Skin Condition Assessment Method and Intelligent Beauty Terminal Based on Skin Detection Function Head

Abstract

The present invention relates to the technical fields of beauty intelligent terminals, multimodal deep learning and transfer learning, image processing, etc. It provides a skin condition evaluation method and a beauty intelligent terminal based on a skin detection function head. By integrating a multispectral imaging module, an electrical impedance analysis module, and a temperature and humidity acquisition module in the same detection head, synchronous acquisition of visible and near-infrared multi-channel skin images and impedance spectra is achieved. The method aligns multimodal data based on a unified timestamp, and introduces dark current correction, wavelet denoising, and environmental normalization in the preprocessing stage to construct a unified feature vector. A multimodal cross-attention model fine-tuned by transfer learning fuses the features of each modality and then outputs four indices: water-oil balance, pigment deposition, texture roughness, and inflammation risk. If a high-risk state is detected, subsequent physiotherapy is automatically restricted and a voice warning is issued, forming a closed loop of detection, evaluation, and protection, improving the safety and evaluation accuracy of intelligent beauty.

Description

Technical Field

[0001] The present invention relates to the technical fields of beauty intelligent terminals, multimodal deep learning and transfer learning, image sensing, generation and enhancement, etc., and particularly relates to a skin condition evaluation method and an intelligent beauty terminal based on a skin detection function head. Background Art

[0002] With the rapid development of the home beauty and skin care market, a large number of portable beauty terminals have emerged on the market. Most of these devices adopt a single sensing means: either analyze skin color and blemishes through a visible light camera, or measure skin moisture content by means of bioelectrical impedance at a fixed frequency point. Due to the single information source and limited quantization dimensions, users often need to alternately use multiple devices to obtain a relatively comprehensive skin condition reference, which is not only cumbersome and costly to operate, but also difficult to keep the measurement conditions consistent among different devices, resulting in discrete results and difficult to compare. In addition, existing devices generally ignore the influence of environmental factors - the skin spectrum and impedance characteristics change significantly with temperature and humidity; without real-time environmental compensation, the evaluation results of the same user in winter and summer may be completely opposite, seriously weakening the credibility of the data. More critically, most products are in a fragmented state in the detection and physiotherapy chain: the detection function and physiotherapy output are independent of each other, and it is impossible to dynamically adjust the light energy or radio frequency power according to the real-time skin condition. When the user is in a high-risk state such as inflammation or allergy, high-power physiotherapy may still be triggered, bringing potential hidden dangers of secondary injury. On the other hand, traditional algorithms are mostly based on handcrafted features or shallow models, which are difficult to fully explore the coupling relationship among spectrum, electricity and environment, with limited evaluation dimensions and poor generalization ability; once different skin colors or different lighting environments are used, the result accuracy drops significantly.

[0003] To sum up, the following technical problems exist in the prior art: the single sensing modality leads to insufficient skin condition evaluation dimensions, and it is difficult to simultaneously quantify water-oil, pigment, texture and inflammation; multi-device or multi-step operations increase the use cost, the measurement conditions are difficult to synchronize, and the results lack comparability; the lack of environmental compensation and time synchronization causes the data to be affected by temperature, humidity and instantaneous physiological fluctuations, with poor stability; the detection and physiotherapy are separated, and it is impossible to reduce the power in real time to protect in high-risk scenarios; the depth of traditional algorithms is insufficient, it is difficult to fuse cross-modal features, and the evaluation accuracy and adaptability need to be improved. Summary of the Invention

[0004] In view of the above deficiencies in the prior art, the present invention provides a skin condition evaluation method and a beauty intelligent terminal based on a skin detection function head to form a closed loop of detection, evaluation and protection, and improve the safety and evaluation accuracy of intelligent beauty.

[0005] In a first aspect, the present invention provides a skin condition evaluation method based on a skin detection function head, which runs on the main control module of a beauty terminal. The skin condition evaluation method includes:

[0006] Drive the multi-spectral imaging module in the skin detection function head to collect skin reflection grayscale images frame by frame for each narrowband channel with a preset number of channels within a preset wavelength range, stack the obtained skin reflection grayscale images according to the time stamp T0 to generate a multi-spectral skin image of narrowband channels with a preset number of channels, and drive the impedance analysis module to apply a stepped sweep constant current signal with a preset signal frequency and peak current to the detection site within a preset window based on T0, synchronously measure the AC voltage, and calculate the real part and imaginary part of the impedance at each frequency point of the AC voltage using a phase-locked algorithm to obtain an impedance spectrum vector;

[0007] After the temperature and humidity acquisition module in the skin detection function head outputs the skin-contact temperature and relative humidity at a preset acquisition frequency, select the temperature and humidity data with the time stamp closest to T0 and correlate it with the data of the multi-spectral skin image and the data of the impedance spectrum vector, call the data preprocessing unit to perform dark current correction and flat field correction on the multi-spectral image to obtain clean multi-spectral data, perform wavelet band-pass denoising on the impedance spectrum vector to obtain clean impedance data, standardize the temperature and humidity data, the clean multi-spectral data, and the clean impedance data and splice them into a unified feature vector F;

[0008] Input the feature vector F into a multi-modal cross-attention model fine-tuned by transfer learning to fuse the features of each modality, and output four indices including the water-oil balance index, the pigment deposition index, the texture roughness index, and the inflammation risk index; obtain the four indices for display control, and when it is recognized that the inflammation risk index exceeds a preset threshold, trigger the voice interaction module of the beauty terminal to broadcast a warning and limit the subsequent physiotherapy output.

