JPH09179985A - Inspection method and classifying method for image defect - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decide whether or not a defect is good by applying fractal dimensions as indexes of complicatedness to a calculated singular area and setting a threshold according to its size. SOLUTION: For the manufacturing process of an industrial product for uniformly applying, for example, a photosensitive material or dye, those methods are used to automatically inspect coating unevenness of paint and a fine flaw. For example, an optical means such as a CCD camera 2 is used to input a luminance distribution image of the surface of a uniform body to be inspected to a preprocessing part 4 and multi-gradational image data obtained by digital conversion are analyzed to detect the singular part (defect part = A) of the inspected body 1. This methods have a preprocessing means 4 which extracts an unsteady area of high possibility of defect presence from the input image, a calculating means 5 for fractal dimensions in the extracted image area, and a means which can binarizes the calculated fractal dimensions on the basis of a changeable set threshold, and the defect part and nondefect part of the extracted area are decided with its binarization output.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image defect inspection method and an image defect classification method, and more particularly, to a product that is uniform among industrial product defects using image processing. In regard to the determination of the quality of the product and the inspection of the nature of the defect, which is expected to be a defect in a part that has a (imagewise) unique change such as a scratch in an object or uneven coating of paint, for example, a dye of an optical disk The present invention is suitable for application to an automatic inspection device for detecting coating unevenness defects, OPC coating defects, and the like.

[0002]

2. Description of the Related Art The present invention provides means for performing a quality judgment (defect inspection) of a product to be inspected from two-dimensional image data of the product to be inspected obtained by using an image input device such as a CCD. However, in a conventional general defect detection method, a multi-valued image taken by a CCD is binarized, and the quality is judged by its area, perimeter, shape, or the like. However, as in the inspection of the OPC coating process, defects such as uneven coating with low contrast and minute projections lose their image information at the time of binarization, and therefore they are processed with multi-valued images. However, since the amount of calculation is very large, for example, in the automatic inspection apparatus for the OPC coating process, a high-frequency portion is extracted and binarized by using a Laplacian filter (second-order differential high-pass filter) as preprocessing. By labeling, the processing area in the multi-valued image is limited to the peculiar area of the image, and the quality of the defective portion is determined using the standard deviation which is the statistic of the multi-valued image in the label area.

[0003]

However, when the conventional multi-valued image data as described above is analyzed using statistical means, it is difficult to evaluate the geometrical shape and the amount of change in brightness. The optimum quality determination processing has not been performed for various defects that actually exist. In addition to the purpose of making a good / bad judgment, it is also highly expected to feed back the defect occurrence status to the previous process, but it is difficult at present to classify this defect type.

On the other hand, a fractal dimension has recently received attention as a measure of image complexity. In a basically uniform image like the above-mentioned object, only the defective portion has a very complicated change in luminance behavior, and the fractal dimension of the defective portion becomes larger than that of the non-defective portion. In some cases, the fractal dimension can be used to identify defects by itself, but in practice, it can be used in combination with a conventional detection system such as standard deviation to achieve a greater effect (high-precision defect detection and classification). It is considered possible.

[0005]

According to a first aspect of the present invention, for example, in a manufacturing process of an industrial product for the purpose of uniformly applying a photoconductor or a dye, an inspection of coating unevenness and minute scratches is performed. In order to perform automatically, for example, by inputting a uniform luminance distribution image of the surface of the object to be inspected by using an optical means such as a CCD camera, the multi-gradation image data obtained by digital conversion is analyzed. In the appearance defect inspection method for detecting a peculiar portion (defect portion) of the inspection object, a pre-processing unit for extracting a non-steady region having a high possibility of existence of a defect from the input image, and the extracted image region within the extracted image region. The present invention is characterized in that it has a fractal dimension calculating means and a means for binarizing the calculated fractal dimension with a threshold value capable of changing the setting, and discriminates a defective portion and a non-defective portion of the extraction region by the binarized output. So, The fractal dimension, which is an index of roughness, is applied to a singular area calculated by a method such as that described above, and a threshold is set according to the size of the singular area, so that the quality of defects can be determined. is there.

According to a second aspect of the present invention, in the image defect inspection method according to the first aspect, a plurality of basic amounts for calculating the fractal dimension, such as a minimum, a small, an intermediate, and a relatively large amount, are used when calculating the fractal dimension. The feature quantity is the fractal dimension of the fractal dimension, and the storage means has a fractal dimension database for each type of defect to be detected and classified, and the defect type is estimated by comparing the database with the calculated fractal dimension. By having a means, the type of defects in the corresponding image area is estimated and classified, and by changing the basic unit when calculating the fractal dimension from the image, the same calculation processing system can be used. Classification is performed according to the complexity of defects.

According to a third aspect of the present invention, in the image defect inspection method according to the first aspect, there is provided a plurality of fractal dimension calculating means including at least a fractal dimension relating to a luminance component in the X or Y direction of the image, and further detecting and classifying. For each defect type to be provided, a storage method having a calculation method and a fractal dimension database is provided, and by having a means for estimating the defect type by comparing the database and the calculated fractal dimension, the corresponding image area The feature is that the type of defect is estimated and classified, and therefore, when calculating the fractal dimension from an image, by providing a plurality of calculation methods, it is possible to use the fact that a change occurs depending on the shape behavior of the defect. The shape is used for classification.

