KR101969346B1 - Apparatus for classifying skin conditions, and apparatus for generating skin condition classification model used on the same apparatus and method thereof - Google Patents

Abstract

An apparatus and method for generating a skin condition classifier and a skin condition classification model used in the apparatus are disclosed.
The structure storage unit of the skin condition classification model generation apparatus includes a plurality of classifiers including a learned dictionary for a plurality of skin images, a sparse representation extracted from the dictionary, and a classifier generated from the rare expression, Save the structure. The structure generating unit generates the plurality of classifier structures by using a sparse coding algorithm for data of a plurality of skin images and stores them in the structure storing unit. The final model generator generates a Boosting Sparse Coding Model, which is a final skin condition classification model, by applying an adoboosting algorithm to a plurality of classifiers of the structure storage unit. Using this final skin condition classification model, the skin condition classifier can perform more accurate skin condition classification on a new image.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a skin condition classification apparatus, a device for generating a skin condition classification model used in the apparatus,

The present invention relates to a skin condition classification apparatus, an apparatus and a method for generating a skin condition classification model used in the apparatus.

Image classification technology is a technique of classifying an input image into one of predetermined categories. The definition of this technology is very simple, but it has become one of the key issues in the field of computer vision that has various applications.

On the other hand, various methods for analyzing the state of the skin have been proposed in the fields of make-up counseling and medical care. The skin condition varies according to various environmental factors such as age and seasonal factors. Even if the same person is the same person, the skin condition can be changed according to the situation. Therefore, it is necessary to grasp the skin condition in a specific situation, Accurate skin condition analysis is required in order to properly manage the skin according to the skin condition.

In general, classification of the skin condition is performed by analyzing and classifying the skin image obtained through close-up of the skin. However, unlike other images, the skin image obtained through close-up photography is subtle in the change of 'color' and 'corner point'. Therefore, even if the same subject is photographed at the same place, the existing image feature extraction algorithm such as SIFT (Scale-Invariant Feature Transform), which is an algorithm for recognizing that the same subject is photographed in the same place, There are difficulties in extracting features of an image. For example, when the skin image feature is extracted from the skin image of FIG. 1 (a) as shown in FIG. 1 (b), it can be seen that the feature is extracted only where the pixel due to the fine hair changes.

Accordingly, a sparse coding algorithm for analyzing an image on a patch-by-patch basis to construct a dictionary containing information on oil / water / horny and representing an input image as a weight of bases is obtained through close-up photography It may be more effective in extracting features such as oil / moisture / keratin from the skin image.

However, in such a rare coding algorithm, assuming that the bases are independent among the bases in constructing the dictionary, they are not assumed. Hence, the sparse vector representation for the image is not unique, which leads to poor classification accuracy. Also, in the case of input image data, which is influenced by environmental factors (light, shadow) in the process of obtaining data, the error rate with respect to the sparse vector becomes large, so that the feature vector obtained by the weight of the bases does not accurately represent the original image data . As a result, it has a problem that it can not cope with various situations because it uses the bases of the dictionaries already configured.

To overcome these problems, an improved sparse coding algorithm with new constraints has been proposed. However, the proposed improved sparse coding algorithm uses local information of the image as in the case of Graph regularized sparse coding (GraphSC) or Discriminative sparse coding (DisScMM). For this reason, there is a problem in that an object and a background are not distinguished from each other in an image, and data having an ambiguous edge of an image shape is obtained rather than a conventional rare coding algorithm.

Therefore, there is a need for a classification technique that can improve the accuracy of the classification of skin images over the prior art.

SUMMARY OF THE INVENTION The present invention provides a skin condition classifier capable of improving the accuracy of skin condition classification, an apparatus for generating a skin condition classification model used in the apparatus, and a method thereof.

According to an aspect of the present invention,

A structure storage unit for storing a plurality of classifier structures constructed using a sparse coding algorithm for data of a plurality of skin images, each classifier structure comprising a dictionary learned for a plurality of skin images, A sparse representation extracted from a dictionary and a classifier generated from the sparse representation; And a skin condition classification model for a new skin image inputted for skin condition classification and a plurality of classifier structures stored in the structure storage section to classify the skin condition of the new skin image, Wherein the skin condition classification model is a Boosting Sparse coding model that applies an adaboosting algorithm to a plurality of classifiers stored in the structure storage unit and outputs a skin condition classification model, Coding Model).

Here, the skin condition classification model may include a process of generating a classifier structure including a dictionary, a rare expression, and a classifier by applying a sparse coding algorithm to data of a plurality of skin images repeatedly a predetermined number of times, Generating a skin condition classification model that generates a skin condition classification model that is a boosting sparse coding model by applying an adoboosting algorithm to a plurality of classifiers included in each of the plurality of classifiers, Device.

In addition, the skin image is a skin image in which a label for moisture, oil, keratin, and moisture content is determined.

