ACU-YOLO deep learning method for rapid human body acupoint recognition

Abstract

The invention discloses an ACU-YOLO deep learning method for rapid human body acupoint recognition, and belongs to the technical field of human body acupoint detection, and the method comprises the following steps: S1, collecting a human body chest acupoint data set; s2, preprocessing and labeling the human body front chest acupoint data set in the step S1, and then dividing the human body front chest acupoint data set into a training set for training, a verification set for verification and a test set for testing; s3, a YOLOv10 network structure model is established, and a YOLOv10 network framework is improved, so that an ACU-YOLO network structure model can be obtained; s4, training the ACU-YOLO network structure model in the step S3 by using the training set in the step S2 to obtain a human chest acupoint detection model; and S5, testing the test set by using the human chest acupoint detection model in the step S4. According to the invention, the detection speed can be effectively improved, and higher-degree accuracy of human chest acupuncture points can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of human acupoint detection, and in particular relates to an ACU-YOLO deep learning method for rapid human acupoint recognition. Background Art

[0002] With the rapid development of artificial intelligence technology, deep learning has shown great potential in the field of human acupoint recognition. In the past, acupoint treatment mainly relied on the experience and professional knowledge of doctors to find and locate acupoints. It took several years to train an excellent doctor, and the number of patients treated was limited. There were also problems such as low efficiency and unstable accuracy. In the past decade, deep learning algorithms such as convolutional neural networks (CNNs) have achieved remarkable results in tasks such as image recognition and target detection. Applying deep learning to acupoint recognition can learn the characteristics and patterns of acupoints from a large amount of data, thereby improving the accuracy and efficiency of recognition.

[0003] However, current deep learning models still face some challenges in acupoint recognition applications. Since the location and characteristics of human acupoints may vary from individual to individual, especially the chest acupoints require higher accuracy, which puts higher requirements on the generalization ability of the model. In addition, acupoint recognition requires high-precision positioning to ensure the effectiveness and safety of treatment.

[0004] Therefore, there is a need for an ACU-YOLO deep learning method for rapid human acupoint recognition that can accurately and quickly identify the locations of acupoints on the front chest. Summary of the invention

[0005] The purpose of the present invention is to provide an ACU-YOLO deep learning method for rapid human acupoint recognition, which can effectively improve the detection efficiency and accuracy of human chest acupoint detection.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] An ACU-YOLO deep learning method for rapid human acupoint recognition includes the following steps:

[0008] Step S1: Collecting a data set of acupuncture points on the human chest;

[0009] Step S2: preprocessing and labeling the human chest acupoint dataset in step S1, and then dividing the human chest acupoint dataset into a training set for training, a verification set for verification, and a test set for testing;

[0010] Step S3: Build a YOLOv10 network structure model, improve the YOLOv10 network framework, and obtain the ACU-YOLO network structure model;

[0011] Step S4: using the training set in step S2 to train the ACU-YOLO network structure model in step S3 to obtain a human chest acupoint detection model;

[0012] Step S5: Using the human chest acupuncture point detection model in step S4, the test set is tested.

[0013] A further improvement of the technical solution of the present invention is that step S2 includes the following steps:

[0014] Step S201: pre-processing the human chest acupuncture point data, and adjusting the size of each image in the human chest acupuncture point data set to 1920*1200;

[0015] Step S202: using the image tool Labelme to annotate each image in the human chest acupoint dataset preprocessed in step S201, obtain an annotation file in JSON format, and convert the format to obtain the dataset required for YOLOv10 training;

[0016] Step S203: marking acupoints, including Daheng point SP15, Tianshu point ST25, Mangshu point KI16, Shenque point RN8, Zhongting point RN16, Tiantu point RN22, Zuoruzhong point ST171, and Youruzhong point ST172;

[0017] Step S204: Divide the data set required for ACU-YOLO training into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0018] A further improvement of the technical solution of the present invention is that step S3 includes the following steps:

[0019] Step S301: Build a YOLOv10 network structure model, including an input layer, a Backbone structure for extracting image features, a Neck structure for further processing and fusing features, and an output layer Head; the Backbone structure includes a Conv structure, a C2f structure, a ScDown structure, and an SPPF structure, and the Neck structure includes a C2f structure and a C2fCIB structure;

[0020] Step S302: Add the PSMASH attention mechanism module to the Backbone structure;

[0021] Step S303: adding MSBlock multi-scale blocks to the Neck structure;

[0022] Step S304: Integrate the improved Backbone and Neck structures to obtain the ACU-YOLO network structure model.

