A traditional Chinese medicine constitution identification method and system based on deep learning

Abstract

The application relates to the technical field of constitution identification, and particularly provides a traditional Chinese medicine constitution identification method and system based on deep learning, which comprises the following steps: acquiring tongue image information of a user and diet habit information of the user in a first preset time period, and acquiring tongue normal feature sets of the user; a tongue feature extraction model based on deep learning is used to extract features from the tongue image information, so as to obtain real-time tongue feature sets; for each real-time tongue feature, a target fusion weight corresponding to the real-time tongue feature is determined according to the similarity between the real-time tongue feature and a corresponding tongue normal feature and the diet habit information; all real-time tongue features are subjected to feature fusion according to all target fusion weights, so as to obtain fused tongue features, and then constitution identification is performed according to the fused tongue features, so as to obtain current constitution information; the method can enable a feature fusion strategy to perceive and process the influence of recent diet habits of a user on tongue features.

Description

Technical Field

[0001] This application relates to the field of constitution identification technology, and more specifically, to a method and system for identifying traditional Chinese medicine constitution based on deep learning. Background Technology

[0002] Users conduct daily health monitoring and physical fitness assessment on a remote health management platform. The platform provides a function for physical fitness identification based on tongue images. Specifically, users can obtain physical fitness identification results by collecting tongue images through the platform's mobile application on their personal devices and uploading the tongue images to the platform.

[0003] The specific process for constitution identification by the remote health management platform is as follows: extracting multiple tongue features (such as tongue color, tongue shape, coating color, coating texture, and tongue size) from the tongue image; fusing the extracted multiple tongue features to obtain fused features; and identifying constitution based on the fused features. Existing technologies employ fixed feature fusion strategies to fuse multiple tongue features (e.g., assigning fixed fusion weights to each tongue feature). Since tongue features are influenced by a user's recent dietary habits—for example, frequent consumption of spicy and oily foods may result in a thick, greasy tongue coating or a reddish tongue, while consuming cold foods may lead to a white, slippery tongue coating—existing fixed feature fusion strategies cannot perceive or process the impact of recent dietary habits on tongue features. Therefore, existing technologies suffer from the problem of misjudging a user's constitution when short-term, non-constitutional changes occur in tongue features due to recent dietary habits. For instance, a reddish tongue caused by frequent consumption of spicy foods might be misjudged as a sign of excessive heart fire, resulting in low accuracy and reliability in constitution identification.

[0004] Currently, there is no effective technical solution to the above-mentioned problems. It should be noted that the information disclosed in this section is only for understanding the background of the present invention and therefore may include information that does not constitute prior art. Summary of the Invention

[0005] The purpose of this application is to provide a TCM constitution identification method and system based on deep learning, which can effectively solve the problem that fixed feature fusion strategies cannot perceive and process the influence of the user's recent dietary habits on the tongue appearance features when the tongue appearance features undergo short-term non-constitutional changes due to factors such as recent dietary habits, thus leading to misjudgment of the user's constitution.

[0006] Firstly, this application provides a method for identifying traditional Chinese medicine constitution based on deep learning, which includes the following steps:

[0007] S1. Obtain the user's tongue image information and the user's dietary habits information within a first preset time period, and obtain the user's normal tongue image feature set;

[0008] S2. Use a deep learning-based tongue image feature extraction model to extract features from the tongue image information to obtain a real-time tongue image feature set. Each real-time tongue image feature in the real-time tongue image feature set corresponds to a normal tongue image feature in the normal tongue image feature set.

[0009] S3. For each real-time tongue image feature, determine the target fusion weight corresponding to the real-time tongue image feature based on its similarity to the corresponding normal tongue image feature and dietary habit information.

[0010] S4. Perform feature fusion on all real-time tongue image features according to all target fusion weights to obtain fused tongue image features, and then perform constitution identification based on the fused tongue image features to obtain current constitution information.

[0011] This application provides a deep learning-based method for TCM constitution identification. By determining the target fusion weight corresponding to the real-time tongue image feature based on the similarity between the real-time tongue image feature and the corresponding normal tongue image feature, as well as dietary habit information, the feature fusion strategy can perceive and process the influence of the user's recent dietary habits on the tongue image feature. Therefore, this application can effectively solve the problem that when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, the fixed feature fusion strategy cannot perceive and process the influence of the user's recent dietary habits on the tongue image feature, leading to misjudgment of the user's constitution. This effectively improves the accuracy and reliability of constitution identification.

[0012] Secondly, this application also provides a deep learning-based TCM constitution identification system, which includes:

[0013] The data acquisition module is used to acquire the user's tongue image information and the user's dietary habits information within a first preset time period, and to acquire the user's normal tongue image feature set;

[0014] The feature extraction module is used to extract features from tongue image information using a deep learning-based tongue image feature extraction model to obtain a real-time tongue image feature set. Each real-time tongue image feature in the real-time tongue image feature set corresponds to a normal tongue image feature in the normal tongue image feature set.

[0015] The fusion weight confirmation module is used to determine the target fusion weight corresponding to each real-time tongue image feature based on its similarity to the corresponding normal tongue image features and dietary habit information.

[0016] The constitution identification module is used to perform feature fusion on all real-time tongue image features according to the fusion weights of all targets to obtain fused tongue image features, and then perform constitution identification based on the fused tongue image features to obtain the current constitution information.

[0017] This application provides a deep learning-based TCM constitution identification system. This system determines the target fusion weight corresponding to the real-time tongue image feature based on the similarity between the real-time tongue image feature and the corresponding normal tongue image feature, along with dietary habit information. This allows the feature fusion strategy to perceive and process the influence of the user's recent dietary habits on the tongue image feature. Therefore, this application effectively solves the problem that a fixed feature fusion strategy cannot perceive and process the influence of the user's recent dietary habits on the tongue image feature when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, leading to misjudgments of the user's constitution. This effectively improves the accuracy and reliability of constitution identification.

[0018] As can be seen from the above, the TCM constitution identification method and system based on deep learning provided in this application can determine the target fusion weight corresponding to the real-time tongue image feature by determining the similarity between the real-time tongue image feature and the corresponding normal tongue image feature and dietary habit information. This enables the feature fusion strategy to perceive and process the influence of the user's recent dietary habits on the tongue image feature. Therefore, this application can effectively solve the problem that when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, the fixed feature fusion strategy cannot perceive and process the influence of the user's recent dietary habits on the tongue image feature, thus leading to misjudgment of the user's constitution. This effectively improves the accuracy and reliability of constitution identification. Attached Figure Description

[0019] Figure 1 A flowchart of a deep learning-based method for identifying traditional Chinese medicine constitutions, provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the structure of a TCM constitution identification system based on deep learning, provided in an embodiment of this application.

