Cross-skin color spectral calibration method, system and storage medium based on transfer learning

Abstract

The application discloses a cross-skin-color spectrum calibration method and system based on transfer learning, and a storage medium, relates to the technical field of image processing, and comprises the following steps: analyzing and obtaining the correlation between at least two specified features and storing the specified features as mark features; obtaining a pre-training model and adding a full connection layer to the pre-training model to form an intermediate model; freezing the layer parameters of the pre-training model, training the intermediate model, and analyzing the output accuracy; if the output accuracy exceeds a set threshold, using the intermediate model as a target model; if the output accuracy does not exceed the set threshold, confirming whether the pre-training model has the mark features through data analysis; if the mark features exist, confirming the channel where the mark features are located based on the data analysis, and strengthening the channel weight; if the mark features do not exist, unfreezing the set layer parameters of the pre-training model; training the corrected intermediate model until a target model is obtained; and the above strategy enables the generated target model to have better performance on the cross-skin-color spectrum calibration result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a cross-skin-color spectral calibration method and system based on transfer learning and a storage medium. BACKGROUND

[0002] Hyperspectral imaging (HSI) is an advanced method combining spectral analysis and imaging technology, which can simultaneously obtain spatial information and continuous spectral information of a target. In the field of skin detection, hyperspectral technology can be used for skin disease diagnosis, skin quality analysis, moisture / oil detection, melanin analysis, etc.

[0003] The inventors have found that in practical applications, when using hyperspectral images to detect and analyze the skin quality of a user, the skin color itself often interferes, for example, for black skin, the spectral reflectance change caused by the difference in skin color will affect the accuracy of various applications, including medical diagnosis and face recognition, etc., for example, when the skin color is too dark, it will affect the system's judgment of skin melanin. However, if a hyperspectral-based skin quality intelligent detection model is trained separately for different skin colors, on the one hand, it is difficult to collect training data samples under various skin color conditions, and on the other hand, a large amount of computing resources and time are needed to train the model, which is obviously not suitable in terms of timeliness and economy.

[0004] Transfer learning is a method of machine learning, whose core idea is to use the knowledge trained on one task (source domain) to improve the learning performance of another related but different task (target domain), in short, it is to use an already trained hyperspectral-based skin quality intelligent detection model, keep the previous convolutional layers and pooling layers unchanged, only fine-tune the weights and bias coefficients of the last fully connected layer (or replace the fully connected layer in specific cases) or part of the intermediate layers, significantly reducing the training time and data requirements of the new model. Due to the lack of hyperspectral image data samples of specific colors such as black skin, how to further reduce the data amount required by transfer learning while ensuring the recognition accuracy of the new model is a problem that needs to be solved. SUMMARY

[0005] To address the challenges of obtaining hyperspectral image data samples of different skin tones in existing technologies, which leads to poor performance of hyperspectral-based intelligent skin texture detection models and insufficient data samples during transfer learning, thus hindering further performance improvement of the new model, this application aims to provide a cross-skin tone spectral calibration method based on transfer learning. This method effectively improves training speed while ensuring the accuracy of the new model's output by introducing marker features and adjusting the feature weight coefficients related to these marker features in the pre-trained model. To implement the aforementioned cross-skin tone spectral calibration method based on transfer learning, this application also aims to propose a cross-skin tone spectral calibration system based on transfer learning. Finally, it proposes to protect a computer-readable storage medium, the specific scheme of which is as follows:

[0006] A cross-skin color spectral calibration method based on transfer learning, comprising:

[0007] Based on skin images of specific skin tones and their corresponding hyperspectral image data, analyze and obtain the correlation between at least two specified features and store them as marker features;

[0008] Obtain a pre-trained model generated based on the source data, and then add a fully connected layer with a set target task output to form an intermediate model;

[0009] Freeze the parameters of each layer of the pre-trained model, train the intermediate model, and analyze the output accuracy:

[0010] If the output accuracy exceeds the set threshold, the intermediate model is used as the target model.

[0011] If the output accuracy does not exceed the set threshold, data analysis will be used to confirm whether the pre-trained model possesses the aforementioned signature features.

