Ear acupoint recognition and positioning method

Abstract

The invention relates to the technical field of acupoint data processing, in particular to an ear acupoint recognition and positioning method, designs a novel ear acupoint detection pen which integrates a miniature camera and an electric signal probe, and provides a method based on a converter and a deep convolutional network on the basis of the novel ear acupoint detection pen. The ear image and the ear electric signal data are dynamically fused according to different positions of the ear acupuncture points, so that the ear acupuncture points can be identified and positioned more quickly, flexibly and accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of acupoint data processing, and particularly relates to a method for identifying and locating auricular acupoints. Background Art

[0002] Auricular point therapy originated from traditional Chinese medicine and has been widely applied in modern medical practice. The auricle is regarded as a microcosm of the human body, on which there are reflection areas corresponding to various organs and tissues of the whole body. By stimulating these specific areas, the functions of corresponding zang-fu organs can be regulated to achieve the purpose of preventing and treating diseases. However, there are more than 90 auricular acupoints, which are densely distributed and have significant individual differences. In addition, the auricular structure is special, with complex concave and convex curved surfaces, folds and hidden areas, and accurate positioning has always been a difficult point in practice.

[0003] Currently, there are mainly the following methods for auricular acupoint positioning:

[0004] Traditional manual positioning: Physicians locate acupoints based on anatomical landmarks and experience, which is highly subjective, relies on the professional level of the operator, and is time-consuming and laborious.

[0005] Electrical signal measurement method: Based on the characteristic that the skin resistance in the acupoint area is lower than that of the surrounding tissues, the acupoints are detected by measuring the resistance change (or current telephone). There are already various resistance-type auricular acupoint positioning detection pens on the market. When the probe touches the acupoint, the device will prompt the operator through beeping or an indicator light. However, such devices have two main disadvantages: one is that the user does not know which specific acupoint is being touched and must refer to the acupoint atlas at the same time; the other is that the accuracy of relying solely on the resistance measurement method is limited and is easily affected by factors such as skin humidity and contact pressure. Almost all portable auricular acupoint positioning pens on the market currently adopt this method.

[0006] Image positioning method: For example, the patent with the application number CN202411396308.2 proposes to locate acupoints by processing binocular auricular images through a convolutional neural network. However, this method is mainly applied to head-mounted devices. After the device is fixed on the head, the camera can only collect images at a fixed angle and it is difficult to cover the acupoints in areas such as the inner ear, lacking flexibility.

[0007] The ear, as a special area of the human body, has the following characteristics: one is that the structure is complex, including multiple anatomical areas such as the concha, helix, antihelix, tragus, triangular fossa, and earlobe, forming a complex three-dimensional structure; the second is that the acupoints are widely distributed, not only on the outer ear plane, but also extending to the inner ear, edge and deep positions; the third is that there are significant individual differences, and the ear shapes, sizes and angles of different people are different. These characteristics make it difficult for traditional single-modal detection methods to meet the comprehensive and accurate auricular acupoint positioning requirements.

[0008] From a market perspective, we believe that there is currently a lack of portable devices in the market that can quickly, flexibly, and accurately identify and locate ear acupoints.

[0009] From a technical perspective, there is currently no product or technology in the market that integrates a current probe and a micro camera in the same ear acupoint positioning detection pen. This integration method is of particularly important significance for acupoint positioning in the ear with its special structure. Since ear acupoints are distributed in multiple areas such as the outer side, inner side, and edge, the design of the ear acupoint positioning detection pen enables the operator to flexibly adjust the detection angle and position to cover all areas of the ear; and the combination of multi-modal information of electrical signals and image signals can make up for the limitations of a single type of signal and provide a more comprehensive and accurate basis for acupoint judgment.

