Infrared image meridian thermal path positioning method based on deep learning

Abstract

The invention provides an infrared image meridian thermal path positioning method based on deep learning, and the method comprises the steps: firstly, collecting infrared image data, constructing a MIRAGE-YOLO model, inputting an infrared image into the MIRAGE-YOLO model, and outputting skeleton key points; secondly, inferring acupoint coordinates according to skeleton key points, and adjusting the transverse positions of the acupoints through the shoulder width to ensure that the acupoints are aligned with the body contour of the subject; sequentially connecting the deduced acupuncture points along a preset meridian template, and constructing a continuous anatomical trajectory; a bilinear sampling method is adopted to extract the temperature value of each acupuncture point on the track, a temperature curve with continuous space is generated, and quantitative characteristics of the temperature curve are calculated; and finally, outputting an infrared image, a pseudo-color ren vein temperature distribution diagram and a temperature curve which are overlapped with the key points and the acupoint marks, and meanwhile, blurring a non-anatomical background. According to the method, the traditional Chinese medicine meridian theory and the data analysis technology are fused, and heat diagnosis capable of being explained in repeatability and anatomy is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of the intersection of infrared thermal imaging and traditional Chinese meridian theory, and in particular to an infrared image meridian heat path positioning method based on deep learning, which is especially suitable for realizing acupoint positioning, meridian trajectory reconstruction and temperature curve extraction through infrared thermal imaging (IRT) technology, and can be applied to the fields of traditional Chinese medicine clinical diagnosis, physiological state monitoring and medical image analysis. BACKGROUND

[0002] As a non-invasive detection technology, infrared thermal imaging (IRT) technology can obtain the spatially resolved temperature distribution of the human skin surface, and has shown application value in clinical scenarios such as inflammation monitoring, wound assessment, pressure ulcer diagnosis and physiological monitoring. However, existing IRT analysis methods mostly rely on pixel-level thermal visualization or heuristically defined regions of interest, generally ignore anthropometric landmarks, lack systematic solutions to ensure inter-subject consistency and intra-test repeatability, resulting in limited application in anatomically interpretable and quantitatively robust thermal assessment.

[0003] Traditional Chinese meridian theory provides a standardized anatomical framework, believing that physiological processes can be explained through structured meridian paths connecting acupoints of the human body, and the spatial distribution of acupoints follows the "bone degree and inch method" - mapping acupoint positions based on the size ratio of human bones, which provides a natural anatomical prior for spatially structured analysis. However, for a long time, this bone degree ratio model has not been integrated into the modern infrared thermal imaging paradigm, resulting in qualitative assessment of temperature analysis based on meridian theory, lacking anatomical accuracy and method repeatability.

[0004] To fill the above method gap, it is necessary to accurately locate the skeletal landmarks in the infrared image and project the anatomical prior into the temperature distribution. Although existing human pose estimation techniques can extract key points in the visible light and infrared domains, mainstream architectures include high-resolution convolutional networks, Transformer regression modules, and hierarchical coarse-to-fine filtering mechanisms, but these models are mostly optimized for tasks such as action recognition and human tracking, and their adaptability in anatomically aligned thermal measurements (especially in meridian path reconstruction scenarios) has not been fully explored, and the method system is still not perfect.

[0005] Current infrared image acupoint positioning and meridian temperature analysis faces three core challenges: first, infrared images inherently lack clear anatomical landmarks, making it difficult to align thermal data with underlying structures; second, most temperature extraction methods are limited to isolated key points or rough regions, ignoring the continuity of physiological pathways; third, there is a lack of comprehensive framework integrating measurement repeatability, spatial resolution, and uncertainty quantification, which severely limits the application of IRT in pathway-level thermal analysis, including conventional health assessment and TCM diagnosis research. Existing infrared pose estimation methods either rely on RGB-IR domain transfer and paired data supervision (low clinical feasibility) or are designed for coarse skeletal joints, lacking the resolution and spatial prior required for acupoint-level detection; traditional pose estimation models do not incorporate domain-specific anatomical knowledge of TCM acupoints (such as meridian distribution patterns), making it difficult to meet the needs of clinical-level acupoint automatic detection and physiological constraint alignment. SUMMARY

[0006] To address the technical problems of existing infrared thermal imaging temperature analysis lacking anatomical consistency, meridian positioning accuracy, and method repeatability, the present invention proposes an infrared image meridian thermal pathway positioning method based on deep learning, which integrates TCM meridian theory and data analysis technology to achieve accurate anatomical positioning and meridian reconstruction.