[0012] If it exists, the channel containing the marker feature is confirmed based on the above data analysis, and the weight of the above channel is strengthened;

[0013] If it does not exist, unfreeze the parameters of the last layer of the pre-trained model;

[0014] The above-mentioned modified intermediate model is trained and the accuracy of the output results is verified. If the output accuracy does not exceed the set threshold, the pre-trained model is unfrozen layer by layer until the target model is obtained.

[0015] If the output accuracy exceeds the set threshold, the current intermediate model will be used as the target model.

[0016] By the technical solution, first, a specified feature in a determined relationship, i.e., a global feature, is introduced, and it is determined whether the feature already exists in a pre-training model, if it exists, it is strengthened through an attention mechanism to improve the accuracy of the output of the new model, and if it does not exist, it is optimized and trained through the way of layer-by-layer unfreezing the pre-training model parameters, and the above-mentioned phased training strategy enables the generated target model to have better performance on new tasks.

[0017] Further, the association relationship between the at least two specified features is analyzed and obtained and stored as a landmark feature, including:

[0018] A hyperspectral image with a specific skin color is obtained, and hyperspectral image data of a location of a task target and hyperspectral image data of an associated target region associated with the task target are determined;

[0019] The above-mentioned hyperspectral image data is normalized in a set two-dimensional coordinate system to convert into a feature array carrying a relative position relationship feature;

[0020] The association relationship between the task target and the associated target region is analyzed and obtained through a clustering algorithm to form the landmark feature;

[0021] The task target region includes a skin region whose spectral reflectance is affected by skin color difference;

[0022] The associated target region is defined as a skin region whose distribution position has a specific association relationship with the task target region.

[0023] Through the above technical solution, the association relationship between the melanin aggregation region and the normal skin region on the skin can be established, and since the skin colors of the two skin regions are the same when the above-mentioned hyperspectral image data is collected, the influence of skin color difference on the above-mentioned association relationship can be reduced, i.e., the above-mentioned landmark feature is not affected by skin color difference. Since the above-mentioned association relationship is far apart in space, the above-mentioned landmark feature is usually ignored or weakened when training based on source data and using a neural network, and thus by enhancing the above-mentioned landmark feature, the influence of skin color difference on the output result of skin hyperspectral data can be weakened in reverse.

[0024] Further, the cross-skin color spectral calibration method further includes:

[0025] A plurality of skin images with a specific skin color are obtained, and the positions of the task target region and its associated target region are determined;

[0026] The above-mentioned skin images are translated, rotated, or scaled, and the colors of the task target and its associated target region are synchronously adjusted to generate a plurality of training extended images.

[0027] By the above technical solution, the number of samples used for target model training can be maximized without damaging the landmark features, which helps to improve the performance of the target model generated by training.

[0028] Further, the data analysis confirms whether the pre-trained model has the landmark feature, including:

[0029] Calculate the correlation between different spatial position features in the pre-trained model to determine whether there are position-dependent features:

[0030] If there are features with position-dependent relationships, the input image is cropped and spliced, the relative position relationship between the task target and its associated target is adjusted, the change of feature response is observed, and whether the pre-trained model has the landmark feature is determined through the above feature response change.

[0031] Through the above technical solution, it can be known whether the pre-trained model already has the landmark feature, which is convenient for determining the parameter adjustment object and the channel of the pre-trained model.

[0032] Further, the association between the at least two specified features includes:

[0033] The position association relationship between the two specific hyperspectral image data position coordinates; and

[0034] The numerical association relationship between the hyperspectral image data in the two specific position coordinate regions.

[0035] Through the above technical solution, it is helpful to diversify the specified features and facilitate precise fine-tuning of the pre-trained model parameters.

[0036] A cross-skin color spectrum calibration system based on transfer learning, comprising:

[0037] A landmark feature acquisition unit configured to analyze and acquire the association between at least two specified features based on specific skin color skin images and their corresponding hyperspectral image data, and store it as a landmark feature;

[0038] An intermediate model generation unit configured to obtain a pre-trained model trained based on source data, and add a full connection layer with a set target task output to form an intermediate model;

[0039] A target model generation unit configured to train a target model based on the intermediate model and the source data;

[0040] The target model generation unit includes:

[0041] A first execution subunit configured to freeze the parameters of each layer of the pre-trained model, train the intermediate model and analyze the output accuracy.

