WO1996007158A2

WO1996007158A2 - Method for detecting imperfections on a surface inspected

Info

Publication number: WO1996007158A2
Application number: PCT/FI1995/000469
Authority: WO
Inventors: Pertti Kontio
Original assignee: Rautaruukki Oy
Priority date: 1994-09-01
Filing date: 1995-09-01
Publication date: 1996-03-07
Also published as: FI944023A; FI944023A0; FI100366B; WO1996007158A3

Abstract

The invention relates to a method for detecting imperfections on a surface inspected. The method comprises the steps of providing and segmenting image information to form blobs (BL1-BL4); feature extraction for the blobs; clustering of the blobs, in which the image blobs are grouped into several clusters (A, B), each of which contains one or more blobs (BL1-BL4). The clustering is continued by further grouping the clusters (A, B) to form one or more higher-level clusters (AB) comprising lower-level clusters (A, B); and the method also comprises classification in which clusters are compared with imperfection prototypes and one or more clusters are defined that achieve sufficient classification certainty in relation to an imperfection prototype. In the invention, the clusters formed (A, B) are subjected to feature extraction, and the classification is conducted such that in addition to lower-level clusters (A, B), a higher-level cluster (AB) comprising these lower-level clusters (A, B) is compared with the imperfection prototypes.

Description

Method for detecting imperfections on a surface inspected

The invention relates to a method for detecting imperfections on a surface inspected, the method comprising the steps of providing pixels-containing image information about the surface inspected; segmentation of the image information, in which blobs are formed from the pixels; feature extraction for the blobs; clustering of the blobs on the basis of the feature extraction, the image blobs being grouped in the clustering into several clusters, each of which contains one or more blobs; and the method also comprising continuation of clustering by further grouping the clusters to form one or more higher-level clusters comprising lower-level clusters; and the method also comprising classification in which clusters are compared with imperfection prototypes and one or more clusters are defined that achieve sufficient classification certainty in relation to an imperfection prototype.

With regard to terminology, it is pointed out that a so-called blob is a group of adjacent pixels of the same brightness. A cluster, on the other hand, comprises one or more blobs. The present invention relates to computer vision systems utilizing digital image processing. It is applicable especially to detection of surface defects or other imperfections on band- or sheet-like products in quality control. In particular, the method can be applied in steel industry e.g. for checking a rolled steel band. A steel band rolled in cold-rolling mills is 1.5 to 2 m in length and may run forward at a rate of 10 to 1000 m/min. The band that is being rolled may comprise surface defects or other imperfections that must be detected and identified, or such imperfections may form thereon during the rolling. At least 50 different types of defects are known, e.g. a splinter, slag seam, scratch and hole. A splinter, for example, may vary in length from a few centimetres to several metres. In addition to the defects, the imperfections include e.g. numbers and other marks, differences in polish, and spots of grease, which are usually not regarded as imperfections impairing the quality of the product. Quality control is conventionally carried out by people trained for the purpose, who inspect by the naked eye whether there are any surface defects in the passing band. As man's ability to reliably inspect the surface of a band reduces as the rate of the band rises to above 100 m/min, computer vision systems must be used in quality control for the inspection. Computer vision systems based on digital image processing can be used for quality control up to band rates of 500 to 600 m/min. Both the currently used methods and the method of the present application comprise the steps of collecting pixels-containing image information; segmentation of the image information consisting of pixels into so-called blobs; feature extraction for the blobs; clustering, or grouping, of the blobs; and finally classification in which clusters are compared with known imperfection prototypes to define what best corresponds to a known imperfection prototype, the prototypes thus representing known defect models. The present method concerns the final steps of the method, i.e. clustering and classification.

It is characteristic of the metal surfaces inspected that an imperfection, in practice a defect, on a metal surface produces a splintered image. The defect on the surface may appear uniform, but produces, however, a splintered image. The inspection equipment should be able to detect a group of related imperfection areas in a splintered imperfection. Clustering is used for forming a sensible imperfection, i.e. a potential defect, and in classification the defect that best corresponds to a known defect, or an imperfection prototype, is selected from among these potential defects.