[0009] In a second aspect, the present invention provides an intelligent beauty terminal, including:

[0010] A skin detection function head detachably installed on a function head connection seat at the front end of a handle assembly of the beauty terminal, and a multi-spectral imaging module, an impedance analysis module, and a temperature and humidity acquisition module are integrated inside the skin detection function head;

[0011] A main control module is arranged inside the handle assembly, runs the above-mentioned skin condition assessment method based on the skin detection function head, and the main control module supplies power to the multi-spectral imaging module, the impedance analysis module, and the temperature and humidity acquisition module through electrical contacts in the function head connection seat and completes data interaction.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] The present invention provides a skin condition assessment method and a beauty intelligent terminal based on a skin detection function head. By driving a multispectral imaging module in the skin detection function head to sequentially capture skin reflection grayscale images frame by frame in a preset wavelength band range according to narrowband channels with a preset number of channels, stacking the obtained skin reflection grayscale images according to the time stamp T0 to generate a multispectral skin image of narrowband channels with a preset number of channels, and driving the impedance analysis module to apply a stepped sweep constant current signal with a preset signal frequency and peak current to the detection part within a preset window based on T0, synchronously measuring the alternating current voltage and using a phase-locked algorithm to calculate the real part and imaginary part of the impedance at each frequency point of the alternating current voltage to obtain an impedance spectrum vector. After the temperature and humidity acquisition module in the skin detection function head outputs the skin contact point temperature and relative humidity at a preset acquisition frequency, select the temperature and humidity data with the time stamp closest to T0 and associate them with the data of the multispectral skin image and the data of the impedance spectrum vector. Call the data preprocessing unit to perform dark current correction and flat field correction on the multispectral image to obtain clean multispectral data, perform wavelet band-pass denoising on the impedance spectrum vector to obtain clean impedance data, standardize the temperature and humidity data, the clean multispectral data, and the clean impedance data and splice them into a unified feature vector F. Input the feature vector F into a multi-modal cross-attention model fine-tuned by transfer learning to fuse the features of each modality, and output four indexes including a water-oil balance index, a pigment deposition index, a texture roughness index, and an inflammation risk index; obtain the four indexes for display control and when it is recognized that the inflammation risk index exceeds a preset threshold, trigger the voice interaction module of the beauty terminal to broadcast a warning and limit subsequent physiotherapy output, so as to realize the synchronous acquisition of visible and near-infrared multi-channel skin images and impedance spectra, form a detection, evaluation, and protection closed loop, and improve the safety and evaluation accuracy of intelligent beauty. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. Some specific embodiments of the present invention will be described in detail later with reference to the drawings in an exemplary rather than restrictive manner. The same reference numerals in the drawings denote the same or similar components or parts. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0015] Figure 1 is a flowchart of a skin condition assessment method based on a skin detection function head according to an embodiment of the present invention;

[0016] Figure 2 is a schematic architecture diagram of an intelligent beauty terminal according to an embodiment of the present invention;

[0017] Figure 3It is a schematic diagram showing a state of the assembly of the skin detection function head and the handle assembly of the intelligent beauty terminal according to an embodiment of the present invention.

[0018] Description of the reference numerals:

[0019] 1. Handle assembly; 10. Touch display screen;

[0020] 2. Skin detection function head. Specific embodiments

[0021] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below 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 the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] Embodiment 1

[0023] See Figures 1 - 3 , this embodiment provides a skin condition assessment method based on a skin detection function head. The skin condition assessment method runs on the main control module of the beauty terminal. The skin detection function head 2 is detachably installed on the function head connection seat at the front end of the handle assembly 1 of the beauty terminal. The multi-spectral imaging module, the impedance analysis module, and the temperature and humidity acquisition module are integrated inside the skin detection function head; the main control module is arranged inside the handle assembly. The main control module supplies power to the multi-spectral imaging module, the impedance analysis module, and the temperature and humidity acquisition module through the electrical contacts in the function head connection seat and completes data interaction.

