Deep learning-based skin soothing efficacy evaluation method and system

Abstract

The invention relates to the technical field of computer vision and artificial intelligence, and discloses a skin soothing efficacy evaluation method and system based on deep learning, and the method comprises the steps: carrying out the sensitive hierarchical data preprocessing, and dividing the skin into different hierarchies according to the sensitivity; constructing a hierarchical contrast learning framework, and learning effective feature representation from the unlabeled data; hierarchical knowledge distillation flow construction is carried out to realize knowledge transmission from low sensitivity to high sensitivity; sparse annotation reinforcement learning is carried out, and limited annotation data are utilized to the maximum extent through active learning and semi-supervised learning; according to the method, the problem of individualized soothing evaluation under the condition of data scarcity is solved, accurate evaluation of people with different sensitivities is realized, evaluation accuracy and data utilization efficiency are improved, labeling cost is reduced, and reliable technical support is provided for research and development of skin care products and formulation of individualized skin care schemes.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and artificial intelligence technology, and more specifically, to a skin soothing efficacy evaluation method and system based on deep learning. Background Art

[0002] As people become more aware of skin health, skin soothing efficacy evaluation technology is increasingly being used in cosmetics research and development, dermatology, and personal care. Traditional skin soothing efficacy evaluation relies primarily on subjective expert scoring or simple instrumental measurements. These methods often struggle to objectively quantify the soothing effect, and the evaluation results are significantly influenced by the evaluator's experience and environmental factors. With the development of computer vision and artificial intelligence technologies, image analysis-based skin evaluation methods have gradually emerged. These methods assess soothing effects by capturing microscopic changes on the skin surface. However, these methods typically use uniform evaluation criteria and fail to fully account for individual differences in people with different sensitivities.

[0003] In recent years, deep learning technology has made significant progress in the field of medical image analysis, providing a new technical path for the accurate evaluation of skin soothing effects. Skin image analysis methods based on deep neural networks can automatically extract skin features and establish a mapping relationship between features and soothing effects through model training. However, such methods face the challenge of data scarcity in practical applications, especially for skin data of highly sensitive populations. Due to the difficulty in collecting data and the small sample size, the model training effect is poor. In addition, existing deep learning evaluation methods often mix data from people with different sensitivities and fail to fully utilize the knowledge associations between different levels, resulting in insufficient accuracy and pertinence in the evaluation results.

[0004] In terms of knowledge transfer and model optimization, techniques such as knowledge distillation and contrastive learning offer potential solutions to data scarcity. Knowledge distillation transfers knowledge from complex models to simpler ones, achieving model compression and knowledge transfer. Contrastive learning, on the other hand, can learn effective feature representations from unlabeled data. However, the application of these techniques in the field of skin soothing efficacy evaluation is still insufficient. In particular, there is a lack of hierarchical learning frameworks tailored to the characteristics of different sensitivity groups, making it impossible to effectively address the data scarcity and knowledge sharing issues in personalized soothing evaluation.

[0005] Therefore, a technical solution is needed that can adapt to the characteristics of people with different sensitivities and provide accurate evaluation under conditions of data scarcity. Summary of the Invention

[0006] The present invention provides a skin soothing efficacy evaluation method and system based on deep learning, which solves the problem in related technologies of how to build an accurate skin soothing efficacy evaluation system based on individual differences of people with different sensitivities when data is scarce, while overcoming the problems in existing technologies of ignoring individual differences, redundant model parameters and insufficient cross-level knowledge sharing.

[0007] The present invention provides a skin soothing efficacy evaluation method based on deep learning, comprising the following steps: Perform skin sensitivity stratification data preprocessing, including collecting skin image data, extracting sensitivity assessment indicators, and using the sensitivity stratification neural network model to divide the user's skin into different sensitivity levels; Based on the stratified skin data, a hierarchical contrastive learning framework is constructed. Specific data enhancement strategies are designed for each sensitivity level, and a multi-view contrastive learning network is constructed to form a hierarchical specific representation space. By utilizing the layer-specific representation space, we construct a hierarchical knowledge distillation flow, achieve knowledge transfer from low sensitivity to high sensitivity through the inter-layer knowledge distillation network, use the hierarchical knowledge transfer mechanism to fuse knowledge from different layers, and dynamically adjust the adaptive weights; Combining the output of the knowledge distillation flow, we build a sparse annotation reinforcement learning module to achieve sparse annotation sample selection, semi-supervised learning framework construction, and annotation efficiency optimization. Based on the trained model, individualized relief evaluation results are generated, including calculation of multi-dimensional relief indicators, identification of sensitive characteristic markers and generation of personalized evaluation reports.

