WO2014076360A1

WO2014076360A1 - Measurement of structural properties

Info

Publication number: WO2014076360A1
Application number: PCT/FI2012/051132
Authority: WO
Inventors: Matti-Paavo SARÉN
Original assignee: Metso Automation Oy
Priority date: 2012-11-16
Filing date: 2012-11-16
Publication date: 2014-05-22

Abstract

Apparatus for organizing structural properties (150) wherein the apparatus comprises an unsupervised classifier (120). The unsupervised classifier (120) receives a plurality of images of structural properties (150), each image having a single structural property (150), and organizes the images of the structural properties (150) with respect to each other such that the more similarity in predetermined features the images of the structural properties (150) have the closer to each other the images are associated. The unsupervised classifier (120) is configured to output the images of the structural properties (150) for teaching a supervised classifier (122).

Description

Measurement of structural properties Field

The invention relates to a measurement of structural properties. Background

To ensure good paper quality it is important to know the properties of the wood fibers used in paper making. Important properties include fiber length, fiber thickness, fiber wall thickness, fiber fibrillation and existence of vessels, for example.

The properties can be measured in an automated manner by cap- turing images of objects, in which case this method is suitable for industrial circumstances. Features of a fiber or a fiber-like particle may be measured by using a line camera and a thin circular capillary tube. The fibers move one by one in a thin tube from which the line camera captures an image.

Images of objects such as the fibers may be classified in different groups on the basis of their structural features in order to define quality of the product. For example, a supervised classifier such as a multi-layer perceptron may be taught to distinguish different object from each other. However, the teaching is time consuming and burdensome to complete because a user has to classify thousands or even millions of objects for different representative classes for teaching the supervised classifier each object and its class one by one in order to make the supervised classifier to learn the classes and to classify objects by itself. Because of the difficulty of the user input, the teaching is often insufficient or in the worst case flawed.

Hence, there is a need for a better measurement. Summary

The following presents a simplified summary of the invention in order to provide a basic understanding of some aspects of the invention. Its purpose is to present some concepts of the invention in a simplified form as a prelude to the more detailed description that is presented later.

An aspect of the invention relates to apparatus of claim 1 .

An aspect of the invention relates to a method of claim 13.

An aspect of the invention relates to apparatus of claim 24.

Although the various aspects, embodiments and features of the invention are recited independently, it should be appreciated that all combina- tions of the various aspects, embodiments and features of the invention are possible and within the scope of the present invention as claimed.

The present solution provides advantages. Classification of the images can be done automatically to help labeling and to simplify teaching of a supervised classifier. A properly taught supervised classifier works then reliably.

Brief description of the drawings

In the following the invention will be described in greater detail by means of exemplary embodiments with reference to the attached drawings, in which

Figure 1 shows an example of imaging and classification configuration;

Figure 2 shows an example of a measurement chamber for structural properties in suspension;

Figure 3 shows examples of structural properties,

Figure 4 shows an example of representation of ordered images of structural properties in suspension,

Figure 5 presents an example of a flow chart of an unsupervised classifying, and

Figure 6 presents an example of a flow chart of teaching.

Description of embodiments

Exemplary embodiments of the present invention will now be described more fully hereinafter with reference to the accompanying drawings, in which some, but not necessarily all embodiments of the invention are shown. Indeed, the invention may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Although the specification may refer to "an", "one", or "some" embodiment(s) in several locations, this does not necessarily mean that each such reference is to the same embodiment(s), or that the feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments.

In general, to classify structural properties, an unsupervised classifier receives a plurality of images of structural properties. Each image typically has a single structural property. Then the unsupervised classifier organizes the images of the structural properties with respect to each other such that the more similarity in predetermined features the images of the structural properties have, the closer to each other the images are associated. The closeness or strength of the association may represent or refer to a symbolic or real distance in an organized map, for example. In this way, the unsupervised classifi- er 120 may reduce the dimensionality of a multidimensional feature data of the structural properties. The unsupervised classifier outputs the images of the structural properties in order to teach a supervised classifier.

