Intelligent scalp detection system and detection method thereof

Abstract

The invention discloses an intelligent scalp detection system and a detection method thereof, relates to the technical field of intelligent scalp detection, and aims at obtaining shooting conditions and definition evaluation through real-time environment perception and quality detection, solving hair occlusion and image noise by using deep learning segmentation and directional enhancement, showing a scalp deep structure by means of multi-modal fusion and splicing, and improving the detection accuracy. A diagnosis and optimization strategy is output in comprehensive analysis and self-adaptive closed-loop feedback; through software algorithm level expansion, online training and multi-label pathological expansion can be performed on multi-user long-period data, and adaptive evolution and group statistical analysis are realized; according to the method, the image acquisition stability and segmentation accuracy can be effectively improved, the adaptability of detection to complex illumination and multi-band information is enhanced, the risk of pathological misjudgment such as grease, inflammation and fungi is reduced, and personalized nursing and wide health management application are supported.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent scalp detection, and specifically to an intelligent scalp detection system and its detection method. Background Art

[0002] In daily life, many users hope to know their scalp health status at any time and place, so as to timely intervene in potential problems such as dandruff, sebum secretion imbalance, hair follicle blockage, and even fungal infections. In this process, users may use handheld or wearable intelligent detection devices, combined with a macro camera to take close-up photos of the scalp, so as to obtain relatively fine image data. The actual use scenarios include various indoor and outdoor environments such as homes, barbershops, and clinics, and the lighting conditions vary from dim to strong direct sunlight; in addition, there are also significant differences in hair density, color, and curl degree among some people, which makes the scalp area more likely to be blocked by hair, and users may move quickly or shift the angle during combing or shooting, further increasing the difficulty of image acquisition. Facing these complex and variable environmental factors, how to ensure both the clarity and integrity of scalp images and also take into account usability has become an important challenge for intelligent scalp detection systems.

[0003] After retrieval, in the Chinese invention patent with the authorized announcement number CN113761974B, a method for monitoring the scalp is disclosed, which is applied to an intelligent hair dryer including a shooting device, and includes: when the intelligent hair dryer is running, using the shooting device to take pictures of the user's head to obtain multiple head photos; identifying the head areas corresponding to the multiple head photos respectively; identifying the relevant parameters of each head area based on the head photos corresponding to each head area; and analyzing the scalp conditions of each head area based on the relevant parameters of each head area and the historical relevant parameters corresponding to each head area.

[0004] Combined with the above application scenarios and existing technologies: Since the reliability of subsequent algorithm analysis needs to be ensured during the scalp acquisition process, when the external light is poor or the focal length cannot be stabilized, the resolution and contrast of scalp images are often difficult to meet the standards, resulting in obvious deviations in recognition or diagnosis results.

[0005] Specifically, if the camera fails to focus adequately on the scalp area or is subject to improper user operations (such as rapid movement, occlusion by long curly hair, etc.), problems such as blurred images, distorted edges, or loss of hair follicle details are highly likely to occur. This will not only cause a sharp decline in the accuracy of automatic detection of oil concentration, hair follicle density, and minor inflammation, but may also misjudge more serious disease risks such as potential fungal infections and sebaceous gland dysfunction, resulting in users missing the best intervention opportunity or receiving incorrect care advice. From the perspective of industrial application and user experience, insufficient acquisition quality in the above scenarios will directly affect the effectiveness of subsequent multimodal fusion and feature analysis, increase the error rate of algorithm discrimination, and weaken the objective reflection of the actual scalp condition, thus failing to meet the requirements of refined detection and personalized health management.

[0006] Therefore, the present invention provides an intelligent scalp detection system and its detection method. Summary of the Invention

[0007] (I) Technical Problems to be Solved Aiming at the deficiencies of the prior art, the present invention provides an intelligent scalp detection system and its detection method. By real-time environmental perception and quality detection, it obtains shooting conditions and clarity evaluation, uses deep learning segmentation and directional enhancement to solve hair occlusion and image noise, utilizes multimodal fusion and stitching to display the deep structure of the scalp, and outputs diagnosis and optimization strategies in comprehensive analysis and adaptive closed-loop feedback. Finally, through software algorithm expansion, it can perform online training on long-term data of multiple users and multi-label pathology expansion, realizing adaptive evolution and population statistical analysis. This solution effectively improves the stability of image acquisition and the accuracy of segmentation, enhances the adaptability of detection to complex lighting and multi-band information, and reduces the risk of misjudging pathologies such as oil, inflammation, and fungi, thus solving the technical problems recorded in the background art.

[0008] (II) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent scalp detection method, including that when it is detected that the light intensity and distance meet the available threshold, the environmental perception module combines with the image quality detection unit to calculate the mapped environmental score in real time and output the environmental and preliminary quality data; When the environmental and preliminary quality data indicates that imaging is available and the mapped environmental score is higher than the corresponding threshold, the deep learning segmentation model performs hair and scalp separation on the captured image and removes the noise area, calls the PDE-based enhancement component for denoising and super-resolution correction, and outputs the segmented and enhanced image data; If the segmented and enhanced image data stably presents scalp features, the multimodal fusion module reads visible light and near-infrared images, matches the overlapping areas through optimal transformation and performs local compensation to generate multimodal fusion image data; The intelligent analysis model calls the global health score based on the environment, preliminary quality data, and segmentation results to discriminate pathological abnormalities, outputs the analysis result data, and simultaneously transmits back the low-confidence regions for pre-sequence segmentation and environment-adaptive optimization iteration; When the user generates a large amount of analysis result data during long-term use and expands pathological recognition, multi-label detection is combined to update the model parameters by federated learning, and the global distribution function is used to reveal the coupling between fungal or sebaceous gland abnormalities and environmental factors at the population level, and corresponding processing strategies are pushed to the user.

[0009] Preferably, the environment perception module collects the light intensity, the distance value of the camera from the scalp surface, the tilt angle value of the camera relative to the scalp normal direction, and the motion smoothness, and defines a mapped environment score to reflect the usability of the shooting environment.

