Human body meridian acupoint detection method and device, electronic equipment and storage medium

Abstract

The application provides a human meridian and acupoint detection method and device, electronic equipment and storage medium, and relates to the technical field of computer vision.The method comprises the following steps: obtaining a to-be-detected human image corresponding to a user; inputting the to-be-detected human image into a meridian and acupoint detection model to obtain a meridian and acupoint detection result output by the meridian and acupoint detection model; the meridian and acupoint detection model is obtained based on human image samples and heat map true values corresponding to the human image samples; the Gaussian kernel standard deviation corresponding to the heat map true value is adaptively determined based on different geometric shapes of human body regions in the human image samples and key point densities in the human image samples; and the meridian and acupoint detection model is used for decoupling each meridian group region in the to-be-detected human image in spatial position and channel respectively.The application can adaptively adjust a receptive field and perform double decoupling in spatial position and channel respectively, effectively reduces false detection and missed detection, and improves acupoint detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method, device, electronic device, and storage medium for detecting acupoints along human meridians. Background Technology

[0002] Traditional Chinese medicine meridian therapy, as a primary non-drug treatment method, is based on the core theory that "the meridian system is the channel for the circulation of Qi and blood." It adjusts the functions of the internal organs by stimulating acupoints, achieving a dynamic balance of Qi and blood throughout the body. With the development of computer vision, using two-dimensional images to detect meridians has become a promising research method.

[0003] In existing technologies, keypoint detection methods based on heatmap regression are generally used for acupoint detection. These methods typically generate heatmaps using a fixed Gaussian kernel and represent all keypoints with a single heatmap. However, the receptive field size in the heatmap is controlled by preset parameters, making it difficult to adapt to moderate changes in the target area. For example, since acupoints along the human meridian possess rich feature and topological information, if the receptive field is small, the model can only learn feature information within a small area of ​​each acupoint. When this area is occluded, global contextual information is easily lost, leading to false positives and false negatives. If the receptive field is large, it leads to feature confusion among dense acupoints. Furthermore, in situations with high density of acupoints along the human meridian, a single heatmap causes mutual influence between different keypoints, resulting in false positives and false negatives, further reducing detection accuracy. Summary of the Invention

[0004] This invention provides a method, device, electronic device, and storage medium for detecting acupoints along human meridians, which solves the defects of the prior art in detecting acupoints using a fixed Gaussian kernel and a single heatmap, resulting in false detections and missed detections, leading to low detection accuracy.

[0005] This invention provides a method for detecting acupoints along human meridians, comprising the following steps.

[0006] Obtain the human image to be detected corresponding to the user; The human image to be detected is input into the meridian acupoint detection model to obtain the meridian acupoint detection results output by the meridian acupoint detection model. The meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values. The Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples. The meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0007] According to the method for detecting acupoints along human meridians provided by this invention, the true value of the heat map is determined based on the following steps: Based on the different geometric shapes of the human body region in the human body image sample, determine the first Gaussian kernel standard deviation corresponding to each geometric shape; The second Gaussian kernel standard deviation is determined based on the area and number of key points of the human image sample; the area and number of key points of the human image sample are used to characterize the key point density corresponding to the human image sample. Determine the optimal Gaussian kernel standard deviation based on all first Gaussian kernel standard deviations and second Gaussian kernel standard deviations; Based on the optimal Gaussian kernel standard deviation, the true value of the heatmap corresponding to the human image sample is determined.

[0008] According to the method for detecting acupoints along human meridians provided by the present invention, the step of determining the optimal Gaussian kernel standard deviation based on all first Gaussian kernel standard deviations and second Gaussian kernel standard deviations includes: The largest Gaussian kernel standard deviation among all the first Gaussian kernel standard deviations and the second Gaussian kernel standard deviations is determined as the optimal Gaussian kernel standard deviation.

[0009] According to the method for detecting acupoints along human meridians provided by the present invention, the acupoint detection model includes a feature extraction module, a heatmap decoupling module, and an acupoint detection module. The step of inputting the human image to be detected into the meridian acupoint detection model and obtaining the meridian acupoint detection results output by the meridian acupoint detection model includes: The human image to be detected is input into the feature extraction module to obtain a convolutional feature map output by the feature extraction module; the feature extraction module is used to determine the global features corresponding to the human image to be detected, and to perform convolution on the global features to obtain the convolutional feature map; The convolutional feature map is input into the heatmap decoupling module to obtain the fusion features corresponding to each of the meridian group regions output by the heatmap decoupling module; the heatmap decoupling module is used to decouple the convolutional feature map in terms of spatial location and channel respectively; the fusion features are used to characterize the spatial locality and semantic discriminativeness simultaneously possessed by each of the meridian group regions. The fusion features corresponding to each meridian group region are input into the acupoint detection module to obtain the meridian acupoint detection results output by the acupoint detection module.

[0010] According to the method for detecting acupoints along human meridians provided by the present invention, the heat map decoupling module includes a spatial recalibration module, a channel recalibration module, and a fusion module; The step of inputting the convolutional feature map into the heatmap decoupling module to obtain the fused features corresponding to each meridian group region output by the heatmap decoupling module includes: The convolutional feature map is input into the spatial recalibration module to obtain the spatial features corresponding to each meridian group region output by the spatial recalibration module; the spatial features are used to characterize the foreground localization of each meridian group region; the spatial recalibration module is used to determine the spatial features corresponding to each meridian group region based on the convolutional feature map and the spatial mask corresponding to each meridian group region; the spatial mask is used to characterize the spatial coordinates and feature similarity corresponding to each meridian group region; The convolutional feature map is input into the channel recalibration module to obtain the channel features corresponding to each meridian group region output by the channel recalibration module; the channel features are used to characterize the semantic information of each meridian group region; the channel recalibration module is used to determine the channel features corresponding to each meridian group region based on the convolutional feature map and the instance features of each meridian group region. The spatial features and channel features corresponding to each meridian group region are input into the fusion module to obtain the fusion features corresponding to each meridian group region output by the fusion module.