The supervised classifier, in turn, receives the images from the unsupervised classifier. The supervised classifier also has at least one label available, each label labeling a unique cluster of the images. The supervised classifier learns the at least one labeled cluster. The teaching of the supervised classifier enables the supervised classifier to organize and classify further images of structural properties on the basis of the teaching.

Let us now look at the embodiments more closely. Figure 1 presents an example of a measurement configuration for measuring structural properties. The structural property may refer to solid matter and to its structure. However, a structural property may be material structure which causes or can be observed as a difference in a refraction index, in an attenuation coefficient or in a reflection coefficient with respect to surrounding matter. Such differences make a structural property to be distinguished from its background in an image. Different structures may have different shapes which may be distinguished by their structural features. A structural property may be a fiber, a vessel, a scleroid particle or texture, a part of a fiber, a flock, texture of slurry, un- desired particles or the like. The predetermined structural features may refer to such as a length, a thickness, a thickness of a substructure, a curl, a twist, branching, flocking of the structural property, opacity, texture, for example. Each such structural feature is measurable from an image of the structural property.

In an embodiment of Figure 1 the configuration has a measurement chamber 100 which may have suspension having structural properties 150. In this example the structural properties may be fibers, their parts or the like and the suspension may be pulp. Suspension may have liquid as medium wherein solid particles are dispersed. The measurement chamber 100 may be a part of a hydraulic circuit 102 where a pump 104 may pump the suspension round the hydraulic circuit 102 and through the measurement chamber 100 such that at different moments different structural properties 150 may be in the measure- ment chamber 100. The hydraulic circuit 102 may or may not be connected to an outside source. If the outside source is available, it may input new suspension matter to the hydraulic circuit 102. Alternatively, the suspension in the chamber 100 may not circulate but it may still be mixed.

In an embodiment, the measured suspension may comprise sewage water.

An optical source 106 may output optical radiation whose wavelength range is wholly or partly from ultraviolet light to infrared light which may comprise wavelengths about 1 nm to 100 μηη. In an embodiment, the used band may be within about 300 nm to 10 μηη. The optical source 106 may comprise a laser, a led (Light Emitting Diode), an incandescent lamp, a halogen lamp, a gas discharge lamp or the like. The optical radiation from the optical source 106 may be directed to the measurement chamber 100. At least on optical component 108 may be used to converge or diverge rays of the optical radiation. The optical component 108 may comprise at least one lens. Additionally or alternatively, the optical component 108 may comprise a mirror. The optical component 108 may collimate the optical radiation coming from the optical source 106 and direct it to the measurement chamber 100 such that the rays of the optical radiation become at least approximately parallel. However, a focused optical radiation or diverging optical radiation may also be used. The optical axis of the optical radiation hitting the measurement chamber 100 may be parallel to the normal of the surface of the measurement chamber 100. However, other angles between the optical axis and the normal may also be used.

Figure 2 shows an example of the measurement chamber 100. The measurement chamber 100 may have two plain windows 200, 202 between which the suspension may flow. The windows 200, 202 are transparent to the optical radiation used in the measurement. An optical pass band of the windows 200, 202 may range from about 190 nm to about 1200 nm or even to about 3800 nm, for example. The material of the windows 200, 202 may comprise glass, quartz, sapphire, zinc selenide, germanium, calcium fluoride, or plastic, or the like for example. A distance D of the plain windows 200, 202 may be from hundreds of micrometers to thousands of micrometers. The distance D may be 3 mm, for example. The windows 200, 202 and the edges of the chamber 100 are leak proof and are attached to the rest of the hydraulic circuit 102 in a leak proof manner such that the suspension may not escape out of the measurement chamber 100. The plain surface area of the windows 200, 202 may vary from a few square millimeters to a few square centimeters, for example.