[0010] Preferably, the image quality detection unit calculates the image quality score based on the received mapped environment score and the initial captured image using gradient distribution or wavelet features; the quality score and the mapped environment score are encapsulated as environment and preliminary quality data, which also includes environmental illumination, focal length, motion smoothness, and image sharpness; when both the quality score and the mapped environment score are lower than the corresponding preset thresholds, the user prompt mechanism is triggered.

[0011] Preferably, using the initial captured image as the image input, a pre-trained deep learning segmentation network is used to distinguish the pixels of the hair and the scalp, combines the pixel-level segmentation result of the deep learning with the environmental quality feedback, defines an objective functional, and obtains a segmentation mask after solving. The scalp area and the non-scalp area are identified based on the segmentation mask, and the segmentation edge is adaptively smoothed or corrected.

[0012] Preferably, a directional enhancement model combining partial differential equations and non-linear response functions is introduced into the scalp area determined by the segmentation mask to output the segmented and enhanced image data.

[0013] Preferably, multi-modal additional image data is collected, including near-infrared bands, multi-spectral channels, or other-angle shootings; The segmented and enhanced image data and the multi-modal additional image data are accurately aligned in the same spatial coordinate system, and an optimal transformation is obtained after constructing a regional matching energy functional: After applying the optimal transformation to the multi-modal additional image data, the segmented and enhanced image data and the aligned multi-modal image data in the same coordinate system are obtained.

[0014] Preferably, the segmented and enhanced image data and the multi-modal additional image data are weighted and superimposed with adaptive weights in the same coordinate system to generate multi-modal fusion image data, including the matching results of visible light, near-infrared, and multi-spectral bands.

[0015] Preferably, the intelligent analysis model performs multi-level feature extraction on the multi-modal fusion image data, including: texture scale decomposition, key point detection, and spectral difference measurement; Define a deep representation mapping to map the pixel and neighborhood information of the multi-modal fusion image data into a high-dimensional feature vector containing multi-modal information representations of regions such as the scalp and hair follicles.

[0016] Preferably, the anomaly recognition model classifies and detects potential lesions or anomalies, and outputs a health score and analysis result data including basic detection results. Combined with the environmental correction term and spatial distribution information, perform cumulative measurement on it in the scalp area in a matrix-weighted manner, and finally obtain the global health score.

[0017] Preferably, after generating the analysis result data, it is transmitted back to the corresponding module in the previous step, including: If obvious local texture distortion or abnormal confidence is detected, the location information of the area and the corresponding environmental factors can be transmitted back to prompt to improve the lighting or maintain stability during the next shooting; For areas with repeated segmentation errors, the analysis result data can be used as training set increment or hard example annotation; If it is found that a specific band contributes insufficiently to the detection, dynamically adjust the band selection strategy or improve the alignment accuracy during subsequent acquisitions.

[0018] Preferably, after obtaining the user's historical usage records, based on the user increment data and environmental indicators, regularly or real-time update the existing deep learning segmentation model or recognition model.

[0019] Preferably, use the updated model parameters to perform multi-label classification or segmentation on other scalp pathologies other than oiliness and inflammation, and output a confidence vector of multiple categories after defining the multi-label discriminant function: If necessary, it can be extended to multi-label segmentation, that is, on the basis of the mask of the aforementioned scalp area, multiple pathological areas are simultaneously labeled to achieve a finer-grained lesion map.

[0020] Preferably, perform population statistics or clustering analysis on the detection data from multiple users. Define a global distribution function to characterize the correlation between different pathological labels and environmental factors. For each user, extract the corresponding pathological label gradient intensity in each environmental factor dimension from the global distribution function, and obtain a standardized feature vector after processing; Select spectral clustering or Gaussian mixture model to cluster the standardized feature vectors. For each cluster, assign a risk label according to the gradient intensity distribution of the cluster center in each pathological dimension. According to the risk tags, the infrared thermal imaging frame rate is preferentially increased for people sensitive to high inflammation, and the ultraviolet spectral band acquisition is automatically enabled for people prone to fungi in humid environments; and corresponding treatment strategies are recommended for people sensitive to high inflammation and people prone to fungi in humid environments.

[0021] An intelligent scalp detection system, including a data acquisition unit, when the detected light intensity and distance meet the available threshold, the environmental perception module combines with the image quality detection unit to calculate the mapped environmental score in real time and output the environmental and preliminary quality data; An image processing unit, when the environmental and preliminary quality data indicates that imaging is available and the mapped environmental score is higher than the corresponding threshold, the deep learning segmentation model performs hair and scalp separation on the captured image and removes the noise area, calls the PDE-based enhancement component to combine denoising and super-resolution correction, and outputs the segmented and enhanced image data; A data fusion unit, if the segmented and enhanced image data stably presents scalp features, the multi-modal fusion module reads the visible light and near-infrared images, matches the overlapping areas through optimal transformation and performs local compensation to generate multi-modal fusion image data; A data analysis unit, the intelligent analysis model calls the global health score based on the environmental and preliminary quality data and the segmentation result to discriminate pathological abnormalities and outputs the analysis result data, and at the same time transmits the low-confidence area back for the previous segmentation and environmental adaptive optimization iteration; A push unit, when a large amount of analysis result data is generated during the long-term use of the user and the pathological recognition is extended, the multi-label detection is combined to update the model parameters by federated learning, and the global distribution function is used to reveal the coupling between fungal or sebaceous gland abnormalities and environmental factors at the population level, and the corresponding treatment strategy is pushed to the user.