According to another aspect of the present invention, there is provided an apparatus for generating a skin condition classification model,

A structure storage unit for storing a plurality of classifier structures including a dictionary learned for a plurality of skin images, a sparse representation extracted from the dictionary, and a classifier generated from the rare expression; A structure generation unit for generating the plurality of classifier structures by using a sparse coding algorithm for data of a plurality of skin images and storing the plurality of classifier structures in the structure storage unit; And a final model generation unit for generating a boosting sparse coding model, which is a final skin state classification model, by applying an adoboosting algorithm to a plurality of classifiers of the structure storage unit.

Here, the structure generation unit may include a dictionary construction unit that constructs a learned dictionary by using an alternating minimization algorithm for the data of the plurality of skin images, and stores the constructed dictionary in a corresponding classifier structure of the structure storage unit, ; A rare expression extracting unit for extracting a rare expression for data of the plurality of skin images using a dictionary formed by the pre-constructing unit and storing the extracted rare expression in a corresponding classifier structure of the structure storing unit; And a classifier generator for generating a classifier by applying a machine learning algorithm to the sparse expression extracted by the sparse expression extractor and storing the generated classifier in a corresponding classifier structure of the structure storage unit.

The pre-configuration unit may include: a patch matrix constructing unit configured to construct a patch matrix from the data of the plurality of skin images; And an intersection minimization unit for constructing a finally learned dictionary by performing learning using the cross minimization algorithm for the patch matrix constituted by the patch matrix constituent unit and storing the constructed dictionary in a corresponding classifier structure of the structure storage unit.

The patch matrix forming unit forms a plurality of patch data from the data of the plurality of skin images, and rearranges each of the plurality of patch data in a matrix form to constitute one patch matrix.

In addition, the cross minimization algorithm performed by the cross minimization performing unit may include: obtaining the dictionary using a Lagrange dual problem in a state in which a sparse vector is fixed, (feature sign search) algorithm to obtain the sparse vector until the convergence condition is satisfied, thereby constructing the finally learned dictionary.

The intersection minimization algorithm performed by the intersection minimization performing unit may be expressed by the following equation

    Here,? Is a dictionary and a is a rare vector

.

Also, the rare expression extracting unit may include a feature sign search performing unit that performs a feature sign search algorithm in advance, which is ultimately configured by the pre-configuration unit; And a sum pooling unit that performs sum pooling on the result data performed by the feature search unit and extracts a rare expression for each skin image and stores the extracted rare expression in a corresponding classifier structure of the structure storing unit, ) Performing unit.

In addition, the machine learning algorithm used by the classifier generator is a SVM (Support Vector Machine) algorithm.

In addition, the structure generating unit generates a plurality of classifiers through a plurality of iterations in a manner of generating one classifier structure including a dictionary, a rare expression, and a classifier through a single generating operation.

In addition, the classifier generator may generate the Adaboost parameter for each classifier by updating the weights of the patches according to the learning error rate for each iterative operation, and the final model generator generates a final skin condition Create a classification model.

The updating of the weights of the patches also includes updating the weights of the patches corresponding to the misclassified rare expressions and updating the weights for the patches initially drawn to generate the dictionary in the repeated operation.

Also, the final model generation unit updates the weights of the patches according to the learning error rate, and calculates the final skin condition classification model by obtaining the adaboasting weight for each classifier using the weights and the error rates of the patches.

)은 다음의 수학식In addition, the final skin condition classification model (

) Is expressed by the following equation

는 가중치이고,

는 분류기이며,

는 이미지이고,

는 지시함수이며, k는 클래스 라벨이고,here,

Is a weight,

Is a classifier,

Is an image,

Is an indicator function, k is a class label,

이며,

Lt;

임

being

.

According to another aspect of the present invention,

CLAIMS What is claimed is: 1. A method for generating a skin condition classification model used by a skin condition classification model generation device to classify a skin condition for a skin image, the method comprising: applying a sparse coding algorithm to data of a plurality of skin images to generate a dictionary, Generating a plurality of classifier structures by repeating a process of generating a classifier structure including a plurality of classifiers by a predetermined number of times; And generating a final skin condition classification model, which is a boosting sparse coding model, by applying an adaboosting algorithm to a plurality of classifiers respectively included in the plurality of classifiers.

Here, before the step of generating the plurality of classifier structures, a step of constructing a plurality of patch data from the data of the plurality of skin images and rearranging each of the plurality of patch data in a matrix form to form one patch matrix .

The generating of the plurality of classifier structures may include constructing a learned dictionary using data of the plurality of skin images using an alternating minimization algorithm and storing the learned data in a corresponding classifier structure; Extracting a sparse representation of the data of the plurality of skin images using a dictionary formed and storing the extracted sparse expressions in a corresponding classifier structure; Generating a classifier by applying an SVM (Support Vector Machine) algorithm to the extracted rare expression, and storing the classifier in a corresponding classifier structure; And constructing the learned dictionary by the predetermined number of times and storing the learned dictionary in a corresponding classifier structure, extracting the rare expression and storing the extracted rare expression in a corresponding classifier structure, and storing the generated classifier in a corresponding classifier structure And generating the plurality of classifier structures by repeating the steps.