[0024] See Figures 1 - 3 , the skin condition assessment method based on the skin detection function head provided in this embodiment includes the following steps:

[0025] S101. Drive the multi-spectral imaging module in the skin detection function head to sequentially collect skin reflection grayscale images frame by frame in a preset band range according to narrowband channels with a preset number of channels, stack the obtained skin reflection grayscale images according to the time stamp T0 to generate a multi-spectral skin image of narrowband channels with a preset number of channels, and drive the impedance analysis module to apply a stepped sweep constant current signal with a preset signal frequency and peak current to the detection part within a preset window based on T0, synchronously measure the AC voltage, and use the lock-in algorithm to calculate the real part and the imaginary part of the impedance at each frequency point of the AC voltage to obtain an impedance spectrum vector;

[0026] S102. After the temperature and humidity acquisition module in the skin detection function head outputs the skin contact point temperature TS and the relative humidity RHS at a preset acquisition frequency, select the temperature and humidity data with the time stamp closest to T0 among the skin contact point temperature TS and the relative humidity RHS, associate them with the data of the multispectral skin image and the data of the impedance spectrum vector, call the data preprocessing unit to perform dark current correction and flat field correction on the multispectral image to obtain clean multispectral data, perform wavelet band-pass denoising on the impedance spectrum vector to obtain clean impedance data, standardize the temperature and humidity data, the clean multispectral data and the clean impedance data, and splice them into a unified feature vector F;

[0027] S103. Input the feature vector F into the multi-modal cross-attention model fine-tuned by transfer learning to fuse the features of each modality, and output four indices including the water-oil balance index, the pigment deposition index, the texture roughness index and the inflammation risk index; obtain the four indices for display control, and when it is recognized that the inflammation risk index exceeds the preset threshold, trigger the voice interaction module of the beauty terminal to broadcast a warning and limit the subsequent physiotherapy output.

[0028] It should be noted that in this embodiment, a multispectral imaging module and an electrical impedance analysis module are deployed simultaneously in the same detection head. Multispectral images can reflect the differences in skin pigment, hemoglobin and collagen absorption in multiple narrowband channels in the visible and near-infrared, and the impedance spectrum can characterize the electrical properties such as the water content of the stratum corneum and the capacitance of the cell membrane; after the two complement each other, four types of physiological indicators, namely water, oil, pigment, texture and inflammation, can be inferred simultaneously in a single detection, getting rid of the dilemma of users going back and forth between multiple devices. In addition, in this embodiment, optical and electrical dual sampling is completed within the same time reference T0, and the sampling time difference is strictly limited by a preset window; then a frame of data with a timestamp closest to T0 is selected in the temperature and humidity acquisition module and bound to the above two types of data. Such a synchronization mechanism ensures that all modalities reflect the same physiological transient, improving the repeatability and credibility of cross-time comparison. In addition, in this embodiment, the synchronous temperature and humidity are included in the unified feature vector F as the third mode, and dark current and flat field correction are performed on the spectrum in the preprocessing stage, and wavelet denoising is performed on the impedance, and then unified and standardized, so as to eliminate the fixed noise of the sensor at the physical level, and introduce temperature and humidity deviation at the algorithm level, so that the subsequent model can automatically learn the coupling relationship between the environment and skin characteristics, and enhance regional versatility. In addition, in this embodiment, the inflammation risk index is output in real time after model reasoning, and the results exceeding the threshold are broadcast and warned, and the subsequent energy output is limited to form a closed loop of detection, evaluation and protection, and eliminate the hidden danger of secondary stimulation. Moreover, this embodiment adopts a multimodal cross-attention model fine-tuned by transfer learning. The model can automatically assign weights to spectral, impedance, and environmental features, and mine deep coupling information; transfer learning uses the prior knowledge pre-trained by large-scale public skin data sets to shorten the local annotation cycle and improve the adaptability to different skin colors and illumination.

[0029] It should also be noted that, in this embodiment, the preset band range may be 400nm-1000nm, the preset number of channels may be 16 narrowband channels; the preset window based on T0 may be a T0±5ms window; in the preset signal frequency and peak current, the signal frequency may be 0.1kHz-500kHz, and the peak current ≤200µA; the preset acquisition frequency may be 10Hz. Among them, 400nm-1000nm covers the absorption peaks of melanin, hemoglobin, and collagen; 16 channels ensure the balance between spectral resolution and hardware cost; 0.1kHz-500kHz sweep energy spectrum covers the change of conductive properties from extracellular fluid to intracellular fluid; 200µA peak current meets the limit of skin safety current; 10Hz temperature and humidity sampling frequency is sufficient to capture rapid fluctuations in indoor environment. The T0±5ms synchronization window ensures that the spectral frame and impedance sweep are in the same cardiac cycle to reduce errors.

[0030] See also Figure 3When obtaining the four indices for display control, it may include: writing the four indices together with the timestamp T0 into the data storage module of the beauty terminal, and controlling the four indices stored in the data storage module to be displayed on the touch display screen 10 of the beauty terminal in a card-style UI. Among them, the card-style UI instead of a list or a line graph can display the levels, colors, and brief suggestions of the four indices at one time on a limited handheld screen, reducing the reading threshold and shortening the decision-making time. The UI card can also carry the interactive logic of click-and-jump, facilitating users to view historical curves or share them to the mobile APP.