[0011] According to the method for detecting acupoints along human meridians provided by the present invention, the spatial mask is determined based on the following steps: The relative coordinate map and feature similarity map corresponding to each meridian group region are determined; the relative coordinate map is used to characterize the relative relationship between the spatial position of the acupoint in each meridian group region and the target center; the feature similarity map is the pixel-wise inner product between the convolutional feature map and the instance features of the corresponding meridian group region; the instance features of each meridian group region are obtained by partitioning the global features. Based on the relative coordinate map and feature similarity map corresponding to each meridian group region, the spatial mask corresponding to each meridian group region is determined.

[0012] According to the method for detecting acupoints of human meridians provided by the present invention, the geometric shape of the human body region includes the minimum circumscribed rectangle, the equivalent circle, and the maximum inscribed circle of the human body region.

[0013] The present invention also provides a human meridian acupoint detection device, comprising the following modules.

[0014] The acquisition module is used to acquire the human image to be detected corresponding to the user; The detection module is used to input the human image to be detected into the meridian acupoint detection model to obtain the meridian acupoint detection results output by the meridian acupoint detection model. The meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values. The Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples. The meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the human meridian acupoint detection method as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the human meridian acupoint detection method as described above.

[0017] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the human meridian acupoint detection method as described above.

[0018] This invention provides a method, device, electronic device, and storage medium for detecting acupoints along human meridians. It adaptively determines the ground truth of a heatmap by considering different geometric shapes of human regions and the density of key points in human image samples. A meridian acupoint detection model is trained using human image samples and the ground truth heatmap. The human image to be detected is input into the trained meridian acupoint detection model. The model decouples each meridian group region in the human image from its spatial location and channel, resulting in the output of the meridian acupoint detection results. In this invention, the standard deviation of the Gaussian kernel is adaptively adjusted based on the geometric shape of the human region and the density of key points. The optimal receptive field corresponding to the ground truth heatmap is determined based on the standard deviation of the Gaussian kernel. This guides the meridian acupoint detection model to fully learn the receptive field kernel and topological relationship information, significantly reducing false positives and false negatives when meridian acupoints are occluded. Meanwhile, the human body image to be detected is partitioned, and different meridian groups are decoupled in terms of spatial location and channel dimension. This effectively reduces interference between key points, the complexity of the global perspective, and the false detection and false negative detection of meridian acupoints, thereby improving the accuracy of meridian acupoint detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for detecting acupoints along human meridians provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the meridian acupoint detection model provided in the embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram showing the comparison results of the model before and after optimization provided in the embodiments of the present invention.

[0023] Figure 4 This is a schematic diagram of the human meridian acupoint detection device provided in an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] To address the problem of low accuracy in existing technologies that use fixed Gaussian kernels and single heatmaps for acupoint detection due to false positives and false negatives, this invention provides a method for detecting acupoints along human meridians. Figure 1 This is a flowchart illustrating the method for detecting acupoints along human meridians provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes steps 110 and 120.

[0027] Step 110: Obtain the human body image to be detected corresponding to the user.

[0028] Specifically, the human body image to be detected can be a supine or prone image of the human body, including at least one meridian group region containing acupoints. Each meridian group region corresponds to a part of the user's body. The meridian group region can include parts of the user's body such as the hands, feet, upper limbs, lower limbs, and torso. For example, the human body image to be detected can be a supine image including acupoints on the user's upper limbs, or a prone image including acupoints on the user's torso and lower limbs.

[0029] The human image to be detected can be taken by any camera or smart device deployed in the TCM meridian therapy room and uploaded to an electronic device.

[0030] Optionally, after acquiring the human body image to be detected, preprocessing operations such as image denoising, brightness adjustment, and size normalization can be performed on the human body image to be detected. The image resolution of the human body image to be detected can be 256×256 or 512×512, and this embodiment of the invention does not limit this.

[0031] Step 120: Input the human image to be detected into the meridian acupoint detection model to obtain the meridian acupoint detection results output by the meridian acupoint detection model; the meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values, and the Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples; the meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0032] Before executing step 120, a training dataset is constructed using human image samples and corresponding acupoint coordinate annotations from the supine human dataset. Based on the acupoint coordinate annotations, different geometric shapes of human regions in the human image samples, and keypoint densities, the optimal Gaussian kernel standard deviation is dynamically determined, and the ground truth heatmap value is determined based on this Gaussian kernel standard deviation. This ground truth heatmap value is a two-dimensional probability distribution map with the same image size as the human image sample. Each pixel value in this ground truth heatmap value represents the probability that an acupoint exists at that pixel. Then, the human image samples are input into the initial meridian acupoint detection model to obtain the predicted heatmap output by the initial meridian acupoint detection model. The initial meridian acupoint detection model is then trained under supervised supervision based on the ground truth heatmap value and the predicted heatmap until the model converges, resulting in the trained meridian acupoint detection model.

