A facial information analysis method, system and storage medium

Through the combination of deep learning models and facial information database, the misdiagnosis problem of traditional facial diagnosis relying on experience is solved, providing more accurate and reliable facial information analysis, and improving the objectivity and comprehensiveness of the diagnosis.

CN118379780BActive Publication Date: 2024-11-26NANJING UNIV OF TRADITIONAL CHINESE MEDICINE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410832436.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-26
Publication Date
2024-11-26
Estimated Expiration
2044-06-26

AI Technical Summary

Technical Problem

Traditional in-person diagnosis relies on the subjective experience of doctors, making it difficult to accurately judge rare symptoms or lack of diagnostic references, which can easily lead to misdiagnosis and lack of objective basis.

Method used

The facial information analysis method based on the deep learning model is adopted, and the facial information vector is obtained through the feature extraction module, the SE module and the semantic analysis module, and the vector matches with the constructed facial information database to generate a list of analysis results.

Benefits of technology

It provides more objective and comprehensive general reference opinions, helps doctors better understand the condition, overcomes the limitations of traditional in-person consultations that rely on personal experience and lack of objective basis, and improves the accuracy and reliability of the diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118379780B_ABST
    Figure CN118379780B_ABST
Patent Text Reader

Abstract

The present application provides a facial information parsing method, system and storage medium, which relate to the field of image processing. The facial information parsing method includes: obtaining a facial image; obtaining a facial information vector based on the facial image and a trained deep learning model; performing vector matching based on the facial information vector and a constructed facial information database to obtain a parsing result list, which can be used as a reference for doctors, so that doctors can make a more comprehensive judgment on the patient's condition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to a facial information parsing method, system and storage medium. Background Art

[0002] Facial diagnosis in traditional Chinese medicine is an important part of the four diagnostic methods of traditional Chinese medicine. By observing the color, luster, shape and other information of the patient's face, it is possible to judge the patient's qi and blood, the functional state of the internal organs, etc., and then guide diagnosis, syndrome differentiation and treatment. However, traditional facial diagnosis mainly relies on the subjective experience of doctors. For doctors with less experience, or conditions with rare symptoms, or doctors cannot remember the corresponding symptoms for a while, if there is a lack of diagnostic reference during diagnosis, misdiagnosis is likely to occur. Summary of the invention

[0003] The present application provides a facial information parsing method, including: obtaining a facial image; obtaining a facial information vector based on the facial image and a trained deep learning model; performing vector matching based on the facial information vector and a constructed facial information database to obtain a parsing result list.

[0004] Furthermore, the deep learning model includes a feature extraction module, an SE module and a semantic parsing module. The facial information vector is obtained based on the facial image and the trained deep learning model, including: performing feature extraction based on the facial image and the feature extraction module to obtain an initial feature map; performing weighted processing based on the SE module and the initial feature map to obtain a weighted feature map; performing semantic parsing based on the semantic parsing module and the weighted feature map to obtain a facial information vector.

[0005] Furthermore, based on the initial feature map and based on the SE module, weighted processing is performed to obtain a weighted feature map, including: performing global average pooling processing on the initial feature map to obtain channel descriptors of each channel; generating corresponding channel weights of each channel according to the channel descriptors; and performing weighted processing on each channel according to the channel weights to obtain a weighted feature map.

[0006] Furthermore, the feature extraction module is established based on the residual network, including an initial convolutional layer, a normalization layer, a ReLU function, a first pooling layer, a residual unit and a second pooling layer. The feature extraction is performed based on the facial image and based on the feature extraction module to obtain an initial feature map, including: based on the facial image, feature extraction is performed based on the initial convolutional layer to obtain a primary visual feature map of the facial image; based on the primary visual feature map, standardization is performed based on the normalization layer to obtain a normalized feature map; based on the normalized feature map, activation is performed based on the ReLU function to obtain an activated feature map; based on the activated feature map, dimensionality reduction is performed based on the first pooling layer to obtain a pooling feature map; based on the pooling feature map, deep-level feature extraction is performed based on the convolutional layer in the residual unit to obtain residual features; based on the pooling feature map and the residual features, feature compression is performed based on the second pooling layer to obtain an initial feature map.