[0010] Therefore, there is an urgent need for a new technical solution that can simultaneously identify and accurately locate acupoints and has sufficient flexibility to cover acupoints in all areas of the ear to solve the pain points of consumers and meet the needs of the market. For this reason, we propose the present invention. Summary of the Invention

[0011] The object of the present invention is to provide an ear acupoint positioning method. By innovatively integrating a micro camera and a micro current measurement module on the same ear acupoint positioning detection pen and combining advanced deep learning algorithms to process multi-modal data, high-precision identification and positioning of ear acupoints are achieved. At the same time, the acupoint name and related information are reported by a buzzer and displayed on the screen, greatly improving the convenience and practicality of clinical applications.

[0012] An ear acupoint positioning method, which is completed by using an ear acupoint positioning system. The system includes a hardware part and a software part. The positioning method includes the following steps:

[0013] Contact the ear with the ear acupoint detection pen, and simultaneously collect tiny current signals and ear images;

[0014] Preprocess the collected tiny current signals and ear images;

[0015] Use a transformer encoder to extract current signal features from the preprocessed tiny current signals;

[0016] Use a deep convolutional network to extract image features from the preprocessed ear images;

[0017] Use a multi-modal dynamic fusion module to fuse the extracted current signal features and image features to obtain fusion features;

[0018] Based on the fusion features, simultaneously identify the type and precise position of the ear acupoints;

[0019] Feed the recognition result back to the user through buzzer reporting and screen display.

[0020] The core innovations of the present invention include:

[0021] For the first time, a micro camera and a micro current measurement module are integrated into the same auricular acupoint positioning detection pen. The multi-modal fusion architecture of the transducer-depth convolutional network is used to jointly process the ear electrical signals and image signals to accurately identify and locate the auricular acupoints, and a buzzer and a screen are equipped to intuitively display the acupoint names and relevant information, enabling users to accurately locate and understand the acupoint information without referring to the acupoint atlas. The same device realizes the multi-modal information collection and real-time feedback of auricular acupoints, enhances the operation flexibility, and can cover the acupoints in various regions of the ear.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. High-precision identification and positioning: Combining the current signal and the image information to give full play to the complementary advantages of the two modalities;

[0024] 2. Comprehensive coverage: For the first time, a current probe and a micro camera are integrated into the same auricular acupoint positioning detection pen. The pen design has high flexibility and can cover various acupoints in the outer, inner and edge regions of the ear;

[0025] 3. Intelligent identification and feedback: Automatically identify the specific acupoint type in contact, and announce the acupoint name and relevant information through the buzzer and display on the screen, without the operator referring to the acupoint atlas at the same time, greatly improving the convenience of use; BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is the overall architecture schematic diagram of the auricular acupoint positioning system of the present invention.

[0027] Figure 2 is the rendering diagram of the auricular acupoint positioning detection pen provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Overall architecture of the system:

[0030] As Figure 1As shown in the figure, the auricular acupoint positioning system of the present invention mainly includes a hardware part and a software part. The hardware part includes an auricular acupoint positioning detection pen and a main control unit. The main control unit can be located in the detection pen or outside the detection pen. When it is located outside the detection pen, it can be a separate hardware or integrated into an intelligent terminal (such as a smart phone). The detection pen can be connected to the external main control unit through a communication unit. The software part includes a data acquisition and preprocessing module, a feature extraction module, a multi-modal fusion module, and an acupoint recognition and positioning module. When the system works, the auricular acupoint positioning detection pen contacts the ear and collects micro-current signals and image data, which are transmitted to the main control unit through a data line or wirelessly. After algorithm processing, the acupoint recognition result is output, and the acupoint name and related information are announced through a buzzer and displayed on the screen to provide intuitive feedback to the user.

[0031] Hardware Design of the Auricular Acupoint Positioning Detection Pen:

[0032] As Figure 2 shown, the auricular acupoint positioning detection pen includes the following main components:

[0033] Probe: Located at the front end of the auricular acupoint positioning detection pen, made of medical-grade stainless steel material, directly contacting the ear and conducting micro-current;

[0034] LED lamp: One or more, arranged at the front end of the auricular acupoint positioning detection pen, providing uniform illumination and reducing the influence of shadows;

[0035] Micro camera: Integrated near the probe, capable of taking clear images of the area contacted by the probe;

[0036] Buzzer: Set on the auricular acupoint positioning detection pen, used to announce the name of the auricular acupoint contacted, with adjustable volume and supporting multi-language announcements;

[0037] Screen: Installed on the upper part of the auricular acupoint positioning detection pen, capable of displaying the acupoint name and other information;

[0038] Button: Set at a convenient operation position on the auricular acupoint positioning detection pen, used to trigger detection, switch functions, adjust volume, etc.