In addition, the crossover minimization algorithm may be performed by using a Lagrange dual problem with the sparse vector fixed, obtaining the dictionary, and using the feature sign search algorithm with the dictionary fixed And the process of obtaining the sparse vector is repeated until the convergence condition is satisfied, thereby constructing the finally learned dictionary.

In addition, in the step of extracting the rare expression and storing it in a corresponding classifier structure, a feature sign search algorithm is performed in the dictionary, and then a sum pooling is performed to obtain a rare expression .

In addition, in the step of generating the plurality of classifier structures, the weights of the patches are updated according to the learning error rate for each iteration to generate the adaboost parameters for each classifier, and in the final skin condition classification model, A final skin condition classification model is created by linear combination of the AdaBoost parameters for each classifier.

According to another aspect of the present invention, there is provided a computer-

A computer-readable recording medium storing a program for performing a method for generating a skin condition classification model, the method comprising: generating a classifier structure including a dictionary, a rare expression, and a classifier by applying a rare coding algorithm to data of a plurality of skin images; Generating a plurality of classifier structures by repeatedly performing a predetermined number of times; And a function of generating a final skin condition classification model, which is a boosting sparse coding model, by applying an adaboosting algorithm to a plurality of classifiers respectively included in the plurality of classifiers.

According to the present invention, the skin condition classification becomes more robust to the user's environment (light, shadow, indoor / outdoor, etc.).

In addition, the feature vector is unique, and the classification error can be reduced.

Therefore, the accuracy with respect to the skin condition classification is improved as compared with the conventional technique.

1 is a view showing an example of skin image feature extraction according to the conventional SIFT algorithm.
2 is a diagram illustrating the concept of re-rendering an image patch as a rare weight vector using bases in a sparse coding algorithm in general.
FIG. 3 is a view showing a schematic configuration of a skin condition classifying apparatus according to an embodiment of the present invention.
4 is a schematic block diagram of an apparatus for generating a skin condition classification model according to an embodiment of the present invention.
5 is a view showing a configuration example of the structure storage unit shown in FIG.
Fig. 6 is a diagram showing a specific configuration of the pre-configuration unit shown in Fig. 4. Fig.
FIG. 7 is a view showing an example in which the patch matrix constitution unit shown in FIG. 6 constitutes a patch matrix.
8 is a diagram showing a specific configuration of the rare expression extracting unit shown in FIG.
FIG. 9 is a view schematically showing a method of updating weights of patches indirectly affecting misclassification according to an embodiment of the present invention.
10 is a flowchart of a method of generating a skin condition classification model according to an embodiment of the present invention.
11 is a specific flowchart of a step of creating the structure shown in FIG.
12 is a diagram showing an example of a skin image data sample input for skin condition classification in the embodiment of the present invention.
13 is a diagram showing an example of a dictionary learned according to the skin condition classification method according to the embodiment of the present invention.
FIG. 14 is a diagram schematically illustrating the concept of a matrix minimizing algorithm according to a skin condition classification method according to an embodiment of the present invention. Referring to FIG.
FIG. 15 is a view showing an example of a classifier structure constructed by the number of iterations according to the skin condition classification method according to the embodiment of the present invention.
16 is a diagram illustrating the accuracy according to the number of repetitions in the skin condition classification method according to the embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when an element is referred to as " comprising ", it means that it can include other elements as well, without excluding other elements unless specifically stated otherwise. Also, the terms " part, " " module, " and " module ", etc. in the specification mean a unit for processing at least one function or operation and may be implemented by hardware or software or a combination of hardware and software have.

First, in the existing sparse coding algorithm, a dictionary is constructed by extracting arbitrary patches from each learning image. At this time, the bases constituting the dictionary reinterpret the patch as a linear combination of bases when a new patch comes in. For example, if a new patch image can be represented with only a few baselines in an overcomplicated dictionary, the weights of all the bases except for the weight of the selected bases will be zero. Thus, the patch image is reinterpreted and represented as a very sparse weight vector. Reference can be made to FIG. 2, which illustrates the concept of re-rendering an image patch as a sparse weight vector using basis in a sparse coding algorithm. Here, a non-zero weight value can be referred to as 'feature extraction' for an image patch, meaning that it is expressed again as a sparse vector.

However, as described above, the existing sparse coding algorithm has a problem that the classification accuracy is low because the sparse vector expression for the image is not unique, and the sparse vector representation is not robust to the user's environment.

Hereinafter, embodiments of the present invention for solving the above problems will be described.

FIG. 3 is a view showing a schematic configuration of a skin condition classifying apparatus according to an embodiment of the present invention.

As shown in FIG. 3, the skin condition classification apparatus 100 according to an embodiment of the present invention receives a close-up skin image as an input for skin condition classification and outputs skin condition classification information of the skin image.

The skin condition classification apparatus 100 includes a structure storage unit 110, a skin condition classification model 120, and a skin condition classification unit 130.

