Hairstylist intelligent recommendation system and method based on multi-modal hairstyle feature analysis and dynamic skill scoring

Abstract

This invention discloses an intelligent hairdresser recommendation system and method based on multimodal hairstyle feature analysis and dynamic skill scoring. The method first constructs a multimodal hairstyle feature database and extracts hairstyle features using a deep neural network that fuses texture, structure, and color modalities. Second, it establishes a hairdresser skill scoring model, calculating hairstyle fidelity through an AI-human hybrid scoring mechanism and a trust management mechanism. Finally, based on the user's selected hairstyle, it comprehensively considers multiple dimensions such as distance, fidelity, price, and user ratings, generating a personalized hairdresser recommendation list through an adaptive ranking algorithm. This invention significantly improves hairstyle recognition accuracy through multimodal feature fusion, balances efficiency and accuracy through a hybrid scoring mechanism, and enhances data reliability through a trust mechanism. This results in a significant increase in user click-through rate and appointment conversion rate, providing users with accurate hairdresser recommendation services.

Description

Technical Field

[0001] This invention relates to the field of mobile internet technology, specifically to an intelligent recommendation system and method for hairdressers based on multimodal hairstyle feature analysis and dynamic skill scoring. Background Technology

[0002] As living standards improve, people are paying increasing attention to their personal image and hairstyle. Currently, consumers primarily find hairdressers through search engines, map applications, social media recommendations, or friend referrals. However, existing technologies have the following problems: First, there is a lack of a precise matching mechanism between hairstyles and hairdressers' skills. Existing hair salon search services only provide basic information about the salons (such as location, price, and user reviews), leaving consumers unable to determine whether the hairdresser at that salon is skilled in the specific hairstyle they want. Consumers need to spend a significant amount of time browsing hairdressers' portfolios and find it difficult to objectively evaluate a hairdresser's mastery of a particular hairstyle.

[0003] Second, hairstyle recognition and evaluation lacks automation. Currently, hairstyle evaluation mainly relies on subjective human judgment, which is inefficient and inconsistent. Although existing image recognition technologies can be applied to fields such as facial recognition and product recognition, hairstyle recognition has its own unique characteristics: hairstyles have complex structural layers, diverse texture variations, and rich color combinations, and are easily affected by factors such as shooting angle, lighting conditions, and occlusion, making it difficult for general image recognition algorithms to achieve ideal recognition results.

[0004] Third, the data sources for evaluating hairdressers' skills are singular and their reliability is difficult to guarantee. Current platform reviews primarily rely on subjective ratings from users after booking appointments. These ratings often have little correlation with the hairdresser's actual skill level and are easily influenced by factors such as service attitude and shop environment. Furthermore, there is a lack of detailed evaluations of the skill differences among hairdressers for different hairstyle types.

[0005] Fourth, recommendation algorithms lack personalization due to their limited dimensionality. Existing merchant recommendation algorithms are mostly based on geographical location or simple rating ranking, failing to comprehensively consider multiple dimensions such as distance, skill matching, price, and user preferences, and also unable to personalize the algorithm based on users' historical behavior.

[0006] Therefore, there is an urgent need for a system and method that can automatically identify hairstyle features, accurately match hairdresser skills, and provide personalized recommendations. Summary of the Invention

[0007] The present invention aims to solve the above-mentioned technical problems and provides a hairdresser intelligent recommendation system and method based on multimodal hairstyle feature analysis and dynamic skill scoring. By combining AI technology with manual review, it can achieve accurate matching between hairstyles and hairdresser skills and provide personalized hairdresser recommendation services.

[0008] To achieve the above objectives, the present invention provides the following technical solution: Technical Solution 1: A hairdresser intelligent recommendation method based on hairstyle feature matching, comprising the following steps: S1: Construct a multimodal feature database of hairstyles, including: S1.1: Collect hairstyle sample images and use a multimodal deep neural network to extract texture features, structural features and color features respectively; S1.2: The features of the three modalities are fused into a hairstyle feature vector through an adaptive feature fusion layer; S1.3: Store the hairstyle feature vector in the hairstyle database and associate it with the hairstyle ID.

[0009] S2: Construct a barber skills assessment model, including: S2.1: Receive the work data uploaded by the hairdresser, which includes data from both user uploads and platform uploads; S2.2: For user-uploaded works, the hairstyle type is automatically identified through a hairstyle recognition algorithm, and the similarity with the target hairstyle is calculated; S2.3: For works uploaded to the platform, after being reviewed by human experts, the hairstyle type will be marked and an authoritative skill score will be assigned; S2.4: A confidence-driven hybrid scoring mechanism is used to calculate the final fidelity score of the work.