[0042] The landmark feature detection subunit is configured to confirm whether the landmark feature exists in the pre-trained model through data analysis.

[0043] The accuracy determination subunit is configured to determine whether the intermediate model output accuracy exceeds a set value.

[0044] The second execution subunit is configured to determine the channel where the landmark feature is located, and to strengthen the channel weight to form a new intermediate model.

[0045] The third execution subunit is configured to thaw the pre-trained model layer by layer to form a new intermediate model for training.

[0046] The target model training output subunit is configured to be data-connected with the first execution subunit, the second execution subunit, the third execution subunit, and the accuracy determination subunit, to obtain the output data of each execution subunit, and if the output accuracy exceeds the set threshold, to take the current intermediate model as the target model and output and store it.

[0047] Further, the cross-skin color spectrum calibration system further includes a material expansion unit configured to obtain multiple skin images of a specific skin color, determine the positions of the task target region and its associated target region, and then perform translation, rotation or scaling on the skin images, and synchronously adjust the colors of the task target and its associated target region to generate multiple training expansion images.

[0048] Through the above technical solution, the number of training materials of the target model can be expanded in a targeted manner, and the accuracy of the target model output can be improved.

[0049] Further, the cross-skin color spectrum calibration system is configured in a cloud server, and the generated target model is loaded in a skin quality detector, and when in use, a processor in the skin quality detector calls program algorithms in a local storage device or a cloud server to realize related function output.

[0050] A computer-readable storage medium having a program algorithm loaded thereon, which is used to implement the cross-skin color spectrum calibration method based on transfer learning as described above when a processor executes the program algorithm.

[0051] The present application at least includes the following beneficial effects:

[0052] By introducing the landmark features existing in the determined relationship, it is determined whether the above landmark features already exist in the pre-trained model, if they exist, the landmark features are strengthened through the attention mechanism, the training and generation of the new model can be completed with less training materials, and the accuracy of the output of the new model can be improved; if they do not exist, the pre-trained model parameters are thawed layer by layer for optimization training, and the new full connection layer and the thawed layer are used for stage-by-stage training, and the above strategy enables the generated target model to have better performance on the new task. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A schematic diagram of the cross-skin color spectral calibration method of the present application;

[0054] Figure 2 A schematic diagram of the method for obtaining landmark features of the present application;

[0055] Figure 3 A schematic diagram of the functional modules of the system of the present application.

[0056] Reference signs: 100, landmark feature acquisition unit; 200, intermediate model generation unit; 300, target model generation unit; 301, first execution subunit; 302, landmark feature detection subunit; 303, accuracy determination subunit; 304, second execution subunit; 305, third execution subunit; 306, target model training output subunit. DETAILED DESCRIPTION

[0057] The embodiments of the present application will be described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0058] In the description of the present specification, the description of the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0059] A cross-skin color spectral calibration method based on transfer learning, as shown in Figure 1 The method mainly includes the following steps:

[0060] S100, based on the specific skin color skin image and its corresponding hyperspectral image data, the correlation between at least two specified features is analyzed and stored as a landmark feature;

[0061] S200, obtaining a pre-training model trained based on source data, and adding a full connection layer with a set target task output to form an intermediate model;

[0062] S300, freezing parameters of each layer of the pre-training model, training the intermediate model, and analyzing an output accuracy rate:

[0063] S310, if the output accuracy rate exceeds a set threshold, using the intermediate model as a target model;

[0064] S320, if the output accuracy rate does not exceed the set threshold, confirming whether the pre-training model has the flag feature through data analysis:

[0065] S3201, if yes, confirming a channel where the flag feature is located based on the data analysis, and strengthening a weight of the channel;

[0066] S3202, if no, unfreezing a parameter of a last layer of the pre-training model;

[0067] S330, training the modified intermediate model and verifying an accuracy rate of an output result:

[0068] S3301, if the output accuracy rate does not exceed the set threshold, unfreezing the pre-training model layer by layer from back to front until a target model is obtained;

[0069] S3302, if the output accuracy rate exceeds the set threshold, using the current intermediate model as the target model and storing the target model.

