Skin evaluation and skin care decision-making system and method based on AI drive

Abstract

The invention discloses a skin evaluation and skin care decision-making system and method based on AI drive, and belongs to the field of big data and artificial intelligence, and the system comprises a data collection module which is used for obtaining a skin image and user information; the skin evaluation module is used for performing skin state evaluation through multi-modal data fusion; the recommendation decision module is used for reinforcing learning to generate a personalized skin care scheme; the resource scheduling module is used for intelligently scheduling offline nursing resources; and the effect tracking module is used for acquiring the post-nursing skin image corresponding to the expected skin improvement period in the expected effect and comparing and analyzing the post-nursing skin image with the pre-nursing skin image to obtain a nursing effect. According to the method, skin state evaluation can be carried out through multi-modal data fusion, and a personalized skin care scheme is generated in combination with reinforcement learning.

Description

Technical Field

[0001] The present invention relates to the field of big data and artificial intelligence technology, and in particular to an AI-driven skin assessment and skin care decision-making system and method. Background Art

[0002] The existing skin care decision-making method relies only on image analysis and does not integrate multiple data such as user living habits, resulting in a single evaluation dimension. When recommending skin care products, fixed rules are generally used to match products, and strategies cannot be optimized based on user dynamic feedback. The existing skin care decision-making system stops at online suggestions and does not achieve intelligent connection with offline care services. Summary of the invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes an AI-driven skin assessment and skin care decision-making system and method, which can assess skin condition through multimodal data fusion, generate personalized skin care plans in combination with reinforcement learning, and intelligently schedule offline care resources considering time and space constraints.

[0004] On the one hand, an embodiment of the present invention provides an AI-driven skin assessment and skin care decision-making system, including: a data acquisition module, used to obtain skin images and user information, the user information including: age, gender, skin quality, work and rest information, nursing time preference and nursing history information; a skin assessment module, used to obtain global skin problem classification probability, local lesion detection results and skin quality multi-dimensional scoring results through a skin image classification model, a skin lesion detection model and a skin quality scoring model, and fuse the user information, perform multi-model feature integration and dynamic weighting, and generate a joint feature vector; a recommendation decision module, used to input the joint feature vector, perform collaborative filtering based on the user historical behavior matrix, and obtain a potential factor vector and group preference distribution through SVD++ decomposition; calculate the product A list of candidate products is obtained based on component similarity; a state vector is constructed based on the latent factor vector and the group preference distribution, and an action vector is obtained based on the state vector through the DDPG algorithm to optimize the list of candidate products; a nursing plan is generated based on the action vector; a resource scheduling module is used to obtain the nurse's profile and real-time status, and recommend nurses based on the nurse-user bipartite graph matching; a scheduling plan is generated based on the appointment request and the resource calendar, and the time period demand and nurse load are obtained, and the real-time service price is obtained based on the dynamic pricing model; an effect tracking module is used to obtain the skin image after care corresponding to the expected skin improvement cycle in the expected effect and compare and analyze the skin image before care to obtain the nursing effect; the nursing cycle of the nursing plan is adjusted based on the nursing effect and the nurse-user bipartite graph is incrementally updated.

[0005] According to some embodiments of the present invention, the data acquisition module includes: an image acquisition unit for acquiring a standardized skin image based on illumination compensation and face positioning; a user input unit for acquiring user information and obtaining a standardized user vector through a rule encoder; the skin evaluation module includes:

[0006] The image classification unit is used to build a skin image classification model based on EfficientNet and obtain the global skin problem classification probability. The formula is: ;in, is the global skin problem classification probability, is the global skin problem classification weight matrix, is the classification bias term, is the image feature vector; the lesion detection unit is used to build a skin lesion detection model based on YOLOv8, locate local skin abnormalities and quantify local lesion detection results, which include lesion location, lesion number and inflammation area; the detection formula is: ;in, is the confidence level, is the probability of each lesion type existing in the prediction box, Represents the predicted bounding box With the true bounding box The degree of overlap between them; a quality scoring unit, used to build a skin quality scoring model based on the ResNet50+SE attention mechanism to obtain a multi-dimensional skin quality scoring result; the skin quality scoring model is: ;in, is the score of the i-th skin quality dimension, wherein the skin quality dimension includes at least one of the following items: fineness, glossiness and firmness, is the activation function, is the weight function of the regression head, k is the feature channel index, d is the total number of ResNet feature channels, Mark the regression task parameters, The feature map extracted by ResNet, is the global average pooling result of the ResNet feature map; the multimodal fusion unit is used to perform multimodal fusion and dynamic weighting on the skin image features and user information based on the first fusion formula to obtain a first joint feature vector; the fusion formula is: ;in, is the first joint eigenvector, and are the image feature weight coefficient and the user feature weight coefficient respectively, is a standardized user vector obtained according to the user information; the image feature weight coefficient is adjusted according to the clarity of the skin image, and the user feature weight coefficient is adjusted according to the completeness of the user information.