[0021] Attached labels: 1. Data acquisition module; 2. Feature extraction module; 3. Fusion weight confirmation module; 4. Body constitution identification module. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] Firstly, such as Figure 1 As shown, this application provides a deep learning-based method for identifying TCM constitution, which includes the following steps:

[0025] S1. Obtain the user's tongue image information and the user's dietary habits information within a first preset time period, and obtain the user's normal tongue image feature set;

[0026] S2. Use a deep learning-based tongue image feature extraction model to extract features from the tongue image information to obtain a real-time tongue image feature set. Each real-time tongue image feature in the real-time tongue image feature set corresponds to a normal tongue image feature in the normal tongue image feature set.

[0027] S3. For each real-time tongue image feature, determine the target fusion weight corresponding to the real-time tongue image feature based on its similarity to the corresponding normal tongue image feature and dietary habit information.

[0028] S4. Perform feature fusion on all real-time tongue image features according to all target fusion weights to obtain fused tongue image features, and then perform constitution identification based on the fused tongue image features to obtain current constitution information.

[0029] The image information in step S1 refers to an image including the user's tongue. This tongue image information is preferably an image collected and uploaded by the user to the remote health management platform. The first preset time period can be a time range set by those skilled in the art according to actual needs; for example, the first preset time period can be the most recent day, the most recent three days, or the most recent week. Dietary habit information refers to the user's dietary situation within a specific time period (the first preset time period). This dietary habit information can include the types and frequencies of food recently consumed. This embodiment obtains dietary habit information by using user input, questionnaires, or querying dietary records connected to the health management platform. This embodiment can provide non-constitutional factors affecting short-term changes in tongue image characteristics for constitution identification by obtaining dietary habit information. The normal tongue image feature set refers to the set of tongue image features collected and extracted from the user in a healthy or stable state. This embodiment can obtain the normal tongue image feature set by collecting the user's tongue image feature sets at different time periods while in a healthy state, and then averaging all the tongue image feature sets. That is, the normal tongue image feature set can reflect the user's normal tongue image state. This embodiment is equivalent to using the set of tongue image features collected and extracted from the user in a healthy or stable state as the benchmark for the user's individual tongue image characteristics.

[0030] The real-time tongue image feature set in step S2 refers to the set of tongue image features extracted from the user's current tongue image image using a deep learning-based tongue image feature extraction model. In other words, the real-time tongue image feature set in this embodiment reflects the user's current tongue image state. Each real-time tongue image feature in the real-time tongue image feature set corresponds to a normal tongue image feature in the normal tongue image feature set. For example, the real-time tongue color feature corresponds to the normal tongue color feature, and the real-time tongue coating feature corresponds to the normal tongue coating feature. The deep learning-based tongue image feature extraction model in step S2 can employ a convolutional neural network (CNN) structure (e.g., ResNet, VGG, or Inception network). This tongue image feature extraction model has been trained on a large amount of tongue image image data and can identify and extract various tongue image features such as tongue color, tongue shape, tongue coating color, and tongue coating texture from the tongue image.

[0031] The similarity in step S3 refers to the degree of difference between real-time tongue image features and corresponding normal tongue image features. Step S3 can be achieved by calculating the Euclidean distance between real-time tongue image features and corresponding normal tongue image features, calculating the cosine similarity between real-time tongue image features and corresponding normal tongue image features, or using a similarity measurement method based on statistical distribution to analyze the similarity between real-time tongue image features and corresponding normal tongue image features. Determining the target fusion weight is the core of this scheme. The target fusion weight refers to the weight coefficient used to combine different real-time tongue image features. This target fusion weight is a dynamically determined weight value based on the similarity between real-time tongue image features and corresponding normal tongue image features, as well as dietary habit information. This embodiment can adjust the contribution of each real-time tongue image feature in constitution identification by dynamically confirming the target fusion weight corresponding to each real-time tongue image feature. Step S3 can determine the target fusion weight corresponding to the real-time tongue image features by querying a pre-constructed mapping table of feature similarity, dietary habits, and fusion weights based on the similarity between real-time tongue image features and normal features, and dietary habit information. Alternatively, Step S3 can determine the target fusion weight corresponding to the real-time tongue image features by inputting the similarity between real-time tongue image features and normal features, and dietary habit information, into a pre-trained fusion weight allocation model. This model can evaluate the contribution of the input real-time tongue image features to the real-time tongue image features based on the similarity between the input real-time tongue image features and normal features, and dietary habits to the real-time tongue image features, and output the corresponding fusion weight based on the evaluation results. Step S3 can also... The target fusion weight for the real-time tongue features is determined by first establishing the fusion weight based on the similarity between real-time tongue features and normal tongue features, and then adjusting the established fusion weight based on dietary habit information. For example, if the real-time tongue features differ significantly from the normal tongue features and the user has recently consumed a lot of spicy and oily food, then the real-time tongue features are considered to be greatly affected by diet, and the real-time tongue features do not fully reflect the user's constitution. Therefore, the target fusion weight for the real-time tongue features is relatively small. Conversely, if the real-time tongue features differ slightly from the normal tongue features and the user has recently consumed a light diet, then a larger target fusion weight can be assigned to the real-time tongue features.

[0032] Step S4 can employ weighted averaging, weighted summation, or a more complex fusion network structure to perform feature fusion on all real-time tongue image features based on all target fusion weights. Specifically, the feature fusion process can be as follows: first, multiply each real-time tongue image feature by its corresponding target fusion weight; then, combine all multiplied real-time tongue image features to form a comprehensive fused tongue image feature vector (fused tongue image characteristics). Step S4 can utilize a classification model (e.g., Support Vector Machine (SVM), Decision Tree, Neural Network, or Bayesian Classifier) ​​to identify constitution based on the fused tongue image features. This classification model, trained on a large amount of labeled constitution data, can output the user's current constitution information (e.g., balanced constitution, Yang deficiency constitution, Yin deficiency constitution, etc.) based on the fused tongue image feature vector. Because step S4 uses dynamically adjusted target fusion weights to fuse real-time tongue image features, this embodiment enables the fused tongue image features to more accurately reflect the user's true constitution state, thereby effectively reducing the interference of dietary habits on constitution identification.