The structure storage unit 110 includes a plurality of classifier structures constructed using a sparse coding algorithm for data of a plurality of skin images. Each classifier structure includes a dictionary learned for a plurality of skin images, A sparse representation extracted from a dictionary, and a classifier generated by a rare representation. This will be described in detail later.

The skin condition classification model 120 may be constructed by applying a multiclass adaboosting algorithm to a plurality of classifier structures stored in the structure storage unit 110 to generate a final skin state classification model called a boosting rare coding model Boosting Sparse Coding Model). At this time, the skin condition classification model 120 may be a processing device capable of performing an algorithm according to a boosting sparse coding model.

The skin condition classifying unit 130 classifies the skin image inputted from the outside for the skin condition classification by using a boosting rare coding model of the skin condition classification model 120 and a plurality of classifier structures stored in the structure storing unit 110, The skin condition for the image is classified and the skin condition classification information corresponding to the result is outputted. Here, the skin image may be a skin image having a label for moisture, oil, keratin, and moisture content (%). In this case, the skin condition classification information output by the skin condition classifier 130 may be classified by each label Information that classifies the condition of the skin is included.

Hereinafter, an apparatus and method for generating a skin condition classification model according to an embodiment of the present invention will be described.

4 is a schematic block diagram of an apparatus for generating a skin condition classification model according to an embodiment of the present invention.

4, an apparatus 200 for generating a skin condition classification model according to an embodiment of the present invention includes a structure storage unit 210, a structure generation unit 220, and a final model generation unit 230.

The structure storage unit 210 is the same as the structure storage unit 110 described with reference to FIG. 3 and includes a plurality of classifier structures 211, 213, and 215.

Each classifier structure 211, 213, and 215 includes a dictionary 2111, a rare expression 2113, and a classifier 2115, as shown in FIG. The dictionary, the rare representation, and the classifier included in each of the classifier structures 211, 213, and 215 will be different from each other.

The structure generating unit 220 generates a plurality of classifier structures 211, 213, and 215 stored in the structure storing unit 210. The structure generating unit 220 generates a classifier structure 211, 213, and 215 using a sparse coding algorithm for data of a plurality of skin images, and the structure generating unit 220 generates a classifier structure A plurality of classifier structures 211, 213, and 215 are generated. For example, the classifier structure 211 is a structure that is first generated by the structure generation unit 220, the classifier structure 213 is a structure that is generated secondarily by repetitive execution of the structure generation unit 220, The structure 215 represents a structure generated by the n-th iteration of the structure generating unit 220. In order to show such a repetition number, 1, 2, and n are used for each classifier structure 211, 213, and 215, respectively.

Specifically, the structure generating unit 220 includes a pre-forming unit 221, a rare expression extracting unit 223, and a classifier generating unit 225.

The pre-construction unit 221 constructs a learned dictionary using data of a plurality of skin images using an alternating minimization algorithm, and generates a corresponding classifier structure of the structure storage unit 210, for example, a classifier structure 221 ). That is, the pre-construction unit 221 repeats the pre-configuration operation by repetition of the structure generation unit 220 so that the dictionary (dictionary 1, dictionary 2, ..., dictionary n) Respectively, in the classifier structures 211, 213, and 215, respectively.

FIG. 6 is a diagram showing a specific configuration of the pre-configuration unit 221 shown in FIG.

As shown in FIG. 6, the pre-configuration unit 221 includes a patch matrix configuration unit 2211 and an intersection minimization performance unit 2213.

As shown in FIG. 7, the match matrix construction unit 2211 constructs a matrix of patches of data of a plurality of skin images. Referring to FIG. 7, a plurality of patches 23 may be generated from the skin image 21, and each patch of the patches 23 may be rearranged in a matrix form to form one patch matrix.

The cross minimization performing unit 2213 performs learning using the cross minima algorithm for the patch matrix constituted by the patch matrix construction unit 2211 to construct the finally learned dictionary and stores the dictionary in the structure storage unit 210. [

Specifically, the intersection minimization unit 2213 performs an intersection minimization algorithm such as Equation (1).

[Equation 1]

Referring to Equation (1), the dictionary Φ is obtained by using a Lagrange dual problem with the sparse vector a fixed, and the sparse vector a is obtained by fixing the Φ (Feature sign search algorithm). The final dictionary is constructed after the above process is repeated until the convergence condition is established.

The sparse expression extracting unit 223 extracts a sparse representation of data of a plurality of skin images using a dictionary constructed in the pre-constituent unit 221 and extracts a corresponding classifier structure of the structure storage unit 210, (221). That is, the sparse expression extracting unit 223 extracts a dictionary (sparse expression 1, sparse expression 2, ..., sparse expression n), which are each formed by repeating the operation of extracting the rare expression by repeated execution of the structure generating unit 220, Are stored in the classifier structures 211, 213, and 215 corresponding to the number of iterations, respectively.

8 is a diagram showing a specific configuration of the rare expression extracting unit 223 shown in FIG.

8, the rare expression extracting unit 223 includes a feature sine search performing unit 2231 and a sum pooling performing unit 2233.