[0022] (III) Beneficial effects The present invention provides an intelligent scalp detection system and its detection method, having the following beneficial effects: Parameters such as environmental and preliminary quality data (EPQD) and mapped environmental score can adaptively optimize the lighting and focus at the shooting source, reduce image distortion caused by poor environment; can effectively suppress hair interference and highlight scalp details; The segmented and enhanced image data is further matched with the near-infrared or multi-spectral images, and the generated multi-modal fusion image data can overcome the limitations of the single-band perspective. By performing local compensation at the severely occluded or complex lighting areas, the scalp surface and deep structures are presented more completely and clearly. This fusion not only utilizes the shooting angle information but also combines the segmentation mask to perform adaptive matching and interpolation on the overlapping or mismatched areas, integrating the advantages of multi-source images into one, and greatly improving the accuracy of downstream recognition.

[0023] The intelligent analysis model uses multi-modal fusion image data to identify various scalp features such as sebum and inflammation, and adaptively tunes the model parameters based on the environmental quality and the uncertainty of the segmentation results. The resulting analysis result data can provide real-time diagnosis and be fed back to the previous steps. This multi-directional feedback mechanism ensures that the system can continuously optimize the segmentation and fusion quality in complex scenarios, reducing the probability of false alarms and missed detections.

[0024] During long-term use, the images collected from users multiple times and the corresponding pathological judgments are used to continuously update the model through online training or federated learning, incorporating multi-label detection, and then achieving accurate identification of complex problems such as fungal infections and sebaceous gland abnormalities in a wider population. It also reveals the potential interaction rules between pathology and environmental factors. The data from different bands, environmental feedback, and the results of the deep model produce a synergistic effect, far exceeding traditional single-band or fixed algorithms in terms of scalp detection accuracy and stability, capable of meeting the personalized needs of diverse users and scenarios and continuously iterating, providing a forward-looking and complete systematic solution for intelligent scalp health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a schematic flow diagram of the intelligent scalp detection method of the present invention; Figure 2 is a schematic structural diagram of the intelligent scalp detection system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Please refer to Figure 1 , the present invention provides an intelligent scalp detection method, including, Step 1. When initially collecting images, the environmental perception module senses the light intensity, irradiation angle, and camera distance, and calls the image quality detection unit to evaluate the clarity and scalp visibility, obtaining environmental and preliminary quality data and mapping environmental scores ; The above Step 1 includes the following contents: Step 101. Environmental parameter collection and comprehensive scoring The environmental perception module respectively collects the light intensity, the distance between the camera and the scalp, and the movement parameters during user combing, calculates the mapping environmental score according to the irradiation angle information , and encapsulates it into the environmental and preliminary quality data; where: The main acquisition quantities of the environmental perception module include light intensity; the distance value of the camera from the scalp surface, the tilt angle value of the camera relative to the scalp normal direction, and the motion smoothness; To comprehensively reflect the usability of the shooting environment, a mapping environment score is defined :

[0028] Among them:

[0029]

[0030] The 2×2 matrix W to be calibrated can be calibrated through experimental data or set based on prior experience; is the index vector; L is the light intensity and d is the distance; is the tilt angle of the camera relative to the scalp normal direction; m is the motion smoothness, is the desired smoothness; Logistic function: , will from map to interval; is the angle correction term, is an adjustable parameter used to control the influence degree of the angle on the shooting quality; is the motion jitter penalty term, is the adjustment coefficient, ; When in use, it can automatically identify unfavorable environments (such as too dark light, too large distance or severe jitter), and reduce the score in time, so that in the subsequent steps, after receiving this score, it can decide whether to prompt the user to correct the shooting method; Step 102, Preliminary Image Quality Detection and User Prompt The image quality detection unit calculates the image quality score using advanced methods such as gradient distribution or wavelet features based on the received mapping environment score and the initial captured image , and this score will be encapsulated with the mapping environment score into the environment and preliminary quality data. Among them, the environment and preliminary quality data EPQD contains environmental illumination, focal length, motion smoothness, image sharpness, and mapping environment score etc.; Finally, it is transmitted to Step 2, specifically as follows: With the mapping environment score and the initial captured image As the input of the image quality detection unit, in the following manner:

[0031] Where: is the initial captured image; represents the coefficient vector obtained by performing multi-scale wavelet transform on the initial captured image at scale s and direction o: Here, is the position of the image in the corresponding wavelet domain or the corresponding spatial domain; represents the vector formed by the local gradient (or partial derivative) of at the same scale s and direction o, which is used to characterize the detail intensity and change trend of the image; is a pre-defined or trained real symmetric weighting matrix (with dimensions matching ), and different s and o can correspond to different weight matrices to highlight or suppress specific frequency band and direction information; S and O respectively represent the scale set and direction set in the wavelet transform. For example, , ; is the image coordinate domain, which can be the mapping area of wavelet coefficients in the spatial or frequency domain; is the mapping environment score is the adjustment coefficient when , and its value can be between 0.1 and 1.0, mainly controlling the amplification or suppression degree of the mapping environment score on the final result of the image quality score; When the quality score and the mapping environment score both fall below the preset threshold, the user prompt mechanism is triggered to recommend that the user adjust the shooting angle or slow down the moving speed. When both are relatively high or showing an upward trend, the subsequent steps can be continued without prompting; When in use, on the premise of obtaining the mapping environment score , performing multi-scale wavelet analysis on the image quality can achieve more stable image sharpness judgment, and at the same time, flexible regulation of different environmental states can be achieved through logarithmic mapping; The quality score and the mapping environment score The combined output of the two provides the environment and preliminary quality data, fundamentally reducing the risk of misjudgment of defective images in subsequent steps. This synergistic effect lies in the ability to capture and quantify external interference factors in real time, and also to mutually verify the actual imaging effect with the help of image features; it also achieves non-linear characterization and adaptive optimization in complex shooting scenarios, providing a more reliable initial input for subsequent segmentation and multi-modal fusion.