[0031] In addition, when inputting the feature vector F into the multi-modal cross-attention model fine-tuned by transfer learning to fuse the features of each modality, it includes: dividing the unified feature vector F into three parts: spectral feature, impedance feature, and environmental feature according to a predefined offset; respectively using three sets of fully connected networks to map the three parts of the features to obtain spectral embedding vectors, impedance embedding vectors, and environmental embedding vectors with the same size, and adding sine-cosine position encoding based on the channel order after each embedding vector to retain the original order information; splicing the spectral embedding vectors, impedance embedding vectors, and environmental embedding vectors with the same size into a vector sequence in a predetermined order, and sending the spliced vector sequence into a multi-layer cross-attention Transformer; in each layer of the Transformer, the network calculates the attention weights between the spectrum and impedance, the spectrum and environment, and the impedance and environment respectively, and dynamically fuses the cross-correlation information of the three modalities through residual connection and layer normalization; setting a CLS marker vector representing the overall information at the very front of the spliced sequence, and after passing through the cross-attention layer, extracting the updated CLS vector as the fused global skin feature; inputting the global skin feature into four parallel regression branches at the same time, each branch consists of two-level fully connected networks and a Sigmoid activation function, and outputs the water-oil balance index, pigment deposition index, texture roughness index, and inflammation risk index respectively, and the value of each index is limited between zero and one. In this embodiment, the multi-modal cross-attention model fine-tuned by transfer learning can fully explore the deep correlation relationship between the spectral, impedance, and environmental features in the same inference, and then give the four skin health indices at the same time, so as to provide a high-precision and real-time evaluation basis for the display, storage, and safe power linkage of the beauty terminal.

[0032] In some preferred embodiments, the multispectral imaging module includes an EEPROM for storing spectral correction coefficients and a reference light source interface. After each power-on, the main control module first reads the spectral correction coefficients in the EEPROM and performs a CRC check; if the check fails or it is detected that the cumulative working duration exceeds a preset threshold, the main control module controls the integrating sphere reference light source built in the skin detection function head to sequentially output a white field and a dark field signal through the reference light source interface, calls the data preprocessing unit to complete the two-point calibration of the spectral correction coefficients, and rewrites the updated spectral correction coefficients back into the EEPROM. At the same time, a calibration progress bar is displayed on the touch display screen of the beauty terminal until the calibration is completed, and then the frame-by-frame acquisition process of the skin reflection grayscale image is entered. It should be noted that in this embodiment, an EEPROM for storing spectral correction coefficients and an integrating sphere two-point calibration mechanism are introduced for the metrological stability of the multispectral channels drifting over time. After the main control module is powered on, it reads the gain and bias matrices in the EEPROM and performs a CRC check. If it is found that the correction coefficients are damaged or the cumulative working duration exceeds the threshold, it automatically calls the integrating sphere white field and dark field calibration to correct the errors caused by LED aging, optical path contamination, or CMOS response attenuation in real time. The updated coefficients are written back to the EEPROM and the progress is displayed in real time on the UI to prevent users from misinterpreting the device as malfunctioning. This can control the absolute error of the band reflectance within 2%, solving the core problem that home devices lack self-calibration and become inaccurate after long-term use. Among them, the integrating sphere two-point calibration refers to: using two extreme conditions, namely the reference white field (high-reflection standard) and the reference dark field (zero optical signal), by measuring the output responses of the multispectral channels in these two states, a gain and bias correction model at the channel level is established, so as to complete the calibration of the spectral sensing system.

[0033] In some preferred embodiments, the electrical impedance analysis module uses replaceable patch electrodes. Before performing impedance acquisition, the main control module applies a detection pulse of 10 kHz and 20 µA to the patch electrodes and measures the open-circuit impedance; when the measured impedance value is higher than the preset aging threshold, the main control module stops the subsequent stepped frequency sweep process, pops up a prompt on the touch display screen indicating that the electrodes need to be replaced due to aging, and broadcasts the replacement prompt content through the voice interaction module. It should be noted that the interface electrochemical impedance of replaceable patch electrodes (such as silver-silver chloride electrodes) increases after multiple uses, which will cause distortion of the high-frequency impedance spectrum. By measuring the open-circuit impedance with a low-amplitude probing pulse and comparing it with the factory-calibrated threshold, the health status of the electrodes can be determined before the formal frequency sweep. If the threshold is exceeded, the acquisition is paused and the user is prompted to replace the electrodes using both voice and UI, avoiding incorrect water-oil index or inflammation index output due to electrode deterioration.