[0010] S3: Receive the user's hairstyle selection request and obtain the selected hairstyle ID and the user's geographical location.

[0011] S4: Perform barber matching and sorting, including: S4.1: Query the set of hairdressers in the hairdresser skills database who can handle this hairstyle; S4.2: Calculate the recommendation score for each barber using the following formula: Score = α×D + β×R + γ×P + δ×U Where: D is the distance normalized score, R is the hairstyle reproduction score, P is the price normalized score, and U is the user rating normalized score; α, β, γ, and δ are adjustable weighting coefficients, and α+β+γ+δ=1; S4.3: Adjust the weight coefficients according to the sorting type selected by the user (default sorting, distance priority, or fidelity priority); S4.4: Sort the recommendations in descending order of score and generate a recommendation list.

[0012] S5: Return the recommendation list to the user's terminal for display.

[0013] Furthermore, the multimodal deep neural network in step S1.1 includes: - TextureNet: Employs a convolutional neural network (CNN) to extract texture details of the hairstyle, such as the curl and layering of the hair strands; - StructureNet: A neural network that uses an attention mechanism to extract the overall structural features of a hairstyle, such as the hairstyle outline and length ratio; - Color Encoder: Uses an encoder-decoder structure to extract the color distribution features of the hairstyle, including the main color tone, gradient effects, etc.

[0014] Furthermore, the adaptive feature fusion layer in step S1.2 employs an attention mechanism, dynamically allocating weights based on the importance of features from different modalities. The fusion formula is as follows: Feature_fused = Attention(Texture, Structure, Color) ⊙ [Texture;Structure; Color] Where [;] represents feature concatenation, ⊙ represents element-wise multiplication, and Attention is the attention weight matrix.

[0015] Furthermore, the confidence-driven hybrid scoring mechanism in step S2.4 includes: S2.4.1: Use a hairstyle recognition algorithm to calculate the AI ​​similarity Sim_AI between the image of the work and the target hairstyle, and output the AI ​​confidence Conf_AI, where Conf_AI∈[0,1]; S2.4.2: Determine the scoring strategy based on confidence level: If Conf_AI ≥ T1 (T1=0.85), then the AI ​​score is used directly; If Conf_AI ∈ [T2, T1) (T2=0.6), then manual review is performed and a mixed score is calculated; If Conf_AI < T2, then manual scoring will be used; S2.4.3: The formula for calculating the mixed score is: Score_final = λ×Score_AI + (1-λ)×Score_human Where λ is the dynamic weight, λ = (Conf_AI - T2) / (T1 - T2), λ=1 when Conf_AI ≥ T1, and λ=0 when Conf_AI < T2.

[0016] Furthermore, step S2 also includes establishing a barber trust model, including: S2.5: Calculate the hairdresser's trust score (Trust_Score) based on the consistency between platform ratings and user-uploaded ratings; Consistency = 1 - |Avg_User_Score - Platform_Score| S2.6: The trust score is dynamically updated using the exponential moving average algorithm; Trust_new = α×Consistency + (1-α)×Trust_old Where α is the learning rate, with a value ranging from [0.05, 0.2]. S2.7: The barber's original skill rating is weighted and adjusted based on the trust score, using the following formula: Score_adjusted = Score_raw×(0.6 + 0.4×Trust_Score).

[0017] Furthermore, the distance normalization score in step S4.2 is calculated using a piecewise function: If distance ≤ user_accept_range: D = 1 - distance / user_accept_range If distance > user_accept_range: D = exp(-(distance - user_accept_range) / 5.0) The user_accept_range is the acceptable distance range for the user, with a default value of 5km.

[0018] Furthermore, the weight adjustment rule in step S4.3 is as follows: - Default sorting: α=0.3, β=0.4, γ=0.1, δ=0.2 - Distance first: α=0.7, β=0.2, γ=0.05, δ=0.05 - Prioritize reduction degree: α=0.1, β=0.7, γ=0.1, δ=0.1 Furthermore, the method also includes a model iterative optimization step: S6: Continuously collect user behavior data (including clicks, appointments, reviews, etc.) and hairdresser work update data; S7: Employs an online learning mechanism to update user preference models and ranking weights in real time; Weight_new = Weight_old + η×∇Loss Where η is the online learning rate, and ∇Loss is the gradient of the loss function; S8: Regularly (weekly or monthly) retrain the model offline using accumulated data; S9: Verify the effectiveness of the new model through A / B testing. If the performance improvement is significant, proceed with a phased rollout. Performance improvement metrics include: click-through rate, appointment conversion rate, and user satisfaction; A new model is considered to have significantly improved when at least two of its metrics increase by more than 5% and the p-value is less than 0.05.

