JPH0793549A - Physical information sorting device using inductive learning and physical information sorting rule generation method using inductive learning - Google Patents

Abstract

PURPOSE:To automatically and accurately generate the sorting rules of feature information relating to optional physical information relating to the sorting technique of the physical information such as a technique for sorting pictures for which skin surfaces are image picked up by the ages of skin. CONSTITUTION:At the time of learning, an inductive learning part 106 lets a picture input part 101 and a feature parameter extraction part 103 extract the sets of plural feature parameter pairs 104 corresponding to the plural skin surfaces for learning belonging to respective sorting classes as the set of positive examples belonging to the respective sorting classes corresponding to the respective plural sorting classes. Then, the inductive learning part 106 generates the sorting rules corresponding to the respective sorting classes for satisfying the positive examples included in the respective sorting classes and not satisfying the positive examples included in the sorting class other than the respective sorting classes by an inductive learning processing. At the time of sorting, a sorting part 108 uses the sorting rules and outputs the sorting classes corresponding to the feature parameter pairs 104 as sorting class recognition output 109.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for classifying physical information, including a technique for classifying an image obtained by picking up an image of a skin surface according to skin age.

[0002]

2. Description of the Related Art A technique for classifying physical information such as image information and audio information is applied to various industrial techniques.

For example, the surface shape of human skin is classified. On the skin surface, there are a number of narrow grooves called skin grooves, hills called cuticles divided by the skin grooves, and pores existing at the intersections of the skin grooves.
They change under physiological influences such as skin metabolism.

Therefore, the classification of the skin surface shape serves as an index for judging the properties of the skin, the skin quality, and changes due to aging thereof, and the index is useful in the fields of skin treatment, diagnosis, beauty hygiene, and the like. Provide information.

In order to classify the skin surface shape, first,
It is necessary to extract the feature information from the skin surface. As a method for extracting features of the skin surface shape, there is a method in which the skin surface is modeled with silicon rubber or the like to create a skin surface replica (negative replica), which a human observes with an optical microscope.

As a second method for extracting the features of the skin surface shape, a skin surface replica is scanned with a stylus using a surface roughness meter, and the height and number of peaks of the undulations are obtained from the undulation value signal obtained as a result. There is a method of determining the degree of unevenness of the skin surface replica surface by obtaining the peak area and the like.

The third method for extracting the characteristics of the skin surface shape is as follows. That is, first, the skin surface or the skin surface replica surface is illuminated from a plurality of directions, for example, three directions. Next, the surface of each illumination is imaged by a television camera through an optical microscope, and the image pickup signal is converted into digital image data to obtain the brightness value (pixel value) of each pixel that constitutes the image for each illumination image. Value). Then, the lightness data thus obtained is subjected to digital image processing to extract geometrical characteristic parameters, which are used as the characteristic information of the skin surface shape.

The fourth method for extracting the features of the skin surface shape is as follows. That is, first, the skin surface or the skin surface replica surface is illuminated from a plurality of directions, for example, three directions. Next, the surface of each illumination is imaged by a television camera through an optical microscope, and the image pickup signal is converted into digital image data to obtain the brightness value of each pixel constituting the image of each illumination image. . Then, based on the information on the illumination direction and the brightness value obtained from each digital image data, the gradient of the skin surface shape at each pixel position is calculated.
It is extracted as a three-dimensional shape of the skin surface. Then, based on the gradient, feature information regarding the skin surface shape is extracted.

The skin surface shapes are classified based on the characteristic information of the skin surface shapes extracted by the various methods as described above.

[0010]

However, in order to classify the skin surface shape on the basis of such characteristic information of the skin surface shape, an expert has heretofore made a trial after thoroughly knowing the characteristics of the characteristic information. By mistake, it was necessary to determine the classification rule for feature information. Therefore, there is a problem that it is difficult to determine the optimum classification rule.

Such a problem is not limited to the classification of the skin surface shape, but in the classification of various physical information such as image information and audio information, when the classification rule of the characteristic information regarding the physical information is determined, There were similar problems.

An object of the present invention is to make it possible to automatically and accurately generate a classification rule of feature information relating to arbitrary physical information.

[0013]

The present invention firstly has a characteristic information extracting means (characteristic parameter extracting section 103) for extracting characteristic information (characteristic parameter set 104) from physical information. This physical information is, for example, image information, and more specifically, image information obtained by imaging the skin surface, for example. In this case, the feature information is, for example, feature information indicating image features, and more specifically, feature information (skin groove parameter 213, pore parameter 215, etc.) related to the shape of the skin surface (skin groove, pore, etc.). .