[0007] To achieve the above purpose, the technical scheme of the present invention is as follows:

[0008] An infrared image meridian thermal pathway positioning method based on deep learning, the steps are as follows:

[0009] S1, data collection and processing: collect the front and back infrared images of the subject, and pre-process the infrared images;

[0010] S2, build MIRAGE-YOLO model: including basic backbone, neck module and detection head; wherein, taking CSPDarknet53 as the basic backbone, using C2PMaSA composite module to replace C2f+PSA module;

[0011] S3, input the pre-processed infrared image into the MIRAGE-YOLO model, and output the skeletal key points;

[0012] S4, according to the "TCM bone degree and inch" method, define any two skeletal key points constitute a body segment, calculate its Euclidean distance , then through the formula inference of the coordinates of any acupoint on the segment , wherein is the clinical anatomical proportion;

[0013] S5, calculate the shoulder width, adjust the transverse position of the acupoint through the shoulder width, and ensure that the acupoint is aligned with the body contour of the subject;

[0014] S6, sequentially connect the inferred acupoints along the preset meridian template to construct a continuous anatomical trajectory ; the temperature values of each acupoint on the trajectory are extracted by using a bilinear sampling method to generate a spatially continuous temperature curve, and quantitative features of the temperature curve are calculated;

[0015] S7, output the infrared image superimposed with key points and acupoint markers, the pseudo-color temperature distribution map of the meridian, and the temperature curve, while blurring the non-anatomical background.

[0016] Preferably, the method for pre-processing the infrared image is: enhancing the thermal contrast of the infrared image by histogram equalization, and unifying the resolution of the infrared image to 384*288 pixels, and then converting the infrared image into a single-channel temperature matrix.

[0017] Preferably, the C2PMaSA composite module includes two parallel attention branches: a standard parallel separation attention (PSA) branch and a newly introduced multi-scale Manhattan self-attention (MASA) branch; the PSA branch and the MASA branch share the same input features, which are first divided into two channel groups using a 1*1 convolution;

[0018] The PSA branch is divided into S groups in terms of channels, and each group is convolved with different kernel sizes to obtain receptive fields of different scales and extract information of different scales; then, the weighting values of the channels in each group are extracted through an SE module, and finally the weighting values of the S groups are normalized by softmax and weighted;

[0019] The MASA branch uses a lightweight multi-resolution pyramid, and each scale is passed through a dedicated Manhattan self-attention module that calculates position bias attention according to horizontal and vertical receptive fields; then, the outputs of all scales are upsampled and fused through a 1*1 convolution layer to form a scale-aware feature representation;

[0020] Finally, the feature fusion output of the two branches is concatenated with the unprocessed residual path, and a 1*1 convolution projection layer is used to restore the original channel size.

[0021] Preferably, the skeletal key points include nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle; for the analysis of the Ren meridian, five key landmark points, i.e., left shoulder, right shoulder, xiphoid, navel, and pubic bone, are extracted.

[0022] Preferably, the method for calculating the shoulder width is: , wherein, is the coordinate of the left shoulder, is the right shoulder coordinate.

[0023] Preferably, the quantitative features of the temperature curve include average temperature, maximum temperature, neighborhood temperature difference, inter-acupoint gradient, local extreme value;

[0024] wherein the average temperature is the arithmetic mean of all acupoint temperatures; the maximum temperature is the maximum temperature in all acupoints and the corresponding acupoint; the neighborhood temperature difference is the difference between the temperature of each acupoint and the average temperature of the adjacent two acupoints; the inter-acupoint gradient is the temperature change between each acupoint on the temperature curve; and the local extreme value is the maximum and minimum of the local temperature in the temperature curve.