[0032] Step 2: When the environment and preliminary quality data indicate that both the illumination and clarity reach the available threshold, the deep learning segmentation model performs hair-scalp pixel segmentation on the captured image, removes the interference regions, and combines the mapped environment score to adaptively adjust the regularization coefficient, perform denoising, and enhance scalp details to suppress hair occlusion, generating the segmented and enhanced image data; The content of Step 2 is as follows: Step 201: Hair-scalp segmentation Using the initial captured image as the image input, a pre-trained deep learning segmentation network is used to distinguish the pixels of hair and scalp, obtaining the probability that each pixel belongs to the scalp category; combining the pixel-level segmentation result of deep learning with the environmental quality feedback to automatically enhance the reliability of the segmentation decision in a suboptimal environment or blurred area, where: Define the following objective functional and obtain the segmentation mask after solution , where the segmentation mask is 1 for the scalp area and 0 for non-scalp areas (including hair and background), realizing the determination of the scalp area by the segmentation mask :

[0033] In the formula: represents the negative log-likelihood term constructed based on the pixel label probability output of the deep learning segmentation network for the initial captured image , which is used to measure the consistency between the segmentation and the model prediction; is the regularization term combined with the environmental and preliminary quality data EPQD, mainly performing adaptive smoothing or correction on the segmentation edge when the mapped environment score is low (such as poor illumination or obvious jitter); specifically as follows: The regularization term is specifically defined as environment-adaptive weighted total variation smoothing, in the form as follows:

[0034] where: is the segmentation mask at pixel The value at is the image coordinate domain to be segmented; represents the gradient vector of the mask in the spatial dimension and is used to measure the change rate of the mask boundary; is a small constant to prevent the denominator from being zero, for example ; is a weight parameter, and its numerical range depends on the experimental tuning results. It is used to balance the importance of the segmentation network signal and the environmental correction information in the segmentation process; When in use, through the weight parameter couple the mapping environmental score in the environment and the preliminary quality data EPQD into the segmentation process, which can moderately smooth or correct the network decision in fuzzy or low-light scenarios, thereby reducing hair and scalp boundary errors; when the environment and the preliminary quality data EPQD have a good score, the weight parameter will play a relatively weakened role and mainly rely on the fine prediction of the network; when the score is poor, the regularization effect will be relatively strengthened to improve the segmentation stability. By reading information such as illumination and jitter in the environment and the preliminary quality data EPQD, the segmentation decision is corrected directionally, which helps to achieve higher accuracy in real complex scenarios.

[0035] Step 202, Image Enhancement Perform key enhancement on the scalp area determined by the segmentation mask to avoid artifacts caused by overprocessing hair pixels, and introduce an orientation enhancement model combining partial differential equations (PDE) and non-linear response functions to output the segmented and enhanced image data , and define the evolution equation as follows:

[0036] In the formula: represents the enhanced image at the evolution time t. At the initial time , it can be set to ; is the divergence operator, is the gradient operator; is the non-linear response function: When (scalp area) and the EPQD indicates low illumination, this function takes a larger value to enhance detail repair; when (hair or background) or the illumination is good, this function takes a smaller value to reduce over-smoothing or enhancement of non-target areas; , is a constant coefficient, and all its values are greater than 0, which is used to control the diffusion strength and fidelity compensation term in the PDE; One ensures that the enhanced result retains the style of the original captured image while compensating for the illumination, and will not be distorted into oversharpening or oversmoothing; When in use, through the segmentation mask shield the hair pixels, and only perform refined processing on the scalp area, ensuring that the segmented area corresponds one-to-one with the enhancement requirements, avoiding unnecessary artifacts, and the non-linear response function Combined with the quality score in the environmental and preliminary quality data EPQD and the mapped environmental score , stronger local enhancement can be applied to areas with uneven illumination or obvious noise in the scalp area, which can improve the visualization effect and recognition accuracy; different from common global filtering or global super-resolution, only the scalp mask area is weighted through the evolution of the environmental and preliminary quality data EPQD, and environmental factors are incorporated, which has higher pertinence and controllability. After completing the above two steps, not only the confusion of hair occlusion for the details of the scalp area is effectively eliminated, but also the recognizability of the scalp texture is improved through targeted enhancement.

[0037] Step 3: When the image data after segmentation and enhancement stably presents the scalp area, the multi-modal fusion module reads the visible light image and near-infrared data, combines the shooting angle and mask information, performs weighted stitching after affine mapping, and applies adaptive weights to multiple scalp lesions such as sebum and inflammation in the overlapping area for real-time recognition, outputs the analysis result data, and feeds back the low-confidence area to the segmentation model and shooting module for adaptive correction; The third step includes the following contents: Step 301, multi-modal data alignment Collect multi-modal additional image data , which may include shooting in the near-infrared band, multi-spectral channels or other angles; Align the image data after segmentation and enhancement and the multi-modal additional image data Precisely in the same spatial coordinate system. To take into account the complex texture of the scalp and possible non-linear distortions, construct the following regional matching energy functional to solve the optimal transformation :

[0038] represents the geometric transformation for mapping the multi-modal additional image data in the target coordinate domain, which can be composed of function families such as affine, bilinear or more advanced thin plate spline (TPS); is the alignment consistency measure, which can be measured based on the Kullback-Leibler divergence or other high-dimensional distribution difference measurement methods of the segmented region indicated by the scalp mask, rather than simple pixel differences; is the regularization term, which is used to limit the drastic distortion of the transformation when the environment is unfavorable (such as large jitter and weak illumination); is the balance coefficient, and its value range is between 0.1 and 5.0; Apply the optimal transformation to the multi-modal additional image data After that, the segmented and enhanced image data and the aligned multi-modal image data in the same coordinate system ; When the environment and the preliminary quality data EPQD indicate high jitter or low illumination, the balance coefficient will be increased , enhancing the transformation smoothness of the geometric transformation , and minimizing the registration failure caused by local extreme changes; when the environment is better, more reliance is placed on the alignment consistency measure for high-precision matching of details; When in use, the non-linear transformation is combined with the distribution information of the pixels within the segmentation mask, enabling higher resolution and finer-level alignment of the multi-modal images in the scalp area: by introducing the feedback of the environment and the preliminary quality data EPQD in the regularization term , automatically adjusting the transformation smoothness according to the shooting environment state to ensure the alignment reliability.