[0007] Furthermore, the method of performing semantic analysis based on the weighted feature map and the semantic analysis module to obtain a facial information vector includes: based on the semantic analysis module, mapping the weighted feature map to a semantic space by nonlinear transformation to obtain a facial information vector.

[0008] Furthermore, the semantic space includes a correspondence between facial features in the weighted feature map and facial information vectors, and the facial information vectors include facial information semantics.

[0009] Furthermore, the facial information database is constructed based on knowledge graph and natural language processing, and the facial information database includes multiple reference vectors and a parsing result corresponding to each reference vector.

[0010] Furthermore, the method performs vector matching based on the facial information vector and a constructed facial information database to obtain a list of parsing results, including: performing vector matching in a constructed facial information database according to the facial information vector to determine a plurality of similar reference vectors and a plurality of similarities corresponding to the plurality of similar reference vectors; based on preset conditions, selecting at least two similar reference vectors from the plurality of similar reference vectors to determine at least two parsing results; and sorting the at least two parsing results according to the similarities corresponding to the parsing results to obtain a list of parsing results.

[0011] The present application also provides a facial information parsing system, including: an acquisition module, which is used to acquire facial images; an image processing module, which is used to acquire facial information vectors based on the facial images and a trained deep learning model; an image parsing module, which is used to perform vector matching based on the facial information vectors and a constructed facial information database to obtain a parsing result list.

[0012] The present application also provides a computer-readable storage medium, wherein the storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes the aforementioned method.

[0013] Compared with the prior art, this application has at least the following beneficial effects:

[0014] 1. By obtaining a list of analysis results based on facial images, deep learning models and facial information databases, and providing the list of analysis results to doctors for reference, doctors can make a more comprehensive judgment on the patient's condition.

[0015] 2. Through the synergy of structured components, the feature extraction module can efficiently extract multi-scale and multi-level visual features from facial images, providing accurate and rich feature input for the subsequent SE module weighted processing and semantic parsing, forming a comprehensive and three-dimensional visual understanding of facial images.

[0016] 3. By introducing the SE attention mechanism, the weights of different facial regions can be adaptively adjusted through back-propagation learning, so that key regions with greater impact on the diagnosis results receive higher weights. By embedding the SE module in the convolutional network, the expressive power of the convolutional features has been significantly improved.

[0017] 4. By mapping the weighted feature map to the semantic space, semantic analysis is performed. The original high-dimensional convolution facial features are mapped to the low-dimensional, structured, and semantic facial diagnosis information vector space, which is very consistent with the thinking mode of TCM facial diagnosis, so that it can be easily matched with the TCM knowledge base in the future. At the same time, the semantic representation discretizes the original continuous value features and can also compress and reduce the noise of facial features.

[0018] 5. By building a facial information database based on knowledge graph and natural language processing, the facial information database is rich in data and of high quality, which can lay an effective foundation for subsequent reference.

[0019] 6. Through vector matching, a mapping relationship between facial information and traditional Chinese medicine knowledge is established, so that a list of analysis results that conform to traditional Chinese medicine theory can be given based on facial appearance information for doctors' reference. This overcomes the limitations of traditional facial diagnosis that relies on doctors' personal experience, rare symptoms cannot be accurately judged, and lacks objective basis. It can better provide doctors with reasonable and well-considered reference opinion information to help doctors better understand the condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings.

[0021] Figure 1 is a flowchart of a facial information parsing method according to some embodiments of this specification;

[0022] Figure 2 is a schematic diagram of a process of obtaining a facial information vector according to some embodiments of this specification;

[0023] Figure 3 is a flowchart of a feature extraction step according to some embodiments of this specification;

[0024] Figure 4 is a flowchart of weighted processing steps according to some embodiments of this specification;

[0025] Figure 5 It is a flowchart of the vector matching steps shown in some embodiments of this specification. DETAILED DESCRIPTION

[0026] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0027] Figure 1 is a flowchart of a facial information parsing method according to some embodiments of this specification, such as Figure 1 As shown, the facial information parsing method may include the following steps.

[0028] Step 110, obtaining a facial image.