Images of the structural properties 150 in the suspension which is between the windows 200, 202 may be captured by a camera 1 10 directed to the windows 200, 202. The camera 1 10 may comprise a CCD (Charge- Coupled Device) or a CMOS (Complementary Metal Oxide Semiconductor) cell (not shown in Figures) which, in turn, may comprise a matrix of detector elements such as pixels for detecting optical radiation and for capturing images with the detected optical radiation. The camera 1 10 may also have at least one optical component (not shown in Figures) for focusing light coming from the measurement chamber 100 to the cell in order to form an image.

The camera 1 10 sends the images it has captured to a processing unit 1 12 which may, in an embodiment, show the images on the screen 1 14 to a user. In an embodiment, the user may input data and control the processing of images through an interface 1 16.

Figure 3 presents two simple examples of images 300, 302. The processing unit 1 12 comprises an unsupervised classifier 120 and a supervised classifier 122. The unsupervised classifier 120 receives a plurality of im- ages 300, 302 of structural properties 150. The images may be in electrical form which may in a typical manner be digital images, for example. Each structural property 150 may be in its own image 300, 302. In an embodiment, the unsupervised classifier 120 may divide an image of having several structural properties 150 in the image into pieces each having only one structural proper- ty. The unsupervised classifier 120 may organize the images 300, 302 of the structural properties 150 with respect to each other. The organizing may mean setting the images distance dependently in a representation such that the more similarity in predetermined structural features the images of the structural properties 150 have the closer to each other the images 300, 302 are placed in the representation. In this way, the unsupervised classifier 120 may reduce the dimensionality of a multidimensional feature data where the number of dimensions may be tens, hundreds, thousands or even more, for example, to a lower level used in the organized representation. The same advantage is also true for a case where no representation is used because the organizing itself can be considered to lower a number of dimensions. The organized representation may be three, two or one dimensional, for example. The organized representa- tion formed by the unsupervised classifier 120 may be suitable for visualizing the structural properties 150 which are mutually proportioned in the organized representation.

In an embodiment, the unsupervised classifier 120 may measure the predetermined features of the structural properties in the images and automatically organize images on the basis of the predetermined features.

Figure 4 illustrates an example of a two-dimensional organized representation 400. In this embodiment, a structural property is a wood fiber and a shape of a fiber is used as a feature argument of the classification. There are 100 potential shapes of different fibers in the example representation 400. The shape of a fiber may depend on several features such as length, width, ratio of length and width, minimum length, maximum length, minimum width, maximum width, ratio thereof or the like. As can be seen the fibers having features similar to each other are organized in two unique clusters 402, 404 of similar images close to each other in the representation 400. Each cluster 402, 404 may define a common area in the organized representation 400 although it is not necessary. For example, vessels are mainly in the lower left area 402 of the organized representation 400. On the other hand, more or less usual fibers are at the middle left area and upper right corner area 404 of the organized represen- tation 400.

In an embodiment, the unsupervised classifier 120 or some other part of the measuring equipment may automatically label the unique clusters 402, 404. The lower left area 402 of the organized representation 400 may be labeled as vessels, for example. Correspondingly, the middle left area and up- per right corner area 404 of the organized representation 400 may be labeled as fibers. In a similar manner, curled fibers, branching fibers, flocks and so on may be labeled in as similar manner. The label may mean a representative name for a cluster. The name may be expressed in an alphanumeric form or in some other symbolic form. A symbolic label may be a graphical sign or mark.

In an embodiment, the processing unit 1 12 may show the organized representation 400 on the screen 1 14 to a user. Instead of automatic labeling, the user may label at least one selected cluster of images and input his/her cluster selection with labeling through the interface 1 16 to the processing unit 1 12 which then may receive the at least one label with a cluster of images as- sociated therewith from the user in response to the shown part of the representation 400. The processing unit 1 12 then feeds the clusters 402, 404 of images 300, 302 with the at least one label to the supervised classifier 122 which is boot-strapped by this sort of supervised teaching with the clusters 402, 404 and their labels.