Next, at the time of learning, a set of a plurality of feature information corresponding to a plurality of learning physical information belonging to each of the plurality of classification classes is associated with each of the plurality of classification classes. The characteristic information extracting means extracts a set of cases, and the classification rules corresponding to the respective classification classes satisfy the positive cases included in the respective classification classes, and correct the positive cases included in the classification classes other than the respective classification classes. It has an induction learning unit (induction learning unit 106) that generates the classification rules that do not satisfy by an induction learning process. Here, the classification class is, for example, the age of a person having a skin corresponding to image information.

Further, at the time of classification, the characteristic information corresponding to the input physical information is extracted by the characteristic information extraction means, and the characteristic information is applied to the classification rule corresponding to each classification class generated by the induction learning means. It has a classification means (classification unit 108) for recognizing the classification class to which the above-mentioned input physical information belongs.

[0016]

The inductive learning means automatically and accurately generates a classification rule for classifying, for example, the age of the imaged skin surface based on the physical information corresponding to the physical information, for example, the characteristic information about the shape of the skin surface. be able to.

[0017]

Embodiments of the present invention will now be described in detail with reference to the drawings. <Overall Configuration> FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

The image input unit 101 includes a magnifying optical system and a CCD.
It is composed of an image sensor, an A / D converter, a semiconductor memory that is a main storage device, a memory device such as a hard disk device or a magneto-optical disk device that is an auxiliary storage device, and a microcomputer that controls these. The image input unit 101 outputs a skin surface image 102 that is digital image data by capturing an image of the skin surface.

Next, the characteristic parameter extraction unit 103 is composed of a memory device and a microcomputer for controlling the memory device. This feature parameter extraction unit 103
Calculates a feature parameter set 104 including a skin groove parameter 213 and a pore parameter 215 from the skin surface image 102 with the configuration shown in FIG. 2 described later.

Next, at the time of induction learning, which will be described later, the switch unit 105 causes the feature parameter set 104 output from the feature parameter extraction unit 103 to be input to the induction learning unit 106. The induction learning unit 106 includes a memory device and a microcomputer that controls the memory device, like the feature parameter extraction unit 103. The induction learning unit 106 obtains a learning skin surface image 102 and a characteristic parameter set 104 whose classification class (described later) is known in advance through the image input unit 101 and the characteristic parameter extraction unit 103, and The classification rule 107 is generated by executing the induction learning process. This classification rule 107 indicates what value each of the characteristic parameters of the characteristic parameter set 104 has, and which classification class the skin surface image 102 corresponding to the characteristic parameter set 104 belongs to. It is a thing.

Finally, at the time of executing the classification described later, the characteristic parameter set 104 output from the characteristic parameter extraction section 103 is input to the classification section 108 by the switch section 105. The classification unit 108 includes the feature parameter extraction unit 1
Like 03, it is composed of a memory device and a microcomputer for controlling the memory device. This classification unit 108
Recognizes a classification class corresponding to the input characteristic parameter set 104 based on the classification rule 107 set by the induction learning unit 106, and outputs a classification class recognition output 109. This classification class is class information indicating which age of the human being the imaged skin surface belongs to. <Description of Feature Parameter Extraction Unit 103> Next, FIG. 2 is a functional block diagram of the feature parameter extraction unit 103 of FIG. This portion includes a portion including a Fourier transform unit 201, a filter processing unit 203, and an inverse Fourier transform unit 204, which performs preprocessing for removing noise from the skin surface image 102 and enhancing the skin groove, and the skin groove. The binarization unit 2 that executes the extraction process of the parameters 213 and the pore parameters 215.
06, a thinning unit 208, a branch point extraction unit 210, a skin groove parameter extraction unit 212, and a pore parameter extraction unit 214.
It consists of a part and.

First, the Fourier transform unit 201 calculates the spatial frequency data 202 by executing a fast Fourier transform process on the skin surface image 102. The spatial frequency data 202 represents information on spatial frequencies and information on phases in the skin surface image 102.

Next, the filter processing unit 203 executes a filtering process for the spatial frequency data 202, leaving the directional component and the spatial frequency component having large values. Then, the inverse Fourier transform unit 204 includes the filter processing unit 20.
By performing a fast inverse Fourier transform on the output of No. 3, the enhanced skin surface image 205 in which the fine noise is removed and the skin groove is enhanced is calculated.

Here, the filter processing unit 203 executes the following filter processing, for example. In the first filter processing, the spatial frequency data 20 below the predetermined frequency
Only 2 is left, and the other spatial frequency data 202 is set to 0. As a result, fine noise having high spatial frequency components can be removed from the skin surface image 102.