[0025] Advantages of the present application:

[0026] 1) Improve the accuracy of clinical diagnosis: The improved MIRAGE-YOLO model designs a double-flow attention mechanism, allowing the model to capture global semantic features and local acupoint structures simultaneously, improving the pose estimation accuracy under challenging thermal conditions. The combination of multi-scale reasoning, direction prior and residual registration significantly improves the robustness and accuracy of the model in infrared key point detection.

[0027] 2) The detection accuracy of the MIRAGE-YOLO model on the standard and self-defined infrared acupoint dataset is significantly improved, and the extracted meridian temperature curve is highly consistent with the expert annotation, which can provide more reliable anatomical positioning basis for traditional Chinese medicine meridian diagnosis and reduce manual annotation errors.

[0028] 3) By performing bilinear sampling along the anatomically reconstructed meridian path, the temperature curve generated by the present application exhibits strong spatial consistency and high repeatability. Comparison with expert annotations shows that the extracted curve maintains anatomical rationality, and the average meridian processing time is reduced by more than 90%, thereby supporting real-time clinical application scenarios.

[0029] 4) Adapt to real-time clinical scenarios: The single meridian processing efficiency is greatly improved compared to traditional methods, and the lightweight design of the model ensures a high frame rate of 32 FPS, meeting the dual demands of speed and accuracy in clinical real-time diagnosis and providing technical support for rapid evaluation.

[0030] 5) Ensure consistency of cross-scene applications: By optimizing the module to effectively deal with the characteristics of infrared images, the robustness of key point detection is improved, ensuring stable and reliable measurement results under different devices and environments, and laying the foundation for establishing a standardized meridian temperature analysis system. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the description are only some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative effort based on the accompanying drawings belong to the protection scope of the present application.

[0032] Figure 1 Flowchart of the present application.

[0033] Figure 2 Names and serial numbers of meridians and acupoints.

[0034] Figure 3 Overall architecture diagram of the MIRAGE-YOLO model.

[0035] Figure 4 Structure diagram of the improved C2PMaSA.

[0036] Figure 5 End-to-end output processing process of the present application.

[0037] Figure 6 Comparison visualization of the baseline YOLOv11 and the YOLOv11-PA proposed in the present application on heat input; (a) YOLOv11 detection result, (b) MIRAGE-YOLO detection result. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0039] As shown in Figure 1 The present application provides an infrared image meridian heat path positioning method based on deep learning, and the specific steps are as follows:

[0040] S1, data collection and processing: collect the front and back infrared images of the subject, and pre-process the infrared images;

[0041] Dataset construction, recruit subjects (20-60 years old, gender balance), in a constant temperature (24.5°C-26.5°C), uniform background environment, using long-wave infrared thermal imager (384*288 pixels, temperature range 25°C-40°C, thermal sensitivity ±0.025°C) to collect front and back images each 1, a total of 1242 samples; According to the proportion of 7:1.5:1.5, it is divided into training set (869 samples), verification set (186 samples) and test set (187 samples), without overlapping of subjects; Acupoint labeling refers to the WHO standard, based on the method of bone degree and inch, 21 acupoints from CV-2 to CV-22 are labeled by 3 Chinese medicine experts. Figure 2 The selected CV acupoints included in the present application are given, as well as their official WHO code, transliteration and English translation. This schematic diagram highlights the representative acupoint positions on the human pulse axis, which runs along the anterior midline of the human body. Although meridians are not visible structures in anatomy, they play an empirical spatial reference role in traditional Chinese medicine and form the basis for thermal trajectory reconstruction.

[0042] The collected original infrared images are subjected to privacy protection (blurring the facial area and non-anatomical background), the thermal contrast is enhanced by histogram equalization, the image resolution is unified to 384*288 pixels, and it is converted into a single-channel temperature matrix (unit: °C).

[0043] S2, constructing the MIRAGE-YOLO model: including the basic backbone, the neck module and the detection head; as shown in Figure 3 , wherein the CSPDarknet53 is used as the basic backbone, and the C2PMaSA composite module is used to replace the C2f+PSA module.