[0033] After training the meridian acupoint detection model and determining the human image to be detected, the human image to be detected is input into the trained meridian acupoint detection model. By drawing on the DCH (Decoupled Heatmap) concept, the meridian acupoint detection model extends the multi-person key point detection method to single-person multi-site acupoint detection. That is, the meridian acupoint detection model divides the human image to be detected into regions, obtaining at least two meridian groups. Acupoint detection is performed independently on each meridian group region to obtain the meridian acupoint detection results. In other words, by performing double decoupling on each meridian group region in both the spatial location and channel dimensions, spatial features are added to the acupoints in each meridian group region, reducing interference between symmetrical parts of the human body. At the same time, for meridian group regions with complex geometric structures and closely arranged acupoints, such as the meridian group regions corresponding to the user's hand or foot, double decoupling better captures subtle differences and avoids misclassification due to overlapping acupoints. In addition, when dealing with changes such as occlusion and rotation, double decoupling can accurately capture visible acupoints and infer the location of occluded or deformed acupoints through contextual information, making it more adaptable and enhancing the robustness and generalization ability of the model, ensuring the accuracy of the obtained meridian acupoint detection results.

[0034] Optionally, the meridian acupoint detection results may include the coordinates, confidence level, acupoint name, and meridian to which the acupoint belongs in the human image to be detected, etc., and the embodiments of the present invention do not limit this.

[0035] Optionally, an initial meridian acupoint detection model can be constructed using the HRNet backbone network, but this embodiment of the invention does not limit this.

[0036] This invention provides a method for detecting acupoints along human meridians. It adaptively determines the ground truth of a heatmap by considering different geometric shapes of human regions and the density of key points in human image samples. A meridian acupoint detection model is trained using the human image samples and the ground truth heatmap. The human image to be detected is input into the trained model, which decouples each meridian group region in the human image based on its spatial location and channel, resulting in the output of the meridian acupoint detection results. In this invention, the standard deviation of the Gaussian kernel is adaptively adjusted by considering the geometric shape of the human region and the density of key points. The optimal receptive field corresponding to the ground truth heatmap is determined based on the standard deviation of the Gaussian kernel. This guides the meridian acupoint detection model to fully learn the receptive field kernel and topological relationship information, significantly reducing false positives and false negatives when meridian acupoints are occluded. Meanwhile, the human body image to be detected is partitioned, and different meridian groups are decoupled in terms of spatial location and channel dimension. This effectively reduces interference between key points, the complexity of the global perspective, and the false detection and false negative detection of meridian acupoints, thereby improving the accuracy of meridian acupoint detection.

[0037] In one embodiment, the true value of the heatmap is determined based on the following steps: Based on the different geometric shapes of the human body region in the human body image sample, determine the first Gaussian kernel standard deviation corresponding to each geometric shape; The second Gaussian kernel standard deviation is determined based on the area and number of key points of the human image sample; the area and number of key points of the human image sample are used to characterize the key point density corresponding to the human image sample. Determine the optimal Gaussian kernel standard deviation based on all first Gaussian kernel standard deviations and second Gaussian kernel standard deviations; Based on the optimal Gaussian kernel standard deviation, the true value of the heatmap corresponding to the human image sample is determined.

[0038] Specifically, on the one hand, at the macroscopic scale of human body image samples, the standard deviation of the first Gaussian kernel corresponding to each geometric shape is determined based on the different geometric shapes of the human body region. The larger the geometric shape corresponding to the human body region, the larger the standard deviation of the first Gaussian kernel corresponding to the shape, which facilitates the initial meridian acupoint detection model to learn a wider range of acupoint features.

[0039] The geometric shape of the human body region includes its minimum bounding rectangle, equivalent circle, and maximum inscribed circle. The minimum bounding rectangle characterizes the span of the human body region in both the horizontal and vertical directions. The standard deviation of the first Gaussian kernel corresponding to the minimum bounding rectangle can be determined based on the width, height, diagonal length, or square root of the area corresponding to the minimum bounding rectangle. The equivalent circle can be a circle with the same pixel area as the human body region. The standard deviation of the first Gaussian kernel corresponding to this equivalent circle can be determined based on its radius or area. The maximum inscribed circle is the largest circle that can be contained within the human body region. The standard deviation of the first Gaussian kernel corresponding to this maximum inscribed circle can be determined based on its area.

[0040] On the other hand, regarding the microscopic distribution of key points (i.e., acupoints) in human image samples, the second Gaussian kernel standard deviation is determined based on the key point density in the human image samples. That is, the human meridian dataset has a constant key point density, meaning that a set of meridians corresponds to the same acupoint density and the same machine viewing distance, and is minimally affected by changes in human body size. Therefore, using equation (1), the second Gaussian kernel standard deviation is calculated based on the area and number of key points in the human image samples. This second Gaussian kernel standard deviation is used to accurately reflect the density of acupoints. By replacing fixed constant settings with dynamic parameters, the spatial information utilization rate of different models trained with multiple meridian versions is optimized. Equation (1) is: .

[0041] Where σ represents the standard deviation of the second Gaussian kernel, S represents the area of ​​the human image sample, and num represents the number of keypoints in the human image sample. The relationship between the standard deviation of the second Gaussian kernel and the number of keypoints is similar to the mapping relationship in the first half of a concave curve. When the keypoint density is low, the receptive field can be appropriately increased to effectively utilize spatial information while ensuring minimal interference between keypoints. When the keypoint density is high, the receptive field can be appropriately decreased, but the decreasing trend of the receptive field gradually flattens to ensure that there are certain feature regions available for guiding training.

[0042] After determining the standard deviations of the first and second Gaussian kernels for different geometries, the optimal Gaussian kernel standard deviation is selected from all the first and second Gaussian kernel standard deviations. Based on the optimal Gaussian kernel standard deviation, the kernel size of the Gaussian kernel is determined to maximize the receptive field within a reasonable range, thereby generating the corresponding heatmap ground truth.