The supervised classifier 122 receives the representation 400 of the images 300, 302 and at least one label of the clusters 402, 404. The processing unit 1 12 provides the supervised classifier 122 with the at least one labeled cluster 402, 404. In this manner, the supervised classifier 122 is taught and thus it learns to distinguish structural properties 150 from each other. The teaching enables the supervised classifier 1 12 later to automatically classify images 300, 302 of structural properties 150 of various samples which have similar structural properties to those of the supervised teaching.

In an embodiment, the at least one label may indicate at least two separate clusters 402, 404 of the images in the representation 400, where each cluster 402, 404 has at least one unique feature of the structural property with respect to each other. A unique feature may be based on a dimension, width, texture or the like.

In an embodiment, the unsupervised classifier 120 may automatically organize the images on the basis of at least one clustering algorithm. The unsupervised classifier 120 may automatically organize the images on the basis of at least one of the following: self-organizing map of neural computing, t- distributed stochastic neighbor embedding, principal component analysis, sammon mapping method, GTM (General Topographic Mapping), LLE (Locally Linear Embedding) mapping, Isomap, agglomerative or hierarchial hierarchal clustering, including single-link-, complete-link-, average-link clustering, clustering error minimization, distance error minimization, K-means clustering, K- method, and graph-based methods like single-or complete link clustering, density based method, density-based spatial cluster of applications with noise (DBSCAN), AUTOCLASS, SNOB, BIRCH, MCLUST, or model based cluster- ing COBWEB or CLASSIT, simulated annealing for clustering, genetic algorithms, Bayesian method, Kernel method, Multidimensional scaling, principal curve, T-SNE, some of their combination or the like.

In an embodiment, the processing unit 1 12 in its unsupervised classifier 120 or in the supervised classifier 122 or in some other part of the pro- cessing unit 1 12 may optimize the number of predetermined features measured from the images of the structural properties for performing the classifica- tion by comparing a reliability of the classification with and without each feature and eliminating a feature not increasing the reliability. The optimization of the features may be automatized or it may require user input.

In an embodiment, the optimization may be performed using a top- bottom algorithm or a bottom-top algorithm. First, a priori information may be used to guess a few features which may be very likely good for classification. Examples of such features may be width and height of a structural property. Then the algorithm may rank the rest of the features in an order where a classification error is minimized according to Tables 1 and/or 2.

In the example of Table 1 the a priori information comprises features width and height which are used in each iteration of the classification. Other features 3, 4, 5 may be any property measurable from the image. The features may comprise thickness of the wall of a fiber, curliness, branching etc.

Table 1

Iteration

1

Width Height feature 3 feature 4 feature 5 right hits

X X 50 %

X X X 52%

X X X 51 %

X X X 59 %

Iteration

2

Width Height feature 3 feature 4 feature 5 right hits

X X X X 59 %

X X X X 67 %

Iteration

3 Width Height feature 3 feature 4 feature 5 right hits

X X X X X 67 %

When only the a priori features are used, the likelihood of right hits is 50 %. When the a priori information is used with feature 3, the likelihood of right hits increases to 52 %. When the a priori information is used with feature 4, the likelihood of right hits becomes to 51 %. When the a priori information is used with feature 5 the likelihood of right hits increases up to 59 %. Now it becomes clear that the classification needs a priori features and feature 5 and new iteration may be performed.

In this example, iteration 2 has two rounds. In the first round, fea- tures 4 and 5 are used with the a priori information and the likelihood of right hits becomes 59 %. Next, the features 3 and 5 are used with the a priori information giving the likelihood of right hits 67 %.

In this example, it can be seen that the last line of each iteration 1 to 3 gives the best reliability for the classification. It also can be seen that iteration 3 has not improved the reliability, and that is why the feature (feature 4) that has been added to the iteration doesn't need to be used in the classification. All in all, the result of the iterations is that feature 4 may be omitted in the classification while the rest of the features may be used.

In an embodiment, the optimization may be performed using a bot- torn-top algorithm an example of which is shown in Table 2.