In known methods, clustering and classification are performed such that the blobs are classified in several parallel clusters, or groups, comprising one or more blobs each, the clusters being mutually independent, i.e. parallel, which means that each cluster comprises only so-called blobs. In known methods of this kind, only clusters that are separate from one another can be defined as defect areas in the classification step that follows the clustering step. This is a notable disadvantage, since especially in the event of a splintered defect, the known methods readily suggest that a certain area is a defect area, although in fact it is only one part of a larger defect area. In these known methods, it is thus possible that the entire defect area may not be covered by the clustering, whereby the defect areas of the surface inspected may be interpreted incorrectly. It may also happen that blobs are combined from too large an area, whereby a plural number of defects are interpreted as a single defect. One known method is disclosed in US 5,274,713, in which classification only aims at identifying large blobs, and no actual clustering even occurs. Another known method is disclosed in EP 048,568, in which parallel and successive blobs are combined by serial clustering.

EP 093,422 discloses a solution comprising clustering of clusters, in which no new features are defined for the already formed clusters but the clusters are classified on the basis of a combination of the classification results obtained on the blob level. Also, variable classification based on the configuration of a cluster in not used in the method. In the method of the publication, classification is based on non-fuzzy rules and not on variable parameters. The method does not show whether a higher-level cluster has a higher classification certainty than the lower-lever clusters that form the higher-level cluster. The intermediate results obtained during the clustering of the clusters are disregarded, if the higher-level cluster produced is sensible. The classifier used in the method is syntactic. The object of the invention is to provide a new kind of method, by which the problems of the prior art are overcome.

The object is achieved with the method of the invention, which is characterized in that before the clustering is continued, the clusters formed are subjected to feature extraction, on the basis of which the clusters are divided into further groups, and that the classification is performed such that in addition to lower-level clusters, a higher-level cluster comprising lower-level clusters is compared with the imperfection prototypes.

The method of the invention is based on the idea that clusters as well as blobs are clustered in the method. Two types of clusters are thus formed in the method: conventional clusters, which consist of blobs, and higher-level clusters, which consist of clusters, which in turn may consist of either blobs or clusters, depending on the level of clustering hierarchy. Hierarchic clustering, i.e. clustering in which totally overlapping clusters are formed, is characteristic of the present invention. Hierarchic clustering means that potential parts of an imperfection area are looked for in an image, and when they have been found, they are combined, or clustered, to better grasp the actual extent of the imperfection. The classification may result either in a higher-level cluster or in a lower- level cluster, which is what a higher-level cluster consists of.

Several advantages are achieved with the method of the invention. Even large imperfection areas can be detected quickly and reliably by the method. In the method, clusters can be combined to form a higher-level cluster, which together with the lower-level clusters can be compared with the known imperfection prototypes to find out which cluster best corresponds to a certain imperfection prototype. The method also helps to avoid a situation where a single, but splintered, area is placed in different imperfection classes. The method thus makes it possible to define, in a controlled and reliable manner, the imperfection class that an imperfection found on the surface inspected belongs to. Based on prototypes, or models, the method is especially well-suited for metal surface inspection, since the method is able to discard in the classification step the clusters and thus also imperfections that the user is not interested in.

In the following the invention will be described in greater detail with reference to the attached drawings, in which fig. 1 is a simplified scheme showing the measurement arrangement used in the method, fig. 2 shows a segmented and 2-level-clustered image of the surface inspected, fig. 3 shows a segmented and 3-level-clustered image of the surface inspected, fig. 4 illustrates imperfection prototypes, fig. 5 shows a max-min classification scheme with respect to three features, and fig. 6 shows the data structures of the clusters.

Fig. 1 shows a surface 1 to be inspected, which in one preferred embodiment is the surface of a band- or sheet-like product, such as a metal product. In the preferred embodiment, the surface 1 may be, for example, a steel band that is rolled in a cold-rolling mill and runs forward at a high speed. The measuring arrangement comprises a lighting means 2, an image production means

3, such as a CCD camera, and an image processing unit

4. The image processing unit comprises a segmentation block 4a, a feature extraction block 4b, a grouping block 4c, or clustering block 4c, and a classifier 4d. Fig. 2 shows an image of the surface inspected, reference number 2a indicating the background. In fig. 1, the direction of travel of the surface 1, which is preferably a steel band, is indicated by an arrow 5.