[0039] The key technical parameters of the auricular acupoint positioning detection pen are shown in Table 1 below.

[0040] Table 1 Technical Parameter Table of the Auricular Acupoint Positioning Detection Pen:

[0041] ;

[0042] One of the core innovations of the present invention is to integrate the current probe and the micro camera into the same ear acupuncture point location detection pen, and to accurately locate the ear acupuncture points by combining the electrical signal and the image signal, which has not been achieved in the prior art. This integrated design is particularly suitable for the detection of ear acupuncture points, because the ear structure is complex and the acupuncture points are distributed in a three-dimensional space with many bumps and wrinkles. The pen-type design can flexibly adjust the angle and position to touch the acupuncture points in various areas of the ear. At the same time, the coordinated collection of current signals and image data provides complementary information for the algorithm, greatly improving the accuracy of acupuncture point location and recognition.

[0043] In addition, the integrated buzzer and screen provide instant feedback of acupoint information, eliminating the need for users to refer to acupoint maps at the same time, greatly improving the convenience of application. This "one-stroke multi-function" design is a breakthrough in traditional auricular acupoint detection technology and has high practical value.

[0044] The pen also includes: Current measurement circuit: generates a stable, tiny AC current, conducts it to the ear through the probe, and measures the resistance change;

[0045] Temperature control unit: built-in temperature sensor and micro heating element to control the probe temperature in the range of 32-34 degrees Celsius, close to human skin temperature;

[0046] Signal processing circuit: including preamplifier, filter and 24-bit ADC, with sampling rate of 1kHz, to process tiny current signals;

[0047] The working process is as follows:

[0048] Data collection and preprocessing:

[0049] 1. Current signal acquisition and preprocessing:

[0050] The acquisition and preprocessing of current signals include the following steps:

[0051] (1) Sampling control: obtaining current data according to the set sampling frequency;

[0052] (2) Real-time filtering: Apply a bandpass filter to remove high-frequency noise and low-frequency drift:

[0053] ;

[0054] in, represents the filtered signal, represents the filter coefficients, represents the filter order, Represents the historical values ​​of the input signal.

[0055] (3)Normalization: The Z-score normalization method is adopted to make the signal mean 0 and the standard deviation 1:

[0056] ;

[0057] where, represents the normalized signal, represents the signal mean, represents the signal standard deviation, represents the input signal value at time t.

[0058] (4)Time-frequency transformation: The short-time Fourier transform (STFT) is applied to extract the time-frequency features of the signal:

[0059] ;

[0060] where j is the imaginary unit, t represents the time step, represents at time and frequency the STFT coefficient at, represents centered on the window function.

[0061] (5)Feature enhancement: The adaptive threshold method is adopted to enhance the features related to acupoints:

[0062] ;

[0063] where, represents the enhanced signal, represents the adaptive threshold, which is dynamically determined according to the signal distribution.

[0064] 2. Image acquisition and preprocessing:

[0065] The image acquisition and preprocessing include the following steps:

[0066] (1)Automatic exposure: Adjust the exposure parameters according to the ambient light conditions;

[0067] (2)Focus detection: Evaluate the image sharpness and prompt the user to adjust the distance;

[0068] (3)Color correction: Apply the color constancy algorithm to adapt to different lighting conditions:

[0069] ;

[0070] where, represents the corrected image, represents the original image, represents the color transformation matrix, which is dynamically calculated according to the image white balance estimation.

[0071] (4)Contrast enhancement: Using an improved CLAHE algorithm:

[0072] ;

[0073] where, represents the enhanced image, is the contrast limit parameter.

