Hand acupuncture point accurate recognition method and system based on HRNet

Abstract

The invention discloses a hand acupuncture point accurate recognition method and system based on HRNet, and belongs to the technical field of image recognition. The method comprises the steps that a hand image data set is acquired and preprocessed, labeling of key acupuncture points of a hand is completed, and the hand image data set is divided into a training set and a test set; introducing a cosine annealing learning rate strategy, a thermodynamic diagram calibration mechanism based on topology conservation and a CBAM attention mechanism, and constructing a model; training the model, and iteratively optimizing the parameters of the model until the difference between the predicted value and the true value is within a preset range; after training is completed, hand images in a test set are input into the model, key acupuncture points of the hand images in the test set are recognized and positioned, acupuncture point coordinates and positioning results are output, and performance evaluation is conducted on the results. According to the invention, the problems of insufficient hand acupoint recognition precision, weak posture adaptability and unstable training in the prior art are solved, and the requirement of millimeter-level positioning of small-scale acupoints can be met.

Description

Technical Field

[0001] This invention relates to the field of image or video recognition technology, and in particular to a method and system for accurate recognition of hand acupoints based on HRNet. Background Technology

[0002] Traditional Chinese medicine (TCM) acupoint therapy now boasts advantages such as low cost, high effectiveness, and high safety, meeting the health needs of modern people. On the other hand, the effectiveness of acupoint therapy is highly correlated with the accuracy of acupoint location. However, traditional acupoint location still relies primarily on manual search, demanding extensive professional knowledge and practical experience, which presents a high barrier to entry and hinders the popularization of TCM acupoint therapy. The hand is the area with the highest concentration of acupoints, which are related to various organs throughout the body. It is convenient to operate, does not infringe on privacy, and has rich image features, making it a key research subject for automatic acupoint recognition.

[0003] Existing automatic hand acupoint localization methods rely on relatively singular criteria, making them ill-suited to the complex structure, flexible postures, and numerous acupoints of the hand. Hand keypoint detection algorithms are mainly divided into model-based detection algorithms and data-driven detection methods. Model-based methods typically involve complex initialization, requiring the predefinition of numerous models for complex gestures. Data-driven methods, on the other hand, use large amounts of data to learn and directly locate hand keypoints, but suffer from weak posture adaptability and difficulty in balancing the practicality of model training and deployment. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for accurate identification of hand acupoints based on HRNet, which solves the problems of insufficient accuracy, weak posture adaptability and unstable training of hand acupoints in the prior art, and can meet the needs of millimeter-level positioning of small-scale acupoints.

[0005] This invention is achieved through the following technical solution: A method for accurate identification of hand acupoints based on HRNet includes the following steps: S1: Obtain the hand image dataset and preprocess it to complete the annotation of key acupoints on the hand, form the sample and label data required for training, and divide the sample and label data required for training into training set and test set according to the preset ratio; S2: An improved HRNet hand acupoint recognition model is constructed by introducing a cosine annealing learning rate strategy, a heatmap calibration mechanism based on topology conservation, and a CBAM attention mechanism. S3: Configure the parameters of the improved HRNet hand acupoint recognition model, input the sample and label data of the training set into the improved HRNet hand acupoint recognition model for training, calculate the difference between the predicted value and the true value through the loss function, and iteratively optimize the parameters of the improved HRNet hand acupoint recognition model until the difference between the predicted value and the true value is within the preset range. S4: After training, input the hand images in the test set into the improved HRNet hand acupoint recognition model. The improved HRNet hand acupoint recognition model identifies and locates the key acupoints in the hand images in the test set, outputs the acupoint coordinates and location results, and evaluates the recognition results through three indicators: average accuracy, proportion of correct key points, and normalized average error.

[0006] The optimized hand image dataset in step S1 includes palm images and back of hand images, and the preprocessing includes image enhancement and coordinate normalization.