The method functions such that pixels- containing image information about the lighted surface to be inspected is produced by the image production means 3. The image consists of a large number of such pixels. The pixels form transverse image sequences, and successive image sequences taken together provide continuous image information. In the image production, the CCD solid state camera 3 can be used, the camera comprising e.g. 4000 pixels transverse across the surface 1. The image information is then segmented, which means that so-called image blobs BL1 to BL4 are formed from the pixels. In practice, blobs BL1 to BL4 are blobs of one or more imperfections, or surface defects. The blobs BL1 to BL4 differ from the surface 1 of the steel band in that they are continuous areas with the same brightness value, which is different from that of the background. Blobs BLl to BL4 are seen in fig. 2. Blobs BLl, BL3 and BL4 differ from the surface as dark imperfections, and blob B2 differs from the surface as a light imperfection, and the surface with no imperfections is distinguished as middletone grey. Segmentation thus defines image blobs BLl to BL4, which in practice are blobs, or elementary units, of the imperfection on the surface. Before or during the segmentation, e.g. compression and filtering can be used. In this respect, reference is made to the Applicant's own Finnish Patents 85,308 and 90,150.

After segmentation, the blobs undergo feature extraction. Feature extraction means that certain preferably geometrical features are defined for the blobs. In feature extraction, one or more of the following features can be defined for blobs BLl to BL4: orientation, length, width, angle, distance between blobs, density of blobs per area. Feature extraction is utilized in the following step, i.e. clustering. In the clustering, the image blobs BLl to BL4 are grouped in several clusters, each of which comprises one or more blobs. The simplest clustering method is a distance-based method. If" two blobs are sufficiently close to each other, they are placed in the same distance-based cluster. Clustering may also be based on orientation of blobs. If two (or more) blobs are elongated, parallel and successive, they form a serial cluster. If the blobs are elongated and parallel but not successive, they form a parallel cluster. If the image comprises only one or few large blobs, they form so-called one-blob clusters. If the image comprises only few small pixels, they form a so- called excess cluster. In fig. 2, blobs BLl and BL2 are clustered, or grouped, in cluster A, and blobs BL3 and BL4 are clustered, or grouped, in cluster B. In fig. 2, clustering in clusters A and B is based on the distance between the blobs.

The method of the invention utilizes so-called hierarchic clustering, which means, as shown in fig. 2, that lower-level clusters, such as clusters A and B, are clustered to form so-called higher-level clusters, such as cluster AB. In order that clusters A, B can be clustered into a higher-level cluster AB, the clusters must be submitted to feature extraction. In the feature extraction, one or more of the following features are defined for the clusters A, B: orientation, length, width, angle, distance between clusters A and B, density of clusters per area. The clusters that form a higher- level cluster are called lower-level clusters. In fig. 2, lower-level clusters A and B are grouped, or clustered, into a higher-level cluster AB. Even this clustering is based on distance, i.e. sufficiently short distance between clusters A and B. The term 'lower-level cluster' does not only mean a cluster consisting of blobs, but a 'lower-level cluster' should be understood as being independent of the level of clustering hierarchy. A lower-level cluster may thus consist of blobs, but it may also consist of clusters that are on an even lower level in the hierarchy. Fig. 3 shows a segmented and 3-level- clustered image of the surface inspected (different area from fig. 2). In fig. 3, the image comprises blobs BL11 to BLl6. When the blobs are submitted to feature extraction and first-level clustering, we obtain clusters H, I and J. Clustering of the blobs into clusters H, I and J is based on the serial form. When the blobs are submitted to feature extraction and second-level clustering, we obtain a higher-level cluster HI. The clustering of clusters H and I into cluster HI is based on the distance or serial form. The resulting new cluster HI is subjected to feature extraction, whereafter third-level clustering is performed to form cluster HIJ, which consists of lower- level clusters HI and J. The formation of cluster HIJ is based on serial form, i.e. successiveness of clusters HI and J. For example, cluster HI is on a higher level than clusters H and I, but on a lower level than the highest cluster HIJ. A cluster that is on a higher level than the clusters it consists of can thus be lower than an even higher cluster that it is part of.