[0043] In one embodiment, determining the optimal Gaussian kernel standard deviation based on all first Gaussian kernel standard deviations and second Gaussian kernel standard deviations includes: The largest Gaussian kernel standard deviation among all the first Gaussian kernel standard deviations and the second Gaussian kernel standard deviations is determined as the optimal Gaussian kernel standard deviation.

[0044] Specifically, after determining all the first Gaussian kernel standard deviations and second Gaussian kernel standard deviations, all the first Gaussian kernel standard deviations and second Gaussian kernel standard deviations are sorted in descending order, and the Gaussian kernel standard deviation that ranks first, that is, the largest Gaussian kernel standard deviation, is determined as the optimal Gaussian kernel standard deviation.

[0045] In this embodiment of the invention, the optimal Gaussian kernel standard deviation is determined from the first Gaussian kernel standard deviation corresponding to all geometric shapes and the second Gaussian kernel standard deviation determined according to the key point density. While ensuring that the key points are not confused due to excessive density, the receptive field is expanded as much as possible. During training, the initial meridian acupoint detection model learns spatial context information and topological relationships over a wider range. When faced with acupoint occlusion scenarios or incomplete input, the trained meridian acupoint detection model can better utilize the features of the surrounding visible parts to accurately predict the location of the occluded acupoints.

[0046] In one embodiment, Figure 2 This is a schematic diagram of the structure of the meridian acupoint detection model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the meridian acupoint detection model includes a feature extraction module, a heatmap decoupling module, and an acupoint detection module; The step of inputting the human image to be detected into the meridian acupoint detection model and obtaining the meridian acupoint detection results output by the meridian acupoint detection model includes: The human image to be detected is input into the feature extraction module to obtain a convolutional feature map output by the feature extraction module; the feature extraction module is used to determine the global features corresponding to the human image to be detected, and to perform convolution on the global features to obtain the convolutional feature map; The convolutional feature map is input into the heatmap decoupling module to obtain the fusion features corresponding to each of the meridian group regions output by the heatmap decoupling module; the heatmap decoupling module is used to decouple the convolutional feature map in terms of spatial location and channel respectively; the fusion features are used to characterize the spatial locality and semantic discriminativeness simultaneously possessed by each of the meridian group regions. The fusion features corresponding to each meridian group region are input into the acupoint detection module to obtain the meridian acupoint detection results output by the acupoint detection module.

[0047] Specifically, after determining the human image to be detected, the image is input into the feature extraction module of the meridian acupoint detection model. This feature extraction module can be the backbone of a convolutional neural network. Through this module, high-dimensional global features containing rich spatial location and semantic information are extracted from the human image. These global features are then convolved to obtain a convolutional feature map, which can be represented as follows: =Conv(F), where F represents the global feature. represents the convolutional feature map, and Conv represents the convolution operation.

[0048] After obtaining the convolutional feature map, it is input into the heatmap decoupling module. The heatmap decoupling module performs joint optimization of spatial recalibration and channel recalibration on the convolutional feature map, achieving dual decoupling in both spatial location and channel dimensions. Then, the features are fused to obtain the fused features corresponding to each meridian group region C. These fused features possess both spatial locality and semantic discriminativeness, providing robust feature representations for subsequent acupoint detection.

[0049] After determining the fusion characteristics, the fusion characteristics of each meridian group region are input into the acupoint detection module. The acupoint detection module then locates the acupoints based on the fusion characteristics to obtain the meridian acupoint detection results.

[0050] In one embodiment, such as Figure 2 As shown, the heatmap decoupling module includes a spatial recalibration module, a channel recalibration module, and a fusion module; The step of inputting the convolutional feature map into the heatmap decoupling module to obtain the fused features corresponding to each meridian group region output by the heatmap decoupling module includes: The convolutional feature map is input into the spatial recalibration module to obtain the spatial features corresponding to each meridian group region output by the spatial recalibration module; the spatial features are used to characterize the foreground localization of each meridian group region; the spatial recalibration module is used to determine the spatial features corresponding to each meridian group region based on the convolutional feature map and the spatial mask corresponding to each meridian group region; the spatial mask is used to characterize the spatial coordinates and feature similarity corresponding to each meridian group region; The convolutional feature map is input into the channel recalibration module to obtain the channel features corresponding to each meridian group region output by the channel recalibration module; the channel features are used to characterize the semantic information of each meridian group region; the channel recalibration module is used to determine the channel features corresponding to each meridian group region based on the convolutional feature map and the instance features of each meridian group region. The spatial features and channel features corresponding to each meridian group region are input into the fusion module to obtain the fusion features corresponding to each meridian group region output by the fusion module.

[0051] Specifically, after determining the global features corresponding to the human body image to be detected, the global features are partitioned according to the human body parts to obtain the instance features corresponding to each meridian group region. Then, the convolutional feature map is input into the spatial recalibration module and the channel recalibration module respectively. The spatial recalibration module uses Equation (2) to calculate the spatial features of each meridian group region based on the spatial mask and convolutional feature map. The generation of this spatial mask integrates spatial coordinates and feature similarity information, and the spatial features provide accurate foreground localization. Equation (2) is: .

[0052] in, This represents the spatial characteristics of the i-th meridian group region. This represents the spatial mask of the i-th meridian group region.

[0053] The channel recalibration module utilizes equation (3) to dynamically modulate the channel dimension based on the instance characteristics of each meridian group region, thereby distributing different meridian group regions into different response channels. Equation (3) is: .