[0070] In step S100, the association relationship between the at least two specified features includes a position association relationship between position coordinates of two specific hyperspectral image data, and a numerical association relationship between the hyperspectral image data in two specific position coordinate regions.

[0071] In the embodiments of the present application, the association relationship between the at least two specified features is analyzed and stored as a flag feature, as shown in the following table: Figure 2

[0072] S110, obtaining hyperspectral image data of skin with a specific skin color, determining hyperspectral image data of a position where a task target is located and hyperspectral image data of an associated target region related to the task target.

[0073] ​The task target area includes a skin area whose spectral reflectance is affected by skin color difference, for example, the hyperspectral image data of a melanin aggregation area is affected by black skin, and the difference between skin colors has a greater impact on the hyperspectral image data of the area. In practical applications, in addition to the impact of black skin on the spectral data of the melanin aggregation area, red skin (such as plateau red) also has an impact on the spectral data of the rosacea aggregation area or the allergic area.

[0074] Unlike the distribution of the lesion area of melanin aggregation and skin surface acne, the distribution of the color of the skin itself is generally uniform, and the distribution of the lesion area generally has a certain regularity, for example, the degree of vascular dilation around rosacea is obviously abnormal, and the vascular dilation area is not completely coincident with the acne area, so that the interference of skin color in the lesion area recognition process can be excluded by analyzing the relative positional relationship between the vascular dilation area and the acne area. In a specific case, the relative positional relationship may have a large interval in space, and the pre-training model generated based on source data of a single skin color often ignores or weakens the existence of the relative positional relationship, causing the pre-training model to be easily affected by skin color when identifying specific skin parameters such as melanin, thereby reducing the universality of related skin detection equipment.

[0075] S120, normalizing the hyperspectral image data in a set two-dimensional coordinate system to convert it into a feature array carrying a relative positional relationship feature.

[0076] S130, analyzing the association between the task target and the associated target area by a clustering algorithm to form the landmark feature.

[0077] Steps S120-S130 first establish a correspondence between the hyperspectral image data and the position coordinate data, and then exclude interference by a clustering algorithm to find the most core and highest reliability relative positional relationship feature. The above process can be implemented based on sample data of different skin colors and is not limited by the amount of sample data. In a specific case, the collection and confirmation of the landmark feature can also be completed by a small amount of hyperspectral image data of skin of a specific skin color.

[0078] In the scheme of the present application, since the application field is user facial skin quality detection, the associated target area is defined as a skin area whose distribution position has a specific association with the task target area, such as a vascular dilation area related to the position of the acne area.

[0079] The landmark feature formed by the above technical solution can inversely weaken the impact of skin color difference on the output result of skin hyperspectral data.

[0080] Further, in step S320, confirming whether the mark feature exists in the pre-trained model through data analysis further includes:

[0081] S321, calculating the correlation between different spatial position features in the pre-trained model to determine whether there are position-dependent features:

[0082] S322, if there are position-dependent features, the input image is cropped and spliced, the relative position relationship between the task target and its associated target is adjusted, and the change of the feature response is observed. Whether the mark feature exists in the pre-trained model is determined through the above feature response change.

[0083] Since the internal features and their relationships cannot be directly accessed or consulted during the training process of the CNN model, the above technical solution can indirectly determine whether the mark feature or the approximate feature already exists in the pre-trained model, and through repeated verification, the channel where the feature is located can be known, which facilitates precise adjustment of the parameters of the related pre-trained model and further shortens the training time of the target model.

[0084] In order to expand the materials for training new models, the cross-skin color spectrum calibration method of the application further includes a data expansion step:

[0085] S3303, obtaining multiple skin images of a specific skin color, and determining the positions of the task target region and its associated target region;

[0086] S3304, performing translation, rotation or scaling on the above skin images, synchronously adjusting the color of the task target and its associated target region, and generating multiple expanded images for training.

[0087] Through the above technical solution, the number of samples used for training the target model can be maximized without damaging the mark feature, which helps to improve the performance of the target model generated by training.