The invention according to claim 4 is characterized in that it has a plurality of fractal calculating means and a database, which combine the inventions according to claims 2 and 3, and thereby classifies defects with higher accuracy. By combining the terms 2 and 3, it is possible to perform more accurate defect classification according to complexity and shape.

The invention of claim 5 is the invention of claim 2 or 3 or 4
In claim 2, the defect types are classified by constructing a neural network using each fractal dimension for each defect calculation means that has been classified in advance as teacher data, and thus the defect types are classified.
In the invention of, each fractal dimension is an input of a neural network trained by a model of each defect,
The defect is classified according to the output.

The invention of claim 6 is characterized in that, for each of the defects classified and estimated by the invention of any one of claims 2 to 5, the quality judgment is performed by using the value of the fractal dimension which significantly represents the defect degree. Therefore, by dynamically changing the pass / fail judgment parameter according to the defect type classified and estimated in the database according to claim 2 or 3 or 4 or in the neutral network according to claim 5, more accurate defect determination can be performed. It's something that you do.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a diagram showing an example of the overall configuration of a system assumed by an image defect inspection method according to the present invention, in which 1 is an inspection object having a defect A, and 2 is A camera 3 for photographing the surface of the object 1 to be inspected, an A / D conversion of the image photographed by the camera 2 to form a multi-valued image of luminance, and an A / D conversion and memory unit for holding in a frame memory, Reference numeral 4 is a pre-processing unit for extracting the peculiar region B of the image, 5 is a feature amount calculation unit for calculating the feature amount in the extraction region, and 6 is a pass / fail determination unit for determining pass / fail of the inspection object from the calculated feature amount. The feature amount calculation unit 5 and the quality determination unit 6 are the fractal dimension application unit.

The pre-processing means 4 can be constituted by a general-purpose image processing apparatus at present, but the subsequent calculation of the fractal dimension and the pass / fail judgment based on the value thereof are performed by dedicated hardware or software. Further, some fractal dimension calculating means may be provided in parallel.

FIG. 2 shows an example of the image extracted by the pre-processing unit 4. When the luminance information is taken as the Z axis and the position information is taken as the XY axes, it can be expressed as shown in FIG. this is,
Since it is expressed three-dimensionally, the three-dimensional surface formed by this brightness is hereinafter referred to as an image surface. If this fractal dimension D of the image surface is taken in another region similarly extracted from the preprocessing means, as shown in FIG. 3, the value of D becomes large in the region including the defect. Have a tendency. If the threshold value T is taken based on the magnitude of this D, it is possible to determine the quality of the product (claim 1). The threshold value for this determination is set in advance by the operator of the inspection device based on the defect data.

Next, a specific example of fractal dimension calculation will be described. Since the center of XY on the surface of the extracted image is located almost at the center of the defect, as shown in FIG. 4, the point inside the circle of radius r with this spatial point as the central point is N
Assuming (r), the correlation dimension ν is defined as a limit where r is sufficiently small.

[0015]

[Equation 1]

As defined by (for example, Order in Chaos, pages 135-139, commercial books, translated by Yoji Aizawa,
(See FIG. 5), when r is changed, N (r) and r take logarithms on both sides of the above formula, and when the prosthesis is taken, a linear tendency is shown as shown in FIG. The slope is obtained by using the least squares method, and the slope represents the correlation dimension.
This correlation dimension is one form of expression of fractal dimension, and the larger the inclination, the more complicated the image in the area and the more the defect is determined (claim 1).

Since the image defect portion is a finite area,
The larger the radius r, the smaller D of the image contained therein. This can be seen from the fact that in FIG. 5, the inclination becomes smaller as r becomes larger. The degree of change in this slope is
Since it depends on the defect, r is obtained as a slope in a minimum region, a slope in the middle, and a slope at a relatively large position, and a database is prepared in advance using typical defect images. By comparing the corresponding fractal dimensions of the extraction areas, it is possible to accurately classify various defects (claim 2). This is relatively easy to implement because various fractal dimensions can be calculated using the same calculation mechanism.

Further, as shown in FIG. 6, X passing through the center point
When the two-dimensional fractal dimension is obtained for each line segment in the direction or the Y direction, each represents the complexity in the X direction and the Y direction. This has a fractal dimension of various defects in the X direction or the Y direction and a database of similar dimensions as in the example of the invention of claim 1,
By comparing with the database, the defect type can be classified. This is not limited to the X and Y directions, and it is possible to identify the defect with higher accuracy by having various fractal dimension calculating means depending on the shape feature of the image surface of the defect (claim 3). . Since this has a plurality of different calculation mechanisms, the processing system becomes more complicated than the method of claim 2. Even in such a different processing system, when the fractal dimension is calculated, the degree of inclination change as in claim 2 is included in the database to enable more accurate defect classification. (Claim 4).

