JPS6340844A - Surface flaw identification - Google Patents

Abstract

PURPOSE:To secure flexibility in application for other processes, by employing a multi-dimensional simultaneous pattern recognition based on a Mahalanobis distance for identification to reduce time and labor required for the building of an identification logic. CONSTITUTION:A flaw data analysis is performed on-line or off-line to determine a mean vector mui and a variance-covariance matrix SIGMAi with respect to a flaw type feature previously selected beforehand whereby an unknown flaw is identified. In this case, a flaw type feature vector (x) is determined to calculate a Mahalanobis distance Di for a known flaw type by the formula I and the flaw type with the distance Di minimized is identified as the unknown flaw type and the unknown flaw is identified to be an indistinct flaw belonging to none of available flaw type when the minimum distance is larger than a set value. In addition, a mean vector muij and a variance-covariance matrix SIGMAij of a grade feature value previously selected are determined for all grades in terms of flaw type beforehand to obtain a grade feature vector x' for the flaw type to which the unknown flaw was identified to belong. A Mahalanobis distance Dij is calculated by the formula II to identify the minimum grade as the unknown flaw grade.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention is directed to the type of surface flaws on an object to be inspected and their degree of harmfulness (
The present invention relates to a method for identifying grades (hereinafter referred to as grades).

[Conventional technology]

A method for identifying defects occurring on the surface of an object to be inspected, such as a rolled steel plate, is to take an image of a surface defect on a steel plate 1 with an ITV camera 2, and then use an A/D converter 3 as shown in FIG. The image is taken into the signal processing device 4 via the
For example, there is a method of extracting feature amounts in the X and Y directions from an image and performing flaw identification using these feature amounts. Dendritic logic is generally used to identify defects using this method (86th Measurement Subcommittee Material, Total 86-5)
-2). For example, as shown in Fig. 3, the dendritic logic performs a step-by-step comparison between a threshold value θ selected in advance for each feature and a feature of unknown flaw data for a plurality of features, as shown in Fig. 3. Ultimately, the purpose is to identify the type of flaw that the flaw data of unknown flaw type succumbs to. In this case, the flow of identification, for example, the special @ used in 5-6-7 in Figure 3
There is no theoretical method for determining the combination and order of mouths (it is usually determined empirically through trial and error. Also, as the number of flaw types increases, the number of characteristic mouths required for identification increases, It is clear that the logic itself is extremely complex.Furthermore, in the production of rolled steel plates, etc., numerous surface inspections are required in each manufacturing process up to the finished product, as well as inspections for each process. The types and properties of surface flaws that occur in various processes make it extremely difficult to apply dendritic logic created for a specific process to other processes.

[Problem that the invention seeks to solve]

As described above, the use of dendritic logic as a method for identifying surface flaws has problems such as increased labor required to construct the identification logic and lack of versatility. The present invention provides a surface flaw identification method that allows easy construction of identification logic and is highly applicable to other processes.

[Means 1 for Solving the Problems] FIG. 1 shows the flaw identification procedure of the present invention. A feature of the means for solving the above problems of the present invention is that multidimensional simultaneous pattern recognition using Mahalanobis distance is used, as will be explained below. The types of surface flaws are classified and established by humans based on their occurrence factors and properties, and even if they are the same type of flaw, it is impossible for them to have exactly the same characteristics. When looking at arbitrary characteristics (for example, length, width, area, etc.), as in the example of the distribution between rooms, the distribution for each flaw type is formed with uniform variations around the average value.

Therefore, the present invention analyzes a large number of defect data online or offline, and calculates the mean value vector pi and variance-covariance matrix Σi for preselected defect type characteristics.
is determined for each type of flaw, and when identifying an unknown flaw, the flaw type characteristic mouth vector , L represents the transpose of the vector) (in FIG. 1). Among these, Mahalanobis distance D
The flaw type for which i is the minimum is identified as the flaw type to which the unknown flaw belongs (Fig. 1 ■). At this time, if the minimum Mahalanobis distance is larger than a preset threshold value θ, the unknown flaw is identified as an unknown flaw that does not succumb to any type of flaw (■ in FIG. 1).