[0088] To implement the above cross-skin color spectrum calibration method based on transfer learning, the embodiment of the application further discloses a cross-skin color spectrum calibration system based on transfer learning, as shown in Figure 3 The system mainly includes a mark feature acquisition unit 100, an intermediate model generation unit 200, a target model generation unit 300, and a material expansion unit.

[0089] The mark feature acquisition unit 100 is configured to analyze and acquire the correlation between at least two specified features based on the specific skin color skin image and its corresponding hyperspectral image data, and store the correlation as a mark feature. The mark feature acquisition unit 100 has a data normalization processing subunit and a clustering algorithm subunit built-in, which is used to acquire the correlation between at least two specified features from the normalized hyperspectral image data.

[0090] The intermediate model generating unit 200 is configured to obtain a pre-trained model trained based on source data, and then add a full connection layer with a set target task output to form an intermediate model.

[0091] The target model generating unit 300 is configured to train a target model based on the intermediate model and the source data.

[0092] The material expansion unit is configured to obtain multiple skin images of a specific skin color, determine the positions of the task target region and the associated target region, and then perform translation, rotation or scaling on the skin images, and synchronously adjust the colors of the task target region and the associated target region to generate multiple expanded images for training.

[0093] In further detail, the target model generating unit 300 includes a first execution subunit 301, a landmark feature detection subunit 302, an accuracy determination subunit 303, a second execution subunit 304, a third execution subunit 305, and a target model training output subunit 306.

[0094] The first execution subunit 301 is configured to freeze the parameters of each layer of the pre-trained model, train the intermediate model, and analyze the output accuracy. The landmark feature detection subunit 302 is configured to confirm whether the pre-trained model has the landmark feature through data analysis. The accuracy determination subunit 303 is configured to determine whether the output accuracy of the intermediate model exceeds a set value. The second execution subunit 304 is configured to determine the channel where the landmark feature is located, and strengthen the weight of the channel to form a new intermediate model. The third execution subunit 305 is configured to thaw the pre-trained model layer by layer to form a new intermediate model for training. The target model training output subunit 306 is data-connected with the first execution subunit 301, the second execution subunit 304, the third execution subunit 305, and the accuracy determination subunit 303, obtains the output data of each execution subunit, and if the output accuracy exceeds a set threshold, stores the current intermediate model as the target model and outputs it.

[0095] In actual application, the execution order of the actions of the above-mentioned functional units can be executed according to the aforementioned cross-skin color spectrum calibration method, or can be selectively executed or exchanged according to actual conditions. For example, when the training materials are sufficient, the related functions of the material expansion unit are not executed.

[0096] A computer readable storage medium, which is loaded with a program algorithm, when a processor executes the program algorithm, is used to implement the cross-skin color spectral calibration method based on transfer learning as described above. The program algorithm is preferably stored in a cloud server, and the processor is configured in a skin texture detector. When in use, the processor calls the program algorithm in the local storage device or the cloud server to realize the related function output.

[0097] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A cross-skin color spectral calibration method based on transfer learning, characterized in that, include: Based on skin images of specific skin tones and their corresponding hyperspectral image data, analyze and obtain the correlation between at least two specified features and store them as marker features; Obtain a pre-trained model generated based on the source data, and then add a fully connected layer with a set target task output to form an intermediate model; Freeze the parameters of each layer of the pre-trained model, train the intermediate model, and analyze the output accuracy: If the output accuracy exceeds the set threshold, the intermediate model is used as the target model. If the output accuracy does not exceed the set threshold, data analysis will be used to confirm whether the pre-trained model possesses the aforementioned signature features. If it exists, the channel containing the marker feature is confirmed based on the above data analysis, and the weight of the above channel is strengthened; If it does not exist, unfreeze the parameters of the last layer of the pre-trained model; The above-mentioned modified intermediate model is trained and the accuracy of the output results is verified. If the output accuracy does not exceed the set threshold, the pre-trained model is unfrozen layer by layer until the target model is obtained. If the output accuracy exceeds the set threshold, the current intermediate model will be used as the target model. This includes analyzing and obtaining the association between at least two specified features and storing them as flag features, including: Acquire hyperspectral images of skin with specific skin tones to determine the location of the mission target and the hyperspectral image data of associated target regions related to the mission target; The hyperspectral image data is normalized in a set two-dimensional coordinate system and transformed into a feature array carrying relative positional relationship features. The association between the task target and related target regions is obtained by clustering algorithm analysis, forming the aforementioned marker features; The target area for the task includes skin areas whose spectral reflectance is affected by differences in skin color. The associated target region is defined as a skin region whose distribution location has a specific association with the task target region.

