Method for accurately judging acupuncture points based on artificial intelligence

Abstract

The invention provides an accurate acupoint judgment method based on artificial intelligence, and the method comprises the steps: 1, carrying out the semantic segmentation of a human body back image, and constructing a high-quality human body back image semantic segmentation data set; 2, ResNet50 + FPN constructs an RPN to generate a candidate area network, then RoIAlign is used for accurately extracting features of a candidate area, the accuracy of subsequent classification and mask is ensured, and the accuracy of acupoint recognition is improved; 3, outputting accurate judgment of the acupuncture points, and calibrating a received prediction result; and step 4, aligning and combining the image data of the two different modes in space, and carrying out comprehensive analysis of multi-mode data and temperature information extraction of a specific area. According to the method, a high-precision bounding box and a segmentation mask can be provided, the robot can accurately recognize the acupuncture points in a complex environment, the possibility of recognition errors is reduced, adaptability is high, and adjustment can be conducted according to physiological features of different individuals.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent robots, and in particular relates to a method for accurately determining acupoints based on artificial intelligence. Background Art

[0002] In traditional Chinese medicine, moxibustion is widely used to treat a variety of diseases and relies on the accurate identification and stimulation of specific acupuncture points on the body. Each person's physiological structure, body shape and acupuncture point location may be different, making standardized acupuncture point positioning difficult. Existing acupuncture positioning methods are mostly manual operations, lacking scientific basis and systematicity. The experience and skill level of acupuncturists directly affect the treatment effect. Traditional acupuncture point positioning lacks objective quantitative standards and is difficult to combine with the diagnostic standards of modern medicine, which affects the promotion and acceptance of acupuncture therapy.

[0003] In order to solve these problems, in recent years, scientists and researchers have begun to explore acupoint identification methods based on modern technology; by combining computer vision, sensor technology and machine learning, they have developed an automated acupoint recognition system to improve the accuracy and efficiency of determination; the goal of this technology is to combine the traditional wisdom of Chinese medicine moxibustion with modern technology to provide more scientific and reliable treatment options. Summary of the invention

[0004] In order to solve the above-mentioned problems existing in the prior art, the present invention proposes a method for accurately determining acupoints based on artificial intelligence to address the problem of inaccurate acupoint identification. This method firstly labels and semantically segments the collected images for acupoints, then trains the data set through the Mask R-CNN algorithm, and finally comprehensively analyzes the locations of acupoints through multimodal fusion alignment of visible light images and infrared images. This method can improve the accuracy of acupoint identification and identify the acupoint locations more accurately.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions:

[0006] A method for accurately determining acupoints based on artificial intelligence, characterized in that it comprises the following steps:

[0007] Step 1: Perform semantic segmentation on the human back image and build a high-quality human back image semantic segmentation dataset. This step includes the following steps:

[0008] S1: Data Collection and Processing:

[0009] Each person's physiological structure, body shape and acupoint location may be different. A large number of human back images are collected, covering different body shapes, skin colors, postures and backgrounds.

[0010] The images are processed uniformly, adjusted to a uniform size of 480*640, and mean filtering is used to remove noise from the images to meet the model input requirements.

[0011] S2: Create a human back acupuncture point dataset:

[0012] Combined with traditional Chinese medicine knowledge, the calibration standard of the back acupoints is determined: the line connecting the left and right Tianzhu acupoints is selected as the dividing line between the human head and neck, the line connecting the Jianyu acupoint and the Jugu acupoint is selected as the dividing line between the human back and the arm, and the line connecting the Yaoyangguan acupoint, the Xiaochang acupoint and the Baoxue acupoint is selected as the dividing line between the human back and the buttocks; since the acupoints on the back of the human body are mainly concentrated in the middle and right sides of the back, the present invention selects the line connecting the Dazhui acupoint and the Yaoyangguan acupoint, the line connecting the Daliao acupoint and the Xiaochang acupoint, and the line connecting the Pohu acupoint and the Baoxue acupoint for semantic segmentation.

[0013] The Tianzhu, Daqiao and Dazhui acupoints in the human back image are calibrated, and then the lines of the acupoint distribution on the human back are drawn according to their positions to preliminarily segment the human back.

[0014] Orange represents the head, dark red represents the neck, blue represents the left side of the back, green represents the arm, red, cyan and light yellow represent the right side of the back, and pink represents below the waist; the acupuncture point on the back of the human body is located on the right side of the back, that is, the red, cyan and light yellow area.