As shown in fig. 2, clustering thus forms conventional clusters A, B consisting of blobs, and in accordance with the invention, also clusters, such as cluster AB, that consist of other clusters.

In fig. 3, conventional clusters H, I and J are formed, and in accordance with the invention, also clusters HI and HIJ consisting of other clusters are formed using hierarchic clustering. The second-level cluster HI consists of clusters H and I. The third-level cluster HIJ consists of the second-level cluster HI and the first-level cluster J. The preferred embodiment shown in fig. 3 can be presented more generally such that in the preferred embodiment of fig. 3, n-level clustering is performed, in which n - 3, 4, ..., and that in the method, one or more clusters are formed that comprise clusters from at least two lower clustering levels. In fig. 3, n = 3. In this preferred embodiment, the potential imperfection area can be widened, which produces an even clearer overall view of the extent of the imperfection.

In one preferred embodiment, re-clustering, or hierarchic clustering, is stopped by a suitably set restriction. The stop may be triggered e.g. by a time limit, number of clustering rounds, sufficient classification certainty achieved, or some other restriction. By this preferred embodiment, the method can be speeded.

When clusters are clustered into higher-level clusters, one or more of the following clustering principles are used in the preferred embodiment: serial form, distance, parallel form, size. The use of several clustering principles makes it possible to provide various imperfection models, whereby this embodiment enhances the reliability of the method. Corresponding clustering principles are preferably also used when blobs are clustered to form first-level clusters, whereby this embodiment simplifies and speeds the implementation of the method. The clustering algorithm used preferably comprises the following steps, preferably in the following order:

BEGINNING

STEP 1

If this is the first round, the blobs that meet the conditions are clustered using serial clustering, otherwise a move to STEP 2 follows.

STEP 2

The clusters that meet the conditions are clustered by serial clustering.

STEP 3

If this is the first round, the yet unclustered blobs that meet the conditions are clustered by distance-based clustering, otherwise a move to STEP 4 follows.

STEP 4 The clusters that meet the conditions are clustered by distance-based clustering.

STEP 5 A check is conducted to find out whether new clusters have been formed. If new clusters have been formed, a return to BEGINNING follows, otherwise a move to STEP 6 is made.

STEP 6

The unclustered blobs that meet the conditions are clustered by parallel clustering.

STEP 7 The unclustered blobs that meet the conditions are clustered by clustering based on the size of the blob.

STEP 8

The unclustered blobs are clustered to form excess clusters.

After the clustering, classification is conducted. Fig. 4 shows imperfection prototypes. Fig. 4 shows imperfection prototypes Cl and C2 as a function of two parameters, i.e. length and width. The imperfection prototypes, such as Cl and C2, are stored in file form in the classifier 4d. The area beyond the known imperfection prototypes Cl and C2 is indicated by X. The samples, which are pixels in the feature space, are indicated by a and b. It is naturally clear that there may be many more imperfection prototypes in the classifier 4d than the two prototypes Cl and C2 mentioned. There may also be more features.

In the classification step, which follows the clustering, the clusters, such as clusters A, B and AB (fig. 2 ) , are compared with the imperfection prototypes Cl and C2, and one or more clusters are defined that achieve sufficient classification certainty in relation to an imperfection prototype Cl or C2, the prototypes being shown in fig. 4. In the invention, classification is conducted by comparing the higher-level cluster AB, as well as the lower-level clusters A and B, with the imperfection prototypes Cl, C2, the higher-level cluster consisting of the lower-level clusters A and B. Of the clusters A, B and AB, the one that resembles one of the predetermined imperfection prototypes Cl or C2 the most is defined, preferably using the max-min method.