On the other hand, redundancy is provided in the identification results, such as the first candidate, the second candidate, etc. to which the unknown defect belongs in order of decreasing Mahalanobis distance Di, and information for final human judgment. It is also possible to do this. In addition, humans do not use the same identification criteria for all types of scratches regarding the degree of harmfulness (hereinafter referred to as grade); they first identify the type of scratch, and then use empirical methods for each type of scratch. identification criteria,
That is, grades are identified using grade features 9 such as size, unevenness, and contrast. When looking at the distribution in an arbitrary feature space for a certain type of flaw, it is distributed with statistical variation around the average value for each grade, similar to the distribution of the above-mentioned flaw type.

Therefore, as in the case of defect type identification, the mean value vector μij1 variance-covariance matrix Σij of the grade characteristic holes selected in advance for each defect type is calculated for all grades, and the defect type to which the unknown defect is identified is determined. Find the grade feature vector x' for identifying the grade in , and calculate the Mahalanobis distance Dij for all grades of that flaw type using the formula (%) (where j represents the grade) (Figure 1). ■). Among these, the grade with the minimum Mahala'nobis distance AltDij is identified as the grade with unknown flaws (Fig. 1 (■)).

〔Example〕

An example in which the method of the present invention is applied to the identification of surface flaws in the final process of manufacturing a stainless steel groove plate will be shown below.

For the nine types of flaws to be identified shown in Table 1, the flaw types are determined for the 11 types of flaw feature values shown in Table 2.

The mean value vector and variance-covariance matrix of the defect type characteristic gold are calculated in advance from data on about 1,600 defects with known grades.
As a result of identifying the type of flaw in the flaw data, a matching rate of 72% was obtained. In this case, the appropriate threshold value θ for determining whether the bottom is unknown is about 36. Regarding grade identification, four grades were identified using the grade feature amount, which is the sum of the base areas, and a 91% match rate was obtained.

Table 1: Flaw types to be identified and their properties Table 2: Flaw feature quantities for flaw identification [Effects of the invention] By using the present invention for surface m identification, the time and effort required to construct identification logic can be significantly reduced. , it has become possible to ensure flexibility for application to other processes.

[Brief explanation of the drawing]

FIG. 1 is a flowchart showing a surface flaw identification method according to the present invention, FIG. 2 is an explanatory diagram showing an example of the configuration of an apparatus for carrying out the present invention, FIG. 3 is a diagram showing dendritic logic for surface 7jff identification, and FIG. FIG. 4 is an explanatory diagram showing an example of the distribution of surface flaws in a multidimensional feature space. ■...Object to be inspected, 2...ITV camera, 3
...A/D converter, 4...signal processing device, 5,
6.7... Branching point of dendritic logic.

Claims (1)

[Claims] In a method for identifying a flaw type and grade using feature quantities extracted by image processing of a surface flaw on an object to be inspected, a flaw area whose flaw type is preselected from a large number of flaw data whose flaw types are known is used. Mean value vector μi and variance-covariance matrix Σi for the feature quantity
is calculated for each type of defect, and the Mahalanobis distance Di=(x-μi)＾tΣ＾-＾1_i(x- μi)(
i is a subscript representing the flaw type), the flaw type that gives the minimum value is identified as the flaw type to which the unknown flaw data belongs, and then the The mean value vector μij and the variance-covariance matrix Σij for the grade feature selected in advance are obtained for each grade of each defect type, and the grade feature vector x' obtained for the unknown defect data is
Mahalanobis distance D for all grades of the defect type to which the defect data to which the defect type is unknown belongs is identified for
ij=(x′-μij)＾tΣij(x′-μij)(
j is a subscript representing a grade), and the grade that gives the minimum value is identified as the grade to which the unknown flaw data belongs.