[0029] The facial image refers to an image of the face of a patient. For example, the patient can upload the captured facial image to the analysis system, or the analysis system can obtain the device permission of the patient to take the photo.

[0030] Step 120, obtaining a facial information vector according to the facial image based on the trained deep learning model.

[0031] The deep learning model is a deep convolutional neural network. Facial images can be input into the deep learning model, and the deep learning model outputs a facial information vector.

[0032] The facial information vector is a vector describing facial information. In some embodiments, the facial information vector may include facial information semantics, which include at least three dimensions, such as complexion, gloss, and morphology. Specifically, complexion includes white, yellow, green, red, and black, facial gloss includes luster, dull, and dark, and facial morphology includes thin, normal, and round. For example, the facial information vector may be a vector composed of three facial information semantics, such as (yellow, dull, and thin).

[0033] Step 130, according to the facial information vector, based on the constructed facial information database, vector matching is performed to obtain a parsing result list.

[0034] The facial information database is a database that stores facial information and corresponding parsing results. In some embodiments, the facial information database is constructed based on knowledge graphs and natural language processing, wherein natural language processing may include standardization of Chinese medicine terms, named entity recognition, and relationship extraction. The facial information database includes multiple reference vectors and parsing results corresponding to each reference vector. In some embodiments, facial diagnosis knowledge can be collected, represented, and organized based on unstructured facial diagnosis knowledge such as ancient Chinese medical literature, clinical diagnosis and treatment guidelines, electronic medical records and other historical data, using knowledge graphs and natural language processing technology to construct reference vectors and their corresponding parsing results. Specifically, the following steps may be included:

[0035] Define the facial diagnosis knowledge ontology and standardize the basic elements such as concepts, relationships, attributes, etc. of the knowledge base of the facial diagnosis knowledge ontology;

[0036] Using natural language processing technology, the original unstructured facial diagnosis knowledge is structured and the key elements of facial diagnosis knowledge are extracted;

[0037] Instantiate key elements into knowledge ontology and construct facial information database.

[0038] Among them, each facial diagnosis knowledge is represented as a knowledge vector, namely the reference vector. Each dimension of the reference vector corresponds to a concept or attribute defined by the ontology (for example, complexion, luster, morphology), and each reference vector corresponds to an analysis result, namely the main symptom (for example, cold syndrome, Yang excess rejection, blood loss, etc.).

[0039] For example, based on the following documents:

[0040] The Yellow Emperor's Classic of Internal Medicine Lingshu says: "Observe the complexion of the disease. White complexion indicates insufficient qi and blood, red complexion indicates heat syndrome, green complexion indicates pain syndrome, yellow complexion indicates dampness syndrome, and black complexion indicates cold syndrome."

[0041] The Treatise on Febrile Diseases says: "If the face is red around the edges, it is a symptom of Yang excess being blocked; if the face is pale, it is a symptom of blood loss; if the face is blue, it is a symptom of hernia; if the face is yellow, it is a symptom of jaundice; if the face is black, it is a symptom of extreme heat."

[0042] "Diagnostics of Traditional Chinese Medicine" says: "A pale and dull complexion indicates deficiency of qi and blood; a sallow complexion indicates spleen deficiency and excessive dampness; a dark and sallow complexion indicates blood stasis; a red complexion indicates excessive heat; and a dark complexion indicates deficiency of kidney yin."

[0043] The analysis result of cold syndrome can be expressed in natural language as dark complexion and dull complexion. After structural processing, it can be expressed as a vector [0, 0, 0, 0, 1, 0, 1, 0]. The first five dimensions are used to represent the white, yellow, green, red, and black complexion, respectively, and the sixth to eighth vectors are used to represent the luster, dull, and dull. When any one of them is satisfied, the value of the vector in that dimension is 1 (for example, the vector in the fifth dimension is 1, indicating a dark complexion, and the vector in the seventh dimension is 1, indicating a dull complexion); when it is not satisfied, the value of the vector in that dimension is 0. For example, deficiency of Qi and blood can be expressed as [1,0, 0, 0, 0, 1, 0, 0, ...], heat syndrome can be expressed as [0, 1, 0, 0, 0, 0, 0, 1, ...], pain syndrome can be expressed as [0, 0, 1, 0, 0, 0, 1, 0, ...], dampness syndrome can be expressed as [0, 0, 0, 1, 0, 0, 0,1, ...], cold syndrome can be expressed as [0, 0, 0, 0, 1, 0, 1, 0, ...], Yang excess and rejection can be expressed as [0,1, 0, 0, 0, 0, 0, 1, ...], and blood loss can be expressed as [1, 0, 0, 0, 0, 1, 0, 0, ...].