[0054] in, This represents the channel characteristics of the i-th meridian group region. This represents the instance features of the i-th meridian group region. This represents the Hadamard product operation channel by channel. In the channel recalibration module, the instance features and channel features of the i-th meridian group region are weighted based on similarity. This preserves the channel responses relevant to the target instance and suppresses irrelevant channels, effectively enhancing the semantic information unique to different instance features.

[0055] After obtaining the spatial and channel features of each meridian group region, the spatial and channel features are input into the fusion module. The fusion module then uses equation (4) to fuse the spatial and channel features, obtaining the fused features of each meridian group region. Equation (4) is: .

[0056] in, Represents the fusion feature corresponding to the i-th meridian group region, and ReLU (Rectified LinearUnit) represents the activation function.

[0057] In one embodiment, the spatial mask is determined based on the following steps: The relative coordinate map and feature similarity map corresponding to each meridian group region are determined; the relative coordinate map is used to characterize the relative relationship between the spatial position of the acupoint in each meridian group region and the target center; the feature similarity map is the pixel-wise inner product between the convolutional feature map and the instance features of the corresponding meridian group region; the instance features of each meridian group region are obtained by partitioning the global features. Based on the relative coordinate map and feature similarity map corresponding to each meridian group region, the spatial mask corresponding to each meridian group region is determined.

[0058] Specifically, during spatial recalibration, a relative coordinate map and a feature similarity map are first constructed. The relative coordinate map reflects the relative relationship between the spatial location of acupoints in each meridian group region and the target center, where the target center is the acupoint center. The feature similarity map reflects the pixel-wise inner product of the convolutional feature map and the instance feature. After combining the relative coordinate map and the feature similarity map, the spatial mask of the i-th meridian group region is generated using Equation (5) through a convolutional layer and a Sigmoid activation function. Equation (5) is: .

[0059] in, This represents the relative coordinate diagram corresponding to the i-th meridian group region. This represents the feature similarity map corresponding to the i-th meridian group region. The feature map is dynamically weighted and convolved in the spatial dimension using this spatial mask to enhance the salient regions of the target instance.

[0060] The following comparative experiments were conducted in the embodiments of the present invention.

[0061] 1. The ground truth values ​​of heatmaps generated using a fixed Gaussian kernel standard deviation, a conventional dynamic Gaussian kernel standard deviation, and the optimal Gaussian kernel standard deviation provided in this embodiment of the invention were compared. The comparison results are shown in Table 1. As shown in Table 1, the model trained using the ground truth values ​​of heatmaps generated using the optimal Gaussian kernel standard deviation provided in this embodiment of the invention has an accuracy approximately 2% higher than the model trained using the ground truth values ​​of heatmaps generated using the conventional dynamic Gaussian kernel standard deviation, and an accuracy approximately 1% higher than the model trained using the ground truth values ​​of heatmaps generated using the fixed Gaussian kernel standard deviation. Furthermore, since AP calculations do not calculate undetected keypoints, the negative impact of missed detections on accuracy is minimal.

[0062] Table 1

[0063] In Table 1, the DGK-common value can be 1, 1.5, 2, 3, etc. σ=S / 256 is the standard deviation of the conventional dynamic Gaussian kernel, which is a Gaussian algorithm based on the default σ size of the image area S / 256, where S represents the image area. In existing technologies, the standard deviation of the Gaussian kernel for generating the true value of the heatmap is specified, with the height and width of the heatmap being one-quarter of the input image resolution. A fixed Gaussian kernel standard deviation can be (H+W) / 128, where H represents the height of the heatmap and W represents the width. The calculated fixed Gaussian kernel standard deviation is generally 1. In this experiment, the standard deviation of the Gaussian kernel was moderately increased based on a fixed standard deviation of 1, using a 3×3 kernel size to effectively capture information from adjacent pixels. The heatmap for each acupoint is centered on the accurate coordinates of the acupoint, encouraging the model to learn accurate acupoint localization. In terms of parameter settings, the study used Gaussian kernel standard deviations with σ values ​​of 1.5, 2, and 3, corresponding to kernel sizes of 7×7, 11×11, and 17×17, respectively. This effectively increased the expansion of the Gaussian distribution, resulting in a wider influence area in the generated heatmap. The comparative experimental results verifying the impact of different Gaussian kernel standard deviation sizes on model performance are shown in Table 2.

[0064] In Table 2, Intersection Over Union (IOU) is used to represent the size, and medium size is used to represent the size between 32×32 and 96×96. As shown in Table 2, the model accuracy improvement is particularly significant when the Gaussian kernel standard deviation is 2 and the kernel size is 11×11. A larger Gaussian kernel standard deviation corresponds to a larger receptive field, but it is not necessarily better to have a larger one. The reasons are: (1) After heatmap regression, there is an NMS step. In general methods, a fixed number of pixels with the highest confidence are selected and retained for subsequent operations. An excessively large Gaussian kernel standard deviation will cause a multiple increase in the amount of computation, but it will not promote the regression of extreme values ​​(i.e., the final key points) more. (2) Spatial information within a certain range can promote the localization regression of key points. Providing more feature information can improve the robustness of the model. However, too much spatial information may lead to overfitting, and as spatial information increases, common information between different key points will be generated. For example, the wrist and elbow key points are on the same side of the shoulder, which may lead to recognition errors.