[0008] In a preferred embodiment, in the step of preprocessing the skin sensitivity stratification data, the sensitivity evaluation index includes: Skin barrier integrity indicators, calculated by image analysis of transepidermal water loss and stratum corneum hydration; Vascular reactivity index, skin erythema index and vascular dilation degree were calculated by image analysis; Inflammatory markers, which identify and quantify inflammation-related features through image analysis; Sensitivity threshold index, by analyzing the speed and intensity of the skin's response to external stimuli; The sensitivity hierarchical neural network model adopts a multi-layer perceptron structure, with the input being the combined feature vector of sensitivity evaluation indicators and the output being the skin sensitivity hierarchical classification result.

[0009] In a preferred embodiment, in the hierarchical contrastive learning framework, the data augmentation strategies designed for different sensitivity levels include: For low-sensitivity levels, stronger color transformation, geometric transformation, and noise addition are used; At the medium sensitivity level, medium-intensity color and geometric transformations are applied to preserve skin texture details; For high-sensitivity levels, slight color adjustments and local area enhancements are used to preserve sensitivity characteristics to the greatest extent possible; The multi-view contrastive learning network adopts a dual-tower structure, consisting of an encoder network and a projection head network. It learns discriminative representations by maximizing the mutual information of different enhanced views of the same image and minimizing the mutual information of different enhanced views of the same image.

[0010] In a preferred embodiment, in the hierarchical knowledge distillation flow: The inter-level knowledge distillation loss function includes the KL divergence term of the output probability distribution of the teacher model and the student model, and the feature distillation loss term between the feature representations of the teacher model and the student model. The two terms are weighted by the trade-off coefficient. The hierarchical knowledge transfer mechanism is implemented through an adaptive weight fusion, which calculates weights based on the correlation scores of each level, adopts a softmax weight distribution method, and adjusts the smoothness of the weight distribution through a temperature parameter.

[0011] In a preferred embodiment, in the sparse annotation reinforcement learning module: The sparsely labeled samples are selected using an uncertainty sampling strategy, where samples with the lowest confidence level predicted by the model are selected for manual labeling. The semi-supervised learning framework uses pseudo-labeling technology and consistency regularization methods to use a small amount of labeled data to guide the learning of a large amount of unlabeled data; The optimization of labeling efficiency is achieved through an active learning cycle, which includes four stages: model training, uncertainty assessment, sample selection and manual labeling, and improves labeling efficiency through iterative optimization.

[0012] In a preferred embodiment, in generating individualized relief assessment results: The calculation of multi-dimensional relief indicators includes: immediate relief index, which evaluates the soothing effect in the short term after product use; Long-lasting relief index, which evaluates the long-term maintenance of the soothing effect after product use; Deep Soothing Index, which evaluates the soothing effect of the product on the deep tissues of the skin; Sensitivity feature marker identification uses a deep learning model to extract visual features related to sensitivity in skin images and quantify their degree of change; The personalized evaluation report generates targeted soothing effect evaluation and improvement suggestions based on the user's sensitivity level and specific sensitive characteristics.

[0013] In a preferred embodiment, the training process of the sensitivity layered neural network model includes: Initial training was performed using an expert-labeled sensitivity classification dataset; Cross-validation method was used to evaluate model performance and optimize hyperparameters; Introducing a class-balanced loss function to solve the problem of uneven distribution of sensitivity levels; Use model integration technology to fuse the prediction results of multiple basic classifiers to improve classification accuracy and robustness.

[0014] In a preferred embodiment, the contrast loss function used in the hierarchical contrastive learning framework is designed based on the InfoNCE form, which achieves effective representation learning by maximizing the similarity between different enhanced views of the same image while minimizing the similarity between enhanced views of other images in the batch; and different temperature parameters are used for different sensitivity levels, and lower temperature parameters are used for high-sensitivity levels to enhance feature discrimination.

[0015] In a preferred embodiment, a skin soothing efficacy evaluation method based on deep learning further includes a temporal soothing evaluation step: Collect skin images of users at multiple time points before, during, and after using soothing products; Construct a temporal feature extraction network to extract the characteristic representation of skin status at different time points; Design a temporal relief scoring function to quantify the trend of relief effect over time; Generate a soothing time-effectiveness curve to visually demonstrate the process and duration of the product's soothing effect; According to different sensitivity levels, the weights and thresholds of temporal evaluation are adjusted to achieve personalized temporal relief evaluation.

[0016] In a preferred embodiment, a skin soothing efficacy evaluation system based on deep learning is used to perform a skin soothing efficacy evaluation method based on deep learning, comprising: The heat map generation module uses gradient-weighted class activation mapping technology to generate heat maps of the areas in the skin image that contribute most to the soothing evaluation results; Feature importance analysis module, which quantifies the contribution of different skin features to soothing evaluation; A marker recognition module that extracts and visualizes skin features specific to different sensitivity levels; An explanation report generation module that intuitively displays the basis for soothing evaluation and key points of skin changes; The visualization interaction module provides a user-friendly interface, supports multi-dimensional viewing and comparison of the explanatory information of the soothing evaluation results, and improves the interpretability and credibility of the evaluation results.