[0039] Step 302, Fusion Compensation and Mosaic Overlay the segmented and enhanced image data and the multi-modal additional image data These two images are weighted and superimposed in the same coordinate system with adaptive weights to generate multi-modal fusion image data ; Among them: To balance the low-frequency consistency (global brightness and contrast) and high-frequency detail complementarity (hair follicle texture and shallow / deep structures), the following weighted fusion model is adopted in the scalp area in this step: Let represent the multi-modal fusion image data at the evolution time t ( is the coordinate on the image plane, such as ), and at the initial moment, it can be set as:

[0040] That is, first use the visible light image as the initial solution, and define the following evolution equation:

[0041] Wherein: is the divergence operator, is the gradient operator; is the image data after segmentation and enhancement in the visible light domain; is the multi-modal additional image data obtained after alignment (such as near-infrared, multi-spectral, etc.); is the scalp region mask weight function, and the value range is ; is the diffusion tensor (or matrix) in the evolution equation, which is used to control how the image is smoothed and gradient diffused during the iteration process, and can adaptively adjust the diffusion intensity of certain regions according to the mapping environment scoring data (EPQD); is the data fidelity coefficient, which also combines EPQD information, and the value can be between 0.01 and 2.0; Wherein: Initialization: Set as ; Iterative evolution: Update sequentially in the discrete time domain until convergence or reaching the preset iteration upper limit; Convergence result: When approaches 0, the stable multi-modal fusion image data is obtained, which contains the matching results of visible light and near-infrared / multi-spectral bands, etc.; For the iterative solution of the above PDE-based multi-modal fusion formula, if you want to accelerate convergence, you can add parameters such as illumination and distance during initialization or iteration to the diffusion tensor and the data fidelity coefficient for dynamic adjustment, so as to approach the ideal fusion state faster; When in use, by referring to the alignment residual and the mapping environment scoring , local anomalies can be processed directionally, minimizing stitching traces and artifacts. Introducing alignment residuals and EPQD information during the fusion process can achieve precise suppression or repair of mismatched regions, significantly improving the final stitching quality.

[0042] Step three finally outputs the multi-modal fusion image data. This fusion result presents comprehensive information of multiple bands or multiple perspectives in the scalp region, overcoming the occlusion and brightness limitations that may be encountered in pure visible light images, and also using additional information such as near-infrared or multi-spectral to reveal deeper structural features of hair follicles and the scalp. Combining the EPQD, segmentation mask, and illumination enhancement information accumulated in the previous steps, it demonstrates strong adaptability and reliability in the multi-modal data stitching and compensation link.

[0043] Step 4: When the multi-modal fusion image data is completed in splicing and has sufficient clarity, the intelligent analysis model imports the environmental and preliminary quality data and the segmentation results, calls the anomaly recognition model to perform real-time recognition on various scalp lesions such as grease and inflammation, and outputs the analysis result data. At the same time, the low-confidence regions are fed back to the segmentation model and the shooting module for adaptive correction; The content of the said Step 4 includes the following: Step 401: Feature parsing and deep representation The intelligent analysis model performs multi-level feature extraction on the multi-modal fusion image data including but not limited to: Texture scale decomposition: Use multi-scale or multi-directional filters (such as Gabor filtering, Curvelet, etc.) to perform hierarchical detection on the scalp texture from coarse to fine; Key point detection: Extract local feature vectors of hair follicle openings, inflammation hotspots or other abnormal points; Spectral difference measurement: For multi-band information, calculate the dispersion of the feature distribution within the scalp area (such as local energy function or similarity measure based on spectral curve fitting) to determine abnormal oil secretion or pigment deposition, etc.; Define the following deep representation mapping to map the pixel and neighborhood information of the multi-modal fusion image data into a high-dimensional feature vector containing multi-modal information representations of areas such as the scalp and hair follicles :

[0044] In the formula: represents the image coordinates; is the local gradient information, used to identify edges or rapidly changing regions; can refer to the pixel response vector under the multi-spectral channel; Unify and encapsulate the spectral, texture and depth features of the scalp surface, laying a foundation for subsequent classification or clustering; When in use, integrate the visible light, near-infrared / multi-spectral and enhanced scalp area information in the same feature mapping, which can more comprehensively reflect the scalp health indicators and achieve multi-modal integrated analysis; fuse various bands with the gradient / texture features of the segmented regions into a unified high-dimensional representation, rather than simple splicing or sequential processing, which can significantly improve the discriminability of complex scalp features. The deep representation mapping performs multi-scale and multi-channel deep encoding on the scalp area features, reducing missed detections or false detections caused by a single scale or a single band.

[0045] Step 402: Anomaly recognition and health scoring After obtaining the high-dimensional feature vector After that, a potential lesion or abnormality is classified and detected by an anomaly recognition model, and a health score and analysis result data are output , including basic test results such as sebum secretion, inflammation, and follicle health; Among them, a deep neural network (such as Transformer or self-attention mechanism) or a hybrid model (such as the combination of random forest and graph neural network) can be used to learn the high-dimensional feature vector ; the anomaly recognition model is obtained by training a convolutional neural network (such as ResNet, EfficientNet) or a vision Transformer (ViT) architecture, and adding an attention mechanism or a graph neural network layer; Combined with the environmental correction term and spatial distribution information, a cumulative measure is performed on it in the scalp area by matrix weighting, and finally a global health score is obtained :

[0046] In the formula: is a deep discriminant function, and is a neural network model which can typically be represented as a multi-layer perceptron (MLP) or a lightweight convolutional network, and the last layer uses Sigmoid (binary classification) or Softmax (multi-classification) activation to output the probability value that the pixel or local belongs to the abnormal (such as inflammation, excessive sebum) or healthy category, or perform other forms of risk measurement; is the high-dimensional feature vector of the multi-modal fusion image at the pixel ; represents the weights of the neural network or algorithm for the recognition / classification task, which have been determined or dynamically updated in the training stage; represents the discriminant function of the coordinate in the feature space is a real symmetric weighting matrix used to adjust the contribution of gradients at different positions to the final integration result, and its size and direction can be dynamically changed according to the environmental correction function : is a real symmetric positive definite matrix (such as 2×2 or n×n) matching the dimension of the gradient vector, which weights the contributions of the gradient components in different directions; is an environmental correction function, which is linked to the previously defined EPQD (environment and preliminary quality data), and converts the environment and preliminary quality data EPQD into the confidence weight for each pixel in the subsequent processing; is an additional global correction or compensation term for overall adjustment outside the integral term; if no global correction or compensation is required, can be regarded as a constant or set to zero.