[0034] In some preferred embodiments, when the temperature and humidity acquisition module detects that the skin contact point temperature is higher than 40°C or the relative humidity is lower than 20%, the main control module automatically reduces the LED duty cycle of the multispectral imaging module by 30% and shortens the exposure time of the CMOS sensor by 25%, and adds a set of environmental compensation vectors {ΔT, ΔRH} to the eigenvector F, where ΔT and ΔRH are the deviation values of the current temperature and humidity from the preset standard environment, respectively, for the multi-modal model skin state evaluation model to dynamically correct the environmental impact during the inference stage. It should be noted that this embodiment proposes an integrated solution of dynamic power reduction, exposure adjustment, and environmental compensation vector for the double interference of LED heating and skin dehydration on the measurement results in a high-temperature and low-humidity environment. First, at the physical level, the LED duty cycle is reduced by 30% and the exposure is shortened by 25% to avoid overheating of the chip and image saturation; then, at the algorithm level, ΔT and ΔRH are injected into the eigenvector F as additional dimensions, so that the multi-modal model skin state evaluation model can learn the mapping between temperature, humidity and skin response during inference, thus realizing domain adaptation, which can not only prevent hardware damage but also ensure the model accuracy, solve the problem of measurement drift caused by environmental changes, and ensure that the pigment index and water-oil index are controlled within an interpretable range under high and low temperature differences.

[0035] In some preferred embodiments, the multispectral imaging module adopts a variable aperture structure with a stepper motor during the acquisition process. The main control module calculates the target exposure value in real time according to the external illuminance detected by the environmental photosensitive unit of the beauty terminal, and adjusts the blade angle of the variable aperture through closed-loop PID control of the stepper motor, so that the average gray value of the skin reflection grayscale image of the preset number of channels is maintained between 128±5, preventing spectral distortion caused by overexposure or underexposure; when the change rate of the external illuminance exceeds the threshold, the main control module suspends the frame-by-frame acquisition process of the skin reflection grayscale image until the PID output is stable. It should be noted that in this embodiment, by using a variable aperture and real-time PID closed-loop control of exposure, the mean value of the 16-channel grayscale histogram can be locked at 128±5, ensuring the stability of the spectral signal-to-noise ratio. The environmental photosensitive unit feeds back the illuminance in real time, and the PID controls the stepper motor to adjust the blade angle to suppress overshoot. If the light changes suddenly and exceeds the threshold, the acquisition is suspended to ensure that the generated multispectral image has no serious exposure drift, thereby improving the stability of the pigment deposition index and the texture roughness index.

[0036] In some preferred embodiments, all data frames from the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module are marked with a 32-bit hardware timestamp before waiting for the splicing of the feature vector F. The main control module uses a frame alignment tolerance of up to 1 ms. If the data delay of any one of the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module exceeds the tolerance, the corresponding frame is discarded and the corresponding module is re-triggered for acquisition until the time stamp differences of the three types of data of the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module meet the tolerance before the feature vector F is spliced. It should be noted that in this embodiment, the 32-bit hardware timestamp and the 1 ms frame alignment tolerance are used to ensure the time domain consistency of the multimodal data. When the main control module retrieves the frame, if the maximum and minimum differences > 1 ms, the frame is discarded and re-acquired to ensure that the spectral frame, the impedance spectrum, and the temperature and humidity data reflect the same cardiac cycle and the same respiratory phase, avoiding the situation where the spectrum is viewed 0.5 seconds ago and the impedance is viewed now, which may lead to distortion in the discrimination of the multimodal model skin state assessment model.

[0037] In some preferred embodiments, the multimodal model skin state assessment model supports federated learning incremental update. After the user authorizes, the main control module periodically performs differential privacy processing on the local gradient update ΔW and uploads it to the federated server. The federated server completes the aggregation to obtain the global model parameter W and pushes it back to the main control module through OTA. After the main control module verifies the model signature, it replaces the old weight and records the version number; if the signature verification fails, it rolls back to the previous stable version and reports the error log. It should be noted that in this embodiment, by adding noise to the upload of ΔW (local gradient), the user's original data is protected from leaving the device; after the server aggregates, the new weight is sent via OTA and can be verified using RSA signature to ensure that the model has not been tampered with. This embodiment enables the model to continuously iterate with the diversity of global user skin colors, ages, and environments, improving the generalization ability and solving the problem of poor adaptability of shallow algorithms. Among them, OTA refers to over-the-air download, which can remotely download and update the software or model parameters on the device through wireless networks (such as Wi-Fi, Bluetooth, NFC, etc.) without the user having to manually disassemble the device or connect to a computer for upgrading.