[0074] (5)Normalization: Normalize the image pixel values to the range [-1, 1]:

[0075] ;

[0076] (6)Spatial transformation: Apply an affine transformation for geometric correction to compensate for the camera viewing angle difference.

[0077] The key technical parameters of the preprocessing are shown in Table 2 below.

[0078] Table 2 Data preprocessing technical parameter table:

[0079] ;

[0080] Transformer-based current signal feature extraction:

[0081] Considering the temporal characteristics of the current signal, the present invention uses an improved transformer encoder to process the current signal features.

[0082] 1. Input embedding: Convert the preprocessed current signal sequence into an embedding representation through a linear mapping:

[0083] ;

[0084] where, represents the embedding sequence, and are the weight and bias parameters of the embedding layer respectively, is the input feature dimension, is the model hidden layer dimension.

[0085] 2. Positional encoding: To retain the temporal information of the signal, use sine / cosine positional encoding:

[0086] ;

[0087] ;

[0088] where, represents the positional encoding value at position in the th dimension. The positional encoding is directly added to the embedding representation:

[0089] , which is a code that combines input embeddings and positional encodings.

[0090] 3. Multi-Head Self-Attention Mechanism: The multi-head self-attention mechanism is used to capture different patterns in the sequence:

[0091] ;

[0092] Among them, the attention calculation for each head is:

[0093] ;

[0094] ;

[0095] Here, represent the query, key, and value matrices respectively, is a learnable parameter matrix, is the dimension of the key vector, and T represents transpose.

[0096] 4. Improved Feed-Forward Network: The gated linear unit is used to replace the traditional feed-forward network to improve the model's expressive power:

[0097] ;

[0098] Among them, represents element-wise multiplication, represents the activation function, are learnable parameters.

[0099] 5. Adaptive Layer Normalization: To handle signal strength variations, conditional layer normalization is used:

[0100] ;

[0101] Among them, and are the mean and standard deviation of the features respectively, and are the scaling and offset parameters generated by the conditional vector .

[0102] The complete Transformer encoder layer can be represented as:

[0103] ;

[0104] ;

[0105] Stack such encoder layers, and finally obtain the high-level feature representation of the current signal.

[0106] The key technical parameters for extracting the current signal features are shown in Table 3 below.

[0107] Table 3 Technical Parameter Table for Extracting Current Signal Features:

[0108] ;

[0109] Image Feature Extraction Based on a Deep Convolutional Network:

[0110] Considering the special nature of ear acupoints (small targets, low contrast, and dense distribution), the present invention designs a dedicated deep convolutional network architecture for image feature extraction. Its main components include:

[0111] 1. Network backbone: ResNet-18 is used as the basic framework, but the following modifications are made:

[0112] Reduce the size of the first convolutional kernel to 3×3, with a stride of 1, to retain more detailed information;

[0113] Increase the number of channels of the shallow features to enhance the representation ability of fine textures;

[0114] Introduce dilated convolution in the deep features to expand the receptive field while maintaining the feature resolution.

[0115] 2. Pyramid feature fusion: To process acupoints of different sizes and shapes, a Feature Pyramid Network (FPN) structure is adopted:

[0116] ;

[0117] Among them, represents the feature pyramid feature of the th layer, represents the feature map of the deep convolutional network of the th layer, represents a 1×1 convolution operation, represents an upsampling operation.

[0118] 3. Attention enhancement module: Introduce channel and spatial attention mechanisms to highlight the feature regions related to acupoints:

[0119] ;

[0120] ;

[0121] ;

[0122] Among them, represents the input feature map, and respectively represent the channel and spatial attention maps, represents the feature map after attention enhancement, represents broadcast multiplication.

[0123] 4. Multi-scale context module: To capture the hierarchical relationship of the ear structure, a multi-scale context aggregation module is designed:

[0124] ;

[0125] ;

[0126] Among them, represents the dilated convolution with a dilation rate of the hollow convolution, represents the multi-scale feature connection, represents the final feature of context enhancement.

[0127] The complete image feature extraction process can be described as:

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] Among them, represents the finally extracted image features, and are the height and width of the feature map respectively, is the number of feature channels.