[0007] According to some embodiments of the present invention, the data acquisition module also includes: a sensor integration unit for acquiring environmental information based on an environmental detection device, wherein the environmental information includes temperature and humidity data and an ultraviolet index; and a biometric unit for acquiring real-time physiological indicators in user information based on a smart wearable device, wherein the real-time physiological indicators include at least a heart rate or a sebum secretion indicator.

[0008] According to some embodiments of the present invention, the multimodal fusion unit is used to perform multimodal fusion and dynamic weighting on skin image features, user information, environmental information and real-time physiological indicators based on a second fusion formula to obtain a second joint feature vector; the second fusion formula is: ;in, is the second joint eigenvector, and are the environmental characteristic weight coefficient and the physiological characteristic weight coefficient, respectively. is a normalized environmental feature vector obtained according to the environmental information, is a normalized physiological feature vector obtained according to the real-time physiological index.

[0009] According to some embodiments of the present invention, the recommendation decision module includes: a collaborative filtering unit, which is used to obtain a user historical behavior matrix R, , in is the user latent factor matrix, is the product potential factor matrix, N is the number of users, M is the number of products, and the potential factor vector is obtained and group preference distribution , Indicates the product category; product matching unit, used to obtain the current skin problem vector output by EfficientNet and product ingredient matrix , calculate the symptom-component match , get a list of candidate products , is the component vector of the jth product; the reinforcement learning strategy unit is used to learn the latent factor vector , Group Preference Distribution , nurse graph embedding vector , environmental feature vector Constructing the state vector , , update the action vector based on the DDPG strategy , the DDPG strategy is: ,in are policy network parameters, is Gaussian exploratory noise, is the value function, E is the mathematical expectation, r is the reward signal at the current moment, is the future reward discount factor, s' is the execution action After transferring to the new state, are the new strategy network parameters corresponding to the new state s'; ,in, is the product usage intensity coefficient, is the weight of the frequency of in-store care, Get a list of candidate products for cost-sensitive adjustment factors , filter the candidate product list according to the action vector ; Solution generation unit, used to obtain a list of candidate products , motion vector and a user taboo rule base, filters the candidate product list and adjusts the concentration threshold of the candidate products according to the product usage intensity coefficient, re-prioritizes the candidate products according to the cost-sensitive adjustment factor and the graph embedding similarity, calculates the nursing cycle according to the in-store care frequency weight and the nurse load; and outputs a nursing plan, which includes a product portfolio, a store cycle and an expected cost.

[0010] According to some embodiments of the present invention, the resource scheduling module includes: a nurse recommendation unit for obtaining nurse files and real-time status, constructing a nurse-user bipartite graph, wherein the node types include nurse nodes based on nurse skill characteristics, user nodes based on skin characteristics, and store nodes based on location or equipment characteristics, and the edge types include service relationship edges based on service ratings or repurchase times and affiliation relationships based on affiliation between nurses and stores; recommending nurses based on nurse-user bipartite graph matching: the nurse graph matching formula is: ,in, For the Layer nurse node embedding, For the Layer user node embedding, W is the graph convolution weight matrix, is an activation function, MEAN is an aggregation function; the spatiotemporal optimization unit is used to generate a scheduling plan based on appointment requests and resource calendars; the dynamic pricing unit is used to obtain time period demand and nurse load, and obtain real-time service prices based on the dynamic pricing model: ,in, The basic service price is is the demand heat of the time period, is the historical demand average, is the demand standard deviation.