[0017] The beneficial effects of the present invention are: The sensitivity layered neural network model enables accurate classification of people with different sensitivities, overcoming the problem of individual differences being ignored in existing technologies. This layered mechanism lays the foundation for subsequent personalized evaluation, allowing the evaluation results to fully consider the characteristics of people with different sensitivities, significantly improving the relevance and accuracy of the evaluation, especially for people with high sensitivity.

[0018] The hierarchical contrastive learning framework constructed can learn effective feature representations from large amounts of unlabeled skin image data, addressing the problem of scarce labeled data. By designing specific data augmentation strategies and contrastive learning tasks for different sensitivity levels, the system can capture the unique skin features of each sensitivity level, forming a more discriminative feature representation space, significantly improving the model's learning efficiency under small sample conditions.

[0019] The innovative hierarchical knowledge distillation flow mechanism enables knowledge transfer from low-sensitivity to high-sensitivity, effectively addressing the issue of insufficient cross-level knowledge sharing. Through the inter-level knowledge distillation network and hierarchical knowledge transfer mechanism, the system is able to transfer general knowledge learned in low-sensitivity levels to high-sensitivity levels, where data is more scarce, while retaining the specific knowledge of each level. This significantly reduces model parameter redundancy and improves the overall system evaluation efficiency.

[0020] The sparsely labeled reinforcement learning module maximizes the use of limited labeled data through active learning and semi-supervised learning techniques, achieving efficient model optimization. This module intelligently selects the most informative samples for labeling and augments training data through pseudo-labeling technology, significantly reducing the cost of acquiring high-quality labeled data while ensuring the reliability and stability of evaluation results, providing more accurate and personalized relief evaluation results for people with different sensitivity levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a skin soothing efficacy evaluation method based on deep learning of the present invention; Figure 2 It is a bar chart comparing the accuracy of the sensitivity stratification evaluation of the present invention; Figure 3 is a bar chart comparing the training data requirements of the present invention; Figure 4 It is a pie chart of the parameter proportion analysis of the model of this solution of the present invention; Figure 5 is a line graph showing the change in accuracy during the hierarchical enhancement effect training process of the present invention; Figure 6 It is a radar chart of the relief index evaluation of different sensitivity levels of the present invention; Figure 7It is a scatter plot of the impact of sparse annotation on model performance of the present invention. DETAILED DESCRIPTION

[0022] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses a method for evaluating skin soothing efficacy based on deep learning, such as Figure 1 As shown, the following steps are included: Step 1: Preprocessing skin sensitivity stratification data, including collecting skin image data, extracting sensitivity evaluation indicators, and using a sensitivity stratification neural network model to divide the user's skin into different sensitivity levels; It includes the following sub-steps: Step 1.1, skin image acquisition and standardization;

[0024] Skin image data was collected using a standardized image acquisition algorithm, and the collected raw skin image data was preprocessed, including image cropping, illumination compensation, color correction, and resolution unification.

[0025] Through these processing, a standardized skin image dataset was generated to ensure consistency and comparability in subsequent analyses.

[0026] It should be understood that, in some embodiments, the normalization process may further include steps such as image enhancement, noise removal, and contrast adjustment to further improve image quality. Step 1.2, sensitivity assessment index extraction;

[0027] A multi-feature extraction network was applied to extract sensitivity-related features from standardized skin images, including key indicators such as erythema index, moisture barrier integrity, capillary visibility, and skin tone uniformity.

[0028] These features together constitute a multidimensional representation vector of skin sensitivity.

[0029] It should be noted that, in an optional embodiment, skin physiological parameter sensing data, such as transepidermal water loss rate, skin pH value, skin temperature and other indicators, can also be combined with image features to form a more comprehensive sensitivity assessment system. Step 1.3, construction and application of sensitivity hierarchical neural network model;

[0030] A sensitivity hierarchical neural network model is constructed to classify the user's skin into different levels according to sensitivity. The neural network model includes a feature encoding layer, a feature fusion layer, and a classification output layer.

[0031] The feature encoding layer consists of multiple convolutional units, each of which contains a convolutional layer, a normalization layer, and an activation function layer; The feature fusion layer uses the attention mechanism to adaptively weight different feature channels; The classification output layer uses a fully connected structure to output the sensitivity classification results.

[0032] The model is trained using limited expert-labeled data to classify user skin sensitivity into mildly sensitive ( ), moderately sensitive ( ) and severe sensitivity ( The division results serve as the basis for the subsequent hierarchical learning architecture.