2. The cross-skin color spectral calibration method based on transfer learning according to claim 1, characterized in that, The cross-skin color spectral calibration method further includes: Acquire multiple skin images of a specific skin tone to determine the location of the target region and its associated target regions; The skin images are translated, rotated, or scaled, and the colors of the target and its associated target regions are adjusted synchronously to generate multiple extended training images.

3. The cross-skin color spectral calibration method based on transfer learning according to claim 2, characterized in that, Data analysis is used to confirm whether the pre-trained model possesses the aforementioned signature features, including: Calculate the correlation between features at different spatial locations in the pre-trained model to determine whether there are features with location dependencies: If features with positional dependencies exist, the input image is cropped and stitched together, the relative positional relationship between the task target and its associated targets is adjusted, and the changes in feature responses are observed. The presence of the aforementioned marker features in the pre-trained model is determined based on these changes in feature responses.

4. The cross-skin color spectral calibration method based on transfer learning according to claim 1, characterized in that, The association between the at least two specified features includes: The positional correlation between the coordinates of two specific hyperspectral image data points; and Numerical correlation between hyperspectral image data in two specific location coordinate regions.

5. The cross-skin color spectral calibration method based on transfer learning according to claim 1, characterized in that, The cross-skin color spectral calibration method is applied to skin detection devices based on hyperspectral image data.

6. A cross-skin color spectral calibration system based on transfer learning, characterized in that, For implementing the cross-skin color spectral calibration method based on transfer learning as described in any one of claims 1-5, comprising: The feature acquisition unit (100) is configured to analyze and acquire the correlation between at least two specified features and store them as feature based on a skin image of a specific skin color and its corresponding hyperspectral image data; Intermediate model generation unit (200) is configured to obtain a pre-trained model generated based on source data, and then add a fully connected layer with a set target task output to form an intermediate model; The target model generation unit (300) is configured to train and generate a target model based on the intermediate model and the source data; The target model generation unit (300) includes: The first execution subunit (301) is configured to freeze the parameters of each layer of the pre-trained model, train the intermediate model, and analyze the output accuracy. The marker feature detection subunit (302) is configured to confirm whether the pre-trained model has the marker feature through data analysis; The accuracy determination subunit (303) is configured to determine whether the accuracy of the intermediate model output exceeds a set value; The second execution subunit (304) is configured to determine the channel where the flag feature is located and strengthen the channel weights to form a new intermediate model; The third execution subunit (305) is configured to unfreeze the pre-trained model layer by layer to form a new intermediate model for training; The target model training output subunit (306) is configured to be data-connected with the first execution subunit (301), the second execution subunit (304), the third execution subunit (305) and the accuracy determination subunit (303) to obtain the output data of each execution subunit. If the output accuracy exceeds the set threshold, the current intermediate model is used as the target model and output and stored.

7. The cross-skin color spectral calibration system based on transfer learning according to claim 6, characterized in that, The cross-skin color spectral calibration system also includes a material expansion unit, configured to acquire multiple skin images of specific skin colors, determine the position of the task target region and its associated target regions, and then translate, rotate or scale the skin images, synchronously adjust the color of the task target and its associated target regions, and generate multiple training expansion images.

8. The cross-skin color spectral calibration system based on transfer learning according to claim 7, characterized in that, The cross-skin color spectral calibration system is configured on a cloud server, and the generated target model is loaded into the skin texture analyzer. When in use, the processor in the skin texture analyzer retrieves program algorithms from the local storage device or the cloud server to implement relevant function outputs.

9. A computer-readable storage medium having a program algorithm loaded thereon, which, when executed by a processor, is used to implement the cross-skin color spectral calibration method based on transfer learning as described in any one of claims 1-5.

Images