[0033] Specifically, this method first acquires the basic data required for constitution identification, including the user's tongue image information, dietary habit information reflecting recent non-constitutional factors, and a set of normal tongue image features serving as an individual benchmark. Then, a deep learning model is used to process the tongue image information to extract a real-time tongue image feature set reflecting the user's current tongue state, and a correspondence is established between real-time features and normal features. Crucially, for each real-time tongue image feature, instead of using a preset fixed weight, a target fusion weight is calculated based on the similarity between the real-time feature and the corresponding normal feature, as well as the user's dietary habit information within a specific time period. This similarity reflects the degree to which the current tongue image feature deviates from the individual's normal state, while dietary habit information provides clues that this deviation may be caused by non-constitutional factors. This embodiment can reduce the weight of tongue image features that are significantly affected by short-term factors such as diet and increase the weight of tongue image features that are less affected by short-term factors such as diet by comprehensively considering both similarity and dietary habit information. Finally, based on these dynamically determined target fusion weights, all real-time tongue image features are weighted and combined to obtain a fused tongue image feature that more accurately reflects the user's physical condition. This fused feature is then used for physical constitution identification to derive the current physical constitution information. The entire process, by dynamically adjusting feature weights, enables the physical constitution identification process to distinguish between constitutional and non-constitutional changes, thus improving the reliability of the identification results.

[0034] As one embodiment, the specific implementation of this application is as follows: The system obtains tongue image information and dietary habit information through tongue image uploaded by the user and dietary questionnaires filled out by the user over the past week, and retrieves a set of normal tongue image features from the user's historical health records or establishes a set of normal tongue image features through the user's initial assessment. A pre-trained convolutional neural network model is used to process the tongue image, extracting vector representations of real-time tongue image features such as tongue color, tongue shape, tongue coating color, and tongue coating texture. For each extracted real-time feature, the Euclidean distance between its feature vector and the corresponding normal feature vector is calculated as a similarity measure. Simultaneously, based on the content of the dietary questionnaire (e.g., whether spicy or oily foods are frequently consumed), a preset rule table is queried to obtain weight adjustment factors for different types of tongue image features (e.g., adjustment factors for tongue color features corresponding to spicy foods). Initial weights can be calculated based on similarity (e.g., the lower the similarity, the higher the initial weight), and then the initial weights are combined with the adjustment factors corresponding to dietary habits to calculate the final target fusion weights. Finally, all real-time tongue image feature vectors are weighted and averaged according to their respective target fusion weights to obtain a fused feature vector, which is then input into a support vector machine classifier to output the user's constitution type.

[0035] The core innovation of this application lies in determining the target fusion weight corresponding to the real-time tongue image feature based on the similarity between the real-time tongue image feature and the corresponding normal tongue image feature, as well as dietary habit information. This enables the feature fusion strategy to perceive and process the impact of the user's recent dietary habits on the tongue image feature. Therefore, this application can effectively solve the problem that when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, the fixed feature fusion strategy cannot perceive and process the impact of the user's recent dietary habits on the tongue image feature, leading to misjudgment of the user's physical condition.

[0036] This application provides a deep learning-based method for TCM constitution identification. By determining the target fusion weight corresponding to the real-time tongue image feature based on the similarity between the real-time tongue image feature and the corresponding normal tongue image feature, as well as dietary habit information, the feature fusion strategy can perceive and process the influence of the user's recent dietary habits on the tongue image feature. Therefore, this application can effectively solve the problem that when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, the fixed feature fusion strategy cannot perceive and process the influence of the user's recent dietary habits on the tongue image feature, leading to misjudgment of the user's constitution. This effectively improves the accuracy and reliability of constitution identification.

[0037] In some preferred embodiments, step S3 includes:

[0038] S31. Obtain the user's sleep quality information within a second preset time period;

[0039] S32. For each real-time tongue image feature, query the pre-built mapping relationship table of tongue image feature similarity and fusion weight according to its similarity with the corresponding normal tongue image feature, so as to obtain the initial fusion weight corresponding to the real-time tongue image feature.

[0040] S33. For each real-time tongue image feature, query the pre-constructed mapping relationship table of feature type, dietary habits and weight adjustment coefficient according to the type of the real-time tongue image feature and dietary habit information to obtain the first fusion weight adjustment coefficient corresponding to the real-time tongue image feature, and query the pre-constructed mapping relationship table of feature type, sleep quality and weight adjustment coefficient according to the type of the real-time tongue image feature and sleep quality information to obtain the second fusion weight adjustment coefficient corresponding to the real-time tongue image feature.

[0041] S34. For each real-time tongue image feature, calculate the target fusion weight corresponding to that real-time tongue image feature based on the corresponding initial fusion weight, the first fusion weight adjustment coefficient, and the second fusion weight adjustment coefficient.