[0011] According to some embodiments of the present invention, the effect tracking module includes: a comparison and analysis unit for obtaining a skin image after care and a skin image before care for comparison and analysis, and calculating an improvement degree. : ; and are the image feature vectors before and after care, respectively. and They are the skin quality score before care and the skin quality score after care; the change detection unit is used to obtain the skin image before care and the skin image after care, and obtain the pixel-level change heat map based on U-Net.

[0012] According to some embodiments of the present invention, the system includes: a graph data management module for managing graph data; including: a graph storage unit for acquiring real-time interaction data and obtaining a dynamic relationship graph based on Neo4j+Cypher; a graph sampling unit for acquiring original graph data and obtaining a training subgraph through sampling; a graph version control unit for obtaining a version snapshot of the graph structure based on a differential storage algorithm and a graph operation log.

[0013] According to some embodiments of the present invention, the system further includes: a model updating module, configured to obtain an incremental data set and obtain updated model weights based on federated learning.

[0014] An AI-driven skin assessment and skin care decision system according to an embodiment of the present invention includes at least the following beneficial effects: the embodiment of the present invention can evaluate skin condition through multimodal data fusion, generate personalized skin care plans in combination with reinforcement learning, and intelligently schedule offline care resources considering time and space constraints.

[0015] On the other hand, an embodiment of the present invention provides an AI-driven skin assessment and skin care decision-making method, comprising the following steps: S100, obtaining skin images and user information, wherein the user information includes: age, gender, skin quality, work and rest information, nursing time preference and nursing history information; S200, obtaining the global skin problem classification probability, local lesion detection results and skin quality multidimensional scoring results through a skin image classification model, a skin lesion detection model and a skin quality scoring model, and fusing the user information, performing multi-model feature integration and dynamic weighting to generate a joint feature vector; S300, inputting the joint feature vector, performing collaborative filtering based on the user historical behavior matrix, and obtaining a potential factor vector and a group preference distribution through SVD++ decomposition; calculating the product The candidate product list is obtained based on the component similarity; and a state vector is constructed based on the latent factor vector and the group preference distribution. Based on the state vector, an action vector is obtained through the DDPG algorithm to optimize the candidate product list; and a care plan is generated based on the action vector; S400, the nurse profile and real-time status are obtained, and nurses are recommended based on the nurse-user bipartite graph matching; a scheduling plan is generated based on the appointment request and the resource calendar, and the time period demand and nurse load are obtained, and the real-time service price is obtained based on the dynamic pricing model; S500, the skin image after care corresponding to the expected skin improvement cycle in the expected effect is obtained and compared with the skin image before care to obtain the care effect; the care cycle of the care plan is adjusted based on the care effect and the nurse-user bipartite graph is incrementally updated.

[0016] An AI-driven skin assessment and skin care decision-making method according to an embodiment of the present invention includes at least the following beneficial effects: The method according to an embodiment of the present invention is used to implement any of the above-mentioned systems, so the method according to an embodiment of the present invention has all the beneficial effects of the above-mentioned systems.

[0017] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0019] Figure 1 A schematic block diagram of modules of a system according to an embodiment of the present invention;

[0020] Figure 2 The figure is a flowchart of a method according to an embodiment of the present invention.

[0021] Reference numerals:

[0022] Data collection module 100 , skin assessment module 200 , recommendation decision module 300 , resource scheduling module 400 , effect tracking module 500 . DETAILED DESCRIPTION

[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0024] In the description of the present invention, "several" means one or more, "more" means two or more, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0025] Reference Figure 1 The embodiment of the present invention proposes an AI-driven skin assessment and skin care decision system, comprising:

[0026] The data acquisition module 100 is used to obtain skin images and user information, and the user information includes: age, gender, skin quality, work and rest information, care time preference and care history information.

[0027] The skin assessment module 200 is used to obtain the global skin problem classification probability, local lesion detection results and skin quality multi-dimensional scoring results through the skin image classification model, skin lesion detection model and skin quality scoring model, and integrate user information, perform multi-model feature integration and dynamic weighting to generate a joint feature vector.