In fig. 3, in addition to the first-level clusters H, I and J, the second-level cluster HI, which consists of clusters H and I, is compared with the imperfection prototypes Cl and C2, and the third-level cluster HIJ, which consists of clusters HI and J, is also compared with the reference prototypes. Of clusters H, I, J, HI and HIJ, the cluster that resembles one of the predetermined imperfection prototypes Cl or C2 the most is defined, preferably by the max-min method, i.e. the cluster that has the highest classification certainty, i.e. membership value, in relation to a reference prototype Cl, C2. In one preferred embodiment, imperfection prototypes Cl and C2 that vary with respect to their defined classification certainty (membership value) are used in the classification, whereby it is possible that at least primarily continuous multistage numerical values are defined for the classification certainties of the clusters in the method, the values showing the correspondence between the clusters and the imperfection prototypes. In the classification, imperfection prototypes Cl and C2 based on the fuzzy set theory are preferably used. This is clarified in fig. 4 by the vertical axis, which indicates the membership, or classification certainty, of each pixel, or sample, of the feature space, such as samples a and b, in a defect class Cl, C2. Within the imperfection prototypes Cl and C2, classification certainty varies between 0.0 and 1.0 such that on the periphery of the imperfection prototypes Cl, C2 the classification certainty is at its smallest, approaching 0.0. In fig. 4, each imperfection prototype Cl, C2 defines a fuzzy possibility distribution, rather than a sharp area. This preferred embodiment enhances the reliability of the process, since the applicant has noticed that imperfection prototypes that are based on the fuzzy set theory or otherwise vary in size are closer to reality than sharp imperfection prototypes with the value of either 0 or 1.

The classification procedure can be divided into two: definition of the membership values, or classification certainties, preferably by a max-min method, and breaking up of the clustering hierarchy, which helps to avoid placing clusters that are part of a single imperfection in different imperfection classes.

The max-min classification scheme of fig. 5 illustrates three features (length, width, angle), two different imperfection prototypes, i.e. imperfection classes Cl and C2, being shown for each. In addition, fig. 5 shows three clusters A, B and AB, which correspond to the situation shown in fig. 2, i.e. A and B are blob clusters and AB is a higher-level cluster, or a so-called cluster-cluster, which consists of clusters A and B. Blob clusters A and B do not have any blobs in common.

The max-min algorithm is used for finding the best correspondence for clusters A, B and AB in the group of different imperfection prototypes Cl, C2. In the first stage, the minimum membership of a cluster in each class of imperfections for which a prototype exists is computed. On the basis of fig. 5, it is then possible to form tables 1 and 2, which show the membership, or classification certainty, of the samples (clusters) in imperfection classes Cl and C2. Table 3, in turn, shows the minimum membership values.

length width angle min

A 0.6 0.0 1.0 ->0.0

B 0.5 0.55 0.45 ->0.45

AB 0.0 0.0 0.0 ->0.0

Table 1: Membership of samples (clusters) in imperfection class Cl.

length width angle min

A 0.0 0.0 0.2 ->0.0

B 0.3 0.25 0.45 ->0.25

AB 1.0 0.8 0.9 ->0.8

Table 2: Membership of samples (clusters) in imperfection class C2.

Cl C2

A 0.0 0.0

B 0.45 0.25

AB 0.0 0.8

Table 3: Minimum membership of samples (clusters) in imperfection classes Cl and C2.

In the second stage of the max-min method, the highest computed minimum for each cluster A, B and AB is selected to find the imperfection class Cl or C2 that the cluster resembles the most. Table 3 shows that the membership, or classification certainty, of cluster AB in relation to imperfection prototype C2 is 0.8; the membership, or classification certainty, of cluster B in relation to imperfection prototype Cl is 0.45; and the membership, or classification certainty, of cluster A is 0.0, i.e. it is not identified and thus belongs to class X. After the first classification step, it thus seems that cluster AB represents class C2, cluster B represents Cl, and A represents class X (unidentified).