[0042] Sleep quality information refers to data reflecting a user's sleep status within a specific time period (a second preset time period, preferably the same as the first preset time period). This embodiment can obtain sleep quality information by analyzing the results of a user-completed sleep questionnaire. Alternatively, it can obtain sleep quality information by assessing sleep quality based on objective data such as sleep duration, the ratio of deep to light sleep, and the number of nighttime awakenings monitored by wearable devices. Furthermore, it can obtain sleep quality information by comprehensively evaluating a combination of the user-completed sleep questionnaire results and objective data such as sleep duration, the ratio of deep to light sleep, and the number of nighttime awakenings monitored by wearable devices. The second preset time period refers to the time range used to assess sleep quality, which can be set to, for example, the most recent day, the most recent three days, or the most recent week. The mapping table for tongue image feature similarity and fusion weights refers to a pre-established data structure used to store the association between the similarity values ​​of tongue image features and the corresponding initial fusion weights. The initial fusion weights are weight values ​​initially determined based on the similarity between real-time tongue image features and normal features. The initial fusion weights reflect the basic influence of the degree to which the feature deviates from the normal state on the weight. Feature type refers to the specific category of tongue image features, such as tongue color, tongue shape, coating color, coating texture, crack features, or teeth mark features. The mapping table for feature type, dietary habits, and weight adjustment coefficients is a pre-established data structure used to store the association between specific feature types, specific dietary habits, and corresponding weight adjustment coefficients. This mapping table reflects the degree of influence of specific dietary habits on specific types of tongue image features. The first fusion weight adjustment coefficient is an adjustment value obtained from the corresponding mapping table based on feature type and dietary habit information, used to correct the initial fusion weight. The mapping table for feature type, sleep quality, and weight adjustment coefficients is also a pre-established data structure used to store the association between specific feature types, specific sleep quality conditions, and corresponding weight adjustment coefficients. This mapping table reflects the degree of influence of specific sleep quality conditions on specific types of tongue image features. The second fusion weight adjustment coefficient is an adjustment value obtained from the corresponding mapping table based on feature type and sleep quality information, used to further correct the initial fusion weight. This embodiment calculates the target fusion weight corresponding to a real-time tongue image feature by multiplying the initial fusion weight, the first fusion weight adjustment coefficient, and the second fusion weight adjustment coefficient.Since a user's recent sleep quality may affect tongue appearance features—for example, continuous sleep deprivation may cause changes in tongue color and coating—this embodiment can effectively eliminate the influence of the user's recent sleep quality on tongue appearance features by determining a second fusion weight adjustment coefficient based on the type of real-time tongue appearance features and sleep quality information, and then adjusting the preliminary fusion weight of the corresponding real-time tongue appearance features based on the second fusion weight adjustment coefficient. Therefore, this embodiment can enable the determined target fusion weight to more accurately reflect the relative importance of tongue appearance features in constitution identification and reduce the interference of tongue appearance changes caused by short-term factors such as sleep deprivation on constitution identification, thereby effectively improving the accuracy and reliability of feature fusion and further improving the accuracy and reliability of constitution identification.

[0043] In some preferred embodiments, step S34 includes:

[0044] S341. Obtain the current health status information input by the user, including the illness status;

[0045] S342. For each real-time tongue image feature, query the pre-built mapping relationship table of feature type, health status and weight adjustment coefficient according to the type of the real-time tongue image feature and the current health status information to obtain the third fusion weight adjustment coefficient corresponding to the real-time tongue image feature.

[0046] S343. For each real-time tongue image feature, calculate the target fusion weight corresponding to that real-time tongue image feature based on the corresponding initial fusion weight, the first fusion weight adjustment coefficient, the second fusion weight adjustment coefficient, and the third fusion weight adjustment coefficient.

[0047] Current health status information describes a user's current physical health condition, particularly including whether the user has any illnesses or discomforts, such as colds, indigestion, or allergies. Users can input their current health status information actively or by filling out medical consultation information. Since many diseases directly or indirectly affect the appearance of the tongue—for example, some infections may cause a thick, greasy tongue coating, and anemia may cause a pale tongue—illness is an important component of current health status information. The mapping table between feature types, health statuses, and weight adjustment coefficients is a data structure that stores the association between different combinations of feature types and health statuses and their corresponding weight adjustment coefficients. The third fusion weight adjustment coefficient is a value obtained by querying a pre-constructed mapping table based on the type of real-time tongue image features and current health status information. This coefficient reflects the degree of deviation of a certain type of tongue image feature (such as tongue color and tongue coating) from its normal appearance under a specific health state, or its change in reference value in constitution identification. For example, if a user suffers from a disease known to cause yellowing of the tongue coating, the corresponding third fusion weight adjustment coefficient may reduce the weight of the tongue coating feature when calculating its fusion weight, to avoid misjudging yellow tongue coating caused by the disease as a manifestation of damp-heat constitution. This embodiment more accurately assesses the true clinical significance of real-time tongue image features in the current context by considering the user's current health status information when calculating the target fusion weight, distinguishing between tongue image changes caused by constitution factors and those caused by short-term health conditions. Therefore, this embodiment can make the calculated target fusion weight more accurately reflect the relative importance of each tongue image feature in constitution identification, thereby further improving the accuracy and reliability of feature fusion, and consequently, further improving the accuracy and reliability of constitution identification.

[0048] In some preferred embodiments, step S343 includes:

[0049] A1. Obtain recent medication information input by the user;

[0050] A2. For each real-time tongue image feature, query the pre-built mapping relationship table of feature type, medication status and weight adjustment coefficient according to the type of the real-time tongue image feature and recent medication information to obtain the fourth fusion weight adjustment coefficient corresponding to the real-time tongue image feature.

[0051] A3. For each real-time tongue image feature, calculate the target fusion weight corresponding to that real-time tongue image feature based on the corresponding initial fusion weight, first fusion weight adjustment coefficient, second fusion weight adjustment coefficient, third fusion weight adjustment coefficient, and fourth fusion weight adjustment coefficient.

[0052] Recent medication information refers to information such as the type, dosage, and duration of medication taken by the user in a recent period. This embodiment can obtain recent medication information through user input via interface, voice input, or import from other health record systems. The mapping table for feature type, medication status, and weight adjustment coefficients refers to a data structure used to store the weight adjustment coefficients for different tongue feature types under different medication conditions. The fourth fusion weight adjustment coefficient is a value obtained from the mapping table based on the tongue feature type and recent medication information. This fourth fusion weight adjustment coefficient reflects the potential influence of recent medication on the weight of a specific tongue feature in constitution identification. Therefore, this embodiment needs to further adjust the initial fusion weight using the fourth fusion weight adjustment coefficient. For example, taking heat-clearing and detoxifying drugs will make the tongue color darker; that is, when the recent medication information indicates that the user has taken heat-clearing and detoxifying drugs, the fourth fusion weight adjustment coefficient indicates a decrease in the initial fusion weight corresponding to the tongue color feature. Conversely, antibiotics will make the tongue coating less prominent; that is, when the recent medication information indicates that the user has taken antibiotics, the fourth fusion weight adjustment coefficient indicates an increase in the initial fusion weight corresponding to the tongue coating feature. This embodiment can effectively identify and quantify the potential impact of recent medication on real-time tongue features by acquiring recent medication information input by the user, determining the fourth fusion weight adjustment coefficient based on the tongue feature type and recent medication information, and then using the fourth fusion weight adjustment coefficient together with the adjustment coefficients of other factors to calculate the target fusion weight. Therefore, this embodiment can make the calculated target fusion weight more accurately reflect the true state of tongue features after excluding short-term medication interference and avoid misjudging short-term tongue changes caused by drugs as physical problems, thereby further improving the accuracy and reliability of feature fusion, and further improving the accuracy and reliability of physical constitution identification.