[0019] Technical Solution 2: A hairdresser intelligent recommendation system based on hairstyle feature matching, comprising: Hairstyle Feature Database Module: Used to store hairstyle sample images and their multimodal feature vectors, which are obtained by fusing texture features, structural features and color features; Hairdresser Skills Database Module: This module stores hairdresser information, work data, and skill scores. The skill scores are calculated using an AI-human hybrid scoring mechanism. Hairstyle recognition and matching module: This module receives the hairstyle ID selected by the user, extracts the corresponding hairstyle feature vector, and matches it against a set of hairdressers in the hairdresser skill database who can handle that hairstyle. Recommendation ranking engine: used to calculate the recommendation score for each barber, supports multiple ranking dimensions (default ranking, distance priority, and fidelity priority), and can be personalized according to user preferences; User interaction module: used to display a list of hairstyles, hairstyle details and a list of recommended hairdressers to users, and to receive user selection and sorting operations; Model Iteration and Optimization Module: This module collects user behavior data and continuously optimizes the recommendation model through a combination of online learning and offline training.

[0020] Furthermore, the system also includes a trust management module, used for: - Calculate the hairdresser trust score based on the consistency between platform ratings and user-uploaded ratings; - Dynamically updates the hairdresser's trust score; - Adjust the weight of the hairdresser's skill rating based on the trust score.

[0021] Furthermore, the system also includes a data acquisition module for: - Receive user-uploaded hairdressing works and perform automatic hairstyle recognition and preliminary scoring; - Receive the hairdresser's works uploaded to the platform and forward them to human experts for review; - Cross-validate data from different sources to ensure data reliability.

[0022] The beneficial effects of this invention are: First, the multimodal hairstyle feature fusion technology significantly improves the accuracy of hairstyle recognition. Compared with single feature extraction methods, the multimodal fusion method of this invention organically combines texture, structure, and color features, improving the recognition accuracy by more than 35%, and can accurately identify complex hairstyles (such as layered cuts, perms, and dyed hair).

[0023] Secondly, the confidence-driven AI-human hybrid scoring mechanism significantly improves scoring efficiency while ensuring accuracy. Works with high confidence are scored directly by AI, while those with low confidence undergo human review. Compared to purely human scoring, this method improves efficiency by over 60%, while maintaining accuracy comparable to human scoring.

[0024] Third, the reliability of hairdresser skill data is effectively improved through dual data source verification and a trust mechanism. Authoritative ratings uploaded to the platform are cross-verified with AI ratings uploaded by users, and a dynamic trust adjustment mechanism further enhances the detection rate of fake ratings by over 80%.

[0025] Fourth, through a multi-dimensional adaptive sorting algorithm, it provides flexible sorting dimension selection and makes personalized adjustments based on users' historical behavior, resulting in a click-through rate increase of over 25% and an appointment conversion rate increase of over 20%.

[0026] Fifth, continuous model optimization is achieved through an incremental model iteration optimization system. Online learning ensures that the model can quickly adapt to changes in user behavior, while offline training ensures that the model fully utilizes all data for learning, resulting in an average monthly performance improvement of 5-10%. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the intelligent hairdresser recommendation method provided in an embodiment of the present invention; Figure 2 This is a diagram of the multimodal hairstyle feature extraction network architecture provided in an embodiment of the present invention; Figure 3 A flowchart of the AI-human hybrid scoring mechanism provided in this embodiment of the invention; Figure 4This is a flowchart of the barber trust model update process provided in an embodiment of the present invention; Figure 5 This is a diagram of the recommendation ranking engine architecture provided in an embodiment of the present invention; Figure 6 This is a system architecture diagram for model iterative optimization provided in an embodiment of the present invention; Figure 7 This is an overall architecture diagram of the intelligent hairdresser recommendation system provided in an embodiment of the present invention. Detailed Implementation

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

[0029] Example 1: Intelligent Hairdresser Recommendation Method like Figure 1 As shown, this embodiment provides a method for intelligent hairdresser recommendation based on hairstyle feature matching, including the following steps: Step S1: Construct a multimodal feature database of hairstyles The goal of this step is to build a comprehensive hairstyle database, with each hairstyle corresponding to a high-precision feature vector.