In the second filter process, a histogram of the spatial frequency data 202 is calculated for each direction in the two-dimensional coordinate space. Next, a predetermined number of directions having typical peak values are calculated from the histogram. Then, centering on each calculated direction, the spatial frequency data 20 within a predetermined range from the data group in that direction is set.
Only 2 is left, and the spatial frequency data 202 in other ranges are set to 0. Since the skin groove generally has a predetermined directionality in many cases, the skin groove can be emphasized by the second filter processing as described above.

After the filter processing unit 203 executes one or both of the above-mentioned first and second filter processes, the inverse Fourier transform unit 204 executes the fast inverse Fourier transform on the output of the filter processing unit 203. By doing so, the emphasized skin surface image 205 is calculated.

Next, the binarizing unit 206 calculates the binarized skin surface image 207 by executing the binarization process on the emphasized skin surface image 205 obtained as described above. In this binarization processing, the emphasized skin surface image 20
For each line of 5, the histogram of the brightness values of the pixels on the line is calculated, and the brightness value of each pixel on the line is converted into a binary value of 0 or 1 by the threshold value calculated from the histogram. It Such processing is executed for all the lines of the emphasized skin surface image 205.

Subsequently, the thinning section 208 calculates the thinned skin surface image 209 by performing the thinning processing on the binarized skin surface image 207. In this thinning process, a predetermined image calculation operator is applied to the binarized skin surface image 207 based on a well-known thinning process, so that a segment due to a skin groove on the binarized skin surface image 207 ( Hereinafter referred to as skin groove segment)
The pixel width of is thinned to one pixel.

The branch point extraction unit 210 extracts the branch points of the thinned skin groove segment on the thinned skin surface image 209 and outputs it as the branch point information 211. Here, the intersections of the thinned skin groove segments are extracted as branch points.

Branch point information 2 extracted as described above
11, the skin groove parameter extraction unit 212 extracts skin groove parameters 213 from the emphasized skin surface image 205, which is the black and white grayscale image output from the inverse Fourier transform unit 204, and the pore parameter extraction unit 214 also extracts pores. The parameter 215 is extracted. And the skin groove parameter 213
And the pore parameter 215 are the characteristic parameter set 1 of FIG.
04.

As the skin groove parameter 213, nine representative parameters (1) to (9) shown below can be adopted. (1) Average depth of skin groove (2) Standard deviation of depth of skin groove (3) Number of skin grooves (4) Average value of skin groove length (5) Standard of skin groove length Deviation (6) Mean value of skin groove width (7) Standard deviation of skin groove width (8) Skin groove occupancy rate (9) Standard deviation in the skin groove direction Each skin groove parameter 213 is as follows. It is calculated. Average value of the depth of the skin groove and standard deviation On the skin surface image 205, the average value of the lightness value at the same pixel position as the pixel position of the thinned skin groove segment present on the thinned skin surface image 209 and Calculated as standard deviation.

It can be assumed that the brightness value of the skin groove corresponds to the depth of the skin groove. In general, as human skin ages, the depth of the skin groove becomes shallow and the skin groove becomes unclear. Therefore, it is important to evaluate the depth of the skin groove as described above. The number of skin grooves is calculated as the number of segments between branch points based on the branch point information 211 on the thinned skin surface image 209.

Generally, the number of sulci on human skin decreases with age, so it is important to evaluate the number of sulci as described above. Average value of standard length of skin groove and standard deviation On the thinned skin surface image 209, the average value and standard deviation of the number of pixels of the segment between the branch points based on the branch point information 211 are calculated.

In general, human skin has a longer skin groove length with aging, so it is important to evaluate the skin groove length as described above. Average value of the width of the skin groove and standard deviation On the skin surface image 205, each position of the skin groove segment is centered on each pixel position on the thinned skin groove segment existing on the thinned skin surface image 209. It is calculated as the average value and the standard deviation of the number of pixels existing in the direction orthogonal to the direction in and having the brightness value equal to or more than the predetermined threshold value.

It is important to evaluate the width of the sulcus as described above, because the width of the sulcus, together with the depth of the sulcus, serves as an index for knowing the aging state of human skin. On the skin groove occupancy- emphasized skin surface image 205, the number of pixels having a lightness value equal to or greater than a predetermined threshold, which is centered on each pixel position on the thinned skin groove segment existing on the thinned skin surface image 209. Is calculated as the ratio of the total sum of the total number of pixels to the total number of pixels of the enhanced skin surface image 205.