[0038] In some preferred embodiments, the main control module also communicates with an intelligent power management unit. The intelligent power management unit monitors the battery power, temperature, and internal resistance in real time. When it detects that the remaining power is less than 15% or the battery temperature is higher than 45°C, the main control module reduces the LED output power of the multi-spectral imaging module by 50%, shortens the sweep frequency step size, and limits the brightness of the touch display screen to 60%. At the same time, a low power or high temperature warning bar is displayed at the top of the UI. If the temperature continues to rise to 55°C, the main control module enters the sleep state. It should be noted that in this embodiment, when the remaining power < 15% or the battery temperature > 45°C, it triggers the reduction of LED power, the reduction of sweep frequency speed, and the reduction of screen brightness. When the temperature > 55°C, it directly enters the sleep state, reducing the instantaneous power consumption and heat generation while ensuring the detection integrity, and preventing the voltage drop caused by the increase of battery internal resistance from affecting the LED spectrum stability or the impedance constant current accuracy. The UI warning bar prompts the user to charge or cool in real time, further improving the usability and safety of the beauty terminal in the daily high-frequency usage scenarios.

[0039] In some preferred embodiments, before each detection process starts, the main control module calls the attitude sensing unit to detect the inclination angle of the skin detection function head relative to the gravity direction. When the inclination angle is greater than the inclination angle threshold or it detects that the acceleration mutation exceeds the acceleration threshold, the main control module pauses the LED lighting of the multi-spectral imaging module, and at the same time prompts the user that the fitting angle is abnormal through the touch display screen and the voice interaction module. It should be noted that before the detection, the attitude sensing unit is used to judge that the inclination angle is greater than the inclination angle threshold. If the inclination angle > 30° or the acceleration > the acceleration threshold, the LED lighting is paused and the user is prompted to correct the posture, thereby solving the problems of oblique illumination of the light spot and poor contact of the impedance electrode caused by the user's excessive holding angle, and also preventing the direct irradiation of the eyes by the high-power LED during a fall. The acceleration threshold can be based on the empirical value of the drop test, and the inclination angle threshold can be based on the result of the optical path geometric simulation.

[0040] In some preferred embodiments, when the inflammation risk index exceeds the preset threshold and all three consecutive detections are high-risk, the main control module automatically generates a comprehensive report file including the detection trends of the last seven days, environmental parameters, and the user's self-reported symptoms, provides a QR code download interface through the touch display screen, and embeds an integrity verification code based on SHA-256 in the comprehensive report file. It should be noted that when the inflammation risk index exceeds the threshold three consecutive times, a seven-day comprehensive report is automatically generated and the SHA-256 verification code is embedded for downloading through the QR code. Among them, the three consecutive judgments can filter out accidental false alarms; the seven-day window covers a skin barrier update cycle, which is convenient for evaluating the curative effect. The report contains environmental data and the user's self-report, plus the SHA-256 verification, which can ensure that the report has not been modified and ensure the objectivity of the evaluation.

[0041] In some preferred embodiments, the multispectral imaging module supports an optional extended band adapter. The extended adapter is built-in with an InGaAs near-infrared detector array in the range of 1000nm - 1700nm and a corresponding filter set. When the main control module detects the insertion of the extended adapter and reads a legal ID through I²C, it extends the band scanning table to 24 channels, adds a near-infrared water absorption peak correction algorithm in the data preprocessing unit, and updates the dimension mapping table of the feature vector F. It should be noted that in this embodiment, by using a pluggable InGaAs near-infrared extended detector, the band is extended to 1700nm, which can detect the second-order absorption peak of water and the collagen scattering peak, improving the evaluation accuracy of deep moisture and collagen. Reading the legal ID through I²C prevents unofficial adapters from damaging the host; the 24-channel scanning table and the near-infrared water correction algorithm are issued simultaneously to ensure consistent data dimensions.

[0042] In some preferred embodiments, the main control module reserves an OTA update interface and provides a firmware upgrade page in the mobile terminal APP. The firmware package is signed with RSA-2048 and transmitted through two-way TLS verification. During the upgrade process, the main control module first writes the new firmware into the standby partition and performs a CRC32 check. After the check passes, it switches the boot partition and restarts. If the boot fails, it automatically rolls back to the old version and records the reason for the exception. The upgrade log is sent to the APP through BLE broadcast for the user to query. It should be noted that in this embodiment, when writing the firmware, it is first written into the standby partition, and then the boot partition is switched after the CRC32 check passes, which can prevent the device from malfunctioning due to an interrupted upgrade; automatic rollback in case of failure and BLE broadcast of the log allow the user to know the status in a timely manner.