[0030] 1.1 Hairstyle Sample Collection We collected sample images of over 5000 common hairstyles, covering various types including short, medium-length, long, straight, curly, layered, permed, and dyed hair. For each hairstyle, we collected image samples from multiple angles and under various lighting conditions to ensure data diversity.

[0031] 1.2 Multimodal Feature Extraction like Figure 2 As shown, a multimodal deep neural network is used to extract the texture, structure, and color features of the hairstyle.

[0032] The TextureNet texture extraction network employs a convolutional neural network architecture, comprising five convolutional layers and three pooling layers, ultimately outputting a 256-dimensional texture feature vector. This network is capable of capturing the microscopic details of hairstyles, such as the curl, layering, and texture of hair strands.

[0033] The StructureNet network, a neural network employing an attention mechanism, captures the overall structural features of a hairstyle through self-attention, outputting a 256-dimensional structural feature vector. This network focuses on macroscopic features such as the hairstyle's outline, length ratio, and layer distribution.

[0034] The ColorEncoder employs an encoder-decoder architecture, extracting the color distribution features of the hairstyle through convolutional and upsampling layers, and outputting a 128-dimensional color feature vector, including information such as the main hue, gradient effect, and color uniformity.

[0035] 1.3 Adaptive Feature Fusion An adaptive feature fusion layer employing an attention mechanism dynamically assigns weights based on the importance of features from different modalities. The fusion formula is as follows: Feature_fused = Attention(Texture, Structure, Color) ⊙ [Texture;Structure; Color] Where [;] represents feature concatenation, ⊙ represents element-wise multiplication, and Attention is the attention weight matrix, which is learned through training. The fused feature vector has a dimension of 640 (256+256+128).

[0036] 1.4 Feature Vector Storage The fused feature vectors for each hairstyle are stored in a hairstyle database, and a mapping relationship between hairstyle IDs and feature vectors is established. The hairstyle database is implemented using a vector database (such as Milvus), which supports efficient similarity retrieval.

[0037] Step S2: Construct a hairdresser skills assessment model The goal of this step is to assess each hairdresser's mastery of different hairstyles.

[0038] 2.1 Receiving Artwork Data We accept work data from two sources: (1) User upload: Users upload photos of their hairstyles after getting a haircut and include information about the hairdresser; (2) Platform upload: Hairdressers actively upload their representative works.

[0039] For user-uploaded works, the system automatically identifies the hairdresser's identity (either selected by the user or confirmed by the platform); for works uploaded to the platform, the system conducts a preliminary review to ensure the works are authentic and valid.

[0040] 2.2 Automatic recognition and rating of user-uploaded works For user-uploaded works, the hairstyle type is automatically identified using a hairstyle recognition algorithm: (1) Extract the feature vector of the artwork image: Use the multimodal feature extraction network trained in step S1; (2) Feature vector retrieval: Perform nearest neighbor retrieval in the hairstyle database to find the K most similar hairstyles (K=5); (3) Calculate similarity: Calculate the cosine similarity between the feature vector of the artwork and the feature vector of the target hairstyle; Sim_AI = cosine_similarity(Feature_work, Feature_target) (4) Calculate confidence: Evaluate the confidence of AI recognition based on the distribution of search results; Conf_AI = 1 - std([Sim1, Sim2, ..., SimK]) Where std is the standard deviation, the smaller the standard deviation, the more certain the identification.

[0041] 2.3 Manual review of uploaded works on the platform The works uploaded to the platform are manually reviewed by professional hairstylists. (1) Verify the authenticity of the work: confirm that the work is the actual work of the hairdresser; (2) Label the hairstyle type: Label the hairstyle ID corresponding to the work based on professional judgment; (3) Skill score: Based on the similarity between the work and the target hairstyle, a skill score of 0-100 is given.

[0042] Works reviewed by humans serve as authoritative data sources for verifying and correcting AI recognition results.

[0043] 2.4 Confidence-Driven Hybrid Rating like Figure 3 As shown, the scoring strategy is dynamically selected based on the confidence level of AI recognition: (1) High confidence level (Conf_AI ≥ 0.85): AI similarity is directly used as the restoration score; Score_final = Sim_AI (2) Moderate confidence level (0.6 ≤ Conf_AI < 0.85): Manual review and calculation of mixed score; Score_final = 0.7×Sim_AI + 0.3×Score_human (3) Low confidence level (Conf_AI < 0.6): Purely manual scoring was used; Score_final = Score_human Score_human is the score given by human review (normalized to [0,1]).