[0007] Furthermore, the cosine annealing learning rate strategy in step S2 is as follows: S211: Divide the learning rate training rounds into three stages: basic learning stage, constraint optimization stage, and fine-tuning stage; S212: During the basic learning phase, the learning rate is maintained at the maximum learning rate within the preset annealing period for training, without triggering annealing; S213: During the constraint optimization phase, the learning rate for each training round is calculated in real time according to equation (1), and constraint optimization is performed according to the calculated learning rate for each training round: (1); in: express Round learning rate, This represents the minimum learning rate during the annealing cycle. This represents the maximum learning rate during the annealing cycle. This indicates the round number from which the counting begins, starting from the annealing phase. This indicates the total number of annealing cycles from start to finish. S214: In the fine-tuning phase, the learning rate is reduced to the minimum learning rate within the preset annealing period for training.

[0008] Furthermore, the heatmap calibration mechanism based on topology conservation in step S2 is as follows: S221: Calculate the pixel response value of each key acupoint based on the Gaussian kernel function (2) to generate a basic heatmap: (2); in: Indicates the first The pixel response value of each acupoint. In the heat map, the first... Line number Liede The pixel response value of each acupoint. This represents the natural exponential function. Indicates the first Normalized coordinates of each acupoint This represents the Gaussian kernel standard deviation. This represents the maximum response value in the heatmap; S222: The samples in the training set are filtered for visibility based on the annotation of key acupoints on the hand. Samples whose visibility markers for both the target acupoint and the acupoint to be constrained are visible proceed to the next step. S223: The high-response region is extracted from the basic heatmap formed by the visibility-filtered samples according to equation (3). If the high-response region satisfies equation (3), it is determined as follows: (3); in: Indicates a high response mask. The base heatmap representing the target acupoint index. This represents the base heatmap of the acupoint index to be constrained. Indicates the response threshold; S224: Calculate the high response region according to the topology rules. If the calculation result is a topology violation, then the high response region is a violation region. S225: Correct the violation area using the high-response region weighted suppression strategy according to equation (4) to obtain the corrected heatmap, and complete the heatmap calibration based on topology conservation: (4); in: This represents the heatmap after correction for the target acupoint. Indicates the constraint weight. This represents the heat map after the acupoints to be constrained have been corrected.

[0009] Furthermore, the topology rules described in step S224 are as follows: like =left, then the target acupoint is expected to be to the left of the acupoint to be constrained. If the violation condition (5) is satisfied, it is judged as a topological violation: (5); in: This represents the function for determining violations. Indicates the target acupoint exist Coordinates along the axial direction, Indicates acupoints to be constrained. exist Coordinates along the axial direction, Indicates the pixel tolerance threshold; like =right, then the target acupoint is expected to be to the right of the acupoint to be constrained. If the violation condition (6) is satisfied, it is judged as a topological violation: (6); like =Above, then the target acupoint is expected to be above the acupoint to be constrained. If the violation condition (7) is satisfied, it is judged as a topological violation: (7); in: Indicates the target acupoint exist Coordinates along the axial direction, Indicates acupoints to be constrained. exist Coordinates along the axis; like =down, then the target acupoint is expected to be below the acupoint to be constrained. If the violation condition (8) is satisfied, it is judged as a topological violation: (8).

[0010] The optimized CBAM attention mechanism in step S2 includes a channel attention submodule and a spatial attention submodule. The channel attention submodule compresses the feature maps generated by the HRNet deep learning network during the extraction and understanding of image content using both global average pooling and global max pooling. This yields vectors reflecting the overall strength of each feature channel and vectors capturing the most significant responses within each channel. These vectors are then analyzed by a fully connected neural network, and the outputs are summed and activated by the Sigmoid function to obtain the final channel attention weight vector. The spatial attention submodule performs average pooling and max pooling on all feature channels to obtain two single-channel feature maps. These two single-channel maps are concatenated along the channel dimension to form a dual-channel feature map, which is then processed by the Sigmoid function to generate a spatial attention weight map.

[0011] The optimized parameters of the improved HRNet hand acupoint recognition model configured in step S3 include: training rounds, batch size, initial learning rate, PCK metric judgment threshold, heatmap Gaussian kernel standard deviation, cosine annealing period, and minimum learning rate decay.

[0012] Furthermore, the formula for calculating the loss function in step S3 is (9): (9); in: Represents the loss function. This represents the loss function for the regression task. This represents the model's predicted heatmap. Indicates the label mask. This represents a heat map after acupoint correction.