[0043] Embodiment 2

[0044] See Figures 1 - 3 , this embodiment provides an intelligent beauty terminal, including:

[0045] A skin detection function head, detachably installed on the function head connection seat at the front end of the handle assembly of the beauty terminal. The skin detection function head internally integrates a multispectral imaging module, an impedance analysis module, and a temperature and humidity acquisition module;

[0046] The main control module is disposed inside the handle assembly and runs the skin condition assessment method based on the skin detection function head. The main control module supplies power to the multi-spectral imaging module, the impedance analysis module, and the temperature and humidity acquisition module through the electrical contacts in the function head connector and completes data interaction. This skin condition assessment method drives the multi-spectral imaging module in the skin detection function head to sequentially collect skin reflection grayscale images frame by frame in a preset band range according to the narrowband channels with a preset number of channels, stacks the obtained skin reflection grayscale images according to the time stamp T0 to generate a multi-spectral skin image of the narrowband channels with a preset number of channels, and drives the impedance analysis module to apply a stepped sweep constant current signal with a preset signal frequency and peak current to the detection site within a preset window based on T0, synchronously measures the AC voltage, and uses a phase-locked algorithm to calculate the real and imaginary parts of the impedance at each frequency point of the AC voltage to obtain an impedance spectrum vector. After the temperature and humidity acquisition module in the skin detection function head outputs the skin contact temperature and relative humidity at a preset acquisition frequency, selects the temperature and humidity data with the time stamp closest to T0, and correlates it with the data of the multi-spectral skin image and the data of the impedance spectrum vector, calls the data preprocessing unit to perform dark current correction and flat field correction on the multi-spectral image to obtain clean multi-spectral data, performs wavelet band-pass denoising on the impedance spectrum vector to obtain clean impedance data, standardizes the temperature and humidity data, the clean multi-spectral data, and the clean impedance data and splices them into a unified feature vector F, inputs the feature vector F into a multi-modal cross-attention model fine-tuned by transfer learning to fuse the features of each modality, and outputs four indexes including the water-oil balance index, the pigment deposition index, the texture roughness index, and the inflammation risk index; obtains the four indexes for display control, and when it is recognized that the inflammation risk index exceeds a preset threshold, triggers the voice interaction module of the beauty terminal to broadcast a warning and limit the subsequent physiotherapy output, so as to realize the synchronous acquisition of visible and near-infrared multi-channel skin images and impedance spectra, form a detection, evaluation, and protection closed loop, and improve the safety and evaluation accuracy of intelligent beauty.

[0047] It should be noted that the above embodiments are only preferred specific embodiments of the present invention, and the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. The protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A skin condition evaluation method based on a skin detection function head, which runs on the main control module of a beauty terminal, is characterized in that, The skin condition assessment method includes: Driving the multi-spectral imaging module in the skin detection function head to sequentially capture skin reflection grayscale images frame by frame in narrowband channels with a preset number of channels within a preset wavelength range, stacking the obtained skin reflection grayscale images according to the timestamp T0 to generate a multi-spectral skin image of narrowband channels with a preset number of channels, and driving the impedance analysis module within a preset window based on T0 to apply a stepped sweep constant current signal with a preset signal frequency and peak current to the detection site, synchronously measuring the AC voltage and using a phase-locked algorithm to calculate the real and imaginary parts of the impedance at each frequency point of the AC voltage to obtain an impedance spectrum vector; After the temperature and humidity acquisition module in the skin detection function head outputs the skin contact point temperature and relative humidity at a preset acquisition frequency, select the temperature and humidity data with the timestamp closest to T0 and associate them with the data of the multi-spectral skin image and the data of the impedance spectrum vector. Call the data preprocessing unit to perform dark current correction and flat field correction on the multi-spectral skin image to obtain clean multi-spectral data, perform wavelet band-pass denoising on the impedance spectrum vector to obtain clean impedance data, standardize the temperature and humidity data, the clean multi-spectral data, and the clean impedance data and splice them into a unified feature vector F; According to a predefined offset, divide the unified feature vector F into three parts: spectral feature, impedance feature, and environmental feature; respectively use three sets of fully connected networks to map the three parts of the features to obtain spectral embedding vectors, impedance embedding vectors, and environmental embedding vectors with the same size, and add sine-cosine position encoding based on the channel order after each embedding vector to retain the original order information; splice the spectral embedding vectors, impedance embedding vectors, and environmental embedding vectors with the same size into a vector sequence in a predefined order, and send the spliced vector sequence into a multi-layer cross-attention Transformer; in each layer of the Transformer, the network calculates the attention weights between the spectrum and impedance, the spectrum and environment, and the impedance and environment respectively, and dynamically fuses the cross-correlation information of the three modalities through residual connection and layer normalization; set a CLS marker vector representing the overall information at the very front of the spliced sequence, and after passing through the cross-attention layer, extract the updated CLS vector as the fused global skin feature; input the global skin feature into four parallel regression branches at the same time, each branch consists of two-level fully connected networks and a Sigmoid activation function, and outputs the water-oil balance index, pigment deposition index, texture roughness index, and inflammation risk index respectively. The value of each index is limited between zero and one. Obtain the four indices for display control and when it is recognized that the inflammation risk index exceeds a preset threshold, trigger the voice interaction module of the beauty terminal to broadcast a warning and limit the subsequent physiotherapy output.