[0053] In some preferred embodiments, step A3 includes:

[0054] A31. Obtain the user's location information, current season information, age information, and exercise habit information;

[0055] A32. For each real-time tongue image feature, query the pre-built mapping relationship table about feature type, region, season, age, exercise habits and weight adjustment coefficient according to the type of the real-time tongue image feature, the user's location information, current season information, age information and exercise habit information, so as to obtain the fifth fusion weight adjustment coefficient corresponding to the real-time tongue image feature.

[0056] A33. For each real-time tongue image feature, calculate the target fusion weight corresponding to that real-time tongue image feature based on the corresponding initial fusion weight, first fusion weight adjustment coefficient, second fusion weight adjustment coefficient, third fusion weight adjustment coefficient, fourth fusion weight adjustment coefficient, and fifth fusion weight adjustment coefficient.

[0057] The user's geographical location refers to the user's primary geographical area of ​​residence or life, such as the south, north, or plateau regions. Differences in climate, soil, and dietary habits across different regions can have a long-term impact on a person's constitution, which is reflected in the tongue's appearance. Current season information refers to the season in which the user is currently undergoing constitution identification, such as spring, summer, autumn, or winter. Seasonal changes significantly affect physiological functions and pathological states, potentially altering the tongue's appearance. The user's age information refers to their actual age. Different age groups have different physiological functions, metabolic levels, and organ functions, leading to variations in constitution. Exercise habits information refers to the frequency, intensity, and type of daily exercise. Good exercise habits promote blood circulation and organ function coordination, while poor exercise habits can lead to qi stagnation, blood stasis, and internal dampness and phlegm, all of which affect constitution formation and tongue appearance. The fifth fusion weight adjustment coefficient is obtained by querying a pre-constructed mapping table based on the comprehensive factors mentioned above, such as region, season, age, and exercise habits. This mapping table can be trained and constructed based on a large amount of clinical data, traditional Chinese medicine theory, and expert experience to reflect the degree of influence of these external environmental and individual physiological factors on the weights of specific tongue features. For example, in hot and humid regions during summer, a yellow and greasy tongue coating may be more common, and its weight may need to be adjusted to avoid overdiagnosing a damp-heat constitution; while in cold regions during winter, a pale tongue may be more common, and its weight may also need to be adjusted accordingly. Finally, the target fusion weight is based on the previous consideration of dietary habits, sleep quality, current health status, and recent medication information, and further incorporates the fifth fusion weight adjustment coefficient, making the weight calculation more comprehensive, precise, and personalized. This embodiment effectively overcomes the limitations of considering only personal lifestyle and health conditions by incorporating regional, seasonal, age, and exercise habit information into the calculation process of the target fusion weight. This additional information provides a more comprehensive user background, enabling the system to understand more deeply the formation of tongue features and their significance in specific contexts. The solution dynamically confirms the fifth fusion weight adjustment coefficient by querying a pre-built mapping table. This coefficient can finely adjust the weight of real-time tongue features based on the user's specific environment and physiological characteristics. Therefore, this solution ensures that the target fusion weight not only reflects the characteristics of the tongue itself but also fully considers the external environment and internal individual differences affecting constitution, making the constitution identification results more context-adaptable and accurate. Because this embodiment fully considers multi-dimensional factors such as region, season, age, and exercise habits, it enables constitution identification to move beyond single-dimensional information and conduct multi-dimensional, personalized assessments. This allows the constitution identification results to better adapt to complex situations in different individuals and environments, reducing misjudgments caused by external environments or individual differences, and thus providing more instructive health management recommendations.As a specific implementation, suppose a user's tongue image, after processing, shows a reddish tongue with a thin yellow coating, initially suggesting a possible damp-heat constitution. Without considering factors such as region, season, age, and exercise habits, the system might calculate a target fusion weight based on dietary habits (e.g., preference for spicy and oily foods), sleep quality (e.g., difficulty falling asleep), current health status (e.g., occasional bitter taste in the mouth), and recent medication information (e.g., no medication used). However, if further information is obtained that the user's location is in a humid and hot southern region, the current season is summer, their age is 25 (young adult), and their exercise habits include regular weekly exercise, the system will query a pre-built mapping table based on this information. For example, in southern summers, a reddish tongue with a thin yellow coating might be more common, but considering their youth and regular exercise habits, the system might assign a relatively low fifth fusion weight adjustment coefficient to this tongue feature to avoid overdiagnosing a damp-heat constitution or to suggest that the degree of damp-heat constitution might be mild. Conversely, if the user's age is 60 (elderly) and their exercise habits indicate a long-term lack of exercise, the same tongue appearance characteristics might be assigned a relatively high fifth fusion weight adjustment coefficient in the southern summer to more accurately reflect their tendency towards a damp-heat constitution. In this way, the fifth fusion weight adjustment coefficient is used to refine the initial fusion weight, so that the final calculated target fusion weight can more accurately reflect the user's true physical condition under specific environmental and individual conditions.

[0058] In some preferred embodiments, step S2 includes:

[0059] S21. Obtain the brightness distribution information of the tongue image;

[0060] S22. Divide the tongue image information into multiple feature extraction regions based on the brightness distribution information;

[0061] S23. For each feature extraction region, determine the model parameters of the tongue image feature extraction model based on deep learning according to the brightness of the feature extraction region;

[0062] S24. For each feature extraction region, use a deep learning-based tongue image feature extraction model to extract features from the region to obtain a regional tongue image feature set, which includes different types of regional tongue image features.

[0063] S25. Integrate the tongue image features of all regions corresponding to the same type into real-time tongue image features to obtain a real-time tongue image feature set.