[0044] By building a facial information database based on knowledge graph and natural language processing, the facial information database is rich in data and of high quality, which can lay an effective foundation for subsequent references.

[0045] Figure 2 is a schematic diagram of a process of obtaining a facial information vector according to some embodiments of this specification, such as Figure 2 As shown, in some embodiments, according to the facial image, based on the trained deep learning model, a facial information vector is obtained, including:

[0046] Step 210, performing feature extraction based on the facial image and the feature extraction module to obtain an initial feature map.

[0047] The feature extraction module can extract features based on the input facial image to obtain an initial feature map composed of multiple features. In some embodiments, the feature extraction module is established based on a residual network (ResNet), including an initial convolution layer, a normalization layer, a ReLU function, a first pooling layer, a residual unit, and a second pooling layer.

[0048] Figure 3 is a flow chart of feature extraction steps according to some embodiments of this specification, such as Figure 3 As shown, the feature extraction step further includes:

[0049] Step 211, based on the facial image, feature extraction is performed based on the initial convolution layer to obtain a primary visual feature map of the facial image. In some embodiments, the convolution kernel of the initial convolution layer can be set to 7x7, the step size is set to 2, and the corresponding padding is set to 3.

[0050] Step 212: normalize the primary visual feature map based on a normalization layer to obtain a normalized feature map. The normalization layer is used to normalize feature data to improve the training efficiency and generalization ability of the model.

[0051] Step 213, according to the normalized feature map, activation is performed based on the ReLU function to obtain an activated feature map. The ReLU function is used to introduce nonlinear processing to enhance the model's ability to recognize complex patterns.

[0052] Step 214, according to the activated feature map, dimensionality reduction is performed based on the first pooling layer to obtain a pooled feature map. In some embodiments, the pooling kernel of the first pooling layer can be set to 3x3, the step size is set to 2, and the padding is set to 1, which is used to further reduce the feature dimension and reduce the computational complexity while retaining key information.

[0053] Step 215, according to the pooled feature map, deep-level feature extraction is performed based on the convolutional layer in the residual unit to obtain residual features. The residual unit includes multiple convolutional layers and skip connections, which are used to extract facial features from different levels. The residual unit realizes the effective transfer and learning of deep features through its internal structure, and the skip connection is used to help avoid the problem of gradient disappearance in the deep network.

[0054] Step 216, based on the pooled feature map and the residual features, feature compression is performed based on the second pooling layer to obtain an initial feature map. The second pooling layer is configured at the end of the feature extraction module to compress the feature information of each channel into a single value, thereby reducing the number of parameters and complexity for subsequent feature processing.

[0055] Through the synergy of structured components, the feature extraction module can efficiently extract multi-scale and multi-level visual features from facial images, providing accurate and rich feature input for the subsequent SE module weighted processing and semantic parsing, forming a comprehensive and three-dimensional visual understanding of facial images.

[0056] Step 220, performing weighted processing based on the initial feature map and the SE module to obtain a weighted feature map.

[0057] The SE module is a module based on the attention mechanism (Squeeze-and-Excitation), which is embedded in the neural network model to perform weighted processing on some features in the initial feature map.

[0058] Figure 4 is a flowchart of weighted processing steps according to some embodiments of this specification, such as Figure 4 As shown, the weighted processing step further includes:

[0059] Step 221, perform global average pooling processing on the initial feature map to obtain channel descriptors of each channel.

[0060] A channel is an independent feature dimension in the data transmission process. In image processing, for example, the original input image usually has three channels: red, green, and blue. In a deep learning network, there are usually multiple channels after each convolutional layer, and each channel contains some features learned from the input data. A channel descriptor is a feature that characterizes the entire channel through some form of summary statistics. In the SE module, the channel descriptor is obtained by global average pooling, that is, taking the average of all elements of the feature map of each channel to produce a single value to describe the global information of the channel. Channel descriptors can help the network understand the global importance of each channel.