[0033] Optionally, the sensitivity level can be divided into two levels or four or more levels according to actual application requirements. For example, extremely sensitive ( ) levels to meet the needs of detailed evaluation of extremely sensitive skin.

[0034] like Figure 2 The figure shows a comparison of the evaluation accuracy of this application and conventional models at different sensitivity levels (mild sensitivity, moderate sensitivity, and severe sensitivity). It can be seen that this application achieves high accuracy at all sensitivity levels, especially at the severe sensitivity level, where the accuracy improves most significantly, from 76.3% to 95.2%, verifying the excellent adaptability of this application to highly sensitive people.

[0035] Step 2: Based on the stratified skin data, a hierarchical contrastive learning framework is constructed. A specific data augmentation strategy is designed for each sensitivity level, and a multi-view contrastive learning network is constructed to form a hierarchical specific representation space. It includes the following sub-steps: Step 2.1, implementation of sensitivity-specific data enhancement strategy; For each sensitivity level Designing specific data augmentation strategies , according to the characteristics of skin with different sensitivities, the original image Perform transformation to generate enhanced image view .

[0036] For mildly sensitive skin, the focus is on enhancing color and brightness changes; for moderately sensitive skin, the focus is on enhancing texture and local features; for severely sensitive skin, the focus is on enhancing microstructure and erythema features.

[0037] In addition, this application can also adopt an adaptive data enhancement strategy to automatically adjust the enhancement parameters according to the image content to avoid feature distortion caused by excessive enhancement.

[0038] like Figure 3 The figure shows a comparison of the amount of training data required for this application and a standalone training model at different sensitivity levels. This application reduces the amount of training data required for the moderate and severe sensitivity levels through a hierarchical knowledge flow mechanism. For the moderate sensitivity level, the training data requirement is reduced from 8,000 to 1,200 (an 85% reduction); for the severe sensitivity level, it is reduced from 5,000 to 750 (an 85% reduction), demonstrating the efficiency of this application in data-scarce scenarios. Step 2.2, multi-view contrastive learning network construction;

[0039] For each sensitivity level , building a contrastive learning network , which is used to extract consistent feature representations from two enhanced views. The network consists of a feature extraction backbone network, a projection head network, and a contrastive learning loss calculation unit.

[0040] The feature extraction backbone network adopts an improved ResNet structure, which contains multiple residual blocks and a feature pyramid structure; the projection head network is composed of a multi-layer perceptron, which maps high-dimensional features to a low-dimensional representation space; the contrastive learning loss calculation unit is designed based on the InfoNCE loss function principle.

[0041] The loss function of contrastive learning is defined as follows: ;

[0042] in, Indicates sensitivity level Contrastive learning loss function; and Respectively represent the sensitivity level of the same skin image The following is obtained after two different data enhancements Hedi The feature representation vector of each view; Indicates that other images in the batch are at the same sensitivity level The feature representation vector below; Represents the cosine similarity function, which is used to calculate the similarity between two feature vectors; Represents the temperature parameter, which controls the smoothness of the characteristic distribution. A value of will make the similarity distribution steeper and enhance the model's ability to distinguish similar samples; Indicates the summation operation of all samples in the batch, including positive and negative sample pairs.

[0043] Cosine similarity function Used to measure the similarity between two eigenvectors, the value range is The closer the value is to 1, the more similar the two vectors are, the closer it is to -1, the less similar the two vectors are, and a value of 0 means the two vectors are orthogonal.

[0044] By optimizing this loss function, the network learns feature representations that are invariant to enhancement transformations and capture the essential characteristics of skin soothing.

[0045] In some embodiments, the feature extraction backbone network can also adopt other efficient network architectures such as EfficientNet or DenseNet to meet application requirements in different computing resource environments. Step 2.3, level-specific representation space construction;

[0046] A contrastive learning network is used to construct a specific representation space for each sensitivity level, into which skin images are mapped, forming feature representations that distinguish different soothing states. These representation spaces reflect the unique characteristics and soothing response patterns of skin with different sensitivities.

[0047] Optionally, the representation space can be further visualized in low dimensions by mapping the high-dimensional representation to a two-dimensional or three-dimensional space through dimensionality reduction techniques such as t-SNE or UMAP, so as to intuitively present the distribution characteristics of different relaxation states.

[0048] like Figure 4 The figure shows the model parameter composition of this application and the number of parameters reduced through hierarchical knowledge distillation technology. The figure shows that through knowledge distillation and parameter sharing, this application reduces 38 million parameters (about 60% of the number of parameters), while retaining the key parameters of the basic feature extraction layer (12 million), hierarchical contrast learning layer (5 million), hierarchical knowledge output layer (3 million), and sensitivity-oriented specialization layer (5 million). This achieves a lightweight model and enables the system to run efficiently on resource-constrained devices.