[0064] The brightness distribution information of a tongue image refers to the distribution of pixel brightness values ​​in the tongue image. This embodiment can use brightness analysis algorithms in image processing technology to obtain the brightness distribution information of the tongue image, such as calculating the brightness histogram, average brightness, and local brightness differences of the tongue image. The feature extraction region refers to several sub-regions obtained by segmenting or dividing the complete tongue image according to the brightness distribution information. Each sub-region has relatively consistent brightness characteristics. This embodiment can use image segmentation or region partitioning techniques such as brightness threshold-based segmentation, region growing, cluster analysis, or mesh generation to divide the tongue image information into multiple feature extraction regions based on the brightness distribution information. The model parameters of a deep learning-based tongue image feature extraction model refer to the set of numerical values ​​used to configure or adjust the internal computation process of the deep learning model. These parameters determine how the model processes input data and extracts features, such as the weights of the convolutional kernel, bias terms, and activation function parameters. This embodiment can determine the model parameters based on the brightness of the feature extraction region by querying a pre-built mapping table of region brightness and feature extraction model parameters. Alternatively, it can determine the model parameters based on the brightness of the feature extraction region by substituting it into a pre-built function calculation formula. A regional tongue image feature set refers to the set of various types of tongue image features (such as tongue color features and tongue coating features of a specific feature extraction region) obtained after feature extraction from a specific feature extraction region using a deep learning-based tongue image feature extraction model. Real-time tongue image features refer to the final representation of a type of feature obtained by integrating tongue image features of the same type extracted from various feature extraction regions of a tongue image. This embodiment can integrate all tongue image features of the same type into real-time tongue image features by feature splicing. Since this embodiment first divides the tongue image information into multiple feature extraction regions based on brightness distribution information, and then dynamically determines the model parameters of the deep learning-based tongue image feature extraction model suitable for each region based on the brightness of each feature extraction region—for example, for regions with low brightness, the model parameters can be adjusted to focus more on the extraction of details in dark areas; for regions with high brightness, the parameters can be adjusted to avoid overexposure or distortion of features in highlight areas—this embodiment can effectively solve the problem of decreased feature extraction accuracy due to uneven illumination of tongue images and enable the tongue image feature extraction model to better adapt to image regions under different lighting conditions. This effectively improves the robustness and accuracy of tongue image feature extraction, making the obtained real-time tongue image feature set more accurate and reliable, and providing high-quality input data for subsequent constitution identification, thereby further improving the accuracy and reliability of constitution identification.

[0065] In some preferred embodiments, step S23 includes:

[0066] S231. For each feature extraction region, determine the preliminary model parameters of the tongue image feature extraction model based on deep learning according to the brightness of the feature extraction region.

[0067] S232. For each feature extraction region, determine the model parameter adjustment coefficients corresponding to different feature extraction types based on the brightness of the feature extraction region and the pre-set combination of feature extraction types.

[0068] S233. For each feature extraction region, calculate the model parameters of the deep learning-based tongue image feature extraction model corresponding to each extracted feature type based on the model parameter adjustment coefficient and the preliminary model parameters.

[0069] Step S24 includes:

[0070] S241. For each feature extraction region, use the model parameters of the deep learning-based tongue image feature extraction model corresponding to different feature extraction types to extract features from the feature extraction region, so as to obtain different types of regional tongue image features, and then integrate all regional tongue image features into a regional tongue image feature set.

[0071] Preliminary model parameters refer to the set of model parameters initially determined based on the brightness of the feature extraction region, serving as the basis for subsequent adjustments for different feature types. The combination of extracted feature types refers to the set of tongue image features to be extracted from the tongue image, such as extracting five types of tongue image features: tongue color, tongue shape, tongue coating color, tongue coating texture, and cracks. Model parameter adjustment coefficients are coefficients determined based on the brightness of the feature extraction region and specific combinations of extracted feature types, used to adjust the preliminary model parameters. These coefficients reflect the amount or proportion of adjustment that should be made to the model parameters when extracting a specific feature type at a specific brightness level. In this embodiment, the model parameter adjustment coefficients corresponding to each extracted feature type can be obtained by querying a pre-built mapping table of region brightness, feature type, and parameter adjustment coefficients based on the brightness of the feature extraction region and the extracted feature type. This embodiment is equivalent to determining appropriate model parameters for each feature type based on the brightness differences in different regions of the tongue image and the different sensitivities of different tongue feature types to brightness. It then uses tongue feature extraction models with different model parameters to extract features from different types of tongue images. Therefore, this embodiment can more accurately capture different types of tongue features in different brightness regions, thereby further improving the robustness and accuracy of tongue feature extraction, and further improving the accuracy and reliability of constitution identification.

[0072] In some preferred embodiments, step S1 includes:

[0073] S11. Obtain the user's tongue image information and the user's dietary habits information within a first preset time period, and obtain the user's normal tongue image feature set;

[0074] S12. Preprocess the tongue image information.

[0075] Preprocessing refers to a series of image processing operations performed on the original tongue image information. This embodiment can effectively improve the image quality of the tongue image information and remove interfering factors by preprocessing the tongue image information, thus providing better data input for subsequent feature extraction. This embodiment can effectively solve the problems of uneven lighting, color deviation, and noise interference caused by differences in the acquisition environment and equipment in the original tongue image by adding a preprocessing step. That is, the preprocessed image has higher quality and is more standardized. Therefore, this embodiment can significantly improve the accuracy of subsequent tongue image feature extraction and avoid misjudgments caused by image quality problems.

[0076] In some preferred embodiments, preprocessing includes illumination correction, color balancing, and noise filtering. Illumination correction adjusts the brightness distribution of the image to compensate for image quality problems caused by uneven lighting or excessive / inadequate brightness during shooting; this can be achieved using histogram equalization, gamma correction, or the Retinex algorithm. Color balancing adjusts the color components of the image to correct color casts caused by ambient light or device settings, ensuring that the image colors are close to reality; this can be achieved using white balance algorithms, gray-world algorithms, or color constancy algorithms. Noise filtering removes randomly distributed outlier pixels from the image to reduce the interference of image noise on subsequent feature extraction; this can be achieved using median filtering, Gaussian filtering, or bilateral filtering. This embodiment can use illumination correction to unify the brightness levels of different images, thereby reducing the impact of illumination differences on the representation of tongue features. This embodiment can use color balance to eliminate color casts in the image, so that color features such as tongue color and coating color can accurately reflect the real situation. This embodiment can use noise filtering to smooth the image and remove interfering noise points, so that details such as tongue texture, cracks, and teeth marks are clearer. Therefore, this embodiment can improve the quality and consistency of the original tongue image, providing a high-quality input foundation for subsequent accurate extraction of real-time tongue features using a deep learning-based tongue feature extraction model.