[0015] Step 2: Mask R-CNN algorithm trains acupoint recognition, which includes the following steps:

[0016] S1: Select backbone network:

[0017] ResNet50 uses residual blocks to alleviate the gradient vanishing problem in deep networks through jump connections. Each residual block consists of two or three convolutions connected by identity mapping, allowing gradients to flow directly through the network, which can learn more complex features.

[0018] FPN processes objects of different sizes by constructing multi-scale feature maps. The network generates feature maps of multiple scales so that it can process large and small objects at the same time. On multiple datasets, FPN significantly improves the performance of object detection models, especially in the detection of small objects.

[0019] In the present invention, the requirement for acupoint recognition accuracy is relatively high, so ResNet50 is combined with FPN to utilize the ability of ResNet50 to extract deep features, and at the same time, FPN is used to generate multi-scale feature maps.

[0020] S2: Building a regional recommendation network RPN:

[0021] RPN receives feature maps from the backbone network (ResNet50 and FPN) and defines multiple anchors on the feature maps. Anchor boxes are predefined bounding boxes with different sizes and aspect ratios that can cover potential targets (acupoints).

[0022] By using a small convolutional network on the feature map, the output of the convolutional layer will be used to generate classification and regression information for the anchor boxes.

[0023] The classification branch outputs the objectness score of each anchor box and predicts whether the anchor box contains the target (acupoint); the regression branch outputs the bounding box regression parameters of each anchor box, which are used to adjust the position and size of the anchor box to better fit the actual acupoint position.

[0024] The classification loss is calculated using cross entropy loss; assuming there is a true label for an anchor box (in Indicates that the anchor box contains the target, Indicates that the anchor box does not contain the target), and the predicted probability output by the model .

[0025]

[0026] in is the total number of anchor boxes, It is The true labels of the anchor boxes, The model predicts The probability of an anchor box.

[0027] Regression loss uses smoothing Loss is calculated; assuming the true bounding box position is and the predicted bounding box location is .

[0028]

[0029]

[0030] is the difference between the true bounding box and the predicted bounding box, is the number of valid anchor boxes.

[0031] The loss function of RPN is the sum of classification loss and regression loss:

[0032]

[0033] in is a hyperparameter used to balance the importance of the two.

[0034] The final output of RPN is the processed candidate region, which will be passed to the subsequent RoIAlign layer for further feature extraction and classification and mask generation.

[0035] During the training process, RPN learns how to generate suitable candidate regions through a large number of annotated images, which correspond to the acupuncture points in the image.

[0036] S3: Use RoIAlign to accurately extract the features of the candidate region to ensure the accuracy of subsequent classification and masking:

[0037] Compared with traditional RoIPooling, RoIAlign avoids quantization errors and can maintain the spatial accuracy of features, which is particularly important in the detection of acupoints.

[0038] The candidate regions generated by RPN are taken as input, and each candidate region is represented by its position on the feature map in the coordinate format .

[0039] According to the coordinates of the candidate region, its position on the feature map is determined; each candidate region is divided into a fixed number of grid cells, and for each grid cell, a bilinear interpolation method is used to obtain the corresponding eigenvalue from the feature map to ensure that the extracted features are aligned with the original image instead of losing information due to downsampling. Finally, the features of all grid cells in each candidate region are spliced ​​together to form a feature vector of fixed size.

[0040] The extracted feature vector is input into the subsequent network layer for target classification and mask generation; the classification branch outputs the category probability of each candidate region, and the mask branch outputs a binary mask to represent the specific shape within the region.

[0041] Based on the results of classification and mask generation, the loss is calculated for backpropagation and model update.

[0042] The formula for mask loss is:

[0043]

[0044] is the number of pixels in each candidate region, It is The true label of each pixel (1 for the target, 0 for the background), The model generates The predicted probability of each pixel.

[0045] Combine the classification loss and mask loss to form the total loss:

[0046]

[0047] During the training process, RoIAlign helps the model accurately extract the features of each candidate region from the feature map, improving the accuracy of classification and mask generation.

[0048] Step 3: Output of accurate acupoint determination, which includes the following steps:

[0049] S1: Receive prediction results and calibrate:

[0050] The system receives the judgment results of acupoints from the deep learning network, including information such as the coordinate position and category of the acupoints.

[0051] The received prediction results are calibrated to ensure that the results are consistent with the actual human anatomy, and the position is adjusted using the standard acupoint map to reduce deviations.

[0052] S2: Generate visualization data:

[0053] The predicted acupoint locations are marked on the original image, and the precise coordinates and names of the acupoints are displayed in the back image.