[0047] During use, the environmental correction function keeps the model cautious in areas with poor lighting, reducing the risk of false alarms; in high-quality areas, it gives full play to the detection ability to achieve anomaly recognition based on environmental adaptation; by non-linearly integrating the classification results of all pixels / blocks in the scalp area, a highly integrated global health score can be output to help users understand the overall health status; coupling environmental information such as lighting and jitter with the depth discrimination model in the same integral functional constructs an environmentally adaptive health score formula, which is more flexible and stable than traditional single classification or simple weighting.

[0048] Step 403, Closed-loop feedback and adaptive model update After generating the analysis result data, these results are all sent back to the corresponding modules in the previous steps to form a closed loop. Typical feedback paths include but are not limited to: Sent back to the environmental perception module in Step 1: If obvious local texture distortion or abnormal confidence is detected in Step 402, the location information of this area and the corresponding environmental factors (lighting or jitter in EPQD) can be sent back to prompt to improve lighting or maintain stability during the next shooting; Sent back to the deep learning segmentation model in Step 2: For areas with repeated segmentation errors, the analysis result data can be used as an increment of the training set or hard example annotation to further improve the segmentation accuracy of the model under fine textures and complex hair qualities; Sent back to the multi-modal fusion module: If it is found in Step 402 that a specific band contributes insufficiently to the detection, the band selection strategy can be dynamically adjusted or the alignment accuracy can be improved during subsequent acquisitions; During use, by continuously flowing the analysis result data back to the previous stage, its own defects can be continuously corrected and it can continuously iterate and evolve: if the user's operating habits cause a lot of jitter or a complex lighting environment, the segmentation model or fusion strategy can also be adaptively adjusted after closed-loop feedback to reduce inaccurate recognition caused by human errors, thereby lowering the user's operation threshold. During subsequent shootings, certain specific bands can be used preferentially according to the instructions of the closed-loop feedback, or higher-resolution re-acquisition can be performed in some key areas.

[0049] Step Five. When the user accumulates a large amount of analysis result data during long-term use, the online training module and the multi-label detection module extract new pathological dimensions such as fungal infection or sebaceous gland abnormality, and iteratively update the model parameters through federated learning and in the global distribution function Environmental factors in the integration , accurately identify more pathological types and push corresponding treatment strategies to users; The step five includes the following contents: Step 501: Long-term data collection and online training Obtaining historical user usage records: including multiple test results and possible subjective feedback from users (such as whether the user marked that there was real inflammation). This part of data is accumulated over a long period of use; Regularly or in real time update the existing deep learning segmentation model or recognition model through online training module; In order to be compatible with federated learning and local online learning scenarios, an iterative update formula is defined. represents the model parameters at time t, Represents newly collected batch data (which contains user feedback or high-confidence samples after automatic screening), you can construct:

[0050] in: is a gradient function or differential mapping, used according to Correct the current model based on segmentation errors, classification errors or environmental factors; is the learning rate or update step size, which can be dynamically adjusted by the illumination and jitter indicators in EPQD; During use, the model can gradually adapt to individual characteristics. During multiple tests, the model will adapt to the user's specific scalp environment, greatly reducing the long-term detection error of the same user. When new symptoms or extreme lighting conditions are detected, new batches of data The difficult examples in will improve the reliability of the model during iterative updates.

[0051] Step 502: Multi-label pathology recognition expansion Perform multi-label classification or segmentation on scalp pathologies other than oil and inflammation (such as fungal infection, sebaceous gland dysfunction, etc.), and define a multi-label discriminant function , whose output is the confidence vector of multiple categories:

[0052] Where: Each component in corresponds to a pathological label (such as fungal sebaceous gland abnormal inflammation secondary infection, etc.), It is a high-dimensional feature vector, or new features are added after online training; With the updated model parameters , more detailed pathological information can be identified; If necessary, it can be extended to multi-label segmentation, that is, multiple pathological regions are marked simultaneously on the basis of the mask of the scalp region mentioned above to achieve a more fine-grained lesion map.

[0053] When in use, the coverage of disease types is expanded, and more possible pathological types are added on the basis of the original oil, inflammation, etc., which can be closer to clinical needs. The traditional indicators for oil or inflammation can still be output, and the core results of the original recognition process will not be damaged; the confidence levels or segmentation masks of multiple pathologies can be output simultaneously, breaking through the traditional binary classification or single-label limit. If the newly emerging pathological types are less in early samples, after accumulating more real samples, the recognition performance for new labels can be rapidly improved.