[0044] In order to enhance the robustness of key point positioning in low resolution and low texture infrared images, the present application proposes a C2PMaSA composite module, which replaces the original C2PSA block in the S feature backbone of YOLOv11, and is specially designed to improve the semantic representation and spatial perception in thermal imaging environment. The module contains two parallel branches-PSA branch (capturing global semantic features, modeling inter-channel dependence) and MaSA branch (constructing multi-resolution pyramid, enhancing the directional perception of acupoints along meridians through Manhattan self-attention). Both attention branches integrate PositionAlignResidual modules, which process input features through lightweight convolution and batch normalization, and connect with attention output residuals to stabilize infrared domain attention weight learning; The branch output is adaptively weighted by the dynamic gate fusion (DGF) module, which can flexibly focus on global / local features without manual parameter tuning. Three parallel detection heads are used, corresponding to different spatial resolutions, to simultaneously realize the regression of coarse skeletal joints and fine-grained acupoints, and output key point heat maps and coordinates.

[0045] a weighted compound loss function is adopted wherein CIoU is a CIoU bounding box regression loss, BCE is a binary cross-entropy loss (object existence, classification), Lkp is a keypoint spatial accuracy loss, Lvis is a keypoint visibility confidence loss, balancing the contribution of each loss by a weight coefficient.

[0046] As shown in Figure 4 , the C2PMaSA compound module includes two parallel attention branches: a standard parallel split attention (PSA) branch and a newly introduced multi-scale Manhattan self-attention (MASA) branch; the PSA branch and the MASA branch share the same input features, which are first divided into two channel groups using a 1x1 convolution. The PSA branch retains the traditional attention structure of YOLOv11, which is very effective in modeling inter-channel dependencies.

[0047] The PSA branch is divided into S groups from the channel, and each group is convolved with different kernel sizes to obtain different scales of receptive fields and extract information at different scales; then, the SE module is used to extract the weighting values of the channels in each group, and finally the weighted values of the S groups are normalized by softmax and weighted.

[0048] The MASA branch focuses on capturing local spatial patterns and directional perception priors, which are crucial for accurate acupoint alignment along the meridian system of the human body. For multi-scale reasoning, the MASA branch uses a lightweight multi-resolution pyramid, each scale of which is passed through a dedicated Manhattan self-attention module that calculates position-biased attention based on horizontal and vertical receptive fields; then, the outputs of all scales are upsampled and fused through a 1x1 convolution layer to form a scale-aware feature representation.

[0049] To ensure spatial alignment and element continuity, both branches are augmented with PositionAlignResidual modules. This component performs lightweight convolution and batch normalization on the input and adds it to the respective attention output. This residual alignment preserves the semantic integrity of the input while stabilizing the learning process of attention weights in the infrared domain.

[0050] The outputs of the two branches are fused through an adaptive fusion module that uses a dynamic gating mechanism to weight the contributions of the PSA and MASA branches. This adaptive fusion ensures that the model emphasizes global or local context based on the specific input without the need for manual tuning. The final fused output is concatenated with the unprocessed residual path and projected through a 1x1 convolution layer to restore the original channel size.

[0051] The multi-scale MaSA module enhances the model's ability to extract spatially-aware features across multiple resolutions, capturing local details and long-range structures through a hierarchical representation with a novel Manhattan distance prior.

[0052] Given the input feature map , three scale representations are generated:

[0053]

[0054] Each scale passes through a Manhattan Self-Attention (MaSA) block. For position (i,j), the attention weight is computed using Manhattan distance:

[0055]

[0056] The attention map for scale s is computed as:

[0057]

[0058] The attention weight is modulated by the following equation:

[0059]

[0060] where: Q is the query matrix, K is the key matrix, V is the value matrix, is a learnable spatial decay coefficient (initialized to 0.2), N defines a 7x7 local neighborhood, is the key dimension.

[0061] The multi-scale outputs are fused by:

[0062]

[0063] Final output Retains detailed spatial details and global structural patterns, showing special robustness in infrared image processing.

[0064] S3, the preprocessed infrared image is input into the MIRAGE-YOLO model, and the output is the skeletal key point; as a geometric anchor point for anatomical modeling. The skeletal key point coordinates include: nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, right ankle; for the analysis of the Ren channel, the left shoulder, right shoulder, xiphoid, navel, and pubic bone are the five key landmark points.