According to the second, third, and fourth aspects of the invention, the comparison and identification become very difficult due to an increase in the number of feature quantities called fractal dimension of the image in the database and classification of defects. Therefore, a back propagation neural network is used. Learning is performed using a defect model in which the output is the type of defect and the input is each fractal dimension, and is classified as the teacher data. The fractal dimensions of the input image are input to the learned neural network to classify defects (claim 5).

By classifying by the above method, using the fractal dimension that remarkably represents each defect type,
As in the first aspect, it is possible to determine the quality. Further, if the types of defects are different, even if the same fractal dimension is basically used, the threshold of the quality judgment can be changed. Therefore, the parameters of the quality judgment (fractal dimension and threshold) are classified for each of the classified and estimated defects. By changing dynamically, it becomes possible to detect defects better than in claim 1 (claim 6). The fractal dimension calculation system representing the defect type and its characteristics and the threshold value thereof are empirically determined. Further, the value used for the determination is not limited to one type. Further, here, "better defect detection" means a detection system in which a defect is surely detected and a non-defective one is less likely to be a defect.

[0021]

According to the first aspect of the present invention, by applying the fractal dimension, which is an index of complexity, to the calculated singular area, a defect is determined by setting a threshold value according to the size. Is possible. In addition, the method of extracting the singular area from the entire image area can be performed at low cost and at high speed, such as edge extraction processing and label processing by a general-purpose image processing device. Easier to build an inspection system,
In addition, it is easy to introduce other features.

According to the second aspect of the present invention, in order to calculate the fractal dimension, a plurality of sizes of basic units to be calculated are set, and a database is created in each of the basic units to classify the defects according to their complexity. Is possible.
Since this is based on the same calculation system, the classification detection system can be configured relatively easily.

According to the third aspect of the present invention, by providing a plurality of fractal dimension calculation methods, the results of each calculation method and defect type are stored in a database.
It becomes possible to classify defects with shape information.

According to the invention of claim 4, by combining the inventions of claims 2 and 3, it is possible to classify the defects with higher accuracy according to the complexity and the shape.

According to the invention of claim 5, claim 2 or 3
Alternatively, in the invention of 4), by inputting each fractal dimension to the neural network learned by each defect model, it becomes possible to classify the defects by its output, and suppress the cost for analyzing and updating the database. be able to.

According to the invention of claim 6, claim 2 or 3
Alternatively, it is possible to determine the defect with higher accuracy by dynamically changing the parameter for pass / fail judgment according to the database of the invention of 4 or the defect type classified and estimated by the neural network of claim 5. .

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of an entire inspection system assumed by an inspection method according to the present invention.

FIG. 2 is a diagram showing an example of a unique portion extraction region in which luminance is plotted on the Z axis.

FIG. 3 is a diagram showing a relationship between a fractal dimension of an extraction region and a quality judgment.

FIG. 4 is a diagram showing an example of counting the number of points in a sphere having a radius r from the center of the image surface.

FIG. 5 is a diagram showing a relationship between a radius r and N (r) contained therein.

FIG. 6 is a diagram showing line segments on the image surface in the XY directions.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Object to be inspected, 2 ... Camera, 3 ... A / D conversion and memory part, 4 ... Pre-processing part, 5 ... Feature amount calculation part, 6 ... Quality determination part.

Claims (6)

[Claims] 1. A multi-gradation image data obtained by inputting a brightness distribution image on the surface of an object to be inspected by using an industrial optical means for uniformly applying a paint and analyzing the image data, In an appearance defect inspection method for detecting a peculiar part (defect part) of an inspection object, a pre-processing unit for extracting a non-steady region having a high possibility of existence of a defect from an input image, and a fractal dimension of the extracted image region. An image defect characterized by having a calculating unit and a unit capable of binarizing the calculated fractal dimension with a threshold whose setting can be changed, and discriminating a defective portion and a non-defective portion of the extraction region by the binarized output. Inspection methods. 2. The image defect inspection method according to claim 1, wherein
When the fractal dimension is calculated, the basic amount for calculating the fractal dimension is a plurality of fractal dimensions as a feature amount, and a storage means having a database of each fractal dimension for each type of defect to be detected and classified, the database and the calculated fractal. An image defect classification method, comprising: estimating a defect type of a corresponding image area by performing classification by comparing the dimension with a defect type. 3. The image defect inspection method according to claim 1,
At least a plurality of fractal dimension calculating means including a fractal dimension relating to a luminance component in the X or Y direction of an image, a storage means having a calculation method and a fractal dimension database for each type of defect to be detected and classified, and its database. And an estimated defect type by comparing the calculated fractal dimension with the calculated fractal dimension, thereby estimating and classifying the defect type of the corresponding image region. 4. An image defect classification method, wherein a defect is classified with higher accuracy by having a plurality of fractal calculation means and a database which are a combination of claims 2 and 3. 5. The method according to claim 2, 3 or 4, wherein the defect types are classified by configuring a neural network using each fractal dimension for each defect calculation means that has been classified in advance as teacher data. Characteristic image defect classification method. 6. An image defect inspection method, characterized in that, for each of the defects classified and estimated according to any one of claims 2 to 5, the quality judgment is performed using a value of a fractal dimension that remarkably represents the defect degree.