[0028] The recommendation decision module 300 is used to input a joint feature vector, perform collaborative filtering based on the user historical behavior matrix, obtain a latent factor vector and a group preference distribution through SVD++ decomposition; calculate the similarity of product ingredients to obtain a candidate product list; and construct a state vector based on the latent factor vector and the group preference distribution, and based on the state vector, obtain an action vector through the DDPG algorithm to optimize the candidate product list; and generate a care plan based on the action vector.

[0029] The resource scheduling module 400 is used to obtain the nurse's profile and real-time status, and recommend nurses based on the nurse-user bipartite graph matching; generate a scheduling plan based on the appointment request and the resource calendar, and obtain the time period demand and nurse load, and obtain the real-time service price based on the dynamic pricing model.

[0030] The effect tracking module 500 is used to obtain the skin image after care corresponding to the expected skin improvement cycle in the expected effect and compare and analyze the skin image before care to obtain the care effect; adjust the care cycle of the care plan based on the care effect and update the nurse-user bipartite graph incrementally.

[0031] In some embodiments, the data acquisition module 100 includes:

[0032] The image acquisition unit is used to obtain a standardized skin image based on illumination compensation and face positioning.

[0033] The user input unit is used to obtain user information and obtain a standardized user vector through a rule encoder.

[0034] In some embodiments, the skin assessment module 200 includes:

[0035] The image classification unit is used to build a skin image classification model based on EfficientNet and obtain the global skin problem classification probability. The formula is:

[0036] ;

[0037] in, is the global skin problem classification probability, is the global skin problem classification weight matrix, is the classification bias term, is the image feature vector. EfficientNet leads multimodal fusion and provides basic skin quality judgment (such as oily / dry).

[0038] The lesion detection unit is used to build a skin lesion detection model based on YOLOv8, locate local skin abnormalities and quantify the local lesion detection results, which include lesion location, number of lesions and inflammation area. The skin lesion detection model quantifies the severity (such as the number / area of ​​acne) and triggers specific care plans (such as local application of acid products). For example, if active inflammation is detected (such as >3 pustules), the weight of anti-inflammatory components is automatically increased in the recommended plan.

[0039] The quality scoring unit is used to build a skin quality scoring model based on the ResNet50+SE attention mechanism to obtain a multi-dimensional skin quality scoring result; it is used to track the effect (Δ score calculation) and guide the cycle adjustment (such as shortening the care interval when the score improves slowly). For example, when the firmness score is low, radio frequency care items are recommended first. The skin quality scoring model of this embodiment is:

[0040] ;

[0041] in, is the score of the i-th skin quality dimension, which includes at least one of the following: fineness, glossiness and firmness, is the activation function, is the weight function of the regression head, k is the feature channel index, d is the total number of ResNet feature channels, Mark the regression task parameters, The feature map extracted by ResNet, is the global average pooling result of the ResNet feature map.

[0042] The multimodal fusion unit is used to perform multimodal fusion and dynamic weighting on the skin image features and the user information based on the first fusion formula to obtain a first joint feature vector; the fusion formula is:

[0043] ;

[0044] in, is the first joint eigenvector, and are the image feature weight coefficient and the user feature weight coefficient respectively, is a standardized user vector obtained according to user information; the image feature weight coefficient is adjusted according to the clarity of the skin image, and the user feature weight coefficient is adjusted according to the completeness of the user information.

[0045] In a specific embodiment, the user detected 2 inflammatory papules (YOLOv8 output), and the ResNet50 oiliness score was 8.5 / 10. Dynamic fusion: YOLOv8 feature weight increased to 0.4 (active lesions detected); generate joint feature vector: [0.87, 6.8, 8.2, 2, 48, 0.85, ...] (including global classification probability, number of lesions, inflammation area, oiliness score, etc.); recommendation decision: collaborative filtering: matching similar users (oily + local inflammation), reinforcement learning: select 1.5% salicylic acid (avoid 2% overstimulation of active inflammation); resource scheduling: give priority to nurses who are good at acne care and have a daily load of <70%. The outputs of the three evaluation models in this embodiment jointly influence the decision through feature-level fusion (non-independent use).

[0046] In some embodiments, the data acquisition module 100 further includes:

[0047] The sensor integrated unit is used to obtain environmental information based on environmental detection equipment, and the environmental information includes temperature and humidity data and ultraviolet index.