[0065] Hardware and software environment setup: The training environment is an Ubuntu 22.0 system equipped with 8 NVIDIA RTX 4090D GPUs (24GB / GPU), a 64-core Intel Xeon Platinum 8481C CPU, and 320GB of RAM. The software is based on PyTorch 2.5.1, Python 3.12, and CUDA 12.4. The inference environment supports CPU / GPU switching, with a GPU inference frame rate of 32 FPS and a CPU inference frame rate of 8 FPS.

[0066] Model parameter settings: The weights of the loss function for MIRAGE-YOLO are set to... , , , , Spatial attenuation coefficient of MaSA module Initialize to 0.2, local neighborhood N is set to 7×7; training batch size is 32, initial learning rate is 1e-4, and training epochs are 1000.

[0067] S4, based on the "Traditional Chinese Medicine Bone Measurement Method," defines any two key skeletal points. Calculate the Euclidean distance of the line segments that form the body. Then through the formula Infer any acupoint on the line segment The coordinates, where For clinical anatomical proportions.

[0068] S5, calculate shoulder width, adjust the lateral position of acupoints based on shoulder width to ensure that acupoints are aligned with the subject's body contour, and improve cross-subject measurement repeatability.

[0069] Deducing the Ren meridian acupoints, based on shoulder width As a morphological normalization reference; secondly, based on the xiphoid process-umbilicus segment, the Zhongwan acupoint (CV-12) was inferred according to the bone ratio. ), Xiawan acupoint (CV-10, ), moisture hole (CV-9, Based on the umbilicus-pubic symphysis line segment, the Guanyuan acupoint (CV-4) was deduced. ), Zhongji acupoint (CV-3, ), Qugu point (CV-2, Finally, the remaining acupoints on the Ren meridian (such as Tanzhong CV-17, Jiuwei CV-15, etc.) are deduced sequentially according to the traditional Chinese medicine bone measurement method, based on the shoulder width. Adjust the horizontal coordinates to ensure alignment with the body's midline.

[0070] S6. Connect the inferred acupoints sequentially along the preset meridian template (from CV-22 to CV-2) to construct a continuous anatomical trajectory. (t represents the meridian number); The temperature values ​​of each acupoint on the trajectory are extracted using a bilinear sampling method to generate a spatially continuous temperature curve, and the quantitative characteristics of the temperature curve are calculated: average temperature... The arithmetic mean of the temperatures of 21 acupoints; the highest temperature. The maximum temperature among 21 acupoints and their corresponding acupoints; neighborhood temperature difference. The difference between the temperature of each acupoint and the average temperature of two adjacent acupoints is represented by the temperature gradient between acupoints on the temperature curve. The local extreme values ​​are the highest and lowest local temperatures on the temperature curve.

[0071] The method of this invention consists of two stages: acupoint localization and temperature feature extraction, as detailed in Table 1. First, candidate regions are estimated by selecting a point located at one-third of the geometric distance between two predefined anatomical landmarks. This pre-guided localization provides a coarse but reliable constraint on the acupoint search space. Subsequently, edge detection and contour fitting are performed within the candidate regions to refine the acupoint center by aligning it with the dominant local contour structure. Once the final center is determined, temperature patches are extracted from the infrared image, and statistical features are calculated: mean temperature (Tmean), maximum temperature (Tmax), and temperature contrast with neighboring regions (TDelta). These quantitative descriptors allow for downstream modeling of thermal properties for diagnostic or analytical purposes.

[0072] Table 1. Acupoint Location and Temperature Feature Extraction

[0073]

[0074] S7, Results Output and Visualization: Outputs a list of Ren meridian acupoint coordinates and temperature curve data (including...). , , The visualization module displays: an infrared image with superimposed key points and acupoint markers; a pseudo-color Ren meridian temperature distribution map (Jet color scale); and a Ren meridian temperature curve (horizontal axis: acupoint codes CV-2 to CV-22, vertical axis: temperature °C).

[0075] Model Performance Validation

[0076] The performance of the proposed method is first evaluated by visualizing its complete processing flow and comparing the anatomical consistency among model variants. These results are crucial for verifying whether the proposed system can maintain the spatial correspondence between anatomical structures and thermal signals under realistic conditions.