[0044] 2.5 Barber Trust Model like Figure 4 As shown, a barber trust model is established to assess the reliability of barber data.

[0045] For each barber, maintain a trust score (initially 0.7). Each time new rating data arrives, calculate consistency and update the trust score: (1) Calculation consistency: Consistency = 1 - |Avg_User_Score - Platform_Score| / 100 Where Avg_User_Score is the average score of user-uploaded works, and Platform_Score is the average score of works reviewed by the platform.

[0046] (2) Update trust level: Trust_new = 0.1×Consistency + 0.9×Trust_old (3) Adjust skill rating: Score_adjusted = Score_raw×(0.6 + 0.4×Trust_Score) Barbers with high trust scores receive a higher weight in skill ratings when recommended; barbers with low trust scores have their skill ratings reduced.

[0047] 2.6 Skills Rating Aggregation For each barber and each hairstyle, calculate the barber's overall skill score for that hairstyle: Skill_Score(hairdresser, hairstyle) = Σ(Score_adjusted_i) / N Where N represents the number of works created by the hairdresser for that hairstyle.

[0048] Step S3: Receive user's hairstyle selection request Users browse the hairstyle library on the app and select their favorite hairstyle. (System data retrieved) (1) Hairstyle ID: The identifier of the hairstyle selected by the user; (2) User location: The user's current geographical location (longitude and latitude) is obtained through GPS.

[0049] Step S4: Barber Matching and Sorting 4.1 Barber Search Search the barber skills database for the set of barbers who can handle this hairstyle: Hairdresser_Set = {h | Skill_Score(h, hairstyle) ≥ Threshold} Threshold is the skill rating threshold (default 0.5).

[0050] 4.2 Multi-dimensional scoring calculation like Figure 5 As shown, calculate the recommendation score for each barber: (1) Distance normalized fraction D: Obtain the location of the barbershop and calculate the straight-line distance between it and the user; If distance ≤ 5km: D = 1 - distance / 5 If the distance is greater than 5km: D = exp(-(distance - 5) / 5) (2) Hairstyle reproduction score R: R = Skill_Score(hairdresser, hairstyle) (3) Price normalization fraction P: The score is calculated based on the hairdresser's price for the haircut; the lower the price, the higher the score. P = (Max_Price - Price) / (Max_Price - Min_Price) Max_Price and Min_Price represent the maximum and minimum prices for that region.

[0051] (4) Normalized score of user rating U: U = User_Rating / 5 User_Rating is the overall user rating (1-5 points) for the hairdresser.

[0052] (5) Overall Recommendation Score: Score = α×D + β×R + γ×P + δ×U Among them α+β+γ+δ=1, default weight: α=0.3, β=0.4, γ=0.1, δ=0.2 4.3 Sorting Weight Adjustment Adjust weights based on the sorting type selected by the user: 1) Default sorting: α=0.3, β=0.4, γ=0.1, δ=0.2 2) Distance priority: α=0.7, β=0.2, γ=0.05, δ=0.05 3) Prioritize reduction degree: α=0.1, β=0.7, γ=0.1, δ=0.1 In addition, personalized adjustments are made based on users' historical behavior: If users frequently choose barbers who are nearby, increase the alpha weight. If users frequently choose barbers with high fidelity to their image, increase the beta weight.

[0053] 4.4 Generating a Recommendation List Sort the hairdressers in descending order of their recommendation scores and select the top 20 to generate a recommendation list. The recommendation list returned to the user includes: 1) Basic information of the hairdresser (photo, name, shop name) 2) Skill rating and accuracy of reproduction 3) Distance information 4) Price Information 5) User ratings 6) Showcase of representative works Step S5: Display the recommendation list The user terminal receives the recommendation list and displays it on the interface. Users can: 1) View barber details 2) View the hairdresser's portfolio 3) Click to book or add to favorites 4) Switch sorting dimensions Steps S6-S9: Iterative Model Optimization like Figure 6 As shown, an incremental model iterative optimization system is established.