[0048] The biometric unit is used to obtain real-time physiological indicators in user information based on the smart wearable device, and the real-time physiological indicators at least include heart rate or sebum secretion indicators.

[0049] In a specific embodiment, the environment: temperature 28°C, humidity 35%, UV=9; physiology: sebum secretion rate 25μg / cm² / min, pH=6.0. The final adaptation recommendation decision is: environmental adaptation: SPF50+ sunscreen is recommended; physiological adaptation: increase the salicylic acid concentration to 2.5% (sebum secretion rate>20 threshold).

[0050] In some embodiments, the multimodal fusion unit is used to perform multimodal fusion and dynamic weighting on the skin image features, user information, environmental information and real-time physiological indicators based on a second fusion formula to obtain a second joint feature vector; the second fusion formula is:

[0051] ;

[0052] in, is the second joint eigenvector, and are the environmental characteristic weight coefficient and the physiological characteristic weight coefficient, respectively. is the normalized environmental feature vector obtained based on environmental information, is the normalized physiological feature vector obtained based on real-time physiological indicators.

[0053] In some embodiments, the recommendation decision module 300 includes:

[0054] Collaborative filtering unit, used to obtain the user's historical behavior matrix R, , in is the user potential factor matrix, is the product potential factor matrix, N is the number of users, M is the number of products, and the potential factor vector is obtained and group preference distribution , Indicates the product category.

[0055] Product matching unit, used to obtain the current skin problem vector output by EfficientNet and product ingredient matrix , calculate the symptom-component match , get a list of candidate products , is the component vector of the j-th product.

[0056] Reinforcement learning policy unit, used to learn , Group Preference Distribution , nurse graph embedding vector , environmental feature vector Constructing the state vector , , update the action vector based on the DDPG strategy , ,in, is the product usage intensity coefficient, is the weight of the frequency of in-store care, Get a list of candidate products for cost-sensitive adjustment factors , filter the candidate product list based on the action vector .

[0057] Solution generation unit, used to obtain a list of candidate products and motion vector (It may also include a user taboo rule library), filter the candidate product list according to the product usage intensity coefficient and adjust the concentration threshold of the candidate products, re-prioritize the candidate products according to the cost-sensitive adjustment factor and the graph embedding similarity, calculate the nursing cycle according to the in-store care frequency weight and the nurse load; output the nursing plan, which includes the product portfolio, store cycle and expected cost.

[0058] In the solution generation unit of this embodiment, a state vector is constructed according to the nurse graph embedding vector, the product graph embedding vector is obtained through the product knowledge graph, the graph embedding similarity is obtained based on the product graph embedding vector and the nurse graph embedding vector, and the user-nurse matching degree is obtained. Then, the product priority is rearranged and the nurse load used in calculating the nursing cycle can be read from the graph node attributes of the nurse graph embedding vector.

[0059] In some embodiments, the resource scheduling module 400 includes:

[0060] The nurse recommendation unit is used to obtain the nurse's profile and real-time status, and construct a nurse-user bipartite graph. The node types include nurse nodes based on nurse skill characteristics, user nodes based on skin characteristics, and store nodes based on location or equipment characteristics. The edge types include service relationship edges based on service ratings or repurchase times and affiliation edges based on the affiliation between nurses and stores. Recommend nurses based on GraphSAGE and nurse-user bipartite graph matching: The nurse graph matching formula is: ,in, For the Layer nurse node embedding, For the Layer user node embedding, W is the graph convolution weight matrix, is an activation function and MEAN is an aggregation function.

[0061] The spatiotemporal optimization unit is used to generate a scheduling solution according to the reservation request and the resource calendar. The spatiotemporal optimization unit of this embodiment is optimized by GAT+constraint programming and GNN matching.

[0062] The dynamic pricing unit is used to obtain time period demand and nurse load, and obtain real-time service prices based on the dynamic pricing model: ,in, The basic service price is is the demand heat of the time period, is the historical demand average, is the demand standard deviation.

[0063] In this example, data preparation is first performed to build a nurse-user bipartite graph (historical service records) and a pre-trained graph neural network embedding (cold start initialization). Then, data is collected for multimodal fusion to generate joint features. Nurses are recommended through graph attention matching, and dynamic pricing and scheduling optimization are performed. Finally, feedback learning is performed: nursing effect data is collected and the graph structure is incrementally updated.