2. The skin condition assessment method based on a skin detection function head according to claim 1, wherein When obtaining the four indices for display control, it includes: writing the four indices together with the timestamp T0 into the data storage module of the beauty terminal, and controlling the four indices stored in the data storage module to be displayed on the touch display screen of the beauty terminal in a card-style UI.

3. The skin condition assessment method based on the skin detection function head according to claim 1, wherein, The preset wavelength range is 400nm - 1000nm, and the preset number of channels is 16 narrowband channels; the preset window based on T0 is the T0 ± 5ms window; among the preset signal frequency and peak current, the signal frequency is 0.1kHz - 500kHz, and the peak current ≤ 200µA; the preset acquisition frequency is 10Hz frequency.

4. The skin condition evaluation method based on a skin detection function head according to claim 1, wherein The multispectral imaging module includes an EEPROM for storing spectral correction coefficients and a reference light source interface. After each power-on, the main control module first reads the spectral correction coefficients in the EEPROM and performs CRC verification; if the verification fails or it is detected that the cumulative working duration exceeds the preset threshold, the main control module controls the integrating sphere reference light source built into the skin detection function head to sequentially output white field and dark field signals through the reference light source interface, calls the data preprocessing unit to complete the two-point calibration of the spectral correction coefficients, and rewrites the updated spectral correction coefficients back into the EEPROM. At the same time, a calibration progress bar is displayed on the touch display screen of the beauty terminal until the calibration is completed, and then the frame-by-frame acquisition process of the skin reflection grayscale image is entered.

5. The skin condition assessment method based on a skin detection function head according to claim 1, characterized in that, The skin detection function head is detachably installed on the function head connection seat at the front end of the handle assembly of the beauty terminal. The multispectral imaging module, the impedance analysis module, and the temperature and humidity acquisition module are integrated inside the skin detection function head; the main control module is arranged inside the handle assembly. The main control module supplies power to the multispectral imaging module, the impedance analysis module, and the temperature and humidity acquisition module through the electrical contacts inside the function head connection seat and completes data interaction.

6. The skin condition assessment method based on a skin detection function head according to claim 1, characterized in that, The impedance analysis module uses replaceable patch electrodes. Before performing impedance acquisition, the main control module applies a detection pulse of 10kHz and 20µA to the patch electrodes and measures the open-circuit impedance; when the measured impedance value is higher than the preset aging threshold, the main control module stops the subsequent stepped sweep frequency process, pops up a prompt for electrode aging and replacement on the touch display screen, and broadcasts the replacement prompt content through the voice interaction module.

7. The skin condition assessment method based on a skin detection function head according to claim 1, wherein, When the temperature and humidity acquisition module detects that the skin contact point temperature is higher than 40°C or the relative humidity is lower than 20%, the main control module automatically reduces the LED duty cycle of the multispectral imaging module by 30% and shortens the exposure time of the CMOS sensor by 25%, and adds a set of environmental compensation vectors {ΔT, ΔRH} to the eigenvector F, where ΔT and ΔRH are the deviation values of the current temperature and humidity from the preset standard environment respectively, for the multi-modal model skin condition assessment model to dynamically correct the environmental impact during the inference stage.

8. The skin condition assessment method based on a skin detection function head according to claim 1, characterized in that, The multispectral imaging module adopts a variable aperture structure with a stepper motor during the acquisition process. The main control module calculates the target exposure value in real time according to the external illuminance detected by the environmental photosensitive unit of the beauty terminal, and adjusts the blade angle of the variable aperture through closed-loop PID control of the stepper motor, so that the average grayscale of the skin reflection grayscale image of the preset number of channels is maintained between 128 ± 5, preventing spectral distortion caused by overexposure or underexposure; when the change rate of the external illuminance exceeds the threshold, the main control module suspends entering the frame-by-frame acquisition process of the skin reflection grayscale image until the PID output is stable.

9. The skin condition assessment method based on the skin detection function head according to claim 1, wherein, Before the data frames from the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module are waiting to be spliced with the feature vector F, they are all marked with a 32-bit hardware timestamp. The main control module uses a frame alignment tolerance of up to 1 ms. If the data delay of any of the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module exceeds the tolerance, the corresponding frame is discarded and the corresponding module is re-triggered for acquisition until the time stamp differences of the three types of data of the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module meet the tolerance before the feature vector F is spliced.

10. An intelligent beauty terminal, characterized in that, Including: A skin detection function head, a function head connection seat detachably installed at the front end of the handle assembly of the beauty terminal. The skin detection function head integrates a multispectral imaging module, an electrical impedance analysis module, and a temperature and humidity acquisition module inside; A main control module, arranged inside the handle assembly, running the skin condition evaluation method based on the skin detection function head according to any one of claims 1-9. The main control module supplies power to the multispectral imaging module, the electrical impedance analysis module, and the temperature and humidity acquisition module through the electrical contacts in the function head connection seat and completes data interaction.

Images