[0077] In some preferred embodiments, real-time tongue image features include tongue color features, tongue shape features, tongue coating color features, tongue coating texture features, crack features, or teeth mark features. Tongue color features refer to the color information presented by the tongue body, which can be achieved by analyzing the pixel value distribution, hue, and saturation of the tongue body area using image processing techniques. Tongue shape features refer to the morphological information of the tongue body, which can be achieved by analyzing the outline, size, and edge shape of the tongue body using image processing techniques. Tongue coating color features refer to the color information presented by the tongue coating, which can be achieved by analyzing the pixel value distribution of the tongue coating area using image processing techniques. Tongue coating texture features refer to the texture information of the tongue coating, which can be achieved by analyzing the texture, gloss, and thickness of the tongue coating area using image processing techniques. Crack features refer to the crack information appearing on the tongue surface, which can be achieved by detecting and analyzing the number, depth, and direction of cracks on the tongue surface using image processing techniques. Teeth mark features refer to the indentation information appearing on the edges of the tongue body, which can be achieved by detecting and analyzing the shape and degree of the indentation on the edges of the tongue body using image processing techniques.

[0078] As can be seen from the above, the TCM constitution identification method based on deep learning provided in this application can determine the target fusion weight corresponding to the real-time tongue image feature by determining the similarity between the real-time tongue image feature and the corresponding normal tongue image feature and dietary habit information. This enables the feature fusion strategy to perceive and process the influence of the user's recent dietary habits on the tongue image feature. Therefore, this application can effectively solve the problem that when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, the fixed feature fusion strategy cannot perceive and process the influence of the user's recent dietary habits on the tongue image feature, thus leading to misjudgment of the user's constitution. This effectively improves the accuracy and reliability of constitution identification.

[0079] Secondly, such as Figure 2 As shown, this application also provides a deep learning-based TCM constitution identification system, which includes:

[0080] Data acquisition module 1 is used to acquire the user's tongue image information and the user's dietary habits information within a first preset time period, and to acquire the user's normal tongue image feature set;

[0081] Feature extraction module 2 is used to extract features from tongue image information using a deep learning-based tongue image feature extraction model to obtain a real-time tongue image feature set. Each real-time tongue image feature in the real-time tongue image feature set corresponds to a normal tongue image feature in the normal tongue image feature set.

[0082] The fusion weight confirmation module 3 is used to determine the target fusion weight corresponding to each real-time tongue image feature based on its similarity to the corresponding normal tongue image features and dietary habit information.

[0083] The constitution identification module 4 is used to perform feature fusion on all real-time tongue image features according to the fusion weight of all targets to obtain fused tongue image features, and then perform constitution identification based on the fused tongue image features to obtain the current constitution information.

[0084] This application provides a deep learning-based TCM constitution identification system, comprising a data acquisition module 1, a feature extraction module 2, a fusion weight confirmation module 3, and a constitution identification module 4. This embodiment of the deep learning-based TCM constitution identification system is used to perform the steps in the deep learning-based TCM constitution identification method provided in the first aspect above. The principle of the deep learning-based TCM constitution identification system provided in this embodiment is the same as that of the deep learning-based TCM constitution identification method provided in the first aspect above, and will not be discussed in detail here.

[0085] As can be seen from the above, the TCM constitution identification method and system based on deep learning provided in this application can determine the target fusion weight corresponding to the real-time tongue image feature by determining the similarity between the real-time tongue image feature and the corresponding normal tongue image feature and dietary habit information. This enables the feature fusion strategy to perceive and process the influence of the user's recent dietary habits on the tongue image feature. Therefore, this application can effectively solve the problem that when the tongue image feature undergoes short-term non-constitutional changes due to factors such as recent dietary habits, the fixed feature fusion strategy cannot perceive and process the influence of the user's recent dietary habits on the tongue image feature, thus leading to misjudgment of the user's constitution. This effectively improves the accuracy and reliability of constitution identification.