[0023] A further improvement of the technical solution of the present invention is that in step S302, the attention mechanism module PSMASH replaces the original channel attention with MSCA, which enables the channel attention to have multi-scale performance, based on the CBAM module, wherein MSCA includes deep convolution to aggregate local information, multi-branch deep convolution to capture multi-scale context, and 1x1 convolution to model the relationship between different channels.

[0024] A further improvement of the technical solution of the present invention is that in step S303, the MSBlock multi-scale block adopts a hierarchical feature fusion strategy, replacing the standard 3×3 convolution with an inverted bottleneck layer to enhance the ability of the real-time target detector to extract multi-scale features using large kernel convolution.

[0025] The further improvement of the technical solution of the present invention lies in that the improved Backbone Network in the ACU-YOLO network structure model obtained in step S304 is composed of a Conv structure, a C2f structure, a ScDown structure, an SPPF structure and a PSMASH structure, wherein the PSMASH structure is added after the SPPF; the improved Neck Network is connected to the MSBlock structure after C2f and C2fCIB respectively.

[0026] A further improvement of the technical solution of the present invention is that in step S4, when the YOLOv10 network structure model is trained using the training set, epochs is set to 100, batch is set to 2, imagsz is set to 640, the depth of the model is set to 0.67, the width is set to 0.75, and the maximum number of channels is set to 768.

[0027] A further improvement of the technical solution of the present invention is that: in step S4, a training set and a YOLOv10 network structure model are used for training to obtain a weight file of the trained YOLOv10 network structure model and a mAP value of the YOLOv10 network structure model on the training set; a training set and an ACU-YOLO network structure model are used for training to obtain a weight file of the trained ACU-YOLO network structure model and a mAP value of the ACU-YOLO network structure model on the training set; wherein the trained ACU-YOLO network structure model is a human chest acupuncture point detection model.

[0028] A further improvement of the technical solution of the present invention is that: in step S5, the weight file of the trained YOLOv10 network structure model and the test set are used for testing to obtain the error distance between the predicted position and the actual position of the acupoint to be detected, as well as the average error distance of the acupoints of this model, the maximum error acupoint error distance and the minimum error acupoint error distance;

[0029] The weight file and test set of the trained ACU-YOLO network structure model are used for testing to obtain the error distance between the predicted position and the actual position of the required acupoint detection point, as well as the average error distance of the acupoints of this model, the maximum error acupoint error distance, and the minimum error acupoint error distance.

[0030] A further improvement of the technical solution of the present invention is that in step S301, a key point detection head is applied in the YOLOv10 framework, and the key point detection head acts as a decoder to extract the category information and position information of the acupoints and output the final acupoint prediction results to expand the application direction of acupoint detection.

[0031] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0032] The ACU-YOLO deep learning method for rapid human acupoint recognition of the present invention can obtain the position and category information of all objects by performing one forward propagation on the image, which can effectively improve the detection speed. By adding MSBlock after C2f of the Neck structure and adding MSBlock after C2fCIB of the Neck structure, its hierarchical fusion feature strategy realizes the accurate recognition of the target acupoints. At the same time, the YOLOv10 acupoint detection algorithm of PSMASH and MSBlock is introduced to detect the acupoints of the human chest, thereby realizing a higher degree of accuracy of the acupoints of the human chest. The present invention can be deployed on the hardware platform of an acupuncture robot or a mobile phone terminal, so that the acupoints of the user can be accurately found by opening the terminal, and has good real-time performance and high precision.

[0033] The present invention extracts and processes features based on the YOLOv10 model framework, combines the key point detection head as a decoder to extract the category information and position information of acupoints, and outputs the final acupoint prediction result, so that the current YOLOv10 expands the key point detection function from a single target detection function, and improves the YOLOv10 backbone network and neck network to obtain ACU-YOLO suitable for rapid acupoint detection.