[0077] Figure 5End-to-end processing of the framework proposed by the present invention is illustrated. Starting from an original infrared thermal imaging (IRT) image with low spatial contrast, the model sequentially performs skeleton keypoint detection, acupoint inference, meridian mapping, and thermal profile extraction. The leftmost column shows the original IRT image, where anatomical structures are barely distinguishable due to the lack of texture. In the subsequent images, the detected keypoints, estimated acupoints, and pseudo-color temperature maps are shown. The acupoints (black dots) are positioned along consistent spatial alignments in the expected anatomical regions, even under low-contrast input conditions. The color maps reveal smooth temperature transitions along the constructed acupoint trajectories. This confirms that the present method is able to preserve anatomical authenticity and thermal continuity, which are fundamental for reliable temperature-based diagnostics. Furthermore, by blurring non-anatomical backgrounds and avoiding facial identifiers, privacy protection is ensured, making the method ethically feasible for clinical data acquisition.

[0078] To further assess the accuracy of anatomical positioning, Figure 6 A side-by-side comparison between the baseline YOLOv11-Pose model and the enhanced YOLOv11-PA variant (incorporating the proposed C2PMaSA module) is provided. The baseline model shows inaccurate detection of the ankle joint, leading to misaligned acupoint interpolation and a significantly distorted meridian path. This failure would result in anatomical discordance, especially near the foot, where the meridians deviate from reasonable trajectories. In contrast, the YOLOv11-PA variant can correctly detect the ankle keypoint, allowing acupoints to be placed along smooth and anatomically coherent curves. This comparison highlights the importance of robust keypoint detection in thermal environments. Improved positioning accuracy not only enhances visual consistency but also directly impacts the reliability of meridian path thermal measurements. Since temperature values are interpolated along these paths, any deviation in keypoint placement will propagate downstream, affecting diagnostic accuracy and measurement reproducibility.

[0079] Table 2 presents quantitative results of acupoint detection by different model variants. The proposed C2PMaSA-enhanced models consistently outperform their baseline YOLOv11 counterparts on both the COCO2017 validation set and the thermal meridian dataset collected by the present invention. Specifically, the slim-PA variant achieves the highest mean average precision (mAP) of 0.813 on the thermal dataset, reflecting significant improvements in keypoint localization under low-texture infrared conditions.

[0080] Table 2: Performance comparison of baseline YOLOv11 models and C2PMaSA-enhanced variants on COCO2017 and infrared datasets.

[0081]

[0082] The metrics in Table 2 include mAP@0.5, mAP@0.9, recall (R) and precision (P) across object scales.

[0083] Compared to traditional models, the present method exhibits enhanced performance in overall mAP, as well as region-specific precision and recall metrics for small, medium, and large anatomical targets, as shown in Table 2. This confirms the effectiveness of the multi-scale attention fusion and anatomical alignment strategies of the present invention.

[0084] Table 3 demonstrates comparative evaluations between the proposed MIRAGE-YOLO and several state-of-the-art pose estimation models adapted for infrared images. Although ViTPose based on transformers achieves a competitive mAP of 0.809, its high computational cost and low inference speed significantly limit its applicability in real-time or resource-constrained thermal diagnostic systems. Similarly, HRNet and ED-Pose++ provide moderate accuracy but exhibit high model complexity and low frame rates.

[0085] Table 3 Comparative evaluation with state-of-the-art pose estimation models on the infrared dataset.

[0086]

[0087] In contrast, the proposed model achieves the highest accuracy on the infrared dataset while maintaining real-time performance and lightweight parameter size. This indicates that the present method provides a good trade-off between accuracy, efficiency, and model size, making it particularly suitable for practical deployment in infrared medical imaging applications.

[0088] Furthermore, this balance is achieved through targeted architectural enhancements, such as the C2PMaSA module, which effectively integrates local thermal structures and multi-scale spatial priors without introducing excessive computational overhead. Unlike general-purpose pose models designed for RGB datasets, the present method is tailored to the low-texture, noise-sensitive characteristics of infrared images, enabling real-time anatomically meaningful acupoint detection and temperature analysis.