[0054] Step 4: Multimodal image fusion alignment: the image data of two different modalities are spatially aligned and combined. This step includes the following steps:

[0055] S1: Spatial alignment and coordinate unification of multi-source images:

[0056] Individuals have different physiological characteristics, body shapes, postures and other factors, and the relative positions of acupoints vary. The acupoint positions of different people may differ; therefore, the present invention adopts a multimodal image fusion and alignment method to ensure more accurate identification of acupoints on the back of the human body.

[0057] The human back image and infrared image come from different sensors and have different viewing angles, resolutions, and imaging characteristics. To ensure that images from different sources correspond to the same part in space, the feature point alignment method is used to detect the key point of the human body, the spine, and then the two images are superimposed in a coordinate system.

[0058] S2: Spatial correspondence and matching between infrared temperature data and visible light images:

[0059] The relative position, angle and scaling relationship between the infrared and visible light images are obtained through the camera's internal and external parameter matrix; then the corresponding feature point spine is found in the two images through the feature matching algorithm SIFT to ensure that the same features are correctly matched in the two images and establish the spatial relationship between the infrared and visible light images.

[0060] Using the matched feature points, the coordinate system of the infrared image is converted to the coordinate system of the visible light image; using affine transformation, the infrared image is adjusted to a position aligned with the visible light image to ensure that the temperature data of each pixel is aligned with the corresponding part on the visible light image.

[0061] Finally, the temperature data in the infrared image is superimposed on the corresponding area of ​​the visible light image. Through pseudo-color mapping, the infrared temperature information is displayed in the form of a transparent overlay on the visible light image, making the temperature data intuitively visible.

[0062] S3: Comprehensive analysis of multimodal data and extraction of temperature information in specific areas:

[0063] The acupuncture points identified and marked in the human back image are the target locations for regional temperature extraction. The temperature data of each area is analyzed to extract key temperature information, such as average temperature, maximum temperature, minimum temperature, etc. At the same time, the standard deviation of the temperature is calculated, the temperature distribution and uniformity are analyzed, and it is determined whether there are obvious high and low temperature areas.

[0064] Acupoints are closely related to specific physiological functions or states. Some acupoints may show abnormal temperature changes under certain circumstances (such as inflammation, pain or other pathological conditions); by detecting these temperature changes, auxiliary basis can be provided for the selection and judgment of acupoints.

[0065] Set temperature thresholds to identify temperature areas that deviate significantly from normal values. If the temperature of an area is significantly lower than the surrounding area, the acupoints there may need moxibustion treatment.

[0066] By combining machine vision to identify acupuncture points on the back of the human body with infrared image temperature anomaly detection, the location of the acupuncture points can be determined more accurately.

[0067] Beneficial effects:

[0068] 1. The invention combines the temperature data of images recognized by machine vision and infrared images to achieve a deep fusion of local temperature detection of the human body and acupoint positioning, thereby improving the accuracy of acupoint determination and reducing errors caused by human factors.

[0069] 2. The present invention has strong adaptability and can be flexibly adjusted according to the physiological characteristics, physical differences and individual needs of different individuals, thereby ensuring that it is suitable for various types of usage scenarios and groups, and providing a more personalized and precise service experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of acupoint demarcation and segmentation;

[0071] Figure 2 It is a diagram of the semantic segmentation annotation process;

[0072] Figure 3 It is the process of acupoint identification;

[0073] Figure 4 It is a schematic diagram of infrared image of human back;

[0074] Figure 5 It is a schematic diagram of the mixture of infrared image and visible light image;

[0075] Figure 6 It is a schematic diagram of visible light image temperature display. DETAILED DESCRIPTION

[0076] The following is a further detailed description in conjunction with specific embodiments of the invention.

[0077] First, each step is explained as a whole: a method for accurately determining acupoints based on artificial intelligence, characterized in that it includes the following steps:

[0078] A large number of human back images are collected, covering different body shapes, skin colors, postures and backgrounds; the collected images are processed uniformly, the images are adjusted to a uniform size of 480*640, and mean filtering is used to remove noise in the images to improve image quality to meet the model input requirements.

[0079] Select Labelme software, open the processed image that needs to be labeled, and determine the calibration standard of the back acupoints in combination with traditional Chinese medicine knowledge; select Create point to start labeling the acupoints that need to be labeled, select the left and right Tianzhu points as the dividing line between the human head and neck, select the Bishu point and the Jugu point as the dividing line between the human back and the arm, select the Yaoyangguan point, the Xiaochang point and the Baoxue point as the dividing line between the human back and the buttocks; select the Dazhui point and the Yaoyangguan point, the Daqiao point and the Xiaochang point, and the Pohu point and the Baoxue point for semantic segmentation.