The feature search search performing unit 2231 performs a feature search search algorithm in a dictionary configured by the pre-configuration unit 221 in advance.

The island pulling performing unit 2233 performs island fill-up on the data on which feature sine search has been performed by the feature sine search performing unit 2231 to extract a rare expression for each skin image and stores it in the structure storing unit 210 do. At this time, the dictionary is already learned and uses the one stored in the classifier structure 211, 213, 215 of the structure storing unit 210.

The classifier generating unit 225 generates a corresponding classifier by applying a machine learning algorithm to the rare expression extracted by the sparse expression extracting unit 223 and outputs the corresponding classifier structure of the classifier storage unit 210, (221). That is, the classifier generating unit 225 repeats the classifier generating operation by repeatedly performing the structure generating unit 220 to perform the classifier 1 (classifier 1, classifier 2, ..., classifier n) Respectively, in the classifier structures 211, 213, and 215, respectively.

In the embodiment of the present invention, SVM (Support Vector Machine) algorithm is used as the machine learning algorithm used for classifier generation. This SVM algorithm is a mapping learning algorithm that provides an optimal separation boundary for classifying data. The optimal separation boundary is learned to be located between data of each class. This is well known to those of ordinary skill in the art, and a detailed description thereof will be omitted here.

Next, the final model generation unit 230 generates a final skin condition classification model by synthesizing a plurality of classifiers stored in the classifier structures 211, 213, and 215 of the structure storage unit 210. At this time, the final model generation unit 230 synthesizes a plurality of classifiers using an adaboosting algorithm to generate a final skin condition classification model. Here, the final skin condition classification model to be generated is called a Boosting Sparse Coding Model since it is generated by the adaboing algorithm based on the rare coding algorithm. Meanwhile, the adaboosting algorithm used in the embodiment of the present invention recognizes samples using several weak classifiers at each iteration step, reduces the weight for the recognized samples, And the final classifier consists of the sum of the classifiers generated in each step. Since this is well known, a detailed description is omitted.

In the embodiment of the present invention, a machine learning algorithm (SVM) is applied using dictionaries and rare expressions that are learned every iterative process, and the weights of randomly selected patches are updated according to the learning error rate. Here, two aspects are considered when updating the weights of the patches.

First, for any one rarely misclassified expression, weight update of all applicable patches is considered. Second, since any patch initially drawn to generate the dictionary in each iteration also affects the misclassification, updating of the weights for these is also considered. In both of these methods, the weight updating process occurs independently of each other. In order to adopt both methods, the sum of the weights obtained in each of the two methods is used for the next operation. The adaboost parameter can be obtained for each classifier using the weight and the error rate obtained through this process.

Accordingly, the final model generation unit 230 can generate a final skin condition classification model through linear combination of the obtained Adaboost parameters according to the classification.

The weight updating described above will be described in detail.

)을 [수학식 2]와 같이 구할 수 있다. First, the existing multi-class Ada Boost methodology is used to update the weights of the first case, that is, to update the weights of the patches directly associated with the misclassification. That is, the error rate

Can be obtained as shown in [Equation 2].

&Quot; (2) "

는 패치 가중치이며,

는 분류기이고,

는 지시 함수이다.Here, m is the number of repetitions,

Is the patch weight,

Is a classifier,

Is an indicator function.

)을 통해서 아다부스트 파라미터(

)가 [수학식 3]과 같이 구해질 수 있다.The error rate obtained through the equation (2)

) Through the Adaboost parameter (

) Can be obtained as shown in Equation (3).

&Quot; (3) "

는 피부 클래스 수준이다.here,

Is the skin class level.

)가 구해지면 패치의 가중치 업데이트가 [수학식 4]와 같이 이루어진다.Thus, for each iteration, the Adaboost parameter (

) Is obtained, the update of the weight of the patch is performed as in Equation (4).

&Quot; (4) "

는 실제 피부 클래스 수준이고,

는 예측된 피부 클래스 수준이다. here,

Is the actual skin class level,

Is the predicted skin class level.

The weight of the misclassified patches may be updated through the above process.

We next describe the updating of the weights for the second case, that is, the updating of weights for any patches initially picked to construct a dictionary for each iteration.

If a misclassified patch is included in the initially arbitrarily selected patches (for example, P1 in FIG. 9 (a)), the patch overlapped with the patch P1 (for example, ) To P2) all update the weights. In the example of Fig. 9, the patch P1 is a patch that directly affects the misclassification, and the patch P2 is a patch that indirectly affects the region A1 in Fig. 9 (b). As a result, even if the patch P2 does not belong to the patch P1 which is initially randomly selected, it affects the misclassification indirectly as the region A1 partially included. Therefore, it also updates patches indirectly affecting misclassification. In FIG. 9C, the hatched portion A2 indicates the pixel portion to be actually updated.

First, in (a) of FIG. 9, the update for the patch P1 is updated through the weight update method in the first method described above. That is, [Equation 2], [Equation 3], and [Equation 4] are applied.