[0034] The present invention detects eight acupoints on the front chest, namely, the Daheng point SP15, the Tianshu point ST25, the Bianshu point KI16, the Shenque point RN8, the Zhongting point RN16, the Tiantu point RN22, the Zuoruzhong point ST171, and the Youruzhong point ST172, thereby improving the accuracy of acupoint detection.

[0035] The present invention introduces PSMASH after the SPPF structure output of the Backbone structure. PSMASH considers both channel and space dimensions, thereby reducing the computational complexity.

[0036] The present invention realizes accurate identification of target acupuncture points by adding MSBlock after the C2f structure of the Neck structure and adding MSBlock after the C2fCIB structure of the Neck structure through a hierarchical fusion feature strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the ACU-YOLO deep learning method for rapid human acupoint recognition in the present invention;

[0038] Figure 2 It is the principle diagram of the PSMASH attention mechanism in the present invention;

[0039] Figure 3 It is the MSBlock principle diagram of the present invention;

[0040] Figure 4 It is a network diagram of ACU-YOLO in the present invention;

[0041] Figure 5 This is the training result diagram of ACU-YOLO in the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with embodiments:

[0043] like Figure 1 As shown, the present invention provides an ACU-YOLO deep learning method for rapid human acupoint recognition, comprising the following steps:

[0044] Step S1: Collecting a data set of acupuncture points on the human chest;

[0045] Step S2: preprocessing and labeling the human chest acupoint data set in step S1, and then dividing the human chest acupoint data set into a training set for training, a verification set for verification, and a test set for testing, specifically comprising the following steps:

[0046] Step S201: pre-processing the human chest acupuncture point data, adjusting the size of each image in the human chest acupuncture point data set to 1920*1200;

[0047] Step S202: using the image tool Labelme to annotate each image in the human chest acupoint dataset preprocessed in step S201, obtain an annotation file in JSON format, and convert the format to obtain the dataset required for YOLOv10 training;

[0048] Step S203: marking acupoints, respectively marking Daheng point (SP15), Tianshu point (ST25), Mangyu point (KI16), Shenque point (RN8), Zhongting point (RN16), Tiantu point (RN22), Zuoruzhong point (ST171), and Youruzhong point (ST172);

[0049] Step S204: Divide the data set required for ACU-YOLO training into a training set, a validation set, and a test set in a ratio of 8:1:1;

[0050] Step S3: Build a YOLOv10 network structure model, and improve the YOLOv10 network framework to obtain an ACU-YOLO network structure model that can improve detection capability and detection speed, specifically including the following steps:

[0051] Step S301: constructing a YOLOv10 network structure model, including: an input layer, a Backbone structure for extracting image features, a Neck structure for further processing and fusing features, and an output layer Head, wherein the Backbone structure includes a Conv structure, a C2f structure, a ScDown structure, and an SPPF structure, and the Neck structure includes a C2f structure and a C2fCIB structure; in order to expand the application scope of the YOLOv10 algorithm, a key point detection head is applied in the YOLOv10 framework, and the key point detection head is used as a decoder to extract the category information and position information of the acupoints and output the final acupoint prediction result, so that it expands the application direction of acupoint detection, and enables YOLOv10 to expand the key point detection function from a single target detection function;

[0052] Step S302: Add the PSMASH attention mechanism module to the Backbone structure;

[0053] Specifically, Figure 2 As shown in the figure, the attention mechanism module PSMASH replaces the original channel attention with MSCA based on the traditional CBAM module, so that the channel attention has multi-scale performance. MSCA consists of three parts: deep convolution aggregates local information, multi-branch deep convolution captures multi-scale context, and 1x1 convolution models the relationship between different channels. Convolutional attention encodes context information more effectively than the self-attention in the Transformer deep learning model, so that the attention mechanism module PSMASH can optimize the learning of specific category feature information in the feedforward neural network, effectively improving the detection and classification performance;

[0054] After adding the attention mechanism module PSMASH to SPPF in Backbone Network, it is output to NeckNetwork. The improved backbone network can strengthen the attention to important feature information and suppress the attention to secondary information through the spatial and channel attention mechanism modules, so that the acupoint feature information can be extracted more effectively. The trained model is more robust for the recognition of acupoints on the human chest.