[0061] Step 222: Generate corresponding channel weights for each channel according to the channel descriptor.

[0062] In some embodiments, channel weights can be generated based on a small neural network, including two fully connected layers. The first fully connected layer reduces the dimension of the channel descriptor (usually by a dimensionality reduction ratio, such as reduction=16), followed by a ReLU activation function to provide nonlinear processing capabilities, and then the second fully connected layer increases the dimension back to the original number of channels. The last one outputs the channel weight of each channel through a Sigmoid activation function.

[0063] Step 223, weighted processing is performed on each channel according to the channel weight to obtain a weighted feature map.

[0064] In some embodiments, the training of the SE module can be performed based on the training of the entire deep learning model. For example, based on the gradient descent method, it is obtained by back propagation and error gradient calculation of the entire deep learning model. In this way, the SE module automatically learns which channels are more important to the final task, so it can dynamically adjust its attention allocation to different channels. This mechanism enables the network to focus more on the features that have the greatest impact on the final task, thereby improving overall performance and accuracy.

[0065] The features extracted by traditional convolutional neural networks treat different regions equally, making it difficult to highlight the key regions in facial diagnosis. This application introduces the SE attention mechanism, which can adaptively adjust the weights of different facial regions through back-propagation learning, so that the key regions with greater impact on the diagnosis results receive higher weights. By embedding the SE module in the convolutional network, the expressive power of the convolutional features has been significantly improved.

[0066] For example, the SE module can automatically focus on the key facial areas in the face diagnosis, reduce the interference of non-key areas, and thus extract facial information features with stronger discrimination and more targetedness. For example, to determine whether two facial images A and B are "pale", the facial areas of the two images are very similar, but there are subtle differences in the skin color of the forehead and cheeks of the two images.

[0067] For traditional convolutional neural networks, the overall features extracted may be difficult to distinguish between the two images, because the two images are highly similar in facial areas such as eyebrows, nose, and mouth, and these areas contribute little to the judgment of "pale face". However, by embedding the SE module, it is possible to automatically discover the importance of areas such as the forehead and cheeks for judgment and give these areas higher attention weights.

[0068] In some embodiments, the channel weight ranges from 0 to 1. For example, the SE module learns the importance of the features of the forehead and left and right cheek regions for the diagnosis of "pale face", and therefore gives higher channel weights to the channels corresponding to these regions (for example, set to 0.8 and 0.7, respectively). The corresponding channel weights of irrelevant regions such as the nose and mouth are adaptively reduced (for example, set to 0.2 and 0.1, respectively).

[0069] Step 230, performing semantic analysis based on the weighted feature map and a semantic analysis module to obtain a facial information vector.

[0070] In some embodiments, the semantic parsing module may map the weighted feature map to the semantic space through a nonlinear transformation according to the weighted feature map to obtain a facial information vector.

[0071] The semantic space is a vector space containing various facial diagnosis-related semantics. In some embodiments, the semantic space includes the correspondence between facial features and facial information vectors in the weighted feature map. Facial features can be features represented in the form of channel features, such as RGB and grayscale. By mapping facial features to the semantic space, the specific numerical values ​​of facial features can be converted into descriptions that can be semantically understood. For example, facial features with RGB of (200, 180, 50) and brightness of 80 can be converted into facial information vectors (yellow, dark) through the correspondence between facial features and facial information vectors in the semantic space.

[0072] By mapping the weighted feature map to the semantic space for semantic analysis, the original high-dimensional convolution facial features are mapped to the low-dimensional, structured, and semantic facial diagnosis information vector space, which is very consistent with the thinking mode of TCM facial diagnosis, thus facilitating subsequent matching with the TCM knowledge base. At the same time, the semantic representation discretizes the original continuous value features and can also compress and reduce the noise of facial features.