[0013] Furthermore, in step S4, the formula for calculating the average accuracy is (10), the formula for calculating the proportion of correct key points is (11), and the formula for calculating the normalized average error is (12): (10); (11); (12); in: Indicates average precision. Indicates the first The accuracy value of each prediction result. To preset the accuracy threshold, For indicator functions, Indicates the correct proportion of key points. Indicates the predicted keypoint coordinates. Represents the actual coordinates of key points. Indicates the distance threshold. Represents Euclidean distance. This represents the normalized average error. This represents the total number of acupoints. For the set of real coordinates, To predict the set of coordinates, Indicates the first The actual coordinates of each acupoint Indicates the first Predicted coordinates of each acupoint. This is a normalized reference value.

[0014] A hand acupoint accurate recognition system based on HRNet is used to execute a hand acupoint accurate recognition method based on HRNet as described in any of the above, which includes a preprocessing module, a model building module, a model optimization module and a performance evaluation module; The preprocessing module is used to acquire and preprocess hand image datasets, complete the annotation of key acupoints on the hand, form the sample and label data required for training, and divide the sample and label data required for training into training set and test set according to a preset ratio. The model building module is used to introduce the cosine annealing learning rate strategy, the topology conservation-based heatmap calibration mechanism, and the CBAM attention mechanism into the HRNet hand acupoint recognition model to build an improved HRNet hand acupoint recognition model. The model optimization module is used to input the sample and label data of the training set into the improved HRNet hand acupoint recognition model for training. It calculates the difference between the predicted value and the true value through the loss function, and iteratively optimizes the parameters of the improved HRNet hand acupoint recognition model until the difference between the predicted value and the true value is within a preset range. The performance evaluation module is used to input hand images from the test set into the improved HRNet hand acupoint recognition model, identify and locate key acupoints in the hand images from the test set, output acupoint coordinates and positioning results, and evaluate the recognition results using three indicators: average accuracy, proportion of correct key points, and normalized average error.

[0015] Beneficial effects of the invention: This invention provides a method and system for accurate hand acupoint recognition based on HRNet. It introduces a cosine annealing learning rate strategy, a heatmap calibration mechanism based on topology conservation, and a CBAM attention mechanism to construct an improved HRNet hand acupoint recognition model. The cosine annealing learning rate strategy is a phased model training strategy based on cosine annealing, which can dynamically adjust the learning rate and improve the stability and convergence of model training. The heatmap calibration mechanism based on topology conservation corrects the heatmap through formal topological constraint rules, ensuring the spatial rationality of acupoint location. The CBAM attention mechanism improves the HRNet feature extraction architecture, enhancing the acupoint feature capture capability through channel and spatial attention mechanisms. These three mechanisms work together to solve the problems of insufficient accuracy, weak posture adaptability, and unstable training in existing hand acupoint recognition technologies. It can meet the needs for millimeter-level positioning of small-scale acupoints, bringing great convenience to product development and practical applications in related fields. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the process of this invention. Detailed Implementation

[0017] A method for accurate identification of hand acupoints based on HRNet, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps: S1: Obtain the hand image dataset and preprocess it to complete the annotation of key acupoints on the hand, form the sample and label data required for training, and divide the sample and label data required for training into training set and test set according to the preset ratio; Specifically, the hand image dataset includes palm and back of hand images. The coordinates of key acupoints on the hand must strictly follow the COCO annotation specifications to indicate their visibility. Key acupoints that do not exist can be labeled as 0, key acupoints that exist but are not visible can be labeled as 1, and key acupoints that exist and are visible can be labeled as 2.

[0018] Specific preprocessing methods include image enhancement and coordinate normalization. Image enhancement can be achieved by randomly adjusting image attributes, while coordinate normalization can unify the image size to a preset specification. Both of these preprocessing methods are existing technologies.

