordinates.

[0081] In one embodiment, the final coordinate calculation first involves obtaining the predicted offsets (Δx_pred, Δy_pred) through forward propagation of the network. Then, the offsets are scaled to the original image coordinate scale: Δx_final = Δx_pred × W × S, Δy_final = Δy_pred × H × S. Here, W and H are the image width and height, and S is the scaling factor (default 0.1). The final acupoint coordinates are: P_final = (x_initial + Δx_final, y_initial + Δy_final).

[0082] Figure 2 This is a system architecture diagram of a deep learning-based automatic acupoint localization system according to the present invention. Figure 3 As shown, the present invention provides an automatic acupoint localization system based on deep learning, comprising:

[0083] The data acquisition and preparation module acquires a dataset of skin images of the target area and annotates the images in the dataset. The annotation includes the annotation of the bounding box of the acupoint area and the coordinate annotation of the anatomical key points related to the acupoint.

[0084] The region detection model training module trains a target detection neural network model based on a dataset labeled with the bounding boxes of the acupoint regions. The model is used to locate the region of interest containing the target acupoint from the input image.

[0085] The key point detection model training module uses a dataset labeled with the coordinates of the anatomical key points to train a key point detection neural network model. The model is used to accurately detect the coordinates of predefined anatomical key points from the region of interest.

[0086] The acupoint localization module inputs the image to be localized into the trained region detection model to obtain the region of interest; inputs the region of interest into the trained key point detection model to obtain the coordinates of the anatomical key points; and calculates the final coordinates of the target acupoint based on the coordinates of the anatomical key points through a predefined mapping rule.

[0087] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Figure 3 As shown, an electronic device according to one embodiment of the present invention includes one or more input devices 1000, one or more output devices 1000, one or more processors 3000, and a memory 4000.

[0088] In one embodiment of the present invention, the processor 1000, input device 2000, output device 3000, and memory 4000 may be connected via a bus or other means. The input device 2000 and output device 3000 may be standard wired or wireless communication interfaces.

[0089] The processor 1000 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0090] The memory 4000 can be a high-speed RAM or a non-volatile memory, such as a disk storage device. The memory 4000 is used to store a set of computer programs, and the input device 2000, output device 3000, and processor 1000 can call the program code stored in the memory 4000.

[0091] The computer program stored in memory 4000 includes program instructions that, when executed by a processor, cause the processor to perform the steps of the method as described in the above embodiments.

[0092] An embodiment of the present invention also provides a computer-readable storage medium. This computer-readable storage medium may be a high-speed RAM memory or an unstable memory, such as a disk storage device. The computer-readable storage medium can be connected to an external computing device or network to read a set of computer programs stored therein. The computer programs stored in the computer-readable storage medium include program instructions that, when executed by a processor, cause the processor to perform the steps of the method as described in the above embodiments.

[0093] Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based automatic human acupoint positioning method, characterized in that, The method comprises the following steps: S1, data collection and preparation, obtaining a skin image dataset of a target part, and labeling the images in the dataset, the labeling including acupoint region bounding box labeling and acupoint related anatomical key point coordinate labeling; S2, region detection model training, training a target detection neural network model according to the dataset labeled with the acupoint region bounding box, the model being used to locate a region of interest containing a target acupoint from an input image; S3, key point detection model training, training a key point detection neural network model using the dataset labeled with the anatomical key point coordinates, the model being used to accurately detect the coordinates of the predefined anatomical key points from the region of interest; S4, acupoint positioning, inputting a to-be-positioned image into the trained region detection model to obtain a region of interest; inputting the region of interest into the trained key point detection model to obtain the coordinates of the anatomical key points; and based on the coordinates of the anatomical key points, calculating the final coordinates of the target acupoint through a predefined mapping rule. 2.The automatic human acupoint positioning method based on deep learning according to claim 1, wherein, In step S2, the trained region detection model is used to infer the training image set to generate predicted candidate regions; according to the sample weight value, a high-quality sample subset is selected from the training image set; the high-quality sample subset and its labeled anatomical key point coordinates are used to train the key point detection model. 3.The automatic human acupoint positioning method based on deep learning according to claim 2, wherein, The calculation formula of the sample weight value is: sample weight = IoU * C, wherein IoU is the intersection over union of the predicted region and the real region, and C is the class confidence of the region detection model for the predicted region. 4.The automatic human acupoint positioning method based on deep learning according to claim 3, wherein, The key point detection model is trained through iterative optimization, wherein: initially, an initial version of the region detection model is trained using the entire training image set; the initial version of the region detection model is used to perform the screening process to obtain a first-generation high-quality sample subset, and a first-generation key point detection model is trained; the first-generation key point detection model with better performance is used to assist in labeling or generate more difficult-to-distinguish negative samples for training the next-generation region detection model with better performance; the above process is iteratively performed until the performance of the region detection model and the key point detection model converges.

5. The automatic positioning method of human body acupoints based on deep learning according to any one of claims 1 to 4, characterized in that, The region detection model is a model based on the YOLO architecture, and the key point detection model is a model based on the RTMPose architecture. 6.The automatic human acupoint positioning method based on deep learning according to claim 1, wherein, The specific steps of calculating the final coordinates of the target acupoint based on the coordinates of the anatomical key points through a predefined mapping rule include: calculating the initial coordinates of the target acupoint through a predefined mathematical rule according to the coordinates of the anatomical key points, the predefined mathematical rule being determined according to the measurement method in traditional Chinese medicine literature; cutting a local image region centered on the initial coordinates; inputting the initial coordinates and / or the local image region into an adaptive fine-tuning network; the adaptive fine-tuning network outputs a coordinate offset based on the input image features; adding the initial coordinates and the coordinate offset to obtain the final coordinates.

7. The automatic human acupoint positioning method based on deep learning according to claim 6, characterized in that, The adaptive fine-tuning network is a lightweight multilayer perceptron or convolutional neural network, which is trained by a mean square error loss function, and the training target is to minimize the difference between the coordinate offset predicted by the network and the real artificial fine-tuning offset.

8. An acupoint positioning system, characterized in that, The method comprises the following steps: a data acquisition and preparation module acquires a skin image dataset of a target part and labels images in the dataset, the labeling including acupoint region bounding box labeling and acupoint related anatomical key point coordinate labeling; a region detection model training module trains a target detection neural network model according to the dataset labeled with the acupoint region bounding box, the model being used to locate a region of interest containing a target acupoint from an input image; a key point detection model training module trains a key point detection neural network model by using the dataset labeled with the anatomical key point coordinates, the model being used to accurately detect coordinates of predefined anatomical key points from the region of interest; an acupoint positioning module inputs a to-be-positioned image into the trained region detection model to obtain a region of interest, inputs the region of interest into the trained key point detection model to obtain coordinates of the anatomical key points, and calculates final coordinates of a target acupoint based on the coordinates of the anatomical key points through a predefined mapping rule.

9. A computer device, comprising: The computer readable storage medium stores a computer program, and the computer program comprises program instructions, which, when executed by a processor, cause the processor to perform the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program comprises program instructions, which, when executed by a processor, cause the processor to perform the steps of the method according to any one of claims 1 to 7.

Images