Table 2

Iteration

1

Width Height feature 3 feature 4 feature 5 right hits

X X X X X 67 %

X X X X 61 % X X X X 62 %

X X X X 59 %

X X X X 67 %

X X X X 61 %

Iteration

2

Width Height feature 3 feature 4 feature 5 right hits

X X X X 67 %

X X X 60 %

X X X 61 %

X X X 59 %

X X X 67 %

In example of Table 2, no a priori information is used. In the first iteration, all the features are used in the first round and the likelihood of right hits is 67 %. Next, each feature is omitted in the classification one by one. When width is omitted in the classification, the likelihood of right hits becomes 61 %. Then height is omitted in the classification and the likelihood of right hits is 62 %. When feature 3 is omitted, the likelihood of right hits results in 59 %. When feature 4 is out of the features in calculation, the likelihood of right hits is 67 %. Finally, the likelihood of right hits is 61 % without feature 5. Here, it is clear that feature 4 should be dropped out of the features used in classifica- tion.

In iteration 2, feature 4 isn't used and the rest of the features are omitted one by one in the classification to find out which of them is not needed in the classification. In the first round, all the features width, height, feature 3 and feature 5 are used and the likelihood of right hits becomes 67 %. Next, height, feature 3 and feature 5 are used and the likelihood of right hits results in 60 %. When height is omitted, the likelihood of right hits is 61 %. When fea- ture 3 isn't used in calculation, the likelihood of right hits remains as low as 59 %. Finally, when feature 5 is omitted, the likelihood of right hits results in 67 %. Hence, it is clear that features 4 and 5 are not needed in this classification.

With both of these algorithms an "importance factor" may be associ- ated with each feature. In the first example it refers to an implication how much an addition of a feature improves classification reliability. In the second example it refers to an implication how much a removal of a feature decreases right hits in classification. From these factors it may be possible to evaluate what features may be included in classification and/or what features may be omitted.

In an embodiment, the supervised classifier 122 may analyze images of structural properties on the basis of at least one statistical classifier. The statistical classifier may comprise fuzzy-clustering, decision tree, multi-layer perceptron network, principal component method, hybrid realization, some combination thereof or the like. In an embodiment, the statistical classifier may be a multi-layer perceptron. A multi-layer preceptron, which needs supervised teaching, is a neural network with feedforward. It comprises a plurality of node layers, the nodes in one layer being connected to the nodes of the next layer.

When the supervised classifier 122 has been taught to distinguish different structural properties from each other, the supervised classifier 122 may be used to perform unaided classification of structural properties of new samples. For example a taught supervised classifier 122 may be used to divide structural properties of a sample of pulp in two or three groups or clusters such that one group contains fibers, another group contains vessels. The third group may contain other kind of structural properties such as fines, for instance. Then the processing unit 1 12 may measure statistical measures of fibers in the sample without vessels. The statistical measures may be mean, median, variance, and standard deviation of a length of fibers, for example. Because the vessels can be separated away from the measurement, the result is reliable and helps in control of the quality of paper made of the suspension, for in- stance. In general, the reliable grouping of structural properties and the measurement of their physical measures improves the quality control of a manufacturing process.

Figure 5 presents an example of a flow chart of an image organizing method. In step 500, the unsupervised classifier receives a plurality of images of structural properties, each image having a single structural property. In step 502, the images of the structural properties are organized with respect to each other such that the more similarity in predetermined features the images of the structural properties have the closer to each other the images are associated. In step 504, the unsupervised classifier outputs the images of the structural properties for teaching a supervised classifier.

Figure 6 presents an example of a flow chart of an image teaching method. In step 600, the organized images are received from an unsupervised classifier by a supervised classifier while having at least one label available for the supervised classifier, each label labeling a unique cluster of the images. In step 602, the supervised classifier is taught with the at least one labeled clus- ter of the images and the at least one label for enabling the supervised classifier to classify further images of structural properties on the basis of the teaching.

In an embodiment, the aspects of the invention may be realized as software and a computer or a set of computers of the processing system or a web service system connected to the Internet.