[0019] S2: An improved HRNet hand acupoint recognition model is constructed by introducing a cosine annealing learning rate strategy, a heatmap calibration mechanism based on topology conservation, and a CBAM attention mechanism. Specifically, the cosine annealing learning rate strategy is as follows: S211: Divide the learning rate training rounds into three stages: basic learning stage, constraint optimization stage, and fine-tuning stage; S212: During the basic learning phase, the learning rate is maintained at the maximum learning rate within the preset annealing period for training, without triggering annealing; S213: During the constraint optimization phase, the learning rate for each training round is calculated in real time according to equation (1), and constraint optimization is performed according to the calculated learning rate for each training round: (1); in: express Round learning rate, This represents the minimum learning rate during the annealing cycle. This represents the maximum learning rate during the annealing cycle. This indicates the round number from which the counting begins, starting from the annealing phase. This indicates the total number of annealing cycles from start to finish. S214: In the fine-tuning phase, the learning rate is reduced to the minimum learning rate within the preset annealing period for training.

[0020] The learning rate is used to regulate the model's ability to learn overall hand features and key acupoints. Multiple simulation tests were conducted based on different training rounds, and the optimal learning rate training rounds were ultimately divided into the following three stages.

[0021] Specifically, the total number of training rounds can be preferably 65.

[0022] Basic learning phase: This can be set to 1 to 15 rounds. In rounds 1 to 15, the learning rate is maintained at the maximum learning rate within the preset annealing cycle, and annealing is not triggered. This allows the model to quickly learn and distinguish basic features such as the outline and acupoints of the hand. If cosine annealing is used to reduce the learning rate at the beginning of training, the large sample size of the model will result in a very slow learning speed, making it difficult to even master the basic features of the hand in the early stages.

[0023] Constraint optimization phase: This phase can be set to 16 to 45 rounds. The first 15 rounds maintain the maximum learning rate to complete the basic learning, but to further improve the accuracy of acupoint recognition, the learning rate needs to be reduced. Therefore, starting from the 16th round, cosine annealing is introduced to gradually reduce the learning rate between the maximum learning rate and the preset minimum learning rate. By reducing the learning rate, the model can adjust parameters more finely and stably, thereby improving the positioning accuracy and preparing for the final fine-tuning.

[0024] Fine-tuning stage: This stage can be set to 46 to 65 rounds. This part reduces the learning rate to the preset minimum learning rate, which can further calibrate the acupoint prediction to sub-pixel accuracy and eliminate the final fluctuations in parameters. It is a necessary process to give full play to the high-resolution architecture advantage of the HRnet model.

[0025] By employing the cosine annealing learning rate strategy described above, the learning rate can be dynamically adjusted, thereby improving the stability and convergence of model training.

[0026] Specifically, the heatmap calibration mechanism based on topology conservation is as follows: S221: Calculate the pixel response value of each key acupoint based on the Gaussian kernel function (2) to generate a basic heatmap: (2); in: Indicates the first The pixel response value of each acupoint. In the heat map, the first... Line number Liede The pixel response value of each acupoint. This represents the natural exponential function. Indicates the first Normalized coordinates of each acupoint This represents the Gaussian kernel standard deviation. This represents the maximum response value in the heatmap; Since the acupoints on the hand are marked with real coordinates, the response value of each pixel position within the preset size range is calculated, and all the calculated pixel response values ​​are combined into a matrix according to their positions. This matrix is ​​the final basic heat map of the acupoint.

[0027] S222: The samples in the training set are filtered for visibility based on the annotation of key acupoints on the hand. Samples whose visibility markers for both the target acupoint and the acupoint to be constrained are visible proceed to the next step. The marking of key acupoints on the hand is only applicable to the target acupoints. and acupoints to be constrained Topology violation detection is only performed when all visibility markers according to the COCO annotation specification are visible. If any acupoint is not visible, the rule detection is skipped, which can avoid invalid constraints.

[0028] S223: The high-response region is extracted from the basic heatmap formed by the visibility-filtered samples according to equation (3). If the high-response region satisfies equation (3), it is determined as follows: (3); in: Indicates a high response mask. The base heatmap representing the target acupoint index. This represents the base heatmap of the acupoint index to be constrained. Indicates the response threshold; The basic heatmap ensures that constraints are applied only to "potential key point areas with high confidence" in the heatmap. If the area is not a high-response area, the rule detection is skipped, further avoiding invalid constraints.

[0029] The response threshold here The optimal value is 0.6.