[0086] In the embodiments provided in this application, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of the above units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another robot, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0087] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0088] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0089] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1.A deep learning-based traditional Chinese medicine (TCM) constitution identification method, characterized in that, The deep learning-based traditional Chinese medicine constitution identification method comprises the following steps: S1, obtaining tongue image information of a user and diet habit information of the user in a first preset time period, and obtaining a tongue normal feature set of the user; S2, using a deep learning-based tongue feature extraction model to extract features from the tongue image information to obtain a real-time tongue feature set, each real-time tongue feature in the real-time tongue feature set corresponding to a tongue normal feature in the tongue normal feature set; S3, for each real-time tongue feature, determining a target fusion weight corresponding to the real-time tongue feature according to a similarity between the real-time tongue feature and the corresponding tongue normal feature and the diet habit information; S4, fusing all the real-time tongue features according to all the target fusion weights to obtain a fused tongue feature, and then identifying a constitution according to the fused tongue feature to obtain current constitution information; Step S3 comprises: S31, obtaining sleep quality information of the user in a second preset time period; S32, for each real-time tongue feature, querying a pre-constructed mapping relationship table about tongue feature similarity and fusion weight according to a similarity between the real-time tongue feature and the corresponding tongue normal feature to obtain an initial fusion weight corresponding to the real-time tongue feature; S33, for each real-time tongue feature, querying a pre-constructed mapping relationship table about feature type, diet habit and weight adjustment coefficient according to a type of the real-time tongue feature and the diet habit information to obtain a first fusion weight adjustment coefficient corresponding to the real-time tongue feature, and querying a pre-constructed mapping relationship table about feature type, sleep quality and weight adjustment coefficient according to the type of the real-time tongue feature and the sleep quality information to obtain a second fusion weight adjustment coefficient corresponding to the real-time tongue feature; S34, for each real-time tongue feature, calculating a target fusion weight corresponding to the real-time tongue feature according to the corresponding initial fusion weight, the first fusion weight adjustment coefficient and the second fusion weight adjustment coefficient. 2.The deep learning-based traditional Chinese medical constitution identification method according to claim 1, characterized in that, Step S34 comprises: S341, obtaining current health status information input by the user, the current health status information including a disease condition; S342, for each real-time tongue feature, querying a pre-constructed mapping relationship table about feature type, health status and weight adjustment coefficient according to a type of the real-time tongue feature and the current health status information to obtain a third fusion weight adjustment coefficient corresponding to the real-time tongue feature; S343, for each real-time tongue feature, calculating a target fusion weight corresponding to the real-time tongue feature according to the corresponding initial fusion weight, the first fusion weight adjustment coefficient, the second fusion weight adjustment coefficient and the third fusion weight adjustment coefficient. 3.The deep learning-based traditional Chinese medical constitution identification method according to claim 2, characterized in that, Step S343 comprises: A1, obtaining recent medication information input by the user; A2, for each real-time tongue feature, querying a pre-constructed mapping relationship table about feature type, medication condition and weight adjustment coefficient according to a type of the real-time tongue feature and the recent medication information to obtain a fourth fusion weight adjustment coefficient corresponding to the real-time tongue feature; A3. For each real-time tongue appearance feature, a target fusion weight corresponding to the real-time tongue appearance feature is calculated according to the corresponding initial fusion weight, the first fusion weight adjustment coefficient, the second fusion weight adjustment coefficient, the third fusion weight adjustment coefficient, and the fourth fusion weight adjustment coefficient. 4.The deep learning-based traditional Chinese medical constitution identification method according to claim 1, characterized in that, Step S2 comprises: S21. Obtain brightness distribution information of the tongue appearance image information; S22. Divide the tongue appearance image information into a plurality of feature extraction regions according to the brightness distribution information; S23. For each feature extraction region, determine model parameters of a tongue appearance feature extraction model based on deep learning according to the brightness of the feature extraction region; S24. For each feature extraction region, perform feature extraction on the feature extraction region by using the tongue appearance feature extraction model based on deep learning to obtain a regional tongue appearance feature set, the regional tongue appearance feature set comprising regional tongue appearance features of different types; S25. Integrate all regional tongue appearance features of the same type into a real-time tongue appearance feature to obtain a real-time tongue appearance feature set. 5.The deep learning-based traditional Chinese medical constitution identification method according to claim 4, characterized in that, Step S23 comprises: S231. For each feature extraction region, determine preliminary model parameters of a tongue appearance feature extraction model based on deep learning according to the brightness of the feature extraction region; S232. For each feature extraction region, determine model parameter adjustment coefficients corresponding to different extraction feature types according to the brightness of the feature extraction region and a pre-set extraction feature type combination; S233. For each feature extraction region, calculate model parameters of a tongue appearance feature extraction model based on deep learning corresponding to each extraction feature type according to the model parameter adjustment coefficients and the preliminary model parameters; Step S24 comprises: S241. For each feature extraction region, perform feature extraction on the feature extraction region by using model parameters of a tongue appearance feature extraction model based on deep learning corresponding to different extraction feature types to obtain regional tongue appearance features of different types, and then integrate all the regional tongue appearance features into a regional tongue appearance feature set. 6.The deep learning-based traditional Chinese medical constitution identification method according to claim 1, characterized in that, Step S1 comprises: S11. Obtain tongue appearance image information of a user and dietary habit information of the user within a first preset time period, and obtain a tongue appearance normal feature set of the user; S12. Preprocess the tongue appearance image information. 7.The deep learning-based traditional Chinese medical constitution identification method according to claim 6, characterized in that, The preprocessing comprises illumination correction, color balance, and noise filtering. 8.The deep learning-based traditional Chinese medical constitution identification method according to claim 1, characterized in that, The real-time tongue appearance feature is a tongue color feature, a tongue shape feature, a fur color feature, a fur quality feature, a crack feature, or a tooth mark feature. 9.A Chinese medical constitution identification system based on deep learning, characterized in that, The deep learning-based traditional Chinese constitution identification system comprises: a data acquisition module configured to obtain tongue appearance image information of a user and dietary habit information of the user within a first preset time period, and obtain a tongue appearance normal feature set of the user; a feature extraction module configured to perform feature extraction on the tongue appearance image information by using a tongue appearance feature extraction model based on deep learning to obtain a real-time tongue appearance feature set, each real-time tongue appearance feature in the real-time tongue appearance feature set corresponding to one tongue appearance normal feature in the tongue appearance normal feature set; and a real-time tongue appearance feature set determination module configured to determine a target fusion weight corresponding to each real-time tongue appearance feature in the real-time tongue appearance feature set according to an initial fusion weight corresponding to the real-time tongue appearance feature, a first fusion weight adjustment coefficient, a second fusion weight adjustment coefficient, a third fusion weight adjustment coefficient, and a fourth fusion weight adjustment coefficient corresponding to the real-time tongue appearance feature, and determine a final real-time tongue appearance feature set by integrating all real-time tongue appearance features of the same type into a real-time tongue appearance feature. The fusion weight confirmation module is configured to determine, for each real-time tongue appearance feature, a target fusion weight corresponding to the real-time tongue appearance feature according to similarity of the real-time tongue appearance feature to a corresponding tongue appearance normal feature and the eating habit information; The constitution recognition module is configured to perform feature fusion on all the real-time tongue appearance features according to all the target fusion weights to obtain a fused tongue appearance feature, and then perform constitution recognition according to the fused tongue appearance feature to obtain current constitution information; The step of determining, for each real-time tongue appearance feature, a target fusion weight corresponding to the real-time tongue appearance feature according to similarity of the real-time tongue appearance feature to a corresponding tongue appearance normal feature and the eating habit information comprises: S31, obtaining sleep quality information of a user in a second preset time period; S32, for each real-time tongue appearance feature, querying a pre-constructed mapping relationship table about tongue appearance feature similarity and fusion weight according to similarity of the real-time tongue appearance feature to a corresponding tongue appearance normal feature to obtain an initial fusion weight corresponding to the real-time tongue appearance feature; S33, for each real-time tongue appearance feature, querying a pre-constructed mapping relationship table about feature type, eating habit and weight adjustment coefficient according to the type of the real-time tongue appearance feature and the eating habit information to obtain a first fusion weight adjustment coefficient corresponding to the real-time tongue appearance feature, and querying a pre-constructed mapping relationship table about feature type, sleep quality and weight adjustment coefficient according to the type of the real-time tongue appearance feature and the sleep quality information to obtain a second fusion weight adjustment coefficient corresponding to the real-time tongue appearance feature; S34, for each real-time tongue appearance feature, calculating a target fusion weight corresponding to the real-time tongue appearance feature according to the corresponding initial fusion weight, the first fusion weight adjustment coefficient and the second fusion weight adjustment coefficient.

Images