[0080] After labeling and preliminary segmentation, the images are saved as json files. The coordinates of all targets in a photo are saved in shapes. Each element in the list corresponds to a target, where Label records the category name of the target; Points records the coordinate information of a target.

[0081] Next, we use the high-precision model hrnet18_ocr64_human of the eiseg software to perform semantic segmentation on this dataset; orange represents the head, dark red represents the neck, blue represents the left side of the back, green represents the arm, red, cyan, and light yellow represent the right side of the back, and pink represents the area below the waist; the acupuncture points on the back of the human body are located on the right side of the back, that is, the red, cyan, and light yellow areas.

[0082] ResNet50 uses residual blocks to alleviate the gradient vanishing problem in deep networks through skip connections. Each residual block consists of two or three convolutions connected by identity mapping, allowing the gradient to flow directly through the network.

[0083] In the FPN network structure, the feature matrix generated by each CONV is recorded as C2, C3, C4, and C5, and then a 1*1 convolution is performed on the obtained feature map to adjust the channel, and then the feature map is fused from C5 to C2; ResNet50 is combined with FPN, using the ability of ResNet50 to extract deep features, and at the same time, multi-scale feature map generation is achieved through FPN.

[0084] RPN receives feature maps from the backbone network (ResNet50 and FPN) and defines multiple anchors on the feature maps. Anchor boxes are predefined bounding boxes with different sizes and aspect ratios that can cover potential targets (acupoints).

[0085] By using a small convolutional network on the feature map, the output of the convolutional layer will be used to generate classification and regression information for the anchor boxes.

[0086] The classification branch outputs the objectness score of each anchor box and predicts whether the anchor box contains the target (acupoint); the regression branch outputs the bounding box regression parameters of each anchor box, which are used to adjust the position and size of the anchor box to better fit the actual acupoint position.

[0087] The classification loss is calculated using cross entropy loss; assuming there is a true label for an anchor box (in Indicates that the anchor box contains the target, Indicates that the anchor box does not contain the target), and the predicted probability output by the model .

[0088]

[0089] in is the total number of anchor boxes, It is The true labels of the anchor boxes, The model predicts The probability of an anchor box.

[0090] The regression loss is calculated using smooth L1 loss; assuming the true bounding box position is and the predicted bounding box location is .

[0091]

[0092]

[0093] is the difference between the true bounding box and the predicted bounding box, is the number of valid anchor boxes (only those with true targets are included).

[0094] The loss function of RPN is the sum of classification loss and regression loss:

[0095]

[0096] in is a hyperparameter used to balance the importance of the two.

[0097] After generating the candidate regions, the non-maximum suppression algorithm is used to remove the anchor boxes with large overlaps and only retain the candidate regions with the highest scores, ensuring that the candidate regions output in the end are unique and accurate.

[0098] The final output of RPN is the processed candidate region, which will be passed to the subsequent RoIAlign layer for further feature extraction and classification and mask generation.

[0099] During the training process, RPN learns how to generate suitable candidate regions through a large number of annotated images, which correspond to the acupuncture points in the image.

[0100] The candidate regions generated by RPN are taken as input, and each candidate region is represented by its position on the feature map in the coordinate format .

[0101] According to the coordinates of the candidate region, its position on the feature map is determined; each candidate region is divided into a fixed number of grid cells, and for each grid cell, a bilinear interpolation method is used to obtain the corresponding eigenvalue from the feature map to ensure that the extracted features are aligned with the original image instead of losing information due to downsampling. Finally, the features of all grid cells in each candidate region are spliced ​​together to form a feature vector of fixed size.

[0102] The extracted feature vector is input into the subsequent network layer for target classification and mask generation; the classification branch outputs the category probability of each candidate region, and the mask branch outputs a binary mask to represent the specific shape within the region.

[0103] Based on the results of classification and mask generation, the loss is calculated for backpropagation and model update.

[0104] The formula for mask loss is:

[0105]

[0106] is the number of pixels in each candidate region, It is The true label of each pixel (1 for the target, 0 for the background), The model generates The predicted probability of each pixel.

[0107] Combine the classification loss and mask loss to form the total loss:

[0108]

[0109] During the training process, RoIAlign helps the model accurately extract the features of each candidate region from the feature map, improving the accuracy of classification and mask generation.