In the second step of classification, the clustering hierarchy is broken up, and the clusters, or samples, are placed in the final imperfection classes. The best classification result is here achieved with cluster AB, which belongs to class C2 with a certainty of 0.8 and is thus accepted as such. The second best result is achieved with B, which belongs to class Cl with a certainty of 0.45, but since B is part of cluster AB and AB has already been identified with a greater certainty, B is rejected. What remains is A, but cluster A is rejected since it is also part of cluster AB and since A has been defined as unidentified.

In order that it might be possible to break up the clustering hierarchy, in a method according to one preferred embodiment of the invention, dependence data INDL is added to the data structure of a higher-level cluster to indicate what lower-level clusters the higher-level cluster consists of. Correspondingly, dependence data INDH is added to the data structure of a lower-level cluster to indicate what higher-level cluster the lower-level cluster belongs to. Fig. 6 shows the following data structures for clusters: DATA (A), DATA (B), DATA (AB), DATA (H), DATA (I), DATA (J), DATA (HI), DATA (HIJ). Dependence data is indicated by INDL and INDH. For example, dependence data INDL = (A, B) for cluster AB indicates that cluster AB consists of clusters A and B. Further, dependence data INDH = (AB) for clusters A and B indicates that clusters A and B belong to the higher-level cluster AB. In the situation depicted in fig. 4, cluster HI would thus have both dependence data INDH = (HIJ) and dependence data INDL = (H,I). In the preferred embodiment of the invention, dependence data is utilized such that if two or more clusters (such as clusters B and AB) with a sufficiently high classification certainty are detected in the classification, and if the dependence data of the clusters show that the clusters are a higher-level cluster AB and a lower-level cluster B belonging thereto, the classification result of the cluster with a lower classification certainty is rejected in the classification, i.e. in this case the classification result of cluster B. The above-described preferred embodiment makes it possible to take cluster hierarchy into account, i.e. in practice to break it up, such that no contradictions occur in the classification logic in the classification.

An imperfection on the surface inspected may be splintered such that only separate parts of the imperfection, rather than the whole, can be 'seen' during the classification. It is also true that smaller objects are more difficult to classify since they are more general in outline. To avoid problems like this, in one preferred embodiment a classification difference factor is added to the classification certainty of the higher-level cluster to increase the classification certainty value of the higher-level cluster in relation to the classification certainties of the lower-level clusters belonging to the higher-level cluster. The classification difference factor that artificially increases the classification certainty can be indicated by CLD, and its value may vary between 0 and 1. If in the above example the classification certainty of cluster AB had been lower than that of cluster B, and if to the classification certainty of imperfection AB a classification-certainty-increasing factor CLD had been added that is higher in value than the difference of the classification certainties of clusters B and AB, then the method would have defined a higher classification certainty for cluster AB than for cluster B, thanks to classification difference factor CLD.

In a method according to one preferred embodiment of the invention, comparison with the imperfection prototypes is conducted in the classification in such an order that the clusters are first compared with the most significant imperfection prototypes, and then with less significant imperfection prototypes. This is implemented by arranging the imperfection prototypes Cl, C2, ... in the classifier 4d such that the imperfection prototype of the imperfection belonging to the most significant, i.e. the most harmful, imperfection class is placed first, and the others follow in the order of importance. In the preferred embodiment, if the classification shows that a cluster belongs to two or more imperfection prototypes with essentially the same classification certainty, the imperfection prototype that the cluster was first compared with is selected to be the imperfection prototype corresponding to the cluster. In these preferred embodiments, the classifier indicates the most significant, i.e. the most harmful, imperfections slightly before the less harmful imperfections. In the event of a 'draw', the first imperfection class, i.e. the most harmful one, is selected. The embodiment improves the usefulness, reliability and speed of the method. Although the invention has been described above with reference to the examples illustrated in the attached drawings, it is to be understood that the invention is not limited thereto but can be modified in many ways within the inventive idea disclosed in the attached claims.