[0054] S6: Data Collection Continuously collect user behavior data: 1) Click data: User click behavior when viewing hairdresser details 2) Appointment data: The hairdressers actually booked by the user. 3) Evaluation data: User feedback after getting a haircut 4) Dwell Time: How long a user stays on the page. Collect updated data on hairdressers' work: New works manually uploaded by the hairdresser New works uploaded by users S7: Online Learning The model is updated online using mini-batch gradient descent: 1) Constructing the loss function: Constructing a ranking loss based on user clicks and reservation behavior; 2) Calculate the gradient: Calculate the gradient using the 100 most recent behavioral data points; 3) Parameter update: Weight_new = Weight_old + η×∇Loss Where η = 0.01 is the online learning rate.

[0055] Online learning ensures that the model can quickly adapt to changes in user preferences.

[0056] S8: Offline Retraining Perform a full offline model retraining once a week: 1) Data cleaning: Removing abnormal and duplicate data; 2) Data augmentation: Enhance hairstyle images by rotating, cropping, and changing colors; 3) Model training: Train a new model from scratch using all historical data; 4) Performance evaluation: Evaluate the model performance on the validation set.

[0057] S9: A / B Testing and Release After the new model is trained, perform A / B testing: 1) Traffic allocation: Allocate 10% of users to the new model and 90% of users to the old model; 2) Metrics Collection: Collect metrics such as click-through rate, appointment conversion rate, and user satisfaction; 3) Statistical test: Use t-test to determine whether the performance improvement is significant (p<0.05); 4) Canary release: If performance is significantly improved, gradually increase the traffic of the new model (10%→50%→100%). 5) Rollback mechanism: If the new model is abnormal, it will be rolled back to the old model immediately.

[0058] Through incremental model iteration optimization, the model performance improves by an average of 5-10% per month.

[0059] Example 2: Intelligent Hairdresser Recommendation System like Figure 7 As shown, this embodiment provides an intelligent hairdresser recommendation system based on hairstyle feature matching, including the following modules: 1. Hairstyle Feature Database Module The hairstyle feature database module is implemented using the Milvus vector database, storing feature vectors for over 5000 hairstyles. This module provides the following functions: 1) Hairstyle feature vector storage: Stores 640-dimensional hairstyle feature vectors; 2) Similarity retrieval: Quickly retrieve the most similar hairstyles based on cosine similarity; 3) Feature vector update: Supports online updates of hairstyle features.

[0060] 2. Hairdresser Skills Database Module The barber skills database module is implemented using a PostgreSQL relational database, storing barber information, work data, and skill ratings. This module provides the following functionalities: 1) Hairdresser Information Management: Stores hairdresser ID, name, shop location, price, user rating, etc.; 2) Work Data Management: Stores work images, hairstyle IDs, rating sources, timestamps, etc. 3) Skills Rating Management: Stores the skills ratings for each hairdresser for each hairstyle; 4) Trust Management: Store the barber's trust score.

[0061] 3. Hairstyle Recognition and Matching Module The hairstyle recognition and matching module is implemented using Python and TensorFlow, and provides the following functions: 1) Hairstyle feature extraction: Load the pre-trained multimodal feature extraction network to extract hairstyle feature vectors; 2) Hairstyle matching: Matching a set of hairdressers based on feature vectors; 3) Confidence estimation: Evaluate the confidence level of AI recognition.

[0062] 4. Recommendation Ranking Engine The recommendation ranking engine is implemented using Python and FastAPI, and provides the following features: 1) Multi-dimensional rating calculation: Calculate scores for four dimensions: distance, fidelity, price, and user rating; 2) Weight Adjustment: Dynamically adjust weights based on sorting type and user preferences; 3) Sorting algorithm: Sort in descending order of comprehensive score; 4) Personalized adjustment: Personalize recommendation results based on user's historical behavior.

[0063] 5. User Interaction Module The user interaction module is implemented in the mobile app using React Native and provides the following features: 1) Hairstyle Gallery Display: Showcases thumbnails and names of various hairstyles; 2) Hairstyle Details: Showcases detailed information and reference images of the hairstyle; 3) Recommended List: Displays a list of recommended hairdressers, with support for switching sorting dimensions; 4) Hairdresser Details: Displays detailed information about the hairdresser, their portfolio, ratings, etc. 5) Appointment function: Allows users to book hairdressers.

[0064] 6. Model Iterative Optimization Module The model iterative optimization module includes an online learning engine and an offline training engine: 1) Online learning engine: Receives user behavior data in real time and updates model parameters using mini-batch gradient descent; 2) Offline training engine: Full retraining is performed weekly using accumulated data to generate new models; 3) A / B testing framework: Perform A / B testing on the new model to evaluate performance improvements; 4) Canary release system: Controls the traffic allocation of new models and supports gradual release and rollback.