[0133] The key technical parameters of image feature extraction are shown in Table 4 below.

[0134] Table 4 Image Feature Extraction Technical Parameter Table:

[0135] ;

[0136] Multi-modal dynamic fusion module:

[0137] In view of the special requirements of auricular point location, the present invention proposes a multi-modal dynamic fusion module, which can dynamically adjust the importance of different modalities according to the characteristics of the input data to achieve adaptive fusion.

[0138] The design of the multi-modal dynamic fusion module is as follows:

[0139] 1. Feature alignment and representation:

[0140] (1)Current feature reshaping: Reshape the current features extracted by the converter by performing temporal aggregation and dimensional adjustment:

[0141] ;

[0142] Among them, is the attention weight at time step , calculated through the self-attention mechanism:

[0143] ;

[0144] The obtained aggregated feature is then converted into a representation compatible with the image features through a mapping layer:

[0145] ;

[0146] Among them, represents the aligned current feature, and are learnable parameters.

[0147] (2)Image feature integration: The image feature map extracted by the deep convolutional network is converted into a global feature representation through global pooling and mapping:

[0148] ;

[0149] ;

[0150] Among them, represents the pooled image feature, represents the aligned image feature, and are learnable parameters.

[0151] 2. Dynamic fusion weight calculation:

[0152] (1)Modal specificity evaluation: First, evaluate the reliability and information content of each modal feature:

[0153] ;

[0154] ;

[0155] Among them, respectively represent the scores of the current and image features, are learnable parameters.

[0156] (2)Cross-modal attention: Calculate the mutual correlation between the two modalities:

[0157] ;

[0158] ;

[0159] Among them, respectively represent the influence of current features on image features and the influence of image features on current features, represents the activation function, represents the feature connection operation.

[0160] (3)Dynamic fusion weight calculation: Calculate the fusion weight based on the modal score and cross-attention:

[0161] ;

[0162] ;

[0163] Among them, respectively represent the fusion weights of current and image features.

[0164] 3. Feature fusion and enhancement:

[0165] (1)Global feature fusion: Use the calculated dynamic weights to fuse the global features:

[0166] ;

[0167] Among them, represents the fused global feature.

[0168] (2)Spatial feature retention: Simultaneously retain the spatial features extracted by the deep convolutional network for subsequent localization tasks:

[0169] ;

[0170] Among them, represents the image features retaining spatial information.

[0171] (3)Attention-guided spatial feature enhancement: Use the global features to enhance the spatial feature map:

[0172] ;

[0173] ;

[0174] Among them, represents the spatial attention map, represents broadcasting the global features to the feature map size, represents the spatial features after attention enhancement.

[0175] 4. Final output:

[0176] ;

[0177] Among them, represents the final fused feature representation, which includes two parts: global feature and spatial feature, and is used for subsequent classification and localization tasks respectively.

[0178] The key technical parameters of the multi-modal dynamic fusion module are shown in Table 5 below.

[0179] Table 5 Technical parameter table of the multi-modal dynamic fusion module:

[0180] ;

[0181] The multi-modal dynamic fusion module solves the problem of effective fusion of two heterogeneous modal data, namely current signal and image. Different from simple concatenation or fixed-weight fusion, the multi-modal dynamic fusion module can dynamically adjust the contribution weights of each modality according to different ear regions and acupoint characteristics, and proves its strong adaptability in simulation experiments. For example, for the acupoints on the inner ear, due to poor visual visibility, the system will automatically increase the weight of the current signal; while for the acupoints on the exposed parts such as the earlobe, the system relies more on image features. This adaptive fusion mechanism is particularly suitable for acupoint localization in the complex three-dimensional structure of the ear and is one of the important innovations of the present invention.

[0182] Acupoint recognition and localization:

[0183] To simultaneously achieve acupoint classification and precise localization, the present invention designs a prediction head based on multi-task learning. The specific implementation is as follows:

[0184] 1. Classification branch: Using the global fused feature for acupoint category prediction:

[0185] ;

[0186] ;

[0187] Among them, represents the logits of category prediction, and the logits are the original prediction values output by the classifier that have not been transformed by softmax, is the number of acupoint categories, represents the probability distribution of each category, and are learnable parameters.