[0030] S224: Calculate the high response region according to the topology rules. If the calculation result is a topology violation, then the high response region is a violation region. Specifically, the topology rules are as follows: like =left, then the target acupoint is expected to be to the left of the acupoint to be constrained. If the violation condition (5) is satisfied, it is judged as a topological violation: (5); in: This represents the function for determining violations. Indicates the target acupoint exist Coordinates along the axial direction, Indicates acupoints to be constrained. exist Coordinates along the axial direction, This represents the pixel tolerance threshold, which is preferably 2; like =right, then the target acupoint is expected to be to the right of the acupoint to be constrained. If the violation condition (6) is satisfied, it is judged as a topological violation: (6); like =Above, then the target acupoint is expected to be above the acupoint to be constrained. If the violation condition (7) is satisfied, it is judged as a topological violation: (7); in: Indicates the target acupoint exist Coordinates along the axial direction, Indicates acupoints to be constrained. exist Coordinates along the axis; like =down, then the target acupoint is expected to be below the acupoint to be constrained. If the violation condition (8) is satisfied, it is judged as a topological violation: (8).

[0031] High-response regions are identified as having topological violations and marked as such, so that targeted constraint calculations and heatmap corrections can be performed on them in subsequent steps.

[0032] S225: Correct the violation area using the high-response region weighted suppression strategy according to equation (4) to obtain the corrected heatmap, and complete the heatmap calibration based on topology conservation: (4); in: This represents the heatmap after correction for the target acupoint. Indicates the constraint weight. This represents the heat map after the acupoints to be constrained have been corrected.

[0033] Through the above processing, the potentially biased acupoint heatmap predicted by the model can be calibrated into a more accurate, continuous final coordinate image that conforms to anatomical constraints, ensuring the spatial rationality of acupoint location.

[0034] Specifically, the CBAM attention mechanism includes a channel attention submodule and a spatial attention submodule. The channel attention submodule compresses the feature maps generated by the HRNet deep learning network during the extraction and understanding of image content using two methods: global average pooling and global max pooling. This results in a vector reflecting the overall strength of each feature channel and a vector capturing the most significant response within each channel. The vector reflecting the overall strength of each feature channel and the vector capturing the most significant response within each channel are then analyzed by a fully connected neural network. The output results are then summed and activated by the Sigmoid function to obtain the final channel attention weight vector. The role of the channel attention submodule is to determine which types of features are more important. The final generated channel attention weight vector is a vector between 0 and 1. The closer the weight value is to 1, the more important the corresponding feature channel is, and its information will be enhanced in subsequent processing. Conversely, the channel with a weight value closer to 0 is suppressed, thereby reducing the interference of redundant or irrelevant features.

[0035] The spatial attention submodule performs average pooling and max pooling on all feature channels to obtain two single-channel feature maps. The two single-channel maps are then concatenated along the channel dimension to form a dual-channel feature map. After processing by the Sigmoid function, a spatial attention weight map is generated.

[0036] The spatial attention submodule determines which locations in the feature map are more important. Two single-channel feature maps are generated: one reflects the average activity of each location across all channels, while the other highlights the maximum salience of each location in any given channel. These two single-channel maps are concatenated along the channel dimension to form a dual-channel feature map. After processing with the sigmoid function, a spatial attention weight map is generated. Each pixel value in the spatial attention weight map ranges from 0 to 1; higher values ​​indicate a more critical spatial region, potentially containing acupoints, and these regions will receive higher attention weights from the model in subsequent calculations.

[0037] The core of HRNet is a parallel multi-resolution sub-network that performs cross-resolution information exchange and fusion in each fusion operation. Inserting CBAM after fusion means focusing attention on the new fused feature map, and then integrating the output features, which are obtained by multiplying channel attention and then spatial attention. The features integrated by CBAM can be used more accurately for subsequent heatmap generation, thereby improving the accuracy of acupoint localization and the robustness of the model.