[0065] Table 2

[0066] 2. A comparison was made between MSRA, UDP, Gauss, and the DCH (Heatmap Decoupling) method used in this embodiment of the invention. MSRA, a heatmap-based regression method proposed by Microsoft Research Asia for human pose estimation, introduces an unbiased parameter in its heatmap generation. During encoding, it checks whether keypoint encoding is performed only on a single instance and selects between biased and unbiased Gaussian heatmaps based on the unbiased parameter. During decoding, after finding the maximum value in the heatmap as the predicted keypoint coordinates, coordinate correction is performed based on the unbiased parameter. This method improves the efficiency and guidance of heatmap generation and further enhances the accuracy of keypoint localization by introducing additional correction steps, which aligns with the single-person acupoint detection scenario in this study. UDP, a unified distribution prediction mechanism proposed by MSRA, provides an alternative combined heatmap option besides Gaussian heatmaps, in addition to introducing unbiased data processing. This type of heatmap generates binary label maps as well as X and Y axis offset maps. This combination method not only provides the location information of key points, but also includes the offset relative to the predicted location, further improving the positioning accuracy. The comparison results are shown in Table 3.

[0067] Table 3 shows that the model training guided by the heatmap generated by the DCH heatmap decoupling method has the highest accuracy. Although the MSRA method is limited to a single instance, there is still redundancy in the heatmap confidence region. UDP requires additional processing of the offset map during encoding and decoding, and the fine-tuning steps during decoding are also relatively complex, which may lead to some error accumulation, thus affecting the acupoint detection results. In the meridian conditioning scenario under occlusion, since DCH has an independent heatmap representation for each coordinate dimension, it can theoretically achieve higher spatial resolution. Even with a small heatmap size, it can maintain good localization accuracy and has a strong adaptability to changes such as occlusion and rotation, which is particularly important for pose estimation under non-ideal conditions.

[0068] Table 3

[0069] 3. To verify the effectiveness of the optimal Gaussian kernel standard deviation and heatmap decoupling module in the embodiments of the present invention, a set of ablation experiments were conducted in the embodiments of the present invention to detect the model accuracy of each module individually and in combination when all other network configurations are the same. The results of the ablation experiments are shown in Table 4.

[0070] Table 4

[0071] As shown in Table 4, the average AP of the basic model was 84.52% and the average AR was 82.04% on the supine human dataset; and the average AP was 78.04% and the average AR was 76.86% on the prone human dataset.

[0072] Adding the DGK module (Optimal Gaussian Kernel Standard Deviation) did not significantly change the model's AP, but greatly improved its AR, achieving over 5% performance improvement on both datasets. The DGK module primarily seeks the optimal receptive field by dynamically adjusting the confidence region size of the generated heatmap, enhancing the model's understanding of the correlation between acupoints in the treatment scene. This improvement significantly boosts the model's AR because it better captures occluded or difficult-to-detect keypoint locations. Meanwhile, the AP remained unchanged, indicating that while the DGK module is very effective in detecting occluded acupoints, it did not significantly improve the accuracy of these detections. However, it did not cause a decrease in detection precision and overall still significantly enhanced the model's robustness to occlusion.

[0073] After adding the DCH module, the model's detection accuracy (AP) improved by approximately 1%-2% across various datasets, and its accuracy (AR) improved by about 2%. By decoupling the channel heatmaps, the feature representation of each keypoint became more independent and explicit. This not only improved the model's ability to distinguish between different keypoints but also significantly reduced the probability of false detections.

[0074] When both are applied together to guide training for heatmap generation, the AP of the optimized model shows a further improvement compared to adding the DCH module alone, generally exceeding the base model by about 2%. This indicates that decoupling the acupoints from the dynamic receptive field has a positive impact on performance. Meanwhile, although AR is lower than adding the DGK module alone, the reduction is less than 1%. Overall, when the idea of ​​decoupling the dynamic receptive field and acupoints is applied together, they do not become mutually exclusive, bringing improvements in accuracy and robustness to the acupoint detection model under occlusion conditions. Figure 3 This is a schematic diagram comparing the model before and after optimization according to an embodiment of the present invention. To more intuitively reflect the recovery capability of the optimized model for missed detections and false detections, the visualized detection results of the basic model and the optimized model are shown below. Figure 3 As shown, before optimization, the basic model has a small range of missed detections and false detections, such as... Figure 3 The left leg. In this case, the optimized meridian acupoint detection model in this embodiment of the invention can accurately detect acupoint locations even under occlusion.

[0075] The following describes the human meridian acupoint detection device provided by the present invention. The human meridian acupoint detection device described below can be referred to in correspondence with the human meridian acupoint detection method described above.

[0076] This invention provides a device for detecting acupoints along human meridians. Figure 4 This is a schematic diagram of the structure of the human meridian acupoint detection device provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the human meridian acupoint detection device 400 includes: an acquisition module 410 and a detection module 420.

[0077] The acquisition module 410 is used to acquire the human body image to be detected corresponding to the user.

[0078] The detection module 420 is used to input the human image to be detected into the meridian acupoint detection model to obtain the meridian acupoint detection results output by the meridian acupoint detection model; the meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values, and the Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples; the meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0079] This invention provides a human meridian acupoint detection device. It adaptively determines the ground truth of a heatmap by considering different geometric shapes of human regions and the density of key points in human image samples. A meridian acupoint detection model is trained using the human image samples and the ground truth heatmap. The human image to be detected is input into the trained meridian acupoint detection model. The model decouples each meridian group region in the human image from its spatial location and channel, resulting in the meridian acupoint detection results output by the model. In this invention, the standard deviation of the Gaussian kernel is adaptively adjusted based on the geometric shape of the human region and the density of key points. The optimal receptive field corresponding to the ground truth heatmap is determined based on the standard deviation of the Gaussian kernel. This guides the meridian acupoint detection model to fully learn the receptive field kernel and topological relationship information, significantly reducing false detections and missed detections when meridian acupoints are occluded. Meanwhile, the human body image to be detected is partitioned, and different meridian groups are decoupled in terms of spatial location and channel dimension. This effectively reduces interference between key points, the complexity of the global perspective, and the false detection and false negative detection of meridian acupoints, thereby improving the accuracy of meridian acupoint detection.