The computer programs may be in source code form, object code form, or in some intermediate form, and it may be stored in some sort of carrier, which may be any entity or device capable of carrying the program. Such carriers include a record medium, computer memory, read-only memory, and software distribution package, for example. Depending on the processing power needed, the computer program may be executed in a single electronic digital controller or it may be distributed amongst a number of controllers.

Certain cells such as hardwood tracheids are often considered harmfull from the printability and surface properties point of view. That is why there is need to measure their amount and grinding degree. However, the dimensions and structures of tracheids vary strongly among different wood species. At least one embodiment of what is presented above allows classification and identification of different structural properties including tracheids.

Another example could refer to enrichment of different minerals. When various minerals are found in a mine the minerals may be enriched and/or removed by flotation. In order to perform flotation successfully and effectively, the ratio of minerals and their quality should be known. At least one embodiment of what is presented above used in a particle analyzer allows classification and identification of different structural properties including vari- ous mineral particles. Different mineral particles may be distinguished from each other by their unique features in manner taught above. It will be obvious to a person skilled in the art that, as technology advances, the inventive concept can be implemented in various ways. The invention and its embodiments are not limited to the examples described above but may vary within the scope of the claims.

Claims

1 . Apparatus for organizing structural properties wherein the apparatus comprises an unsupervised classifier;

the unsupervised classifier is configured to receive a plurality of im- ages of structural properties, each image having a single structural property, and organizing the images of the structural properties with respect to each other such that the more similarity in predetermined features the images of the structural properties have the closer to each other the images are associated; and

the unsupervised classifier is configured to output the images of the structural properties for teaching a supervised classifier.

2. Apparatus for organizing structural properties, wherein the apparatus comprising a supervised classifier configured to receive the images from the unsupervised classifier of claim 1 and having the at least one label availa- ble, each label labeling a unique cluster of the images; and

the supervised classifier being configured to learn the at least one labeled cluster for enabling the supervised classifier to classify further images of structural properties on the basis of the teaching.

3. The apparatus of claim 1 , wherein the apparatus further compris- es at least one camera, a user interface and a screen, and the at least one camera is configured to capture the images of the structural properties;

the apparatus is configured to show at least a part of the images to a user on the screen;

the apparatus is configured to receive the at least one label from the user through the interface in response to the shown images, and feed the images with the at least one label to the supervised classifier.

4. The apparatus of claim 1 , wherein the unsupervised classifier is configured to organize images in a representation of one, two or three dimen- sions.

5. The apparatus of claim 4, wherein the unsupervised classifier is configured organize the images such that the more similarity in the predetermined features the images of the structural properties have, the closer to each other the images are placed in the representation.

6. The apparatus of claim 1 , wherein the at least one label is configured to indicate at least two separate clusters of the images, each cluster having at least one unique feature of the structural property with respect to each other.

7. The apparatus of claim 1 , wherein the unsupervised classifier is configured to organize the images on the basis of at least one clustering algorithm.

8. The apparatus of claim 7, wherein the unsupervised classifier is configured to organize the images on the basis of at least one of the following: self-organizing map of neural computing, t-distributed stochastic neighbor embedding, principal component analysis, sammon mapping method, GTM (General Topographic Mapping), LLE (Locally Linear Embedding) mapping, Isomap, agglomerative or hierarchial hierarchal clustering, including single-link-, complete-link-, average-link clustering, clustering error minimization, distance error minimization, K-means clustering, K-method, and graph-based methods like single-or complete link clustering, density based method, density-based spatial cluster of applications with noise (DBSCAN), AUTOCLASS, SNOB, BIRCH, MCLUST, or model based clustering COBWEB or CLASSIT, simulated annealing for clustering, genetic algorithms, Bayesian method, Kernel method, Multi- dimensional scaling, principal curve, T-SNE.

9. The apparatus of claim 1 , wherein the unsupervised classifier is configured to measure the predetermined features of the structural properties in the images and organize the images on the basis of the predetermined features.

10. The apparatus of claim 9, wherein the apparatus is configured to optimize the number of predetermined features measured from the images of the structural properties for classification in the supervised classifier by comparing a reliability of the classification with and without at least one feature and eliminating a feature not increasing the reliability.