[0038] S3: Configure the parameters of the improved HRNet hand acupoint recognition model, input the sample and label data of the training set into the improved HRNet hand acupoint recognition model for training, calculate the difference between the predicted value and the true value through the loss function, and iteratively optimize the parameters of the improved HRNet hand acupoint recognition model until the difference between the predicted value and the true value is within the preset range. The optimized, configured improved HRNet hand acupoint recognition model parameters include: training epochs, batch size, initial learning rate, PCK metric judgment threshold, heatmap Gaussian kernel standard deviation, cosine annealing period, and minimum learning rate decay.

[0039] Specifically, the formula for calculating the loss function is (9): (9); in: Represents the loss function. This represents the loss function for the regression task. This represents the model's predicted heatmap. Indicates the label mask. This represents a heat map after acupoint correction.

[0040] The calculation results are used to guide the model during training, which can learn the correct distribution of acupoint heatmaps in the hand area in a focused and accurate manner, thereby ignoring irrelevant background and achieving high-precision acupoint localization.

[0041] S4: After training, input the hand images in the test set into the improved HRNet hand acupoint recognition model. The improved HRNet hand acupoint recognition model identifies and locates the key acupoints in the hand images in the test set, outputs the acupoint coordinates and location results, and evaluates the recognition results through three indicators: average accuracy, proportion of correct key points, and normalized average error.

[0042] Specifically, the formula for calculating the average accuracy is (10), the formula for calculating the proportion of correct key points is (11), and the formula for calculating the normalized average error is (12): (10); (11); (12); in: Indicates average precision. Indicates the first The accuracy value of each prediction result. To preset the accuracy threshold, For indicator functions, Indicates the correct proportion of key points. Indicates the predicted keypoint coordinates. Represents the actual coordinates of key points. Indicates the distance threshold. Represents Euclidean distance. This represents the normalized average error. This represents the total number of acupoints. For the set of real coordinates, To predict the set of coordinates, Indicates the first The actual coordinates of each acupoint Indicates the first Predicted coordinates of each acupoint. This is a normalized reference value.

[0043] A model with a high average precision (NME) value means that it predicts most acupoints accurately and reliably. A high proportion of correct key points indicates that the model can reliably find the approximate correct location of most acupoints. The smaller the NME value, the more accurate the model's localization.

[0044] A hand acupoint accurate recognition system based on HRNet is used to execute a hand acupoint accurate recognition method based on HRNet as described in any of the above, which includes a preprocessing module, a model building module, a model optimization module and a performance evaluation module; The preprocessing module is used to acquire and preprocess hand image datasets, complete the annotation of key acupoints on the hand, form the sample and label data required for training, and divide the sample and label data required for training into training set and test set according to a preset ratio. The model building module is used to introduce the cosine annealing learning rate strategy, the topology conservation-based heatmap calibration mechanism, and the CBAM attention mechanism into the HRNet hand acupoint recognition model to build an improved HRNet hand acupoint recognition model. The model optimization module is used to input the sample and label data of the training set into the improved HRNet hand acupoint recognition model for training. It calculates the difference between the predicted value and the true value through the loss function, and iteratively optimizes the parameters of the improved HRNet hand acupoint recognition model until the difference between the predicted value and the true value is within a preset range. The performance evaluation module is used to input hand images from the test set into the improved HRNet hand acupoint recognition model, identify and locate key acupoints in the hand images from the test set, output acupoint coordinates and positioning results, and evaluate the recognition results using three indicators: average accuracy, proportion of correct key points, and normalized average error.