[0054] Step 503, Cross-user big data statistics and population characteristic modeling Perform population statistics or clustering analysis on the detection data from multiple users to mine more representative scalp health patterns and define the global distribution function Characterize the correlation between different pathological labels and environmental factors of, Indicates an optional pathology-environment attention vector (such as the parameters corresponding to a specific pathological label or composite index), where:

[0055] Indicates the index domain or measure space of the user set, that is, there are multiple users u in the federated learning or large-scale application scenario; Indicates the environmental factor domain, which can cover light intensity, humidity, region or other extended scenario variable integrals; is the aggregated analysis result data, which is usually associated with the user u and includes pathological detection results (such as labels or probabilities of inflammation, oil, fungal infection, etc.) and aggregated features of scalp images; is an aggregation function used to couple and map the pathological information of the user and environmental factors : It can internally calculate the complex correlation between pathological labels and environmental variables, such as the relationship between oil secretion and indoor temperature, and the relationship between fungal infection and humidity, etc.; Indicates the gradient vector obtained for this mapping in the composite dimension; is a real symmetric weighted matrix (or tensor), which is closely related to the parameter . Different parameters can specify to emphasize or suppress certain pathology-environment correlation dimensions, such as particularly focusing on the fungus + humidity dimension; This is a global correction or compensation term outside the double integral, and its functions may include: Overall weight increase for extreme scenarios (such as ultra-high humidity, rare pathological labels) to avoid being masked by sparse data; One-time deduction or compensation for certain common factors (such as the unified inherent error of equipment). If there is no such requirement, it can be regarded as a constant or set to 0.

[0056] Obtain the global distribution function After that, for each user , extract the corresponding pathological label gradient intensity from the global distribution function in each environmental factor dimension , splice them into a vector , perform column-wise normalization on all to eliminate the dimensional differences of different environmental factors and obtain ; Select spectral clustering (SpectralClustering) or Gaussian mixture model (GMM) to cluster , and for each cluster, assign risk labels according to the gradient intensity distribution of the cluster center in each pathological dimension: If a cluster has a high gradient in the inflammation-temperature or sebum-humidity dimension, it is labeled as highly inflammation-sensitive; If a cluster has a high gradient in the fungus-humidity or sebaceous gland activity-humidity dimension, it is labeled as prone to fungus in humid conditions; The remaining clusters are named according to their actual gradient characteristics (such as low risk / hair follicle weakening, etc.).

[0057] Prioritize increasing the infrared thermal imaging frame rate for highly inflammation-sensitive people, and automatically enable ultraviolet spectral acquisition for people prone to fungus in humid conditions; Recommend corresponding treatment strategies for highly inflammation-sensitive people and people prone to fungus in humid conditions. For example, recommend anti-inflammatory formulas for highly inflammation-sensitive people and antibacterial lotions for people prone to fungus in humid conditions.

[0058] During use, it is possible to evaluate the distribution of scalp problems in specific regions or specific populations, help develop more targeted care programs, and enable the model to better handle diverse scalp types and environmental variables through group characteristics learned from cross-user data. It can be linked with the online learning module to dynamically update the mapping mode to improve the accuracy of statistics and clustering.

[0059] Through online training or federated learning, the system can accumulate data from different users in the long term and continuously iterate the model parameters. The updated model is used to achieve multi-label detection, expand the coverage of the system for scalp pathology, perform population statistics on the detection results of multiple users and multiple scenarios, form a more general understanding of the pathological distribution, and feedback it to steps 501 and 502 again. The model and detection scope can be continuously optimized. While maintaining the accurate detection and analysis capabilities constructed in the first four steps, it has the potential for sustainable evolution facing long cycles, multiple scenarios, and large populations.

[0060] Please refer to Figure 2 , the present invention provides an intelligent scalp detection system, including, A data acquisition unit. When the detected light intensity and distance meet the available thresholds, the environmental perception module combines with the image quality detection unit to calculate and output the mapped environmental score and environmental and preliminary quality data in real time. An image processing unit. When the environmental and preliminary quality data indicates that imaging is available and the mapped environmental score is higher than the corresponding threshold, the deep learning segmentation model performs hair and scalp separation on the captured image and removes the noise area, calls the PDE-based enhancement component for denoising and super-resolution correction, and outputs the segmented and enhanced image data. A data fusion unit. If the segmented and enhanced image data stably presents scalp features, the multi-modal fusion module reads visible light and near-infrared images, matches the overlapping areas through optimal transformation and performs local compensation to generate multi-modal fusion image data. A data analysis unit. The intelligent analysis model calls the global health score to judge pathological abnormalities based on the environmental and preliminary quality data and the segmentation results, and outputs the analysis result data. At the same time, it feeds back the low-confidence regions for pre-segmentation and environment adaptive optimization iteration. A push unit. When a large amount of analysis result data is generated during the long-term use of the user and the pathological recognition is expanded, the model parameters are updated by federated learning in combination with multi-label detection, and the global distribution function is used to reveal the coupling between fungal or sebaceous gland abnormalities and environmental factors at the population level, and corresponding treatment strategies are pushed to the user.

[0061] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0062] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0063] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0064] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0065] As described above, only the specific implementation manners of the present application are provided, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application, and all should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent scalp detection method, characterized in that: including, when the detected light intensity and distance meet the available threshold, the environmental perception module combines with the image quality detection unit to calculate the mapped environmental score in real time and output the environmental and preliminary quality data; if the mapped environmental score is higher than the corresponding threshold, the deep learning segmentation model performs hair and scalp separation on the captured image, eliminates the noise area, calls the PDE-based enhancement combined with denoising and super-resolution correction, and outputs the segmented and enhanced image data; if the segmented and enhanced image data stably presents scalp features, the multi-modal fusion module reads the visible light and near-infrared images, matches the overlapping areas through optimal transformation and performs local compensation to generate multi-modal fusion image data; the intelligent analysis model calls the global health score based on the environmental and preliminary quality data and the segmentation results to discriminate pathological abnormalities, and outputs the analysis result data. At the same time, it transmits the low-confidence areas back for the previous segmentation and environmental adaptive optimization iteration; combines multi-label detection to update the model parameters by federated learning, and uses the global distribution function to reveal the coupling between fungal or sebaceous gland abnormalities and environmental factors at the population level, and pushes the corresponding treatment strategies to the user.

2. The intelligent scalp detection method according to claim 1, wherein: the environmental perception module collects the light intensity, the distance value from the camera to the scalp surface, the tilt angle value of the camera relative to the scalp normal direction, and the motion smoothness, and thereby defines the mapped environmental score to reflect the usability of the shooting environment.