In general, since the density of the sulcus of human skin decreases with age, it is important to evaluate the sulcus occupancy rate as described above. Standard deviation in the skin groove direction The thinned skin groove segment present on the thinned skin surface image 209 is approximated by a polygonal line and calculated as the standard deviation of the inclination of the polygonal line.

Generally, human skin loses radial homogeneity with respect to the sulcus with aging, and the sulcus tends to flow in a certain direction. Therefore, it is not possible to evaluate the sulcus direction as described above. is important.

Next, as the pore parameters 215, the following four typical parameters (10) to (13) can be adopted. (10) Number of pores (11) Average value of depth of pores (12) Standard deviation of depth of pores (13) Average value of area of pores Each pore parameter 215 is calculated as follows. . The number of pores is calculated as the number of pairs (at least two pairs of branch points) in which the distance between the branch points based on the branch point information 211 is equal to or less than a predetermined value. The average value and standard deviation of the depth of the pores are calculated as the average value and the standard deviation of the lightness values of the pixels corresponding to the center of gravity of the set of branch points. The average value of the area of the pores is calculated as the average value and the standard deviation of the number of pixels having the brightness value equal to or more than a predetermined threshold value with the center of gravity of the set of branch points as the center.

Generally, the pores serve as an index for knowing the aging condition of human skin, and, for example, the area of the pores tends to increase with age, so it is important to evaluate the pores as described above. is there.

As described above, the characteristic parameter extraction unit 103 in FIG. 1 can calculate the characteristic parameter set 104 as the skin groove parameter 213 and the pore parameter 215. <Description of Induction Learning Unit 106> Next, the induction learning unit 1 in FIG.
06 will be described in detail.

The induction learning unit 106, via the image input unit 101 and the characteristic parameter extraction unit 103, learns the skin surface image 10 whose classification class (described later) is known in advance.
2 and the corresponding feature parameter set 104,
On the other hand, by executing the induction learning process as shown below, the classification rule 107 is generated.

The classification rule 107 is generated as a set of inequalities indicating the range of values that each characteristic parameter of the characteristic parameter set 104 can take. Inductive learning processing refers to a learning method for generating general rules from a set of cases. In the example of this embodiment, the classification rule 1 is calculated from the set of feature parameter sets 104 extracted from the skin surface image 102.
07 is generated as a general rule.

3 to 6 are operation flowcharts showing the induction learning process. First, in step 301 of FIG. 3, the following processing is executed. That is, first, the image input unit 101 captures, for each classification class, a plurality of learning skin surface images 102 for a plurality of skin surfaces belonging to the classification class. Next, the characteristic parameter extraction unit 103 calculates a characteristic parameter set 104 for each skin surface image 102. Each characteristic parameter set 104 is a positive case in the classification class. That is, a plurality of positive cases corresponding to a plurality of learning skin surface images 102 are obtained for one classification class. The above processing is similarly executed for a plurality of classification classes. Here, the classification class is, for example, class information indicating in what age the imaged skin surface belongs to a human.

Subsequently, steps 302 (FIG. 3) to 324.
The process (FIG. 6) is a process executed by the induction learning unit 106. First, in FIG. 3, in step 302, a plurality of positive cases belonging to one unclassified classification class are selected from the calculation results of step 301.

Next, in step 303, step 30
From the positive cases belonging to the currently processed classification class selected in 2, a predetermined number of arbitrary positive case sets obtained by selecting arbitrary two cases are selected. When the number of selectable positive case sets is less than the predetermined number, the maximum selectable number of positive case sets is selected. The number of selected pairs is m.

In step 304, it is judged whether or not the number m of positive cases selected in step 303 is 0, that is, whether or not there are no more positive cases to be selected in the currently processed classification class. It

If the determination in step 304 is YES,
The process moves to step 323 of FIG. 6 to be described later, and the generation of the classification rule corresponding to the currently processed classification class ends.

If the determination in step 304 is NO, then in step 305, the minimum generalization process is executed for each of the m positive case sets selected in step 303.
In this process, as shown in FIG. 7, for each of the two positive cases # 1 and # 2 included in the positive case set, that is, for the two characteristic parameter groups # 1 and # 2, the characteristic parameters f ₁ and f are calculated.
_2, ..., each data pair regarding f _n (v _1,
v ₁ ′), (v ₂ , v ₂ ′), ..., (v _n ,
For v _n ′), each inequality is generated with the smaller value as the lower limit value and the larger value as the upper limit value. And
The set of these inequalities is calculated as one classification rule corresponding to the positive case set. Here, the characteristic parameters f ₁ , f ₂ , ..., F _n are, for example, (1) to
It is a predetermined one that is appropriately selected from the skin groove parameter 213 or the pore parameter 215 shown in (13).