[0089] Overall, these results validate the effectiveness and practicality of the present method, highlighting its potential for robust thermal pose estimation under various scales and imaging conditions.

[0090] The above description is merely preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of protection of the present application.

Claims

1. A method for locating meridian thermal paths in infrared images based on deep learning, characterized in that, The steps are as follows: S1, Data Collection and Processing: Acquire infrared images of the front and back of the subject and preprocess the infrared images; S2, Construct the MIRAGE-YOLO model: including the basic backbone, neck module and detection head; among which, CSPDarknet53 is used as the basic backbone, and the C2f+PSA module is replaced by the C2PMaSA composite module; S3: Input the preprocessed infrared image into the MIRAGE-YOLO model and output the skeletal key points; S4, based on the "Traditional Chinese Medicine Bone Measurement Method", defines any two key points of the skeleton. Calculate the Euclidean distance of the line segments that form the body. Then through the formula Infer any acupoint on the line segment The coordinates, where For clinical anatomical proportions; S5, calculate the shoulder width, and adjust the lateral position of the acupoints based on the shoulder width to ensure that the acupoints are aligned with the subject's body contour; S6 connects the inferred acupoints sequentially along a preset meridian template to construct a continuous anatomical trajectory. The temperature values ​​of each acupoint on the trajectory were extracted using the bilinear sampling method to generate a spatially continuous temperature curve, and the quantitative characteristics of the temperature curve were calculated. S7 outputs an infrared image with superimposed key points and acupoint markers, a pseudo-color Ren meridian temperature distribution map, and a temperature curve, while blurring the non-anatomical background.

2. The infrared image meridian thermal path localization method based on deep learning according to claim 1, characterized in that, The method for preprocessing infrared images is as follows: enhance the thermal contrast of infrared images through histogram equalization, unify the resolution of infrared images to 384×288 pixels, and then convert the infrared images into a single-channel temperature matrix.

3. The infrared image meridian thermal path localization method based on deep learning according to claim 1, characterized in that, The C2PMaSA composite module includes two parallel attention branches: the standard parallel split attention PSA branch and the newly introduced multi-scale Manhattan self-attention MASA branch; the PSA branch and the MASA branch share the same input features, which are first divided into two channel groups using a 1×1 convolution. The PSA branch divides the channels into S groups, and each group undergoes convolution with different kernel sizes to obtain receptive fields of different scales and extract information at different scales. Then, the SE module extracts the channel weights of each group, and finally, the weights of the S groups are normalized by softmax and weighted. The MASA branch uses a lightweight multi-resolution pyramid, where each scale is passed through a dedicated Manhattan self-attention module that calculates positional bias attention based on the horizontal and vertical receptive fields; then, the outputs at all scales are upsampled and fused through 1×1 convolutional layers to form scale-aware feature representations. Finally, the feature fusion outputs of the two branches are concatenated with the unprocessed residual path and the original channel size is recovered through a 1×1 convolutional projection layer.

4. The infrared image meridian thermal path localization method based on deep learning according to claim 1, characterized in that, The key skeletal points include the nose, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle; for the analysis of the Ren meridian, five key landmarks are extracted: left shoulder, right shoulder, xiphoid process, umbilicus, and pubis.

5. The infrared image meridian thermal path localization method based on deep learning according to claim 1, characterized in that, The method for calculating the shoulder width is as follows: ,in, The coordinates of the left shoulder The coordinates are for the right shoulder.

6. The infrared image meridian thermal path localization method based on deep learning according to claim 4, characterized in that, The quantitative characteristics of the temperature curve include average temperature, maximum temperature, neighborhood temperature difference, gradient between acupoints, and local extreme values. Among them, average temperature The arithmetic mean of the temperatures of all acupoints; the highest temperature. This represents the maximum temperature value and corresponding acupoint among all acupoints; neighborhood temperature difference. The difference between the temperature of each acupoint and the average temperature of two adjacent acupoints is represented by the temperature gradient between acupoints on the temperature curve. The local extreme values ​​are the highest and lowest local temperatures on the temperature curve.

Images