Then, when the update is executed, the number of pixels (for example, the hatched portion A2 in FIG. 9C) overlapping with all the possible patches is calculated through the expression (5).

&Quot; (5) "

Here, p is the dimension of the patch, patch_d is the patch (P1), and pixel is each of the pixels of the patch (P2).

Thereafter, when the number of pixels is determined for each overlapping patch, the weight is updated by the number of pixels as in Equation (6).

&Quot; (6) "

은 먼저 구한 패치(P1)의 가중치이고,

는 업데이트될 패치(P2)의 가중치이다.here,

Is the weight of the patch P1 obtained first,

Is the weight of the patch P2 to be updated.

On the other hand, in the weight updating according to the first method and the second method described above, the weight obtained by adding the two weights becomes the input data for the next update in order to consider all situations. The combined weights are used in the first iteration in the next iteration and in the recursive execution in updating the weights according to the second.

), 즉 최종 모형은 다음의 [수학식 7]과 같다.On the other hand, the final classifier (final classifier), which is the final skin condition classification model generated by the final model generator 230

), That is, the final model is expressed by the following equation (7).

&Quot; (7) "

는 가중치이고,

는 분류기이며,

는 이미지이고,

는 지시함수이며, k는 클래스 수준이고,here,

Is a weight,

Is a classifier,

Is an image,

Is an indicator function, k is a class level,

이며,

Lt;

이다.

to be.

As described above, the final skin classification state model generated by the final model generation unit 230 can be used in the skin condition classification apparatus 100 shown in FIG.

Hereinafter, a method for generating a skin condition classification model according to an embodiment of the present invention will be described.

10 is a flowchart of a method of generating a skin condition classification model according to an embodiment of the present invention.

Referring to FIG. 10, first, data of a plurality of skin images used for skin condition classification learning is configured as a matrix of patches (S110). Here, the step (S110) corresponds to the part described above with respect to the patch matrix forming section 2211 in Fig.

Then, a rare-coding algorithm is applied to the patch matrix to generate a structure 241 including a dictionary, a rare expression, and a classifier (S120). The step S120 corresponds to the portion described for the structure generating unit 220 in FIG.

The step S120 is repeated several times (S130) and the iterative process may be determined by a rare coding algorithm used to generate the structure or may be set by the user.

When the repetition of step S120 is completed (S130), a plurality of classifiers of the structure generated through the step S120 are combined to generate a final skin condition classification model (S140). The step S140 corresponds to the portion described for the final model generation unit 230 in FIG.

Hereinafter, the step of creating a structure (S120) will be described in detail with reference to FIG.

12 is a specific flowchart of the structure generation step (S120) shown in FIG.

Referring to FIG. 12, learning is performed by applying an intersection minimization algorithm to the patch matrix generated in step S110 to construct a finally learned dictionary (S121). The dictionary thus constructed is stored in the structure 241. The step S121 corresponds to the portion described above with respect to the pre-configuration unit 221 in Fig.

Thereafter, a rare expression for data of a plurality of skin images is extracted using the dictionary constructed in step S121 (S122). The thus extracted rare expression is stored in the structure 241. The step S122 corresponds to the portion described for the rare expression extracting unit 223 in FIG.

Next, a corresponding classifier is generated by applying a machine learning algorithm, for example, an SVM algorithm, to the rare expression extracted in the step S122. The generated classifier is stored in the structure storing unit 210. The step S123 corresponds to the part described above with respect to the classifier generating unit 225 in Fig. Meanwhile, in step S123, the weights of the patches described above may be updated according to the learning error rate. Since this has been described in detail above, it is omitted here.

In order that the objects, features, and advantages of the present invention become more apparent and understandable, specific embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Example

<Learning part>

1. Data collection

The data were collected by using a parameter measuring machine and the user 's moisture, oil, cellular structure, and moisture percentage were collected and skin data were recorded using Smart Scope.

The data collected are as follows.

For each collected data, a label was recorded for moisture, oil, cell tissue, and percent moisture information. For example, 11132 indicates a moisture level of 1, an oil level of 1, a cell tissue state level of 1, and a moisture content of 32%. An example of image data and labels is shown in FIG.

2. From the collected image data, collect it with an arbitrary 6 ㅧ 6 ㅧ 3 size patch. The collected patches are converted into vectors, and a patch matrix is formed by collecting the patches.

3. Apply the cross minima algorithm to the patch matrix to learn the dictionary. At this time, the algorithm initializes the initial value at random.

4. A sparse representation of the input image is obtained using a learned dictionary of 160,000 bases as shown in FIG. The input image is transformed into a matrix of patches, and a rare expression is constructed using the learned dictionary and feature sign search algorithm. At this time, the dimension of the rare expression is expressed by the number of converted patches and the number of bases of the dictionary. Then, through proper sum pooling, the image data expressed as a vector of 1 ㅧ 3072 (32 ㅧ 32 ㅧ 3) is represented as a vector of 1 ㅧ 12800 and becomes a very rare vector. The concept of matrix operation of a cross minima algorithm applied to a patch matrix is shown in Fig. 14, n is the patch number, m is the patch dimension, and d is the base number.