In step 306, each of the m classification rules corresponding to the m positive case sets calculated in step 305 belongs to any classification class other than the currently processed classification class. It is checked for positive cases (which we call negative cases for the classification class currently being processed).

In step 307, the classification rule including at least one negative case is excluded based on the result of the check in step 306. In step 308, it is determined whether or not there is a classification rule that remains unexcluded in step 307.

If the determination in step 308 is NO,
The m classification rules corresponding to the m positive case sets calculated in step 305 are not appropriate as the classification rules of the classification class currently being processed. Therefore, step 311
Then, after removing the positive cases forming the m positive case sets selected in step 303 from the set of positive cases belonging to the currently processed classification class, step 30 is performed again.
3 is executed.

If the determination in step 308 is yes,
In step 309, based on the result of the processing in step 307, for each of the classification rules remaining without being excluded,
The number of positive cases included in each classification rule and belonging to the currently processed classification class is calculated.

In step 310, the classification rule with the largest number of positive cases included is calculated as a provisional classification rule based on the result of the processing in step 309. By the processing up to this point, the classification rule corresponding to the currently processed classification class can be represented by the provisional classification rule corresponding to a representative positive case set that belongs to the classification class and consists of two positive cases. . Then, as shown below, the classification rule corresponding to the classification class currently being processed is obtained by adding three positive cases to the above two positive cases plus another positive case belonging to that classification class. It is expressed by a provisional classification rule corresponding to a typical positive case set consisting of cases.

First, following step 310 in FIG.
In step 312, a predetermined number of arbitrary positive cases belonging to the currently processed classification class and other than the plurality of positive cases corresponding to the provisional classification rule calculated in step 310 are selected. Positive cases are selected. When the number of selectable positive cases is less than the predetermined number, the maximum selectable number of positive cases is selected. The number of selected positive cases is m. Then, by combining each of the m positive cases with the provisional classification rule calculated in step 310, m sets are selected.

In step 313, it is determined whether or not the number m of positive cases selected in step 312 is 0, that is, whether or not there are no more positive cases to be selected in the currently processed classification class. It

If the determination in step 313 is yes,
In step 314, the provisional classification rule obtained up to now is stored as one partial classification rule of the classification class currently processed, and then step 3 in FIG. 6 described later.
23, and the generation of the classification rule corresponding to the currently processed classification class ends.

If the decision result in the step 313 is NO, in a step 315, the minimum generalization process is executed for each of the m sets selected in the step 313. In this process, as shown in FIG. 8, first, with respect to the positive cases included in the set, that is, the feature parameter set and the provisional classification rules included in the set, the feature parameters f ₁ , f ₂ ,. , V _n of the data value and the lower limit value (v ₁ , α ₁ ), (v ₂ , α ₂ ), ..., (v
_n , α _n ), the smaller value is set as the lower limit value, and a pair of data value and upper limit value (v ₁ , β ₁ ), (v for each of the feature parameters f ₁ , f ₂ , ..., F _n ) ₂ ,
β ₂ ), ..., (V _n , β _n ), each inequality whose upper limit value is the larger value is generated. Then, the set of these inequalities is calculated as one new classification rule corresponding to the set.

In step 316, similarly to step 306, it is checked whether or not each of the new m classification rules calculated in step 315 includes a negative case. In step 317, similarly to step 307, the classification rule including at least one negative case is excluded based on the result of the check in step 316.

In step 318, similarly to step 308, it is determined whether or not there is a classification rule that remains without being excluded in step 317. If the determination in step 318 is no, the process moves to step 321 in FIG. 5 described later, and 1 corresponding to the classification class currently processed.
The generation of one partial classification rule (a partial classification rule described later) is completed.

If the determination in step 318 is yes,
In step 319, as in the case of step 309, based on the result of the processing of step 317, for each classification rule that remains unexcluded, a positive case included in each classification rule and belonging to the currently processed classification class. Is calculated.

Similar to step 310, in step 320, the classification rule with the largest number of positive cases included is calculated as a new provisional classification rule, based on the result of the process of step 318.

By the processing up to this point, the classification rule corresponding to the currently processed classification class is tentatively changed by the classification rule corresponding to the representative positive case set consisting of three positive cases belonging to the classification class. It can be expressed.