[0049] Step 3: Using the layer-specific representation space, a hierarchical knowledge distillation flow is constructed. Knowledge transfer from low sensitivity to high sensitivity is achieved through the inter-layer knowledge distillation network. The hierarchical knowledge transfer mechanism is used to fuse knowledge from different layers and dynamically adjust the adaptive weights. It includes the following sub-steps: Step 3.1, constructing the knowledge distillation network between levels; Construct an inter-level knowledge distillation network to distill low-sensitivity layers of knowledge is transferred to the high sensitivity level The network consists of a teacher network, a student network, a feature adapter, and a loss calculation module.

[0050] The teacher network is a pre-trained model of the low-sensitivity layer, and the student network is a model to be trained of the high-sensitivity layer; the feature adapter is responsible for mapping the feature space of the teacher network to the feature space of the student network; the loss calculation module integrates the distribution distillation loss and the feature distillation loss.

[0051] The knowledge distillation loss function is defined as follows: ;

[0052] in, Indicates sensitivity level Knowledge distillation loss function; stands for KL divergence, which is a measure of the difference between two probability distributions; Represents the teacher network (i.e., the low sensitivity layer The output probability distribution of the pre-trained model of Represents the student network (i.e., high sensitivity level The output probability distribution of the model to be trained); The weight parameter representing the balanced distribution distillation loss and feature distillation loss; Represents the teacher network's response to the input image Extracted feature vectors; Represents the student network's response to the input image Extracted feature vectors; Represents the square of the L2 norm, which is used to calculate the Euclidean distance between two feature vectors.

[0053] The first distribution distillation loss is used to quantify the difference between the output distribution of the teacher network and the output distribution of the student network; the second feature distillation loss uses the mean square error to calculate the difference between feature vectors to maintain the consistency of the representation space.

[0054] Through the combined effect of these two losses, effective knowledge transfer from the low-sensitivity level to the high-sensitivity level is achieved.

[0055] It should be noted that, in some embodiments, knowledge distillation can also adopt a relational distillation method, which not only focuses on the output of a single sample, but also considers the relationship information between samples to further enhance the knowledge transfer effect. Step 3.2: Implement the hierarchical knowledge transfer mechanism;

[0056] Implement a hierarchical knowledge transfer mechanism for learning from mildly sensitive ( ) to moderately sensitive ( ), then to severely sensitive ( ). This mechanism includes a hierarchical feature extraction unit, an inter-level knowledge conversion unit, and a feature fusion unit.

[0057] The hierarchical feature extraction unit uses a sensitivity-specific feature extractor Get the features of the current level; knowledge conversion unit between levels It includes a feature transformation network and a feature adaptation network, which are responsible for converting low-level knowledge into a form that can be used by high-level layers; the feature fusion unit uses a gating mechanism to combine the current-level features and the converted knowledge.

[0058] The representation fusion process can be expressed as: ;

[0059] in, Indicates the Level (sensitivity level) for samples The feature representation vector of Representation level A feature extraction network is used to extract features from the image data at the current level; Indicates the level Designed data transformation or enhancement functions to process the original image; Representation sample The original image data; Representation level The parameter set of the feature extraction network; represents the inter-level knowledge conversion function, which is used to convert the knowledge of the low-sensitivity level to the high-sensitivity level; Indicates the Level (lower sensitivity level) for samples The feature representation vector of .

[0060] Inter-level knowledge transfer function The specific implementation is a neural network containing multiple fully connected layers, inputting low-level representations , and outputs a transformed representation that is compatible with the current level representation space. Its internal structure is: first, the input dimension is adjusted through the dimension matching layer, then the representation ability is enhanced through the nonlinear transformation layer, and finally the normalization layer is used to ensure that the output has the same statistical characteristics as the current level representation, thereby achieving a smooth transition and effective fusion of different levels of representation space.

[0061] Optionally, in the process of hierarchical knowledge transfer, a skip connection method can also be used to enable the high-level model to simultaneously acquire multiple low-level knowledge, forming a richer knowledge fusion structure.

[0062] like Figure 5As shown, the hierarchical enhancement effect produced by the hierarchical knowledge flow mechanism is demonstrated. The chart compares the changes in accuracy of independently trained lightly sensitive models, independently trained heavily sensitive models, and heavily sensitive models trained using the hierarchical knowledge flow at different training data ratios. It can be seen that the hierarchical knowledge flow of this application enables the heavily sensitive model to achieve an accuracy of 42.7% in the early stages of training (at 20% data), which is much higher than the 25.3% of independent training; after the training is completed, the accuracy of the heavily sensitive model of the hierarchical knowledge flow reaches 95.2%, which is 18.9 percentage points higher than the independent training and exceeds the independently trained lightly sensitive model (91.2%). This proves that the representation ability of highly sensitive hierarchical models significantly benefits from low-level knowledge. Step 3.3, implementation of adaptive weight adjustment algorithm;

[0063] Implement an adaptive weight adjustment algorithm to dynamically adjust the weights of each loss function according to the amount and characteristics of data at different sensitivity levels.