Claims

1. A method for detecting imperfections on a surface inspected, the method comprising the steps of providing pixels-containing image information about the surface inspected (1); segmentation of the image information, in which blobs (BL1-BL4 or BLl1-BLl6) are formed from the pixels; feature extraction for the blobs (BL1-BL4); clustering of the blobs on the basis of the feature extraction, the image blobs (BL1-BL4) being grouped in the clustering into several clusters (A, B, or H, I, J), each of which contains one or more blobs; continuation of clustering by further grouping the clusters to form one or more higher-level clusters (AB) comprising lower-level clusters (A, B); and classification in which clusters are compared with imperfection prototypes and one or more clusters are defined that achieve sufficient classification certainty in relation to an imperfection prototype (Cl,

C2), c h a r a c t e r i z e d in that before the clustering is continued, the clusters formed (A, B) are subjected to feature extraction, on the basis of which the clusters (A, B) are divided into further groups, and that the classification is performed such that in addition to lower-level clusters (A, B), a higher-level cluster (AB) comprising lower-level clusters (A, B) is compared with the imperfection prototypes (Cl, C2).

2. A method according to claim 1, c h a r a c ¬ t e r i z e d by performing n-level clustering, in which n = 3, 4, ..., and forming one or more clusters (HIJ) that comprise clusters (HI, J) from at least two lower clustering levels.

3. A method according to claim 1 or 2, c h a r a c t e r i z e d in that imperfection prototypes (Cl, C2) that vary with respect to the classification certainty are used in the classification, and that multistage numerical values are defined for the classification certainties of the clusters (A, B, AB) in the method, the values showing the correspondence between the clusters and the imperfection prototypes.

4. A method according to claim 3, c h a r - a c t e r i z e d in that at least primarily continuous multistage numerical values are defined for the classification certainties of the clusters (A, B, AB) in the method, the values showing the correspondence between the clusters and the imperfection prototypes.

5. A method according to claim 3 or 4, c h a r a c t e r i z e d in that imperfection prototypes (Cl, C2) based on the fuzzy set theory are used in the classification.

6. A method according to claim 1, 2, 3 or 4, c h a r a c t e r i z e d in that the clusters (A, B, AB) are provided with a specific data structure [DATA (A), DATA (B), DATA (AB)], that dependence data (INDL) is added to the data structure of a higher-level cluster (AB) to indicate what lower-level clusters (A, B) the higher-level cluster (AB) consists of, and that dependence data (INDH) is added to the data structure [DATA (A), DATA (B)] of a lower-level cluster (A, B) to indicate what higher-level cluster (AB) the lower-level cluster (A, B) belongs to.

7. A method according to claim 6, c h a r ¬ a c t e r i z e d in that if two or more clusters (B, AB) with a sufficiently high classification certainty are detected in the classification, and if the depend¬ ence data of the clusters show that the clusters are a higher-level cluster (AB) and a lower-level cluster (B) belonging thereto, the classification result of the cluster with a lower classification certainty is rejected in the classification.

8. A method according to claim 1, c h a r - a c t e r i z e d in that a classification difference factor (CLD) is added to the classification certainty of the higher-level cluster to increase the classification certainty value of the higher-level cluster (AB) in relation to the classification certainties of the lower-level clusters (A, B) belonging to the higher-level cluster (AB).

9. A method according to claim 1, c h a r ¬ a c t e r i z e d in that comparison with the imperfection prototypes (Cl, C2) is conducted in the classification in such an order that the clusters are first compared with the most significant imperfection prototypes, and then with less significant imperfection prototypes.

10. A method according to claim 9, c h a r - a c t e r i z e d in that if the classification shows that a cluster belongs to two or more imperfection prototypes (Cl, C2) with essentially the same classification certainty, the imperfection prototype that the cluster was first compared with is selected to be the imperfection prototype corresponding to the cluster.

11. A method according to claim 1, c h a r ¬ a c t e r i z e d in that when clusters (A, B) are clustered into higher-level clusters (A, B), one or more of the following clustering principles are used: serial form, distance, parallel form, size.

12. A method according to claim 1, c h a r ¬ a c t e r i z e d in that the surface inspected (1) is the surface of a sheet- or band-like metal product, and that the imperfections detected are surface defects.