[0045] In summary, this invention provides a method and system for accurate hand acupoint recognition based on HRNet. Through a preprocessing module, the original hand images are standardized, data-enhanced, and calibrated using heatmaps based on topology conservation. Constraint rules are constructed based on the inherent anatomical spatial topological relationships of hand acupoints to detect and correct topological violations in the basic heatmap, generating supervised labels that conform to physiological logic to ensure the spatial rationality of acupoint location. Furthermore, by embedding a feature extraction branch of the CBAM attention module, the representation ability of key acupoint features is strengthened. Simultaneously, relying on the high-resolution feature flow and multi-scale fusion characteristics of HRNet throughout the process, both local detail capture and global pose adaptation are considered. A cosine annealing staged training strategy is used to optimize the model. In the early stage, a high learning rate is maintained to quickly learn basic features; in the middle stage, a cosine curve is used to reduce the learning rate to balance convergence and perturbation; and in the later stage, a low learning rate is used for fine-tuning of parameters to avoid the model getting trapped in local optima. Finally, the acupoint coordinates are detected and output through evaluation metrics, achieving accurate recognition of hand acupoints. This solves the problems of insufficient accuracy, weak posture adaptability, and unstable training in existing technologies for hand acupoint recognition, meets the need for millimeter-level positioning of small-scale acupoints, and brings great convenience to product development and practical applications in related fields.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for accurate identification of hand acupoints based on HRNet, characterized in that: Includes the following steps: S1: Obtain the hand image dataset and preprocess it to complete the annotation of key acupoints on the hand, form the sample and label data required for training, and divide the sample and label data required for training into training set and test set according to the preset ratio; S2: An improved HRNet hand acupoint recognition model is constructed by introducing a cosine annealing learning rate strategy, a heatmap calibration mechanism based on topology conservation, and a CBAM attention mechanism. S3: Configure the parameters of the improved HRNet hand acupoint recognition model, input the sample and label data of the training set into the improved HRNet hand acupoint recognition model for training, calculate the difference between the predicted value and the true value through the loss function, and iteratively optimize the parameters of the improved HRNet hand acupoint recognition model until the difference between the predicted value and the true value is within the preset range. S4: After training, input the hand images in the test set into the improved HRNet hand acupoint recognition model. The improved HRNet hand acupoint recognition model identifies and locates the key acupoints in the hand images in the test set, outputs the acupoint coordinates and location results, and evaluates the recognition results through three indicators: average accuracy, proportion of correct key points, and normalized average error.

2. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: The hand image dataset mentioned in step S1 includes palm images and back of hand images, and the preprocessing includes image enhancement and coordinate normalization.

3. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: The cosine annealing learning rate strategy in step S2 is as follows: S211: Divide the learning rate training rounds into three stages: basic learning stage, constraint optimization stage, and fine-tuning stage; S212: During the basic learning phase, the learning rate is maintained at the maximum learning rate within the preset annealing period for training, without triggering annealing; S213: During the constraint optimization phase, the learning rate for each training round is calculated in real time according to equation (1), and constraint optimization is performed according to the calculated learning rate for each training round: （1）； in: express Round learning rate, This represents the minimum learning rate during the annealing cycle. This represents the maximum learning rate during the annealing cycle. This indicates the round number from which the counting begins, starting from the annealing phase. This indicates the total number of annealing cycles from start to finish. S214: In the fine-tuning phase, the learning rate is reduced to the minimum learning rate within the preset annealing period for training.

4. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: The heatmap calibration mechanism based on topology conservation in step S2 is as follows: S221: Calculate the pixel response value of each key acupoint based on the Gaussian kernel function (2) to generate a basic heatmap: （2）； in: Indicates the first The pixel response value of each acupoint. In the heat map, the first... Line number Liede The pixel response value of each acupoint. This represents the natural exponential function. Indicates the first Normalized coordinates of each acupoint This represents the Gaussian kernel standard deviation. This represents the maximum response value in the heatmap; S222: The samples in the training set are filtered for visibility based on the annotation of key acupoints on the hand. Samples whose visibility markers for both the target acupoint and the acupoint to be constrained are visible proceed to the next step. S223: The high-response region is extracted from the basic heatmap formed by the visibility-filtered samples according to equation (3). If the high-response region satisfies equation (3), it is determined as follows: （3）； in: Indicates a high response mask. The base heatmap representing the target acupoint index. This represents the base heatmap of the acupoint index to be constrained. Indicates the response threshold; S224: Calculate the high response region according to the topology rules. If the calculation result is a topology violation, then the high response region is a violation region. S225: Correct the violation area using the high-response region weighted suppression strategy according to equation (4) to obtain the corrected heatmap, and complete the heatmap calibration based on topology conservation: （4）； in: This represents the heatmap after correction for the target acupoint. Indicates constraint weights. This represents the heat map after the acupoints to be constrained have been corrected.