[0073] Figure 5 is a flow chart of the vector matching steps shown in some embodiments of this specification, such as Figure 5 As shown, according to the facial information vector, based on the constructed facial information database, vector matching is performed to obtain a parsing result list, including the following steps:

[0074] Step 310: Perform vector matching in the constructed facial information database according to the facial information vector to determine a plurality of similar reference vectors and a plurality of similarities corresponding to the plurality of similar reference vectors.

[0075] Similarity refers to the degree of similarity between the reference vector and the facial information vector. In some embodiments, the semantic similarity of the degree of similarity between the reference vector and the facial information vector can be calculated by calculating cosine similarity.

[0076] Step 320 : Based on a preset condition, select at least two reference vectors from a plurality of similar reference vectors to determine at least two analytical results.

[0077] The preset condition may be in various forms. For example, multiple similar reference vectors may be sorted based on their corresponding similarities, and K similar reference vectors with the highest similarity ranking may be selected to determine K analysis results, wherein the number K may be manually set. For another example, multiple similar reference vectors with similarities greater than a similarity threshold may be selected, wherein the similarity threshold may be manually set.

[0078] Step 330: sort the at least two analysis results according to the similarities corresponding to the analysis results to obtain a list of analysis results.

[0079] Specifically, if the facial image shows a sallow and dull complexion, the analysis result list may present multiple results sorted by similarity, for example, the most similar one is spleen deficiency and dampness, followed by blood stasis, and finally kidney yin deficiency.

[0080] Through vector matching, a mapping relationship between facial information and TCM knowledge is established, so that a list of analytical results that conform to TCM theory can be given based on facial appearance information for doctors' reference. In this way, the limitations of traditional facial diagnosis, which relies on doctors' personal experience, cannot accurately judge rare symptoms, and lacks objective basis, are overcome. It can better provide doctors with reasonable and well-considered reference information to help doctors better understand the condition.

[0081] In some embodiments, in order to verify the effectiveness of the facial information parsing method described in this application, experiments were conducted on a traditional Chinese medicine facial diagnosis dataset. The dataset contains 1,000 facial images, each of which was analyzed and labeled by a senior traditional Chinese medicine expert. The experiment randomly divided the dataset into a training set (70%), a validation set (10%), and a test set (20%), and used a 10-fold cross-validation method for experimental evaluation, obtaining the test results shown in Table 1.

[0082] Table 1

[0083] algorithm Accuracy Support Vector Machine Model 0.70 Decision Tree Model 0.65 Standard convolutional neural network model 0.78 This application method 0.85

[0084] As shown in Table 1, in the eight-category syndrome differentiation task, the facial information parsing method described in the present application achieved an average accuracy of 85%, which is significantly better than traditional machine learning methods such as the support vector machine (SVM) model (70%) and the decision tree model (65%), and is also better than the standard convolutional neural network (CNN) model (78%) that does not introduce attention mechanism and semantic parsing.

[0085] The facial information parsing method described in this application has an accuracy rate of 85%, which is 7 to 20 percentage points higher than the baseline method, fully demonstrating the effectiveness of innovations such as the TCM knowledge base, attention mechanism, and semantic parsing. Through visual analysis, the facial information parsing method automatically focuses on key areas of the face, can accurately extract facial diagnosis elements such as complexion and luster, and gives a dialectical analysis that conforms to TCM theory, showing good interpretability and clinical practicality, and can provide accurate references for doctors' facial diagnosis process.

[0086] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A facial information analysis method, characterized in that: include: Obtaining facial images; According to the facial image, based on a trained deep learning model, a facial information vector is obtained, wherein the facial information vector includes facial information semantics, and the facial information semantics includes at least complexion, gloss, and morphology; According to the facial information vector, based on the constructed facial information database, vector matching is performed to obtain a parsing result list; The deep learning model includes a feature extraction module, an SE module and a semantic parsing module, and the facial information vector is obtained based on the facial image and the trained deep learning model, including: According to the facial image, based on the feature extraction module, feature extraction is performed to obtain an initial feature map; According to the initial feature map, based on the SE module, weighted processing is performed to obtain a weighted feature map; According to the weighted feature map, based on the semantic parsing module, semantic parsing is performed to obtain a facial information vector; The facial information database is constructed based on the following steps: Define the facial diagnosis knowledge ontology, and standardize the concepts, relationships, and attributes of the knowledge base of the facial diagnosis knowledge ontology; Using natural language processing technology, the original unstructured facial diagnosis knowledge is structured and the key elements of facial diagnosis knowledge are extracted; Instantiate key elements into knowledge ontology and construct facial information database.