[0110] The system receives the judgment results of acupoints from the deep learning network, including information such as the coordinate position and category of the acupoints.

[0111] The received prediction results are calibrated to ensure that the results are consistent with the actual human anatomy, and the position is adjusted using the standard acupoint map to reduce deviations.

[0112] The predicted acupoint information is displayed graphically, the predicted acupoint locations are marked on the original image, and the precise coordinates and names of the acupoints are displayed in the back image.

[0113] The human back image and infrared image come from different sensors and have different viewing angles, resolutions, and imaging characteristics. If they are not aligned, the same position in the two images will not be consistent in pixel coordinates, which will cause data misalignment and data errors. Therefore, it is necessary to ensure that images from different sources correspond to the same part in space. This is achieved by using the feature point alignment method, which detects the key point of the human body, the spine, and superimposes the two images in a coordinate system.

[0114] Using the matched feature points, the coordinate system of the infrared image is converted to the coordinate system of the visible light image; using affine transformation, the infrared image is adjusted to a position aligned with the visible light image to ensure that the temperature data of each pixel is aligned with the corresponding part on the visible light image.

[0115] The temperature data in the infrared image is superimposed on the corresponding area of ​​the visible light image. Through pseudo-color mapping, the infrared temperature information is displayed in the form of a transparent overlay on the visible light image, making the temperature data intuitively visible.

[0116] The acupuncture points identified and marked in the human back image are the target locations for regional temperature extraction; the temperature data of each area is analyzed to extract key temperature information, such as average temperature, maximum temperature, minimum temperature, etc.; at the same time, the standard deviation of the temperature is calculated, the temperature distribution and uniformity are analyzed, and it is determined whether there are obvious high and low temperature areas; by detecting these temperature changes, an auxiliary basis can be provided for the selection and determination of acupuncture points.

[0117] By combining machine vision to identify acupuncture points on the back of the human body with infrared image temperature anomaly detection, the location of the acupuncture points can be determined more accurately.

[0118] The method proposed in the present invention performs semantic segmentation on the area where the acupoints on the back of the human body are concentrated. The method first marks the acupoints on the edge of the back, uses a high-precision model suitable for the human body for mask marking, and uses the Mask R-CNN algorithm to identify the area where the acupoints on the back of the human body are concentrated. Finally, by combining the image recognized by machine vision with the temperature data of the infrared image, the deep fusion of local temperature detection of the human body and acupoint positioning can be achieved, which improves the accuracy of acupoint determination and reduces the errors caused by human factors.

Claims

1. A method for accurately determining acupoints based on artificial intelligence, characterized in that: The following steps are involved: Step 1: Perform semantic segmentation on the human back image and build a high-quality human back image semantic segmentation dataset. This step includes the following two steps: S1: Data Collection and Processing: Each person's physiological structure, body shape and acupuncture point location may be different. A large number of human back images covering different body shapes, skin colors, postures and backgrounds are collected and processed uniformly. S2: Create a human back acupuncture point dataset: According to the acupuncture point location method prescribed by traditional Chinese medicine, the acupuncture points such as Tianzhu, Daqiao and Dazhui on the back of the human body are calibrated, and then the lines of the acupuncture points on the back of the human body are drawn according to the positions to perform preliminary segmentation of the back of the human body; Step 2: Mask R-CNN algorithm trains acupoint recognition, which includes the following steps: S1: Select backbone network: ResNet50 is combined with FPN. ResNet50 extracts deep features, while FPN generates multi-scale feature maps. S2: Building a regional recommendation network RPN: RPN receives feature maps from the backbone network (ResNet50 and FPN) and defines multiple anchors on the feature maps with different sizes and aspect ratios to cover potential targets; S3: Use RoIAlign to accurately extract the features of the candidate region to ensure the accuracy of subsequent classification and masking: The candidate regions generated by RPN are taken as input, and each candidate region is represented by its position on the feature map in the coordinate format ; Step 3: Output of accurate acupoint determination, which includes the following steps: S1: Receive prediction results and calibrate: The system receives the determination results of acupoints from the deep learning network, including information such as the coordinate position and category of the acupoints; Calibrate the received prediction results to ensure that the results are consistent with the actual human anatomy, and use the standard acupoint map to adjust the position and reduce deviation; S2: Generate visualization data: The predicted acupoint information is displayed graphically, the predicted acupoint locations are marked on the original image, and the precise coordinates and names of the acupoints are displayed in the back image; Step 4: Multimodal image fusion alignment: the image data of two different modalities are spatially aligned and combined. This step includes the following steps: S1: Spatial alignment and coordinate unification of multi-source images: Individuals have different physiological characteristics, body shapes, postures and other factors, which lead to changes in the relative positions of acupoints. The acupoint positions of different people may be different. The present invention adopts a multimodal image fusion and alignment method to more accurately identify acupoints on the back of the human body. S2: Spatial correspondence and matching between infrared temperature data and visible light images: The relative position, angle and scaling relationship between the infrared and visible light images are obtained through the camera's internal and external parameter matrix; then the corresponding feature point spine is found in the two images through the feature matching algorithm SIFT to ensure that the same features are correctly matched in the two images and establish the spatial relationship between the infrared and visible light images; S3: Comprehensive analysis of multimodal data and extraction of temperature information in specific areas: The acupuncture points identified and marked in the human back image are the target locations for regional temperature extraction; the temperature data of each region is analyzed to extract key temperature information.