[0188] 2. Localization branch: Using the enhanced spatial feature for coordinate regression:

[0189] ;

[0190] ;

[0191] Among them, represents the positioning feature, represents the predicted heat map, where the position with the largest value corresponds to the center of the acupoint.

[0192] 3. Multi-scale heat map prediction:

[0193] ;

[0194] ;

[0195] Among them, represents the enhanced feature of the th scale, represents the heat map prediction of the corresponding scale, represents the final heat map after fusion.

[0196] 4. Coordinate prediction and refinement:

[0197] ;

[0198] To further improve the positioning accuracy, sub-pixel positioning technology is adopted, which is achieved by performing quadratic interpolation on the area near the peak of the heat map:

[0199] ;

[0200] ;

[0201] Among them, represents the refined coordinate prediction, respectively represent the predicted x coordinate and y coordinate.

[0202] 5. Acupoint information retrieval and display:

[0203] Based on the identified acupoint type, it is announced through the buzzer and displayed on the screen.

[0204] The key technical parameters of the acupoint recognition and positioning module are shown in Table 6 below.

[0205] Table 6 Technical parameter table of the acupoint recognition and positioning module:

[0206] ;

[0207] This module announces the acupoint information through the buzzer and displays it on the screen, enabling users to immediately know the name, function, and treatment suggestions of the acupoints they touch without referring to professional acupoint atlases. This function significantly improves the convenience and practicality of clinical applications and is an important advantage of the present invention over the prior art.

[0208] Multi-task Joint Optimization Strategy:

[0209] To achieve the collaborative optimization of classification and localization tasks, the present invention designs a loss function and training strategy with multiple components:

[0210] 1. Classification Loss: Use the standard cross-entropy loss:

[0211] ;

[0212] where, represents the true label (one-hot encoded) of class , and represents the predicted probability.

[0213] 2. Heatmap Loss: Use the improved mean squared error loss:

[0214] ;

[0215] where, represents the value of the predicted heatmap, represents the true heatmap (usually a two-dimensional Gaussian distribution centered on the true acupoint position), represents the number of positive samples, represents the position weight, defined as:

[0216] ;

[0217] where, is the positive sample weight coefficient, is the positive sample threshold.

[0218] 3. Coordinate Loss: Directly optimize the distance between the predicted coordinates and the true coordinates:

[0219] ;

[0220] where, represents the predicted acupoint coordinates, represents the true acupoint coordinates, the subscript 1 represents the first-order distance, the subscript 2 represents the second-order distance, and the following is the weight coefficient of the L1 loss.

[0221] 4. Joint Loss: Combine the above losses into the final multi-task loss:

[0222] ;

[0223] where, , and respectively represent the weight coefficients of the classification, heatmap, and coordinate losses. In this embodiment, the optimal weights are set to , , .

[0224] 5. Training Strategies:

[0225] (1) Two-stage training: First, pre-train the deep convolutional network part only using image data, and then perform end-to-end fine-tuning using multi-modal data;

[0226] (2) Dynamic weight adjustment: Dynamically adjust the loss weights during training, emphasizing feature learning in the initial stage (higher ), and emphasizing accurate prediction in the later stage (higher );

[0227] (3) Gradually increase the difficulty of training samples from easy to difficult, using obvious acupoint samples in the initial stage and introducing more challenging samples in the later stage;

[0228] (4) Regularization techniques: Apply techniques such as weight decay, Dropout, and batch normalization to prevent overfitting.

[0229] The key technical parameters for multi-task joint optimization are shown in Table 7 below.