[0064] In some embodiments, the effect tracking module 500 includes:

[0065] The comparison and analysis unit is used to obtain the skin image after care and the skin image before care for comparison and analysis, and calculate the improvement degree. : ; and are the image feature vectors before and after care, respectively. and The skin quality scores were before and after care.

[0066] The change detection unit is used to obtain the skin image before and after care, and obtain a pixel-level change heat map based on U-Net.

[0067] In some embodiments, the system of the embodiment of the present invention includes: a graph data management module, which is used to manage graph data; including:

[0068] The graph storage unit is used to obtain real-time interactive data and obtain a dynamic relationship graph based on Neo4j+Cypher.

[0069] The graph sampling unit is used to obtain the original graph data and obtain the training subgraph through sampling.

[0070] The graph version control unit is used to obtain a version snapshot of the graph structure based on the graph operation log based on the differential storage algorithm.

[0071] In some embodiments, the system of the embodiment of the present invention further includes:

[0072] The model update module is used to obtain incremental data sets and obtain updated model weights based on federated learning.

[0073] Reference Figure 2The embodiment of the present invention proposes an AI-driven skin assessment and skin care decision-making method, comprising the following steps:

[0074] S100, obtaining skin images and user information, the user information including: age, gender, skin quality, work and rest information, care time preference and care history information.

[0075] S200, obtain the global skin problem classification probability, local lesion detection results and skin quality multi-dimensional scoring results through the skin image classification model, skin lesion detection model and skin quality scoring model, and fuse user information, perform multi-model feature integration and dynamic weighting to generate a joint feature vector.

[0076] S300, input the joint feature vector, perform collaborative filtering based on the user historical behavior matrix, obtain the latent factor vector and group preference distribution through SVD++ decomposition; calculate the similarity of product ingredients to obtain a list of candidate products; and construct a state vector based on the latent factor vector and group preference distribution. Based on the state vector, obtain the action vector through the DDPG algorithm to optimize the candidate product list; and generate a care plan based on the action vector.

[0077] S400, obtain the nurse's profile and real-time status, and recommend nurses based on the nurse-user bipartite graph matching; generate a scheduling plan based on the appointment request and resource calendar, and obtain the time period demand and nurse load, and obtain the real-time service price based on the dynamic pricing model.

[0078] S500, obtaining the skin image after care corresponding to the expected skin improvement cycle in the expected effect and comparing it with the skin image before care to obtain the care effect; adjusting the care cycle and increment of the care plan based on the care effect to update the nurse-user bipartite graph.

[0079] Although specific embodiments are described herein, those of ordinary skill in the art will recognize that many other modifications or alternative embodiments are also within the scope of the present disclosure. For example, any of the functions and / or processing capabilities described in conjunction with a particular device or component may be performed by any other device or component. In addition, although various exemplary implementations and architectures have been described according to embodiments of the present disclosure, those of ordinary skill in the art will recognize that many other modifications to the exemplary implementations and architectures described herein are also within the scope of the present disclosure.

[0080] It will be appreciated by those skilled in the art that all or some of the steps and systems in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or transient medium). As known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0081] The above is a specific description of the preferred implementation of the present application, but the present application is not limited to the above-mentioned implementation mode. Technical personnel familiar with the field can also make various equivalent deformations or substitutions without violating the spirit of the present application. These equivalent deformations or substitutions are all included in the scope defined by the claims of the present application.