[0064] The overall loss function is defined as: ;

[0065] in, represents the overall loss function, are model parameters; For all sensitivity levels Perform summation; Indicates sensitivity level Contrastive learning loss for learning effective feature representations; Indicates sensitivity level The knowledge distillation loss is used to achieve knowledge transfer from low sensitivity to high sensitivity; Indicates sensitivity level Task-specific loss, used to optimize the performance of a specific evaluation task; Indicates the level Adaptive weight parameters for contrastive learning loss; Indicates the level Adaptive weight parameters of the knowledge distillation loss; Indicates the level Adaptive weight parameters for the task-specific loss.

[0066] These three adaptive weight parameters will be dynamically adjusted according to the data characteristics and training status of different sensitivity levels to balance the contribution of each loss function and achieve the optimal model training effect.

[0067] It should be understood that the adaptive weight adjustment algorithm can also automatically adjust the weight parameters based on the changes in verification performance during the model training process to obtain the optimal training effect.

[0068] like Figure 6 As shown, it shows the multi-dimensional indicator evaluation of the soothing effect of the Benson evaluation system on skin with different sensitivity levels. The chart shows the evaluation results of the three levels of mild sensitivity, moderate sensitivity and severe sensitivity in five dimensions: erythema resolution, moisture restoration, barrier repair, inflammation reduction and redness reduction. It can be seen that there are significant differences in the response of skin at each sensitivity level to soothing products: mildly sensitive skin performs best in moisture restoration, moderately sensitive skin responds well in inflammation reduction, and severely sensitive skin has outstanding effects in inflammation reduction but slower moisture recovery. This verifies that the present application can identify characteristic biomarkers of skin with different sensitivities and provide a scientific basis for personalized soothing solutions.

[0069] Step 4: Combine the output of the knowledge distillation flow to build a sparse annotation reinforcement learning module to achieve sparse annotation sample selection, semi-supervised learning framework construction and annotation efficiency optimization; It includes the following sub-steps: Step 4.1, implementation of sparse labeled sample selection algorithm;

[0070] Implement a sparsely labeled sample selection algorithm to select the most informative samples from a large amount of unlabeled data for manual labeling.

[0071] The selection strategy is based on the principles of uncertainty sampling and diversity preservation, ensuring that limited annotation resources can maximize model performance.

[0072] In some embodiments, sample selection can also be combined with active learning strategies to give priority to labeling samples with high model prediction uncertainty, further improving labeling efficiency. Step 4.2, semi-supervised learning framework construction;

[0073] We build a semi-supervised learning framework for training models by combining a small amount of labeled data with a large amount of unlabeled data. The framework includes a supervised learning branch, a pseudo-label generator, and a consistency regularization module.

[0074] The supervised learning branch uses manually labeled data to train the basic model; the pseudo-label generator generates pseudo-labels for unlabeled data based on the confidence of the current model; and the consistency regularization module ensures that the model's prediction results are consistent under different augmented views.

[0075] The training objective consists of a supervision loss and a consistency regularization term, and is of the form: ;

[0076] in, Indicates sensitivity level The task-specific total loss function of Represents the supervision loss based on manually annotated data, which is used to guide the model to learn from labeled data; Represents the consistency regularization loss based on unlabeled data, which is used to improve model performance by utilizing a large amount of unlabeled data; Represents the balance factor, which is used to adjust the weight ratio between the supervision loss and the consistency regularization loss.

[0077] Consistency regularization loss The calculation method of is to measure the difference between the model prediction results of the same unlabeled sample under different data augmentation conditions. Specifically, it can be calculated using methods such as mean square error or KL divergence. Its purpose is to encourage the model to maintain stable predictions for data perturbations, thereby improving the model's generalization ability and the efficiency of utilizing unlabeled data.

[0078] Optionally, the semi-supervised learning framework can also introduce mean square consistency constraints or virtual adversarial training methods to further enhance the model's ability to learn unlabeled data. Step 4.3, annotation efficiency optimization feedback loop;

[0079] Build a feedback loop to optimize annotation efficiency and dynamically adjust sample selection strategies based on changes in model performance. When newly annotated samples do not significantly improve the model, adjust the selection criteria to focus on areas of model weakness.

[0080] Through this iterative optimization, the number of required labeled samples is minimized while maximizing model performance improvement.

[0081] In addition, the feedback loop can also integrate expert knowledge, allowing human experts to confirm or correct the model's labeling suggestions, forming a human-machine collaborative labeling optimization mechanism.