[0230] Table 7 Technical Parameter Table for Multi-task Joint Optimization:

[0231] ;

[0232] Experimental Results and Performance Evaluation

[0233] The system of the present invention was evaluated on a large-scale multi-modal dataset containing 93 auricular acupoints and 15,360 samples. The main evaluation metrics include:

[0234] 1. Classification accuracy (ACC): The proportion of correctly identifying acupoint types;

[0235] 2. Average Euclidean distance (AED): The average Euclidean distance between the predicted position and the true position;

[0236] 3. Success rate (SR@dth): The proportion of samples where the distance between the predicted position and the true position is within the threshold dth;

[0237] 4. Clinical utility index (CUI): A comprehensive index considering accuracy, localization precision, and response time.

[0238] As shown in Table 8, compared with the prior art, the system of the present invention performs excellently in all metrics.

[0239] Table 8 Performance Comparison Table of Different Methods:

[0240] ;

[0241] As shown in Table 9, the system of the present invention shows different performances at the acupoints in different ear regions, and the dynamic fusion mechanism can adaptively adjust the importance of different modalities according to the characteristics of the acupoints.

[0242] Table 9 Performance of different ear regions:

[0243] ;

[0244] For the acupoints in the inner ear region, due to the poor visual visibility, the system automatically increases the weight of the current signal (0.64 vs 0.36); while for the exposed regions such as the helix, the system relies more on the image features (0.38 vs 0.62). This adaptive fusion ability fully verifies the superiority of integrating the current probe and the micro camera in the same ear acupoint positioning detection pen, which is particularly suitable for acupoint positioning in the complex three-dimensional structure of the ear.

Claims

1. A method for identifying and locating ear acupoints, characterized in that: The following steps are involved: The ear acupuncture point detection pen touches the ear to collect tiny current signals and ear images at the same time; Preprocessing the collected tiny current signals and ear images; The converter encoder is used to extract the current signal characteristics from the pre-processed tiny current signal; Use deep convolutional networks to extract image features from preprocessed ear images; The extracted current signal features and image features are fused using a multimodal dynamic fusion module to obtain fusion features; Simultaneously identify the type and precise location of ear acupuncture points based on fusion features; The recognition result is fed back to the user through buzzer broadcast and screen display.

2. The method according to claim 1, characterized in that The preprocessing of the collected tiny current signal includes: Noise removal: Apply a bandpass filter to remove high-frequency noise and low-frequency drift; Normalization: The Z-score normalization method is used to make the signal mean 0 and the standard deviation 1; Time-frequency transformation: Apply short-time Fourier transform to extract the time-frequency characteristics of the signal; Feature enhancement: Adaptive threshold method is used to enhance features related to acupoints.

3. The method according to claim 1, characterized in that Preprocessing of the acquired ear images includes: Color correction: Apply color constancy algorithm to adapt to different lighting conditions; Contrast enhancement: Use the improved CLAHE algorithm to enhance image details; Normalization: Normalize the image pixel values ​​to the range of [-1, 1]; Spatial transformation: Apply affine transformation for geometric correction to compensate for differences in camera perspectives.

4. The method according to claim 1, characterized in that The current signal features and image features extracted by the multimodal dynamic fusion module are fused, including: current feature reshaping: temporal aggregation and dimensionality adjustment of the current features extracted by the converter; image feature integration: converting the image features extracted by the deep convolutional network into a global representation through global pooling and mapping; modal specificity evaluation: evaluating the reliability and information content of each modal feature; cross-modal attention: calculating the correlation between the two modalities; dynamic fusion weight calculation: calculating the fusion weight based on the modal score and cross-attention; global feature fusion: using dynamic weights to fuse global features; attention-guided spatial feature enhancement: using global features to enhance spatial features.

5. The method according to claim 1, characterized in that The simultaneous identification of the type and precise location of ear acupoints based on fusion features includes: acupoint classification: using global fusion features to predict acupoint categories; heat map generation: using enhanced spatial features to generate a heat map representing the probability of acupoint locations; multi-scale fusion: fusing heat map prediction results of different scales; coordinate prediction: obtaining the initial coordinate prediction by finding the maximum value of the heat map; sub-pixel positioning: achieving more accurate positioning by performing secondary interpolation of the area near the peak of the heat map.

Images