2. The facial information analysis method according to claim 1, characterized in that: The step of performing weighted processing based on the initial feature map and the SE module to obtain a weighted feature map includes: Performing global average pooling processing on the initial feature map to obtain a channel descriptor for each channel; Generate corresponding channel weights of the respective channels according to the channel descriptors; According to the channel weights, weighted processing is performed on the channels to obtain a weighted feature map.

3. The facial information analysis method according to claim 1, characterized in that: The feature extraction module is established based on the residual network, including an initial convolution layer, a normalization layer, a ReLU function, a first pooling layer, a residual unit and a second pooling layer. The feature extraction is performed based on the facial image and the feature extraction module to obtain an initial feature map, including: According to the facial image, feature extraction is performed based on the initial convolutional layer to obtain a primary visual feature map of the facial image; According to the primary visual feature map, normalization is performed based on the normalization layer to obtain a normalized feature map; According to the normalized feature map, activation is performed based on the ReLU function to obtain an activated feature map; According to the activated feature map, dimension reduction is performed based on the first pooling layer to obtain a pooling feature map; According to the pooled feature map, deep feature extraction is performed based on the convolution layer in the residual unit to obtain residual features; According to the pooled feature map and the residual features, feature compression is performed based on the second pooling layer to obtain an initial feature map.

4. The facial information analysis method according to claim 1, characterized in that: The step of performing semantic analysis based on the weighted feature map and the semantic analysis module to obtain a facial information vector includes: Based on the semantic parsing module, the weighted feature map is mapped to the semantic space by nonlinear transformation to obtain a facial information vector.

5. The facial information analysis method according to claim 4, characterized in that: The semantic space includes the correspondence between the facial features in the weighted feature map and the facial information vector, and the facial information vector includes facial information semantics.

6. The facial information analysis method according to claim 1, characterized in that: The facial information database is constructed based on knowledge graph and natural language processing, and the facial information database includes multiple reference vectors and a parsing result corresponding to each reference vector.

7. The facial information analysis method according to claim 6, characterized in that: According to the facial information vector, based on the constructed facial information database, vector matching is performed to obtain a parsing result list, including: According to the facial information vector, vector matching is performed in a constructed facial information database to determine a plurality of similar reference vectors and a plurality of similarities corresponding to the plurality of similar reference vectors; Based on a preset condition, selecting at least two similar reference vectors from the plurality of similar reference vectors, and determining at least two analytical results; At least two of the analysis results are sorted according to the similarities corresponding to the analysis results to obtain a list of analysis results.

8. A facial information analysis system, characterized in that: include: An acquisition module, wherein the acquisition module is used to acquire a facial image; An image processing module, the image processing module is used to obtain a facial information vector according to the facial image based on a trained deep learning model, the facial information vector including facial information semantics, and the facial information semantics at least including complexion, gloss, and morphology; An image analysis module, the image analysis module is used to perform vector matching according to the facial information vector and based on a constructed facial information database to obtain a list of analysis results; The deep learning model includes a feature extraction module, an SE module and a semantic parsing module, and the facial information vector is obtained based on the facial image and the trained deep learning model, including: According to the facial image, based on the feature extraction module, feature extraction is performed to obtain an initial feature map; According to the initial feature map, based on the SE module, weighted processing is performed to obtain a weighted feature map; According to the weighted feature map, based on the semantic parsing module, semantic parsing is performed to obtain a facial information vector; The facial information database is constructed based on the following steps: Define the facial diagnosis knowledge ontology, and standardize the concepts, relationships, and attributes of the knowledge base of the facial diagnosis knowledge ontology; Using natural language processing technology, the original unstructured facial diagnosis knowledge is structured and the key elements of facial diagnosis knowledge are extracted; Instantiate key elements into knowledge ontology and construct facial information database.

9. A computer-readable storage medium, characterized in that: The storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer runs the method as claimed in claim 1.