5. The method for accurate identification of hand acupoints based on HRNet according to claim 4, characterized in that: The topology rules described in step S224 are as follows: like =left, then the target acupoint is expected to be to the left of the acupoint to be constrained. If the violation condition (5) is satisfied, it is judged as a topological violation: （5）； in: This represents the function for determining violations. Indicates the target acupoint exist Coordinates along the axis, Indicates acupoints to be constrained. exist Coordinates along the axis, Indicates the pixel tolerance threshold; like = Right, then the target acupoint is expected to be to the right of the acupoint to be constrained. If the violation condition (6) is satisfied, it is judged as a topological violation: （6）； like =Above, then the target acupoint is expected to be above the acupoint to be constrained. If the violation condition (7) is satisfied, it is judged as a topological violation: （7）； in: Indicates the target acupoint exist Coordinates along the axis, Indicates acupoints to be constrained. exist Coordinates along the axis; like =down, then the target acupoint is expected to be below the acupoint to be constrained. If the violation condition (8) is satisfied, it is judged as a topological violation: （8）。 6. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: The CBAM attention mechanism described in step S2 includes a channel attention submodule and a spatial attention submodule. The channel attention submodule compresses the feature maps generated by the HRNet deep learning network during the extraction and understanding of image content using both global average pooling and global max pooling. This yields vectors reflecting the overall strength of each feature channel and vectors capturing the most significant responses within each channel. These vectors are then analyzed by a fully connected neural network, and the outputs are summed and activated by the Sigmoid function to obtain the final channel attention weight vector. The spatial attention submodule performs average pooling and max pooling on all feature channels to obtain two single-channel feature maps. These two single-channel maps are concatenated along the channel dimension to form a dual-channel feature map, which is then processed by the Sigmoid function to generate a spatial attention weight map.

7. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: The parameters of the improved HRNet hand acupoint recognition model configured in step S3 include: training rounds, batch size, initial learning rate, PCK metric judgment threshold, heatmap Gaussian kernel standard deviation, cosine annealing period, and minimum learning rate decay.

8. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: The formula for calculating the loss function in step S3 is (9): （9）； in: Represents the loss function. This represents the loss function for the regression task. This represents the model's predicted heatmap. Indicates the label mask. This represents a heat map after acupoint correction.

9. The method for accurate identification of hand acupoints based on HRNet according to claim 1, characterized in that: In step S4, the formula for calculating the average accuracy is (10), the formula for calculating the proportion of correct key points is (11), and the formula for calculating the normalized average error is (12). （10）； （11）； （12）； in: Indicates average precision. Indicates the first The accuracy value of each prediction result. To preset the accuracy threshold, For indicator functions, Indicates the correct proportion of key points. Indicates the predicted keypoint coordinates. Represents the actual coordinates of key points. Indicates the distance threshold. Represents Euclidean distance. This represents the normalized average error. This represents the total number of acupoints. For the set of real coordinates, To predict the set of coordinates, Indicates the first The actual coordinates of each acupoint Indicates the first Predicted coordinates of each acupoint. This is a normalized reference value.

10. A hand acupoint accurate recognition system based on HRNet, used to execute the hand acupoint accurate recognition method based on HRNet as described in any one of claims 1 to 9, characterized in that: It includes a preprocessing module, a model building module, a model optimization module, and a performance evaluation module; The preprocessing module is used to acquire and preprocess hand image datasets, complete the annotation of key acupoints on the hand, form the sample and label data required for training, and divide the sample and label data required for training into training set and test set according to a preset ratio. The model building module is used to introduce the cosine annealing learning rate strategy, the topology conservation-based heatmap calibration mechanism, and the CBAM attention mechanism into the HRNet hand acupoint recognition model to build an improved HRNet hand acupoint recognition model. The model optimization module is used to input the sample and label data of the training set into the improved HRNet hand acupoint recognition model for training. It calculates the difference between the predicted value and the true value through the loss function, and iteratively optimizes the parameters of the improved HRNet hand acupoint recognition model until the difference between the predicted value and the true value is within a preset range. The performance evaluation module is used to input hand images from the test set into the improved HRNet hand acupoint recognition model, identify and locate key acupoints in the hand images from the test set, output acupoint coordinates and positioning results, and evaluate the recognition results using three indicators: average accuracy, proportion of correct key points, and normalized average error.

Images