[0082] like Figure 7 As shown, the effect of the proportion of sparsely labeled samples on the model accuracy is demonstrated, verifying the effectiveness of the sparsely labeled reinforcement learning module of this application. The chart shows that with only 10% of sparsely labeled samples, the model accuracy can reach 68.7%; when using 30% of labeled samples, the accuracy reaches 84.6%; when using 60% of labeled samples, the accuracy has reached 93.2%, close to the highest accuracy of 95.2% when using all labeled samples. This shows that this application can effectively use a very small amount of labeled data to achieve high-precision personalized evaluation through the sparse labeled sample selection algorithm and semi-supervised learning framework, greatly reducing the cost of data labeling.

[0083] Step 5: Generate individualized relief evaluation results based on the trained model, including calculation of multi-dimensional relief indicators, identification of sensitivity characteristic markers, and generation of personalized evaluation reports.

[0084] It includes the following sub-steps: Step 5.1: Construction of a multi-dimensional relief index calculation system;

[0085] A multi-dimensional soothing index calculation system was built to calculate multi-dimensional soothing indices based on the user's sensitivity level using the corresponding evaluation model for the input skin image. The system includes an index extraction module, an index quantification module, and an index integration module.

[0086] The indicator extraction module extracts key features such as erythema, moisture, barrier and inflammation from the image; The index quantification module converts the extracted features into standardized relief index scores; The indicator integration module integrates various indicators with different weights according to the characteristics of the sensitivity level to form a comprehensive portrait of the soothing effect.

[0087] It should be noted that, in some embodiments, the multi-dimensional relief index can be further expanded into a time series index, which provides a quantitative description of the dynamic relief process by comparing the changes in the index at different time points before and after relief. Step 5.2, construction of sensitivity characteristic marker identification system;

[0088] A sensitivity marker identification system was constructed to automatically identify characteristic biomarkers of skin with different sensitivities using the hierarchical enhancement effect learned in the model. The system includes a feature mapping module, an attention visualization module, and a marker screening module.

[0089] The feature mapping module projects the skin image features into the representation space; The attention visualization module highlights the areas that contribute most to the decision through gradient-weighted class activation mapping technology; The marker screening module selects characteristic biomarkers according to the characteristics of different sensitivity levels, providing an explanation basis for the evaluation results.

[0090] Optionally, the sensitivity feature markers can be associated with the medical expert knowledge base, mapping the markers discovered by the model with known skin physiological indicators to improve the scientificity and interpretability of the recognition results. Step 5.3, build a personalized evaluation report generation system;

[0091] A personalized evaluation report generation system was built to generate personalized evaluation reports tailored to user sensitivity based on the calculated relief index and identified characteristic markers. The system includes a data integration module, a visualization module, and a recommendation generation module.

[0092] The data integration module integrates soothing indicators and marker information; the visualization module generates intuitive charts to display the evaluation results; and the suggestion generation module provides targeted skin care suggestions based on user sensitivity characteristics and evaluation results.

[0093] The final report includes a quantitative score of the soothing effect, changing trends of key indicators, status changes of characteristic markers, and personalized soothing recommendations.

[0094] It should be understood that personalized evaluation reports can provide different levels of detail based on user needs, from professional scientific research-level detailed reports to consumer-friendly concise reports, to meet the needs of different application scenarios.

[0095] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A skin soothing efficacy evaluation method based on deep learning, characterized in that: The following steps are involved: Perform skin sensitivity stratification data preprocessing, including collecting skin image data, extracting sensitivity assessment indicators, and using the sensitivity stratification neural network model to divide the user's skin into different sensitivity levels; Based on the stratified skin data, a hierarchical contrastive learning framework is constructed. Specific data enhancement strategies are designed for each sensitivity level, and a multi-view contrastive learning network is constructed to form a hierarchical specific representation space. By utilizing the layer-specific representation space, we construct a hierarchical knowledge distillation flow, achieve knowledge transfer from low sensitivity to high sensitivity through the inter-layer knowledge distillation network, use the hierarchical knowledge transfer mechanism to fuse knowledge from different layers, and dynamically adjust the adaptive weights; Combining the output of the knowledge distillation flow, we build a sparse annotation reinforcement learning module to achieve sparse annotation sample selection, semi-supervised learning framework construction, and annotation efficiency optimization. Based on the trained model, individualized relief evaluation results are generated, including calculation of multi-dimensional relief indicators, identification of sensitive characteristic markers and generation of personalized evaluation reports.

2. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: In the step of preprocessing the skin sensitivity stratification data, the sensitivity evaluation indicators include: Skin barrier integrity indicators, calculated by image analysis of transepidermal water loss and stratum corneum hydration; Vascular reactivity index, skin erythema index and vascular dilation degree were calculated by image analysis; Inflammatory markers, which identify and quantify inflammation-related features through image analysis; Sensitivity threshold index, by analyzing the speed and intensity of the skin's response to external stimuli; The sensitivity hierarchical neural network model adopts a multi-layer perceptron structure, with the input being the combined feature vector of sensitivity evaluation indicators and the output being the skin sensitivity hierarchical classification result.

3. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: In the hierarchical contrastive learning framework, data augmentation strategies designed for different sensitivity levels include: For low-sensitivity levels, stronger color transformation, geometric transformation, and noise addition are used; At the medium sensitivity level, medium-intensity color and geometric transformations are applied to preserve skin texture details; For high-sensitivity levels, slight color adjustments and local area enhancements are used to preserve sensitivity characteristics to the greatest extent possible; The multi-view contrastive learning network adopts a dual-tower structure, consisting of an encoder network and a projection head network. It learns discriminative representations by maximizing the mutual information of different enhanced views of the same image and minimizing the mutual information of different enhanced views of the same image.

4. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: In the hierarchical knowledge distillation flow: The inter-level knowledge distillation loss function includes the KL divergence term of the output probability distribution of the teacher model and the student model, and the feature distillation loss term between the feature representations of the teacher model and the student model. The two terms are weighted by the trade-off coefficient. The hierarchical knowledge transfer mechanism is implemented through an adaptive weight fusion, which calculates weights based on the correlation scores of each level, adopts a softmax weight distribution method, and adjusts the smoothness of the weight distribution through a temperature parameter.

5. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: In the sparse annotation reinforcement learning module: The sparsely labeled samples are selected using an uncertainty sampling strategy, where samples with the lowest confidence level predicted by the model are selected for manual labeling. The semi-supervised learning framework uses pseudo-labeling technology and consistency regularization methods to use a small amount of labeled data to guide the learning of a large amount of unlabeled data; The optimization of labeling efficiency is achieved through an active learning cycle, which includes four stages: model training, uncertainty assessment, sample selection and manual labeling, and improves labeling efficiency through iterative optimization.

6. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: Individualized relief assessment results are being generated: The calculation of multi-dimensional relief indicators includes: immediate relief index, which evaluates the soothing effect in the short term after product use; Long-lasting relief index, which evaluates the long-term maintenance of the soothing effect after product use; Deep Soothing Index, which evaluates the soothing effect of the product on the deep tissues of the skin; Sensitivity feature marker identification uses a deep learning model to extract visual features related to sensitivity in skin images and quantify their degree of change; The personalized evaluation report generates targeted soothing effect evaluation and improvement suggestions based on the user's sensitivity level and specific sensitive characteristics.

7. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: The training process of the sensitivity layered neural network model includes: Initial training was performed using an expert-labeled sensitivity classification dataset; Cross-validation method was used to evaluate model performance and optimize hyperparameters; Introducing a class-balanced loss function to solve the problem of uneven distribution of sensitivity levels; Use model integration technology to fuse the prediction results of multiple basic classifiers to improve classification accuracy and robustness.

8. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: The contrastive loss function used in the hierarchical contrastive learning framework is based on the InfoNCE formalism. It achieves effective representation learning by maximizing the similarity between different enhanced views of the same image while minimizing the similarity with the enhanced views of other images in the batch. Different temperature parameters are used for different sensitivity levels, and lower temperature parameters are used for high sensitivity levels to enhance feature differentiation.

9. The skin soothing efficacy evaluation method based on deep learning according to claim 1, characterized in that: It also includes the following steps for evaluating the timing of the event: Collect skin images of users at multiple time points before, during, and after using soothing products; Construct a temporal feature extraction network to extract the characteristic representation of skin status at different time points; Design a temporal relief scoring function to quantify the trend of relief effect over time; Generate a soothing time-effectiveness curve to visually demonstrate the process and duration of the product's soothing effect; According to different sensitivity levels, the weights and thresholds of temporal evaluation are adjusted to achieve personalized temporal relief evaluation.

10. A skin soothing efficacy evaluation system based on deep learning, used to implement the skin soothing efficacy evaluation method based on deep learning according to any one of claims 1 to 9, characterized in that: include: The heat map generation module uses gradient-weighted class activation mapping technology to generate heat maps of the areas in the skin image that contribute most to the soothing evaluation results; Feature importance analysis module, which quantifies the contribution of different skin features to soothing evaluation; A marker recognition module that extracts and visualizes skin features specific to different sensitivity levels; An explanation report generation module that intuitively displays the basis for soothing evaluation and key points of skin changes; The visualization interaction module provides a user-friendly interface, supports multi-dimensional viewing and comparison of the explanatory information of the soothing evaluation results, and improves the interpretability and credibility of the evaluation results.

Images