[0055] Step S303: adding MSBlock multi-scale blocks to the Neck structure;

[0056] like Figure 3 As shown in the figure, the MSBlock multi-scale block is a new type of block with a hierarchical feature fusion strategy. It adopts a hierarchical feature fusion strategy and uses a large kernel convolution instead of a standard 3×3 convolution by inverting the bottleneck layer to enhance the ability of the real-time target detector to extract multi-scale features. Its small Kernel convolution can capture detailed features more finely, that is, it uses a smaller convolution kernel for feature extraction, and the large Kernel convolution can obtain a wider range of contextual information and can capture more complex relationships and patterns between adjacent pixels. In the acupoint recognition task, especially in the deep stage, a larger receptive field is required to effectively capture coarse-grained information;

[0057] Step S304: Integrate the improved Backbone and Neck structures to obtain the ACU-YOLO network structure model;

[0058] like Figure 4 As shown in the figure, it is the ACU-YOLO model diagram. The improved Backbone Network in the ACU-YOLO network structure model is composed of Conv structure, C2f structure, ScDown structure, SPPF structure and PSMASH structure. The Conv structure realizes the basic feature extraction of the image, the C2f structure enhances the fusion and expression ability of the feature, the ScDown structure is responsible for the efficient downsampling of the feature map, the SPPF structure as the feature integration layer, fuses the features of different scales, and the PSMASH structure focuses on the important feature areas, enhancing the feature learning ability of the model. While ensuring the detection accuracy and speed, the number of parameters and the computational cost of the model are effectively reduced;

[0059] The improved Neck Network is connected to the MSBlock structure after C2f and C2fCIB respectively. The hierarchical feature fusion strategy of MSBlock can capture richer multi-scale information from the features extracted by C2f. Since the positions and shapes of acupoint key points are different, multi-scale feature extraction can keenly capture such changes and improve the detection accuracy of acupoint key points of different sizes. The C2fCIB module has optimized and fused the features to a certain extent. On this basis, connecting MSBlock can further enhance the representation ability of the features, so that the model can learn more discriminative features, thereby improving the recognition and positioning accuracy of human acupoint key points. In addition, MSBlock can also help the model better process the complex information around the acupoint key points and enhance the model's detection ability of acupoint key points under complex backgrounds.

[0060] Step S4: using the training set in step S2 to train the ACU-YOLO network structure model in step S3 to obtain a human chest acupoint detection model;

[0061] When using the training set to train the YOLOv10 network structure model, epochs is set to 100, batch is set to 2, imagsz is set to 640, the depth of the model is set to 0.67, the width is set to 0.75, and the maximum number of channels is set to 768;

[0062] Use the training set and the YOLOv10 network structure model for training, and obtain the weight file of the trained YOLOv10 network structure model and the mAP value of the YOLOv10 network structure model on the training set; use the training set and the ACU-YOLO network structure model for training, and obtain the weight file of the trained YOLOv10+PSMASH+MSBlock (ACU-YOLO) network structure model and the mAP value of the ACU-YOLO network structure model on the training set; the trained YOLOv10+PSMASH+MSBlock (ACU-YOLO) network structure model is the human chest acupuncture point detection model, and the training results of the YOLOv10 network structure model and the YOLOv10+PSMASH+MSBlock (ACU-YOLO) network structure model are compared as shown in Table 1:

[0063] Table 1 Training results comparison table

[0064]

[0065] Among them, mAP is the average of the average precision (AP) of multiple categories; AP is an evaluation indicator for a single category, which calculates the average precision (Precision) at different recall levels; mAP is obtained by calculating the AP of each category and then averaging the AP of all categories; the higher the mAP value, the better the overall performance of the model in detecting targets of each category; mAP50 (B) represents the target box (Bo) calculated by IoU when the threshold is set to 0.5 x), mAP50-95(B) represents the average mAP value of the target box (Box) calculated by IoU when the step size is 0.05 and the threshold is set to 0.5 to 0.95; mAP50(P) represents the mAP value of the key point (P) calculated by IoU when the threshold is set to 0.5, mAP50-95(P) represents the average mAP value of the key point (P) calculated by IoU when the step size is 0.05 and the threshold is set to 0.5 to 0.95, FPS is the number of frames per second, that is, the prediction speed;