[0080] Optionally, the human meridian acupoint detection device 400 includes a training module, which is specifically used for: Based on the different geometric shapes of the human body region in the human body image sample, determine the first Gaussian kernel standard deviation corresponding to each geometric shape; The second Gaussian kernel standard deviation is determined based on the area and number of key points of the human image sample; the area and number of key points of the human image sample are used to characterize the key point density corresponding to the human image sample. Determine the optimal Gaussian kernel standard deviation based on all first Gaussian kernel standard deviations and second Gaussian kernel standard deviations; Based on the optimal Gaussian kernel standard deviation, the true value of the heatmap corresponding to the human image sample is determined.

[0081] Optionally, this training module is specifically used for: The largest Gaussian kernel standard deviation among all the first Gaussian kernel standard deviations and the second Gaussian kernel standard deviations is determined as the optimal Gaussian kernel standard deviation.

[0082] Optionally, the geometry of the human body region includes the minimum bounding rectangle, the equivalent circle, and the maximum inscribed circle of the human body region.

[0083] Optionally, the meridian acupoint detection model includes a feature extraction module, a heatmap decoupling module, and an acupoint detection module.

[0084] Optionally, the detection module 420 is specifically used for: The human image to be detected is input into the feature extraction module to obtain a convolutional feature map output by the feature extraction module; the feature extraction module is used to determine the global features corresponding to the human image to be detected, and to perform convolution on the global features to obtain the convolutional feature map; The convolutional feature map is input into the heatmap decoupling module to obtain the fusion features corresponding to each of the meridian group regions output by the heatmap decoupling module; the heatmap decoupling module is used to decouple the convolutional feature map in terms of spatial location and channel respectively; the fusion features are used to characterize the spatial locality and semantic discriminativeness simultaneously possessed by each of the meridian group regions. The fusion features corresponding to each meridian group region are input into the acupoint detection module to obtain the meridian acupoint detection results output by the acupoint detection module.

[0085] Optionally, the heatmap decoupling module includes a spatial recalibration module, a channel recalibration module, and a fusion module.

[0086] Optionally, the detection module 420 is specifically used for: The convolutional feature map is input into the spatial recalibration module to obtain the spatial features corresponding to each meridian group region output by the spatial recalibration module; the spatial features are used to characterize the foreground localization of each meridian group region; the spatial recalibration module is used to determine the spatial features corresponding to each meridian group region based on the convolutional feature map and the spatial mask corresponding to each meridian group region; the spatial mask is used to characterize the spatial coordinates and feature similarity corresponding to each meridian group region; The convolutional feature map is input into the channel recalibration module to obtain the channel features corresponding to each meridian group region output by the channel recalibration module; the channel features are used to characterize the semantic information of each meridian group region; the channel recalibration module is used to determine the channel features corresponding to each meridian group region based on the convolutional feature map and the instance features of each meridian group region. The spatial features and channel features corresponding to each meridian group region are input into the fusion module to obtain the fusion features corresponding to each meridian group region output by the fusion module.

[0087] Optionally, the detection module 420 is specifically used for: The relative coordinate map and feature similarity map corresponding to each meridian group region are determined; the relative coordinate map is used to characterize the relative relationship between the spatial position of the acupoint in each meridian group region and the target center; the feature similarity map is the pixel-wise inner product between the convolutional feature map and the instance features of the corresponding meridian group region; the instance features of each meridian group region are obtained by partitioning the global features. Based on the relative coordinate map and feature similarity map corresponding to each meridian group region, the spatial mask corresponding to each meridian group region is determined.

[0088] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a human meridian acupoint detection method, which includes: acquiring a human image to be detected corresponding to the user; inputting the human image to be detected into a meridian acupoint detection model to obtain the meridian acupoint detection result output by the meridian acupoint detection model; the meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values, and the Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples; the meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0089] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the human meridian acupoint detection method provided by the above methods. The method includes: acquiring a human image to be detected corresponding to a user; inputting the human image to be detected into a meridian acupoint detection model to obtain the meridian acupoint detection result output by the meridian acupoint detection model; the meridian acupoint detection model is trained based on human image samples and the heatmap ground truth corresponding to the human image samples, and the Gaussian kernel standard deviation corresponding to the heatmap ground truth is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples; the meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0091] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the human meridian acupoint detection method provided by the above methods. The method includes: acquiring a human image to be detected corresponding to a user; inputting the human image to be detected into a meridian acupoint detection model to obtain a meridian acupoint detection result output by the meridian acupoint detection model; the meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values, wherein the Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on different geometric shapes of human regions in the human image samples and the density of key points in the human image samples; the meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel.