2. The method for accurately determining acupoints based on artificial intelligence according to claim 1, characterized in that: In step 1, S2, the initial segmentation of the back of the human body is performed: The line connecting the two Tianzhu points on the left and right is selected as the dividing line between the head and the neck of the human body, the line connecting the Jianyu point and the Jugu point is selected as the dividing line between the back and the arm of the human body, and the line connecting the Yaoyangguan point, the Xiaochang point and the Baoxue point is selected as the dividing line between the back and the buttocks of the human body; since the acupuncture points on the back of the human body are mainly concentrated in the middle and right sides of the back, the present invention selects the line connecting the Dazhui point and the Yaoyangguan point, the line connecting the Daqiao point and the Xiaochang point, and the line connecting the Pohu point and the Baoxue point for semantic segmentation; The various areas inside the polygon are marked, with orange representing the head, dark red representing the neck, blue representing the left side of the back, green representing the arms, red, cyan and light yellow representing the right side of the back, and pink representing below the waist; The acupuncture points on the back of the human body are concentrated on the right side of the back, namely the red, cyan and light yellow areas.

3. The method for accurately determining acupoints based on artificial intelligence according to claim 1, characterized in that: In step 2, in S2, when building a regional network: By using a small convolutional network on the feature map, the output of the convolutional layer will be used to generate classification and regression information for the anchor box; The classification loss is calculated using cross entropy loss; assuming there is a true label for an anchor box (in Indicates that the anchor box contains the target, Indicates that the anchor box does not contain the target), and the predicted probability output by the model ; Regression loss uses smoothing Loss is calculated; assuming the true bounding box position is and the predicted bounding box location is ; The loss function of RPN is the sum of classification loss and regression loss:

4. The method for accurately determining acupoints based on artificial intelligence according to claim 1, characterized in that: In step 2, S3, RoIAlign is used to accurately extract the features of the candidate region: According to the coordinates of the candidate region, its position on the feature map is determined; each candidate region is divided into a fixed number of grid cells, and for each grid cell, the corresponding feature value is obtained from the feature map using the bilinear interpolation method to ensure that the extracted features are aligned with the original image instead of losing information due to sampling. Finally, the features of all grid cells in each candidate region are spliced ​​together to form a fixed-size feature vector; The extracted feature vector is input into the subsequent network layer for target classification and mask generation; Based on the results of classification and mask generation, the loss is calculated for backpropagation and model update; The formula for mask loss is: Combine the classification loss and mask loss to form the total loss:

5. The method for accurately determining acupoints based on artificial intelligence according to claim 1, characterized in that: In step 4, S2, the spatial correspondence and matching between the infrared temperature data and the visible light image: Using affine transformation, the infrared image is adjusted to a position aligned with the visible light image, ensuring that the temperature data of each pixel is aligned with the corresponding part on the visible light image; Finally, the temperature data in the infrared image is superimposed on the corresponding area of ​​the visible light image. Through pseudo-color mapping, the infrared temperature information is displayed in the form of a transparent overlay on the visible light image, making the temperature data intuitively visible.

6. The method for accurately determining acupoints based on artificial intelligence according to claim 1, characterized in that: In step 4, S3, comprehensive analysis of multimodal data and extraction of temperature information in specific areas: Set temperature thresholds to identify temperature areas that deviate significantly from normal values. If the temperature of an area is significantly lower than the surrounding area, the acupoints may need moxibustion treatment. By combining machine vision to identify acupuncture points on the back of the human body with infrared image temperature anomaly detection, the location of the acupuncture points can be determined more accurately.

Images