3. The intelligent scalp detection method according to claim 2, wherein: the image quality detection unit calculates the image quality score based on the received mapped environmental score and the initial captured image by using gradient distribution or wavelet features; the quality score and the mapped environmental score are encapsulated as environmental and preliminary quality data, which also contains environmental light, focal length, motion smoothness, and image sharpness; when both the quality score and the mapped environmental score are lower than the corresponding preset thresholds, the user prompt mechanism is triggered.

4. The intelligent scalp detection method according to claim 3, wherein: using the initial captured image as the image input, the pre-trained deep learning segmentation network differentiates the pixels of the hair and the scalp, combines the pixel-level segmentation result of the deep learning with the environmental quality feedback, defines the objective functional and obtains the segmentation mask after solving, identifies the scalp area and the non-scalp area according to the segmentation mask, and adaptively smooths or corrects the segmentation edge.

5. The intelligent scalp detection method according to claim 4, wherein: for the scalp area determined by the segmentation mask, a directional enhancement model based on the combination of partial differential equations and non-linear response functions is introduced to output the segmented and enhanced image data.

6. The intelligent scalp detection method according to claim 5, wherein: collect multi-modal additional image data, including shooting in the near-infrared band, multi-spectral channels, or other angles; precisely align the segmented and enhanced image data with the multi-modal additional image data in the same spatial coordinate system, construct the regional matching energy functional and solve the optimal transformation; After applying the optimal transformation to the multimodal additional image data, the segmented and enhanced image data and the aligned multimodal image data in the same coordinate system are obtained.

7. The intelligent scalp detection method according to claim 6, wherein: The segmented and enhanced image data and the multimodal additional image data are weighted and superimposed in the same coordinate system with adaptive weights to generate multimodal fusion image data, including the matching results of visible light, near-infrared, and multispectral bands.

8. The intelligent scalp detection method according to claim 7, wherein: The intelligent analysis model performs multi-level feature extraction on the multimodal fusion image data, including texture scale decomposition, key point detection, and spectral difference measurement; Define a deep representation mapping to map the pixel and neighborhood information of the multimodal fusion image data into a high-dimensional feature vector containing the multimodal information representation of the scalp and hair follicle regions.

9. The intelligent scalp detection method according to claim 8, wherein: The potential lesions or abnormalities are classified and detected through an anomaly recognition model, and a health score and analysis result data including basic detection results are output; Combined with the environmental correction term and spatial distribution information, a cumulative measure is performed on the scalp area by matrix weighting, and finally a global health score is obtained.

10. The intelligent scalp detection method according to claim 9, wherein: After generating the analysis result data, it is transmitted back to the corresponding modules in the previous steps, including: If obvious distortion of local texture or abnormal confidence is detected, the corresponding region position information and corresponding environmental factors are transmitted back to prompt to improve the lighting or maintain stability during the next shooting; For regions with repeated segmentation errors, the analysis result data can be used as training set increment or hard example annotation; If it is found that a specific band contributes insufficiently to the detection, the band selection strategy or alignment accuracy is dynamically adjusted in subsequent acquisitions.

11. The intelligent scalp detection method according to claim 10, wherein: After obtaining the user's historical usage records, based on the user increment data and environmental indicators, the existing deep learning segmentation model or recognition model is updated regularly or in real time.

12. The intelligent scalp detection method according to claim 11, wherein: Use the updated model parameters to perform multi-label classification or segmentation on other scalp pathologies other than grease and inflammation, and output a confidence vector of multiple categories after defining a multi-label discrimination function; Based on the mask of the aforementioned scalp region, multiple pathological regions are marked simultaneously to generate a finer-grained lesion map.

13. The intelligent scalp detection method according to claim 12, wherein: Perform population statistics or clustering analysis on the detection data from multiple users; Define a global distribution function to characterize the correlation degree between different pathological labels and environmental factors. For each user, extract the corresponding pathological label gradient intensity in each environmental factor dimension from the global distribution function, and obtain a standardized feature vector after processing; Select spectral clustering or Gaussian mixture model to cluster the standardized feature vectors. For each cluster, according to the gradient intensity distribution of the cluster center in each pathological dimension, assign a risk label. According to the risk tags, the infrared thermal imaging frame rate is preferentially increased for people sensitive to high inflammation, the ultraviolet spectral band acquisition is automatically enabled for people susceptible to fungi in humid environments, and corresponding treatment strategies are recommended for people sensitive to high inflammation and people susceptible to fungi in humid environments.

14. Intelligent scalp detection system, characterized in that: Including, a data acquisition unit, when the detected light intensity and distance meet the available threshold, the environmental perception module combines with the image quality detection unit to calculate the mapped environmental score in real time and output the environmental and preliminary quality data; an image processing unit, when the environmental and preliminary quality data indicate that imaging is available and the mapped environmental score is higher than the corresponding threshold, the deep learning segmentation model performs hair and scalp separation on the captured image and removes the noise area, calls the PDE-based enhancement component to combine denoising and super-resolution correction, and outputs the segmented and enhanced image data; a data fusion unit, if the segmented and enhanced image data stably presents scalp features, the multimodal fusion module reads the visible light and near-infrared images, matches the overlapping areas through optimal transformation and performs local compensation to generate multimodal fusion image data; a data analysis unit, the intelligent analysis model calls the global health score based on the environmental and preliminary quality data and the segmentation results to discriminate pathological abnormalities and outputs the analysis result data, and at the same time transmits back the low-confidence regions for the previous segmentation and environmental adaptive optimization iteration; a push unit, when a large amount of analysis result data is generated during the long-term use of the user and the pathological recognition is extended, multi-label detection is combined to update the model parameters by federated learning, and the global distribution function is used to reveal the coupling between fungal or sebaceous gland abnormalities and environmental factors at the population dimension, and the corresponding treatment strategies are pushed to the user.

Images