Claims

1. An AI-driven skin assessment and skin care decision-making system, characterized in that: include: A data acquisition module is used to obtain skin images and user information, wherein the user information includes age, gender, skin quality, work and rest information, nursing time preference and nursing history information; The skin assessment module is used to obtain the global skin problem classification probability, local lesion detection results and skin quality multi-dimensional scoring results through the skin image classification model, skin lesion detection model and skin quality scoring model, and fuse the user information, perform multi-model feature integration and dynamic weighting to generate a joint feature vector; A recommendation decision module is used to input the joint feature vector, perform collaborative filtering based on the user historical behavior matrix, obtain the potential factor vector and group preference distribution through SVD++ decomposition; calculate the product component similarity to obtain a candidate product list; and construct a state vector based on the potential factor vector and group preference distribution, and based on the state vector, obtain an action vector through the DDPG algorithm to optimize the candidate product list; and generate a care plan based on the action vector; Resource scheduling module, which is used to obtain nurses' profiles and real-time status, and recommend nurses based on nurse-user bipartite graph matching; generate scheduling plans based on appointment requests and resource calendars, obtain time period demand and nurse load, and obtain real-time service prices based on dynamic pricing models; The effect tracking module is used to obtain the skin images after care corresponding to the expected skin improvement cycle in the expected effect and compare and analyze the skin images before care to obtain the care effect; adjust the care cycle of the care plan based on the care effect and update the nurse-user bipartite graph.

2. The AI-driven skin assessment and skin care decision system according to claim 1, characterized in that: The data acquisition module comprises: An image acquisition unit, used for acquiring a standardized skin image based on illumination compensation and face positioning; A user input unit, used to obtain user information and obtain a standardized user vector through a rule encoder; The skin assessment module includes: The image classification unit is used to build a skin image classification model based on EfficientNet and obtain the global skin problem classification probability. The formula is: ； in, is the global skin problem classification probability, is the global skin problem classification weight matrix, is the classification bias term, is the image feature vector; Softmax is the activation function; A lesion detection unit is used to build a skin lesion detection model based on YOLOv8, locate local skin abnormalities and quantify local lesion detection results, where the local lesion detection results include lesion location, lesion number and inflammation area; The quality scoring unit is used to build a skin quality scoring model based on the ResNet50+SE attention mechanism to obtain a multi-dimensional skin quality scoring result; the skin quality scoring model is: ； in, is the score of the i-th skin quality dimension, wherein the skin quality dimension includes at least one of the following items: fineness, glossiness and firmness, is the activation function, is the weight function of the regression head, k is the feature channel index, d is the total number of ResNet feature channels, Mark the regression task parameters, The feature map extracted by ResNet, is the global average pooling result of the ResNet feature map; The multimodal fusion unit is used to perform multimodal fusion and dynamic weighting on the skin image features and the user information based on a first fusion formula to obtain a first joint feature vector; the fusion formula is: ； in, is the first joint eigenvector, and are the image feature weight coefficient and the user feature weight coefficient respectively, is a standardized user vector obtained according to the user information; the image feature weight coefficient is adjusted according to the clarity of the skin image, and the user feature weight coefficient is adjusted according to the completeness of the user information.

3. The AI-driven skin assessment and skin care decision system according to claim 2, characterized in that: The data acquisition module also includes: A sensor integrated unit, used to obtain environmental information based on the environmental detection device, wherein the environmental information includes temperature and humidity data and ultraviolet index; The biometric unit is used to obtain real-time physiological indicators in user information based on the smart wearable device, and the real-time physiological indicators at least include heart rate or sebum secretion indicators.

4. The AI-driven skin assessment and skin care decision system according to claim 3, characterized in that: The multimodal fusion unit is used to perform multimodal fusion and dynamic weighting on the skin image features, user information, environmental information and real-time physiological indicators based on a second fusion formula to obtain a second joint feature vector; the second fusion formula is: ； in, is the second joint eigenvector, and are the environmental characteristic weight coefficient and the physiological characteristic weight coefficient, respectively. is a normalized environmental feature vector obtained according to the environmental information, is a normalized physiological feature vector obtained according to the real-time physiological index.

5. The AI-driven skin assessment and skin care decision system according to claim 4, characterized in that: The recommendation decision module includes: Collaborative filtering unit, used to obtain the user's historical behavior matrix R, , in is the user potential factor matrix, is the product potential factor matrix, N is the number of users, M is the number of products, and the potential factor vector is obtained and group preference distribution , Indicates the product category; Product matching unit, used to obtain the current skin problem vector output by EfficientNet and product ingredient matrix , calculate the symptom-component match , get a list of candidate products , is the component vector of the j-th product; Reinforcement learning policy unit, used to learn , Group Preference Distribution , nurse graph embedding vector , environmental feature vector Constructing the state vector , , update the action vector based on the DDPG strategy , ,in, is the product usage intensity coefficient, is the weight of the frequency of in-store care, Get a list of candidate products for cost-sensitive adjustment factors , filter the candidate product list according to the action vector ; Among them, the embedding vector of the nurse graph is the nurse-user bipartite graph; Solution generation unit, used to obtain a list of candidate products and motion vector , filter the candidate product list and adjust the concentration threshold of the candidate products according to the product usage intensity coefficient, re-prioritize the candidate products according to the cost-sensitive adjustment factor and the graph embedding similarity, calculate the nursing cycle according to the in-store care frequency weight and the nurse load; output the nursing plan, which includes the product portfolio, the in-store cycle and the expected cost.