[0066] The training results of the YOLOv10+PSMASH+MSBlock (ACU-YOLO) network structure model are as follows Figure 5 As shown in the figure, the mAP value of the ACU-YOLO network structure model on the training set is higher than that of the YOLOv10 network structure model;

[0067] Step S5: using the human chest acupuncture point detection model in step S4 to test the test set;

[0068] The trained YOLOv10 network structure model weight file and test set are used for testing to obtain the error distance between the predicted position and the actual position of the acupoints to be detected, as well as the average error distance of the acupoints of this model, the maximum error acupoint error distance, and the minimum error acupoint error distance;

[0069] The weight file and test set of the trained YOLOv10+PSMASH+MSBlock (ACU-YOLO) network structure model are used for testing to obtain the error distance between the predicted position and the actual position of the acupoints to be detected, as well as the average error distance of the acupoints, the maximum error acupoint error distance, and the minimum error acupoint error distance of this model;

[0070] The corresponding errors of the detection results of each acupoint by the YOLOv10 network structure model and the YOLOv10+PSMASH+MSBlock (ACU-YOLO) network structure model are shown in Table 2:

[0071] Table 2 Corresponding error table of each acupoint detection result (unit: cm)

[0072]

[0073] The unit of the error value is centimeters. The conversion basis is as follows: the depth camera that collects data collects data at a fixed height, measures the length and width of the area corresponding to the position it photographs, and the image resolution is 1920x1200, and obtains the conversion constant between the pixel distance and the actual distance; the pixel error is obtained by dividing the pixel error by the conversion constant to obtain the centimeter error value (retain two decimal places);

[0074] Table 3 Comprehensive error table of acupoint detection of different models (unit: cm)

[0075]

[0076] Among them, SP15, ST25, KI16, RN8, RN16, RN22, ST171, ST172 are the international numbers of eight chest acupuncture points; the max column corresponds to the error distance of the acupuncture point with the largest error among the eight acupuncture points predicted by the model; the average column corresponds to the average error distance of the eight acupuncture points predicted by the model; the min column corresponds to the error distance of the acupuncture point with the smallest error among the eight acupuncture points predicted by the model; the error of the network proposed in the present invention in acupuncture point detection is smaller than that of the YOLOv10 network, and the accuracy is improved; the actual maximum acupuncture point error is controlled within 1.38cm, which has application value;

[0077] In order to adapt to the acupoint detection task, the detection heads of YOLOv10 and ACU-YOLO in the present invention are both key point detection heads.

[0078] In summary, the ACU-YOLO network structure model established by the present invention uses the improved method PSMASH to replace the PSA attention mechanism in the Backbone structure of the YOLOv10 network structure model, adds the MSBlock module to the Neck part, designs the YOLOv10 key point detection module, and applies it to acupoint recognition. The ACU-YOLO network structure model is used to detect acupoints on the front chest of the human body. Compared with the existing YOLO detection algorithm, the accuracy rate of ACU-YOLO detection is higher, and the detection rate of ACU-YOLO far exceeds other detection algorithms such as MMPose. The maximum acupoint prediction error of the ACU-YOLO network structure model is controlled within 1.38cm, and the detection rate is 108.6FPS, so it has practical application value.

[0079] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. An ACU-YOLO deep learning method for rapid human acupoint recognition, characterized in that The following steps are involved: Step S1: Collecting a data set of acupuncture points on the human chest; Step S2: preprocessing and labeling the human chest acupoint dataset in step S1, and then dividing the human chest acupoint dataset into a training set for training, a verification set for verification, and a test set for testing; Step S3: Build a YOLOv10 network structure model, improve the YOLOv10 network framework, and obtain the ACU-YOLO network structure model; Step S4: using the training set in step S2 to train the ACU-YOLO network structure model in step S3 to obtain a human chest acupuncture point detection model; Step S5: Using the human chest acupuncture point detection model in step S4, the test set is tested.

2. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 1, characterized in that: The step S2 comprises the following steps: Step S201: pre-processing the human chest acupuncture point data, and adjusting the size of each image in the human chest acupuncture point data set to 1920*1200; Step S202: using the image tool Labelme to annotate each image in the human chest acupoint dataset preprocessed in step S201, obtain an annotation file in JSON format, and convert the format to obtain the dataset required for YOLOv10 training; Step S203: marking acupoints, including Daheng point SP15, Tianshu point ST25, Mangshu point KI16, Shenque point RN8, Zhongting point RN16, Tiantu point RN22, Zuoruzhong point ST171, and Youruzhong point ST172; Step S204: Divide the data set required for ACU-YOLO training into a training set, a validation set, and a test set in a ratio of 8:1:

1.

3. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 2, characterized in that: The step S3 comprises the following steps: Step S301: Build a YOLOv10 network structure model, including an input layer, a Backbone structure for extracting image features, a Neck structure for further processing and fusing features, and an output layer Head; the Backbone structure includes a Conv structure, a C2f structure, a ScDown structure, and an SPPF structure, and the Neck structure includes a C2f structure and a C2fCIB structure; Step S302: Add the PSMASH attention mechanism module to the Backbone structure; Step S303: adding MSBlock multi-scale blocks to the Neck structure; Step S304: Integrate the improved Backbone and Neck structures to obtain the ACU-YOLO network structure model.

4. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 3, characterized in that: In step S302, the attention mechanism module PSMASH replaces the original channel attention with MSCA, which enables the channel attention to have multi-scale performance, based on the CBAM module. MSCA includes deep convolution to aggregate local information, multi-branch deep convolution to capture multi-scale context, and 1x1 convolution to model the relationship between different channels.

5. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 4, characterized in that: In step S303, the MSBlock multi-scale block adopts a hierarchical feature fusion strategy, replacing the standard 3×3 convolution with an inverted bottleneck layer to utilize large kernel convolution to enhance the ability of the real-time object detector to extract multi-scale features.

6. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 5, characterized in that: The improved Backbone Network in the ACU-YOLO network structure model obtained in step S304 consists of a Conv structure, a C2f structure, a ScDown structure, an SPPF structure and a PSMASH structure, wherein the PSMASH structure is added after the SPPF; the improved Neck Network is connected to the MSBlock structure after C2f and C2fCIB respectively.

7. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 6, characterized in that: In step S4, when the YOLOv10 network structure model is trained using the training set, epochs is set to 100, batch is set to 2, imagsz is set to 640, the depth of the model is set to 0.67, the width is set to 0.75, and the maximum number of channels is set to 768.

8. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 7, characterized in that: In step S4, the training set and the YOLOv10 network structure model are used for training to obtain the weight file of the trained YOLOv10 network structure model and the mAP value of the YOLOv10 network structure model on the training set; the training set and the ACU-YOLO network structure model are used for training to obtain the weight file of the trained ACU-YOLO network structure model and the mAP value of the ACU-YOLO network structure model on the training set; The trained ACU-YOLO network structure model is a human chest acupoint detection model.

9. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 8, characterized in that: In step S5, the weight file of the trained YOLOv10 network structure model and the test set are used for testing to obtain the error distance between the predicted position and the actual position of the acupoints to be detected, as well as the average error distance of the acupoints of this model, the maximum error acupoint error distance, and the minimum error acupoint error distance; The weight file and test set of the trained ACU-YOLO network structure model are used for testing to obtain the error distance between the predicted position and the actual position of the required acupoint detection point, as well as the average error distance of the acupoints of this model, the maximum error acupoint error distance, and the minimum error acupoint error distance.

10. The ACU-YOLO deep learning method for rapid human acupoint recognition according to claim 3, characterized in that: In step S301, a key point detection head is applied in the YOLOv10 framework. The key point detection head acts as a decoder to extract the category information and location information of the acupoints and output the final acupoint prediction results to expand the application direction of acupoint detection.

Images