After step 320, the process returns to step 312, and steps 312 to 320 are repeated to sequentially increase the number of positive cases used for generating the provisional classification rule representing the classification class currently being processed. You can let it go. As a result, the currently processed classification class is represented by a more accurate provisional classification rule.

Step 318 which is a process in the middle of repeating the series of processes of steps 312 to 320 in FIG. 4 described above.
In the case where it is determined in step 317 that there is no classification rule that remains without being excluded, the process proceeds to step 321 of FIG.

In this case, of the positive cases belonging to the classification class currently being processed, the positive cases excluded in step 317 are not appropriate to be represented by the provisional classification rules up to that point. Therefore, the process for the provisional classification rule obtained up to the process immediately before step 317 is terminated, and the provisional classification rule is one formal classification rule (hereinafter, referred to as a partial classification rule) of the classification class currently being processed. It is said that. Then, another partial classification rule is newly generated for the positive case excluded in step 317. When two or more partial classification rules are generated for one classification class in this way, the classification rule obtained by taking the logical sum of these plural partial classification rules is as follows. The final classification rule 107 (FIG. 1) corresponding to the classification class is set.

In order to realize the above function, step 3
If the determination in step 18 is NO, in FIG. 5, first, in step 321, the provisional classification rule obtained up to the processing immediately before step 317 is one partial classification of the classification class currently processed. Remembered as a rule.

Next, in step 322, the positive cases used for generating the above partial classification rule are removed from the set of positive cases belonging to the classification class currently being processed.
After that, the above-described step 303 of FIG. 3 to 32 of FIG.
The process of 2 is repeated. As a result, one or more partial classification rules can be determined for the classification class currently being processed.

In the above iterative processing, the number m of positive cases selected in step 304 of FIG. 3 or step 313 of FIG. 4 in step 303 of FIG. 3 or step 312 of FIG. When it is determined that there are no more positive cases to be selected in the classification class being processed, the generation of the classification rule corresponding to the classification class currently being processed is terminated as follows.

First, if it is determined in step 304 of FIG. 3 that m = 0, then step 323 of FIG. 6 described below is executed as it is. Also, step 3 in FIG.
If it is determined in step 13 that m = 0, step 31
After step 4 is executed, step 323 is executed. In step 314, the provisional classification rule obtained up to the processing immediately before step 312 is stored as the last one partial classification rule of the classification class currently being processed.

In step 323 of FIG. 6, for the classification class currently being processed, the classification rule obtained by taking the logical sum of one or more partial classification rules stored corresponding to that classification class is The final classification rule 107 corresponding to the classification class is output to the classification unit 108 in FIG. When the number of partial classification rules stored is one, the partial classification rule is directly used as the classification rule 107.
Is output as.

Thereafter, in step 324, it is determined whether there is a classification class that has not been processed yet. If the determination in step 324 is YES, the process returns to step 302 in FIG. 3, and the series of processes described above is executed for the unprocessed classification class, and the classification rule 107 corresponding to that classification class is generated.

There are no unclassified classification classes,
When the determination in step 324 is NO, the induction learning process by the induction learning unit 106 in FIG. 1 ends. By the inductive learning process described above, the classification rule 107 for the feature parameter set 104 obtained by imaging the skin surface given for learning can be automatically and accurately generated. <Description of Specific Example of Induction Learning Process> FIGS. 9 to 15 are diagrams showing an example of the classification rule 107 generated by actually performing the above-described induction learning process on the feature parameter set 104. . Input data f ₁ -shown in (a) of each figure
f ₆ corresponds to the characteristic parameters shown as (1) to (13) described above, and h (i) indicates the classification class. Also, in each figure
The learning result shown as (b) is the feature parameter set 10
The classification rule 107 for the input data of 4 is shown.

Here, the symbol “=>” indicates the logic “if”. In addition, "Otherwise" indicates a case where none of the above inequalities is satisfied. For example, FIGS. 9 and 10 are diagrams showing an example of the classification rule 107 that is automatically generated when the human age is selected as the classification class.

In these figures, as the characteristic parameter set 104, in addition to the skin groove parameter 213 and the pore parameter 215 described above, the number of branch points (f ₅ )
The parameter of wrinkle state (f ₁ ) is also used. The former can be given as one of the branch point information 211 shown in FIG. 2, and the latter can be given by a human inspection based on a predetermined standard.