6. The AI-driven skin assessment and skin care decision system according to claim 1, characterized in that: The resource scheduling module includes: The nurse recommendation unit is used to obtain the nurse's profile and real-time status, and construct a nurse-user bipartite graph. The node types include nurse nodes based on nurse skill characteristics, user nodes based on skin characteristics, and store nodes based on location or equipment characteristics. The edge types include service relationship edges based on service ratings or repurchase times and affiliation edges based on the affiliation between nurses and stores. Recommend nurses based on nurse-user bipartite graph matching: The nurse graph matching formula is: ,in, For the Layer nurse node embedding, For the Layer user node embedding, W is the graph convolution weight matrix, is the activation function, MEAN is the aggregation function; A time-space optimization unit, which is used to generate scheduling plans based on appointment requests and resource calendars; The dynamic pricing unit is used to obtain time period demand and nurse load, and obtain real-time service prices based on the dynamic pricing model: ,in, The basic service price is is the demand heat of the time period, is the historical demand average, is the demand standard deviation.

7. The AI-driven skin assessment and skin care decision system according to claim 1, characterized in that: The effect tracking module includes: The comparison and analysis unit is used to obtain the skin image after care and the skin image before care for comparison and analysis, and calculate the improvement degree. : ; and are the image feature vectors before and after care, respectively. and are the skin quality scores before and after care, respectively, and sign is a symbolic function used to determine the directionality of the care effect; The change detection unit is used to obtain the skin image before and after care, and obtain a pixel-level change heat map based on U-Net.

8. The AI-driven skin assessment and skin care decision system according to claim 6, characterized in that: The system includes: a graph data management module for managing graph data; the graph data management module includes: Graph storage unit, used to obtain real-time interactive data and obtain dynamic relationship graphs based on Neo4j+Cypher; A graph sampling unit, used to obtain the original graph data and obtain the training subgraph through sampling; The graph version control unit is used to obtain a version snapshot of the graph structure based on the graph operation log based on the differential storage algorithm.

9. The AI-driven skin assessment and skin care decision system according to claim 1, characterized in that: The system further comprises: The model update module is used to obtain incremental data sets and obtain updated model weights based on federated learning.

10. An AI-driven skin assessment and skin care decision-making method, used in the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S100, obtaining skin images and user information, wherein the user information includes: age, gender, skin quality, work and rest information, nursing time preference and nursing history information; S200, obtaining the global skin problem classification probability, local lesion detection results and skin quality multi-dimensional scoring results through the skin image classification model, the skin lesion detection model and the skin quality scoring model, fusing the user information, performing multi-model feature integration and dynamic weighting, and generating a joint feature vector; S300, input the joint feature vector, perform collaborative filtering based on the user historical behavior matrix, obtain the potential factor vector and group preference distribution through SVD++ decomposition; calculate the product component similarity to obtain a candidate product list; and construct a state vector based on the potential factor vector and group preference distribution, and based on the state vector, obtain an action vector through the DDPG algorithm to optimize the candidate product list; and generate a care plan based on the action vector; S400, obtain the nurse's profile and real-time status, and recommend nurses based on the nurse-user bipartite graph matching; generate a scheduling plan based on the appointment request and resource calendar, and obtain the time period demand and nurse load, and obtain the real-time service price based on the dynamic pricing model; S500, obtaining the skin image after care corresponding to the expected skin improvement cycle in the expected effect and comparing it with the skin image before care to obtain the care effect; adjusting the care cycle of the care plan based on the care effect and updating the nurse-user bipartite graph.

Images