5. Create a structure containing dictionary, rare expression, and classifier through several iterations. After that, Adaboost method is applied to several classifiers to generate a final skin condition classification model, Boosting Sparse Coding model. A structure constructed by the number of repetitions is shown in Fig. In FIG. 15, the SVM indicates that the classifier generated from the rare expression is a classifier generated by the SVM (Support Vector Machine) algorithm. 15 shows that the number of repetitions is k.

<Analysis part>

6. In the above-mentioned learning process, patches and AdaBoost weights are updated according to the error rate within each structure. When a new input image arrives, the new input image extracts the rare representation value from the learned dictionary of each structure and determines the skin condition classification for the input image through the internal classifier.

7. The greater the number of repetitions, the higher the accuracy.

For example, the accuracy of the analysis part performed for each number of iterations with respect to the number of learning data, the number of verification data, and the oil level is shown in FIG.

Number of learning data: 1800, dimension 32 ㅧ 32 ㅧ 3 pixels

Verification data count: 684, dimension 32 ㅧ 32 ㅧ 3 pixels

label: oil level (1, 2, 3, 4, 5)

It is confirmed from FIG. 16 that the more the number of repetitions, the higher the accuracy.

As described above, according to the embodiment of the present invention, it becomes more robust to the user's environment (light, shadow, indoor and outdoor), and the feature vector becomes unique, thereby lowering the classification error.

Therefore, the accuracy with respect to the skin condition classification is improved as compared with the conventional technique.

The embodiments of the present invention described above are not implemented only by the apparatus and method, but may be implemented through a program for realizing the function corresponding to the configuration of the embodiment of the present invention or a recording medium on which the program is recorded.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, It belongs to the scope of right.

Claims (22)

A structure storage unit for storing a plurality of classifier structures constructed using a sparse coding algorithm for data of a plurality of skin images, each classifier structure comprising a dictionary learned for a plurality of skin images, A sparse representation extracted from a dictionary and a classifier generated from the sparse representation; And
A skin condition classification model for a new skin image input for classification of a skin condition and a plurality of classifier structures stored in the structure storage section to classify the skin condition of the new skin image, A skin condition classifying unit
/ RTI &gt;
The skin condition classification model is a Boosting Sparse Coding Model that is finally generated by applying an adaboosting algorithm to a plurality of classifiers stored in the structure storage unit,
Skin condition classifier. The method according to claim 1,
The skin condition classification model includes:
A plurality of classifiers are generated by repeating a process of generating a classifier structure including a dictionary, a rare expression, and a classifier by a predetermined number of times by applying a sparse coding algorithm to data of a plurality of skin images, And generating a skin condition classification model generating a skin condition classification model, which is a boosting sparse coding model, by applying an adaboosting algorithm to a plurality of classifiers included in the classifier,
Skin condition classifier. 3. The method according to claim 1 or 2,
Wherein the skin image is a skin image having a label for moisture, oil, keratin, and moisture content,
Skin condition classifier. A structure storage unit for storing a plurality of classifier structures including a dictionary learned for a plurality of skin images, a sparse representation extracted from the dictionary, and a classifier generated from the rare expression;
A structure generation unit for generating the plurality of classifier structures by using a sparse coding algorithm for data of a plurality of skin images and storing the plurality of classifier structures in the structure storage unit; And
A final model generation unit for generating a boosting sparse coding model, which is a final skin state classification model, by applying an adoboosting algorithm to a plurality of classifiers of the structure storage unit;
A skin condition classification model generating device for generating a skin condition classification model; 5. The method of claim 4,
The structure generating unit includes:
A dictionary organizing unit configured to organize the learned dictionary by using an alternating minimization algorithm for the data of the plurality of skin images and storing the constructed dictionary in a corresponding classifier structure of the structure storing unit;
A rare expression extracting unit for extracting a rare expression for data of the plurality of skin images using a dictionary formed by the pre-constructing unit and storing the extracted rare expression in a corresponding classifier structure of the structure storing unit; And
Generating a classifier by applying a machine learning algorithm to a sparse expression extracted by the sparse expression extraction unit and storing the generated classifier in a corresponding classifier structure of the structure storage unit;
Wherein the skin condition classification model generating device comprises: 6. The method of claim 5,
The pre-
A patch matrix forming unit for forming a patch matrix from data of the plurality of skin images; And
A crossing minimization unit for constructing a finally learned dictionary by performing learning using the cross minima algorithm for the patch matrix constituted by the patch matrix unit and storing the dictionary in a corresponding classifier structure of the structure storage unit,
Wherein the skin condition classification model generating device comprises: The method according to claim 6,
Wherein the patch matrix forming unit forms a plurality of patch data from the data of the plurality of skin images and rearranges each of the plurality of patch data in a matrix form to form one patch matrix,
Skin condition classification model generation device. The method according to claim 6,
The intersection minimization algorithm performed by the intersection minimization performing unit may be configured to obtain the dictionary using a Lagrange dual problem while the sparse vector is fixed and perform a feature search sign search algorithm to find the sparse vector until the convergence condition is satisfied,
Skin condition classification model generation device. 9. The method of claim 8,
Wherein the intersection minimization algorithm performed by the intersection minimization performing unit comprises:

Here,? Is a dictionary,
a is a rare vector
The skin condition classification model generation device according to claim 1. 6. The method of claim 5,
Wherein the sparse expression extracting unit extracts,
A feature sign search performing unit that performs a feature sign search algorithm in advance, which is finally configured by the pre-configuration unit; And
A sum pooling unit that performs sum pooling on the result data performed by the feature search unit and extracts a rare expression for each skin image and stores the extracted rare expression in a corresponding classifier structure of the structure storing unit; Performance department
Wherein the skin condition classification model generating device comprises: 6. The method of claim 5,
The machine learning algorithm used by the classifier generator is a support vector machine (SVM)
Skin condition classification model generation device. 6. The method of claim 5,
Wherein the structure generating unit generates one classifier structure including a dictionary, a rare expression, and a classifier through a single generating operation, the constructor generating unit generates the plurality of classifiers through a plurality of iterative operations,
Skin condition classification model generation device. 13. The method of claim 12,
Wherein the classifier generator updates the weights of the patches according to the learning error rate for each repetitive operation to generate an adaboost parameter for each classifier,
Wherein the final model generating unit generates a final skin condition classification model through linear combination of the Adaboost parameters for each classifier,
Skin condition classification model generation device. 14. The method of claim 13,
The updating of the weights of the patches,
Updating the weights of the patches corresponding to the misclassified rare expressions and updating the weights for the patches originally selected to generate the dictionary in the repeat operation,
Skin condition classification model generation device. 14. The method of claim 13,
The final skin condition classification model (

) Is expressed by the following equation

here,

Is a weight,

Is a classifier,

Is an image,

Is an indicator function,
k is a class label,

Lt;

being
The skin condition classification model generation device according to claim 1. A method for generating a skin condition classification model used by a skin condition classification model generating apparatus to classify a skin condition of a skin image,
Generating a plurality of classifier structures by repeating a process of generating a classifier structure including a dictionary, a rare expression, and a classifier by a predetermined number of times by applying a sparse coding algorithm to data of a plurality of skin images; And
Generating a final skin condition classification model that is a boosting sparse coding model by applying an adaboosting algorithm to a plurality of classifiers respectively included in the plurality of classifiers;
The method comprising the steps of: 17. The method of claim 16,
Before the step of generating the plurality of classifier structures,
Further comprising constructing a plurality of patch data from the data of the plurality of skin images and rearranging each of the plurality of patch data in a matrix form to form one patch matrix,
How to create a skin condition classification model. 17. The method of claim 16,
Wherein the generating of the plurality of classifier structures comprises:
Constructing a learned dictionary using data of the plurality of skin images using an alternating minimization algorithm and storing the constructed dictionary in a corresponding classifier structure;
Extracting a sparse representation of the data of the plurality of skin images using a dictionary formed and storing the extracted sparse expressions in a corresponding classifier structure;
Generating a classifier by applying an SVM (Support Vector Machine) algorithm to the extracted rare expression, and storing the classifier in a corresponding classifier structure; And
Constructing the learned dictionary by a predetermined number of times and storing the learned dictionary in a corresponding classifier structure, extracting the rare expression and storing the extracted rare expression in a corresponding classifier structure, and storing the classifier in a corresponding classifier structure To generate the plurality of classifier structures
And generating a skin condition classification model. 19. The method of claim 18,
The crossover minimization algorithm is a method of obtaining the dictionary using a Lagrange dual problem while fixing a sparse vector and using the feature sign search algorithm with the dictionary fixed, The algorithm for finding the vector is repeated until the convergence condition is satisfied,
How to create a skin condition classification model. 19. The method of claim 18,
Extracting the sparse representation and storing the sparse representation in a corresponding classifier structure, performs a feature sign search algorithm in the dictionary, and then performs a sum pooling to extract a rare representation of each of the skin images doing,
How to create a skin condition classification model. 19. The method of claim 18,
Wherein the generating of the plurality of classifiers comprises: generating patches for each classifier by updating the weights of the patches according to the learning error rate for each iteration;
Generating a final skin condition classification model by linear combination of the adaboost parameters according to the classifier in the step of generating the final skin condition classification model,
How to create a skin condition classification model. A computer-readable recording medium storing a program for performing a method for generating a skin condition classification model,
A function of generating a plurality of classifier structures by repeating a process of generating a classifier structure including a dictionary, a rare expression, and a classifier by a predetermined number of times by applying a sparse coding algorithm to data of a plurality of skin images; And
A function of generating a final skin condition classification model which is a boosting sparse coding model by applying an adaboosting algorithm to a plurality of classifiers included respectively in the plurality of classifiers
The computer-readable recording medium according to claim 1,

Images