As can be seen from a comparison of FIGS. 9 (b) and 10 (b) with FIGS. 9 (a) and 10 (a), an appropriate classification rule 107 can be automatically generated. Recognize. Figure 1
1 and FIG. 12 are diagrams showing an example of the classification rule 107 that is automatically generated when the number of pores is selected as the classification class. In this case, the number of pores is not used as a characteristic parameter, and each skin surface image 102 is
The age is entered manually. This example can show at what age the number of pores is. FIG. 13 is a diagram showing an example of the classification rule 107 that is automatically generated when the depth of pores is selected as the classification class. In this case, the depth of the pores is not used as the characteristic parameter, and the age is manually input for each skin surface image 102. This example can show at what age the pore depth is.

FIG. 14 is a diagram showing an example of the classification rule 107 which is automatically generated when the depth of the skin groove is selected as the classification class. In this case, the depth of the skin groove is not used as a characteristic parameter, and each skin surface image 10 is not used.
For each two, the age is manually input. This example can show at what age the depth of the sulcus is.

FIG. 15 is a diagram showing an example of the classification rule 107 that is automatically generated when the number of skin grooves is selected as the classification class. In this case, the number of skin grooves is not used as a characteristic parameter, and each skin surface image 102 is not used.
Each time, the age is manually input. This example can show at what age the depth of the sulcus is. <Description of Classifying Unit 108> The process of the classifying unit 108 is very simple, and the set of values of the input characteristic parameter set 104 is used as the classification rule 10 set by the induction learning unit 106.
7, the classification rule 107 that satisfies the set of values of the characteristic parameter set 104 is extracted, and the classification class corresponding to the classification rule 107 is output as the classification class recognition result 109. <Other Embodiments> In the embodiments described above, the present invention is applied to the characteristic parameter set 104 extracted from the skin surface image 102 obtained by imaging the skin surface. It is not limited to.

For example, the present invention can be applied to feature parameter sets extracted from images in various categories. In addition, for example, according to the present invention, a feature parameter set relating to speech obtained by speech analysis, such as a linear prediction coefficient, a PARCOR coefficient, an LSP coefficient, or
It can be applied to a set of cepstral coefficients and the like, and a set of feature parameters such as a set of formant frequencies and amplitude values, which can also generate a classification rule for speech recognition.

Further, for example, the present invention can be applied to a feature parameter set extracted from the undulation value signal obtained by scanning a skin surface replica with a stylus using a surface roughness meter. it can.

As described above, according to the present invention, it is possible to automatically and accurately generate the classification rule of the characteristic parameter set relating to various physical information, and thereby to realize the accurate classification processing (recognition processing). it can.

[0081]

According to the present invention, based on the physical information corresponding to the physical information, for example, the characteristic information on the shape of the skin surface, the classification rule for classifying the age of the imaged skin surface is automatically and accurately determined. It is possible to generate well.

As a result, even a non-specialist can realize a physical information classification system with high recognition accuracy.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

FIG. 2 is a configuration diagram of a characteristic parameter extraction unit.

FIG. 3 is an operation flowchart of induction learning processing (No. 1)
Is.

FIG. 4 is an operation flowchart of induction learning processing (No. 2)
Is.

FIG. 5 is an operation flowchart of induction learning processing (No. 3)
Is.

FIG. 6 is an operation flowchart of induction learning processing (No. 4)
Is.

FIG. 7 is an explanatory diagram (1) of the minimum generalization process.

FIG. 8 is an explanatory diagram (2) of the minimum generalization process.

FIG. 9 is a diagram showing a specific example of induction learning processing (example 1 of generation of classification rules relating to years).

FIG. 10 is a diagram showing a specific example of induction learning processing (example 2 of generation of classification rule relating to age).

FIG. 11 is a diagram showing a specific example of induction learning processing (example 1 of generation of a classification rule regarding the number of pores).

FIG. 12 is a diagram showing a specific example of induction learning processing (example 2 of generation of a classification rule regarding the number of pores).

FIG. 13 is a diagram showing a specific example of induction learning processing (an example of generation of a classification rule relating to the depth of pores).

FIG. 14 is a diagram showing a specific example of induction learning processing (an example of generation of a classification rule regarding a skin groove depth).

FIG. 15 is a diagram showing a specific example of induction learning processing (an example of generation of a classification rule regarding the number of skin grooves).