[0065] 7. Trust Management Module The trust management module provides the following functions: 1) Consistency Calculation: Calculate the consistency between the platform rating and the user-uploaded rating; 2) Trust Level Update: The trust level of the hairdresser is updated using an exponential moving average. 3) Rating adjustment: Adjust the weight of the hairdresser's skill rating based on the level of trust.

[0066] 8. Data Acquisition Module The data acquisition module provides the following functions: 1) Receiving User Submissions: Receiving user-uploaded hairdresser submissions; 2) Platform submissions: We accept representative works uploaded by hairdressers; 3) Data validation: Cross-validate data from different sources; 4) Data storage: Store the verified data in the corresponding database.

[0067] System Deployment Architecture The system in this embodiment is deployed using a cloud-native architecture, as detailed below: 1) Mobile App: Developed using React Native and deployed on iOS and Android app stores; 2) Front-end server: Deployed using Nginx+CDN to handle static resource requests; 3) Business Server: Deployed using FastAPI, it handles business logic requests and supports elastic scaling; 4) AI Inference Server: Deployed using TensorFlow Serving, providing hairstyle recognition and feature extraction services; 5) Database: Deployed on a cloud database service using PostgreSQL and MongoDB; 6) Vector Database: Deployed using Milvus, providing hairstyle feature vector retrieval services; 7) Caching: Redis is used for deployment to cache recommendation results and session data.

[0068] System performance indicators The system in this embodiment has the following performance indicators: 1) Hairstyle recognition accuracy: ≥92% (35% improvement compared to single-modal methods) 2) Hairstyle recognition response time: ≤200ms 3) Recommendation list generation time: ≤500ms 4) System concurrency support: ≥10000 QPS 5) User click-through rate: ≥15% (25% improvement over the benchmark) 6) Appointment conversion rate: ≥5% (20% improvement over the benchmark) 7) Scoring efficiency: More than 60% higher than purely manual scoring. The above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligently recommending hairdressers based on hairstyle feature matching, characterized in that, Includes the following steps: S1: Construct a multimodal feature database of hairstyles, including: S1.1: Collect hairstyle sample images and use a multimodal deep neural network to extract texture features, structural features and color features respectively; S1.2: The features of the three modalities are fused into a hairstyle feature vector through an adaptive feature fusion layer; S1.3: Store the hairstyle feature vector in the hairstyle database and associate it with the hairstyle ID; S2: Construct a barber skills assessment model, including: S2.1: Receive the work data uploaded by the hairdresser, which includes data from both user uploads and platform uploads; S2.2: For user-uploaded works, the hairstyle type is automatically identified through a hairstyle recognition algorithm, and the similarity with the target hairstyle is calculated; S2.3: For works uploaded to the platform, after being reviewed by human experts, the hairstyle type will be marked and an authoritative skill score will be assigned; S2.4: A confidence-driven hybrid scoring mechanism is used to calculate the final fidelity score of the work; S3: Receive the user's hairstyle selection request, and obtain the selected hairstyle ID and the user's geographical location; S4: Perform barber matching and sorting, including: S4.1: Query the set of hairdressers in the hairdresser skills database who can handle this hairstyle; S4.2: Calculate the recommendation score for each barber using the following formula: Score = α×D + β×R + γ×P + δ×U Where: D is the distance normalized score, R is the hairstyle reproduction score, P is the price normalized score, and U is the user rating normalized score; α, β, γ, and δ are adjustable weighting coefficients, and α+β+γ+δ=1; S4.3: Adjust the weight coefficients according to the sorting type selected by the user; S4.4: Sort the recommendations in descending order of score and generate a recommendation list; S5: Return the recommendation list to the user's terminal for display.

2. The method according to claim 1, characterized in that, The multimodal deep neural network in step S1.1 includes a texture extraction network, a structure extraction network, and a color encoder, wherein: The texture extraction network uses a convolutional neural network to extract the texture details of the hairstyle; The structure extraction network uses an attention-based neural network to extract the overall structural features of the hairstyle; The color encoder uses an encoder-decoder structure to extract the color distribution features of the hairstyle.

3. The method according to claim 1, characterized in that, The adaptive feature fusion layer in step S1.2 employs an attention mechanism, dynamically allocating weights based on the importance of features from different modalities. The fusion formula is as follows: Feature_fused = Attention(Texture, Structure, Color) ⊙ [Texture;Structure; Color] Where [;] represents feature concatenation, ⊙ represents element-wise multiplication, and Attention is the attention weight matrix.