[0092] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0093] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting acupoints along human meridians, characterized in that, include: Obtain the human image to be detected corresponding to the user; The human image to be detected is input into the meridian acupoint detection model to obtain the meridian acupoint detection results output by the meridian acupoint detection model. The meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values. The Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples. The meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel. The true value of the heatmap is determined based on the following steps: Based on the different geometric shapes of the human body region in the human body image sample, determine the first Gaussian kernel standard deviation corresponding to each geometric shape; The second Gaussian kernel standard deviation is determined based on the area and number of key points of the human image sample; the area and number of key points of the human image sample are used to characterize the key point density corresponding to the human image sample. The largest Gaussian kernel standard deviation among all the first Gaussian kernel standard deviations and the second Gaussian kernel standard deviations is determined as the optimal Gaussian kernel standard deviation; Based on the optimal Gaussian kernel standard deviation, the true value of the heatmap corresponding to the human image sample is determined; The meridian acupoint detection model includes a feature extraction module, a heatmap decoupling module, and an acupoint detection module. The step of inputting the human image to be detected into the meridian acupoint detection model and obtaining the meridian acupoint detection results output by the meridian acupoint detection model includes: The human image to be detected is input into the feature extraction module to obtain the convolutional feature map output by the feature extraction module; The feature extraction module is used to determine the global features corresponding to the human image to be detected, and to perform convolution on the global features to obtain the convolutional feature map. The convolutional feature map is input into the heatmap decoupling module to obtain the fusion features corresponding to each of the meridian group regions output by the heatmap decoupling module; the heatmap decoupling module is used to decouple the convolutional feature map in terms of spatial location and channel respectively; the fusion features are used to characterize the spatial locality and semantic discriminativeness simultaneously possessed by each of the meridian group regions. The fusion features corresponding to each meridian group region are input into the acupoint detection module to obtain the meridian acupoint detection results output by the acupoint detection module.

2. The method for detecting acupoints along human meridians according to claim 1, characterized in that, The heatmap decoupling module includes a spatial recalibration module, a channel recalibration module, and a fusion module; The step of inputting the convolutional feature map into the heatmap decoupling module to obtain the fused features corresponding to each meridian group region output by the heatmap decoupling module includes: The convolutional feature map is input into the spatial recalibration module to obtain the spatial features corresponding to each meridian group region output by the spatial recalibration module; the spatial features are used to characterize the foreground localization of each meridian group region; the spatial recalibration module is used to determine the spatial features corresponding to each meridian group region based on the convolutional feature map and the spatial mask corresponding to each meridian group region; the spatial mask is used to characterize the spatial coordinates and feature similarity corresponding to each meridian group region; The convolutional feature map is input into the channel recalibration module to obtain the channel features corresponding to each meridian group region output by the channel recalibration module; the channel features are used to characterize the semantic information of each meridian group region; the channel recalibration module is used to determine the channel features corresponding to each meridian group region based on the convolutional feature map and the instance features of each meridian group region. The spatial features and channel features corresponding to each meridian group region are input into the fusion module to obtain the fusion features corresponding to each meridian group region output by the fusion module.

3. The method for detecting acupoints along human meridians according to claim 2, characterized in that, The spatial mask is determined based on the following steps: Determine the relative coordinate map and feature similarity map corresponding to each meridian group region; the relative coordinate map is used to characterize the relative relationship between the spatial position of the acupoint in each meridian group region and the target center; The feature similarity map is the pixel-wise inner product between the convolutional feature map and the instance features of the corresponding meridian group region; the instance features of each meridian group region are obtained by partitioning the global features. Based on the relative coordinate map and feature similarity map corresponding to each meridian group region, the spatial mask corresponding to each meridian group region is determined.

4. The method for detecting acupoints along human meridians according to claim 1, characterized in that, The geometry of the human body region includes the minimum bounding rectangle, the equivalent circle, and the maximum inscribed circle of the human body region.

5. A device for detecting acupoints along human meridians, characterized in that, include: The acquisition module is used to acquire the human image to be detected corresponding to the user; The detection module is used to input the human image to be detected into the meridian acupoint detection model and obtain the meridian acupoint detection results output by the meridian acupoint detection model. The meridian acupoint detection model is trained based on human image samples and the corresponding heatmap ground truth values. The Gaussian kernel standard deviation corresponding to the heatmap ground truth values ​​is adaptively determined based on the different geometric shapes of human regions in the human image samples and the density of key points in the human image samples. The meridian acupoint detection model is used to decouple each meridian group region in the human image to be detected in terms of spatial location and channel. The true value of the heatmap is determined based on the following steps: Based on the different geometric shapes of the human body region in the human body image sample, determine the first Gaussian kernel standard deviation corresponding to each geometric shape; The second Gaussian kernel standard deviation is determined based on the area and number of key points of the human image sample; the area and number of key points of the human image sample are used to characterize the key point density corresponding to the human image sample. The largest Gaussian kernel standard deviation among all the first Gaussian kernel standard deviations and the second Gaussian kernel standard deviations is determined as the optimal Gaussian kernel standard deviation; Based on the optimal Gaussian kernel standard deviation, the true value of the heatmap corresponding to the human image sample is determined; The meridian acupoint detection model includes a feature extraction module, a heatmap decoupling module, and an acupoint detection module. The detection module is specifically used for: inputting the human image to be detected into the feature extraction module to obtain a convolutional feature map output by the feature extraction module; the feature extraction module is used to determine the global features corresponding to the human image to be detected, and to perform convolution on the global features to obtain the convolutional feature map; inputting the convolutional feature map into the heatmap decoupling module to obtain the fusion features corresponding to each of the meridian group regions output by the heatmap decoupling module; the heatmap decoupling module is used to decouple the convolutional feature map in terms of spatial location and channel respectively; the fusion features are used to characterize the spatial locality and semantic discriminativeness simultaneously possessed by each of the meridian group regions; inputting the fusion features corresponding to each of the meridian group regions into the acupoint detection module to obtain the meridian acupoint detection results output by the acupoint detection module.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the human meridian acupoint detection method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the human meridian acupoint detection method as described in any one of claims 1 to 4.

Images