1 1 . The apparatus of claim 2, wherein the supervised classifier is configured to analyze images of structural properties on the basis of at least one statistical classifier.

12. The apparatus of claim 1 1 , wherein the supervised classifier comprises at least one of the following: fuzzy-clustering, decision tree, multilayer perceptron network, principal component method, hybrid realization.

13. A method for organizing structural properties, wherein the meth- od comprises:

receiving, by the unsupervised classifier, a plurality of images of structural properties, each image having a single structural property;

organizing the images of the structural properties with respect to each other such that the more similarity in predetermined features the images of the structural properties have the closer to each other the images are associated; and

outputting, from the unsupervised classifier, the images of the structural properties for teaching a supervised classifier.

14. A method for organizing structural properties, wherein receiving, by a supervised classifier, organized images formed by the method of claim 13 while having at least one label available for the supervised classifier, each label labeling a unique cluster of the images; and

teaching the supervised classifier with the at least one labeled cluster of the images and the at least one label for enabling the supervised classi- fier to classify further images of structural properties on the basis of the teaching.

15. The method of claim 13 or 14, the method further comprising capturing the images of the structural properties by at least one camera;

showing at least a part of the images to a user on a screen; and receiving the at least one label from the user through an interface in response to the shown images, and feeding the images with the at least one label to the supervised classifier.

16. The method of claim 13, the method further comprising organiz- ing the images in a representation of one, two or three dimensions.

17. The method of claim 14, the method further comprising indicat- the at least one label at least two separate clusters of the images, each cluster having at least one unique feature of the structural property with respect to each other.

18. The method of claim 13, the method further comprising organizing, by the unsupervised classifier, the images on the basis of at least one clustering algorithm.

19. The method of claim 18, the method further comprising automatically organizing, by the unsupervised classifier, the images on the basis of at least one of the following: self-organizing map of neural computing, t- distributed stochastic neighbor embedding, principal component analysis, sammon mapping method, GTM (General Topographic Mapping), LLE (Locally Linear Embedding) mapping, Isomap, agglomerative or hierarchial hierarchal clustering, including single-link-, complete-link-, average-link clustering, clustering error minimization, distance error minimization, K-means clustering, K- method, and graph-based methods like single-or complete link clustering, den- sity based method, density-based spatial cluster of applications with noise (DBSCAN), AUTOCLASS, SNOB, BIRCH, MCLUST, or model based clustering COBWEB or CLASSIT, simulated annealing for clustering, genetic algorithms, Bayesian method, Kernel method, Multidimensional scaling, principal curve, T-SNE.

20. The method of claim 13, the method further comprising measuring, by the unsupervised classifier, the predetermined features of the structural properties in the images and organizing images on the basis of the predetermined features.

21 . The method of claim 20, the method further comprising optimiz- ing the number of predetermined features measured from the images of the structural properties for classification by comparing a reliability of the classification with and without each feature and eliminating a feature not increasing the reliability.

22. The method of claim 14, the method further comprising analyz- ing, by the supervised classifier, images of structural properties on the basis of at least one statistical classifier.

23. The method of claim 22, wherein the supervised classifier comprises at least one of the following: fuzzy-clustering, decision tree, multi-layer perceptron network, principal component method, hybrid realization.

24. An apparatus for organizing structural properties, the apparatus comprising:

at least one processor and at least one memory including a computer program code, wherein the at least one memory and the computer program code are configured, with the at least one processor, to cause the apparatus at least to perform:

organizing the images of the structural properties with respect to each other such that the more similarity in predetermined features the images of the structural properties have the closer to each other the images are asso- ciated; and

25. An apparatus for organizing structural properties, the apparatus comprising:

receiving, by a supervised classifier, organized images from the ap- paratus of claim 24 while having at least one label available for the supervised classifier, each label labeling a unique cluster of the images; and

teaching the supervised classifier with the at least one labeled cluster of the images and the at least one label for enabling the supervised classifier to classify further images of structural properties on the basis of the teach- ing.