4. The method according to claim 1, characterized in that, The confidence-driven hybrid scoring mechanism in step S2.4 includes: S2.4.1: Use a hairstyle recognition algorithm to calculate the AI ​​similarity Sim_AI between the image of the work and the target hairstyle, and output the AI ​​confidence Conf_AI, where Conf_AI∈[0,1]; S2.4.2: Determine the scoring strategy based on confidence level: If Conf_AI ≥ T1, then the AI ​​score is used directly; If Conf_AI ∈ [T2, T1), then manual review is performed and a mixed score is calculated; If Conf_AI < T2, then manual scoring will be used; Where T1 and T2 are preset thresholds, 0 < T2 < T1 < 1; S2.4.3: The formula for calculating the mixed score is: Score_final = λ×Score_AI + (1-λ)×Score_human Where λ is the dynamic weight, λ=1 when Conf_AI ≥ T1, λ=(Conf_AI-T2) / (T1-T2) when Conf_AI ∈ [T2, T1), and λ=0 when Conf_AI < T2.

5. The method according to claim 1, characterized in that, Step S2 further includes establishing a barber trust model, including: S2.5: Calculate the hairdresser's trust score (Trust_Score) based on the consistency between platform ratings and user-uploaded ratings; Consistency = 1 - |Avg_User_Score - Platform_Score| S2.6: The trust score is dynamically updated using the exponential moving average algorithm; Trust_new = α×Consistency + (1-α)×Trust_old Where α is the learning rate, with a value ranging from [0.05, 0.2]. S2.7: The barber's original skill rating is weighted and adjusted based on the trust score, using the following formula: Score_adjusted = Score_raw×(0.6 + 0.4×Trust_Score).

6. The method according to claim 1, characterized in that, The distance normalization score in step S4.2 is calculated using a piecewise function: If distance ≤ user_accept_range: D = 1 - distance / user_accept_range If distance > user_accept_range: D = exp(-(distance - user_accept_range) / 5.0) Where user_accept_range is the acceptable distance range for the user.

7. The method according to claim 1, characterized in that, The weight adjustment rule in step S4.3 is as follows: Default sorting: α=0.3, β=0.4, γ=0.1, δ=0.2 Distance first: α=0.7, β=0.2, γ=0.05, δ=0.05 Prioritize reduction degree: α=0.1, β=0.7, γ=0.1, δ=0.

1.

8. The method according to claim 1, characterized in that, It also includes model iterative optimization steps: S6: Continuously collect user behavior data and hairdresser work update data; S7: Employs an online learning mechanism to update user preference models and ranking weights in real time; Weight_new = Weight_old + η×∇Loss Where η is the online learning rate, and ∇Loss is the gradient of the loss function; S8: Regularly use accumulated data to perform full offline model retraining; S9: Verify the effectiveness of the new model through A / B testing. If the performance improvement is significant, proceed with a phased rollout.

9. A hairdresser intelligent recommendation system based on hairstyle feature matching, characterized in that, include: Hairstyle Feature Database Module: Used to store hairstyle sample images and their multimodal feature vectors, which are obtained by fusing texture features, structural features and color features; Hairdresser Skills Database Module: This module stores hairdresser information, work data, and skill scores. The skill scores are calculated using an AI-human hybrid scoring mechanism. Hairstyle recognition and matching module: This module receives the hairstyle ID selected by the user, extracts the corresponding hairstyle feature vector, and matches it against a set of hairdressers in the hairdresser skill database who can handle that hairstyle. Recommendation ranking engine: used to calculate the recommendation score for each barber, supports multiple ranking dimensions, and can be personalized according to user preferences; User interaction module: used to display a list of hairstyles, hairstyle details and a list of recommended hairdressers to users, and to receive user selection and sorting operations; Model Iteration and Optimization Module: This module collects user behavior data and continuously optimizes the recommendation model through a combination of online learning and offline training.

10. The system according to claim 9, characterized in that, It also includes a trust management module, used for: The hairdresser's trust score is calculated based on the consistency between platform ratings and user-uploaded ratings. Dynamically update the barber's trust score; adjust the weight of the barber's skill rating based on the trust score.

11. The system according to claim 9, characterized in that, It also includes a data acquisition module, used for: It receives hairdresser works uploaded by users and performs automatic hairstyle recognition and preliminary scoring; Receive the hairdresser's works uploaded to the platform and forward them to human experts for review; Cross-validate data from different sources to ensure data reliability.

Images