[Explanation of symbols]

 101 image input unit 102 skin surface image 103 feature parameter extraction unit 104 feature parameter set 105 switch unit 106 induction learning unit 107 classification rule 108 classification unit 109 classification class recognition output 201 Fourier transform unit 202 spatial frequency data 203 filter processing unit 204 inverse Fourier Transforming unit 205 Emphasis skin surface image 206 Binarizing unit 207 Binarizing skin surface image 208 Thinning unit 209 Thinning skin surface image 210 Branching point extracting unit 211 Branching point information 212 Skin groove parameter extracting unit 213 Skin groove parameter 214 Pore Parameter extraction unit 215 Pore parameters

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Motohiro Yanai Motohiro Yanai 1050 Shinba-cho, Kohoku-ku, Yokohama, Kanagawa Shiseido Research Institute, Inc. (72) Fumio Mizoguchi 2641 Yamazaki, Noda, Chiba Tokyo University of Science Faculty of Engineering Within

Claims (7)

[Claims] 1. A feature information extraction unit for extracting feature information from physical information, and a plurality of feature information extraction units that correspond to a plurality of classification classes at the time of learning and that correspond to a plurality of learning physical information belonging to the respective classification classes. The feature information extracting means extracts a set of characteristic information of the above as a set of positive cases belonging to each of the classification classes, and the classification rule corresponding to each of the classification classes is satisfied, and the positive cases included in each of the classification classes are satisfied. Inductive learning means for generating a classification rule that does not satisfy a positive case included in a classification class other than each of the classification classes by induction learning processing; and characteristic information corresponding to physical information input at the time of classification. To the input physical information by applying the feature information to the classification rule corresponding to each classification class generated by the induction learning means. Physical information classification apparatus using inductive learning, characterized in that it comprises a recognizing classification means the class class, the. 2. A set of a plurality of characteristic information items corresponding to a plurality of classification classes and corresponding to a plurality of learning physical information items belonging to each classification class is set as a set of positive cases belonging to each classification class. A classification rule corresponding to each of the classification classes, which satisfies the positive cases included in the classification classes and does not satisfy the positive cases included in the classification classes other than the classification classes,
A physical information classification rule generation method using inductive learning, which is generated by inductive learning processing. 3. A feature information extracting means for extracting feature information indicating an image feature from image information which is physical information, and a plurality of classes belonging to each class corresponding to each of the plurality of classes at the time of learning. A set of a plurality of feature information corresponding to the image information for learning is extracted by the feature information extraction means as a set of positive cases belonging to each of the classification classes, and a classification rule corresponding to each of the classification classes, An induction learning means for generating by induction learning processing a classification rule that satisfies the positive cases included in each classification class and does not satisfy the positive cases included in classification classes other than the respective classification classes, and the input image information at the time of classification. By causing the feature information extracting means to extract the corresponding feature information and applying the feature information to the classification rule corresponding to each of the classification classes generated by the induction learning means. Physical information classification apparatus using inductive learning characterized by having a recognizing classification means the classification class image information the input belongs. 4. A set of a plurality of feature information corresponding to each of a plurality of classification classes, which corresponds to a plurality of pieces of learning physical information belonging to each of the classification classes and which indicates an image feature, A positive case that is extracted as a set of positive cases belonging to a classification class and that satisfies the positive rules included in the respective classification classes and that is a classification rule corresponding to the respective classification classes and that is included in a classification class other than the respective classification classes Classification rules that do not meet
A physical information classification rule generation method using inductive learning, which is generated by inductive learning processing. 5. A feature information extracting means for extracting feature information relating to the shape of the skin surface from image information which is physical information obtained by capturing an image of the skin surface, and in learning, corresponding to each of a plurality of classification classes, The feature information extracting means extracts a set of a plurality of feature information corresponding to a plurality of image information for learning belonging to each classification class as a set of positive cases belonging to each classification class, and a classification corresponding to each classification class. A rule, which is a rule that satisfies the positive cases included in the respective classification classes and does not satisfy the positive cases included in the classification classes other than the respective classification classes, by an induction learning means for generating by induction learning processing, and at the time of classification, The feature information corresponding to the input image information is extracted by the feature information extracting means, and the feature information is divided into parts corresponding to the respective classification classes generated by the induction learning means. By fitting the rule, physical information classification apparatus using inductive learning characterized by having a recognizing classification means the classification class image information the input belongs. 6. A plurality of features relating to the shape of the skin surface corresponding to each of the plurality of classification classes and corresponding to image information which is physical information obtained by imaging a plurality of learning skin surfaces belonging to each of the classification classes. A set of information is extracted as a set of positive cases belonging to each classification class, and the classification rules corresponding to each classification class satisfy the positive cases included in each classification class, and the classification classes other than each classification class Classification rules that do not satisfy the positive cases included in
A physical information classification rule generation method using inductive learning, which is generated by inductive learning processing. 7. The induction learning according to any one of claims 1 to 6, wherein the feature information extraction means outputs information given from the outside as one of the feature information. A physical information classification device used or a physical information classification rule generation method using inductive learning.