JPH0676069A - Surface defect detecting device - Google Patents

Abstract

PURPOSE:To automatically discriminate the surface defect of an inspection object by automatically generating the algorithm of surface defect inspection by using a little learning data. CONSTITUTION:The defect of the inspection object is automatically discriminated by picking up the images of the inspection object so that the flaw of an inspection part appears as the brightness and the darkness of images by an image pickup means 10, extracting the statistic information of the brightness and the darkness of the images of the inspection part or information relating to the size of the flaw or the both by a featured value extraction means 14, inputting the information of the images extracted into a neural network 16 and learning the inspected result of an inspector as teacher data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a surface defect detecting apparatus, and more specifically, a defect detecting algorithm is automatically generated by learning using a neural network to automatically judge a defect to be inspected. The present invention relates to a surface defect detection device.

[0002]

2. Description of the Related Art In order to automate a defect inspection of a product which an inspector visually inspects, a TV camera is conventionally used to image an inspection object,
With respect to the entire imaged image, image processing is performed to extract the feature value of the defect that should be a defect from the lightness (lightness) of the image,
The presence or absence of defects was judged. Since the properties of defects are different for each product to be inspected, the conventional image processing method
It is necessary to create an algorithm for defect inspection for each inspection object, and in order to solve this and automatically create an algorithm for defect inspection, the following techniques 1) to 3) using a neural network are used. Are known.

1) Japanese Patent Application Laid-Open No. 2-58182 This technique recognizes a pattern of an image by a neural network using the result of local Fourier transform of a two-dimensional image as a feature amount. Since this method uses the result of local Fourier transform of a two-dimensional image as a feature amount, it is effective in enhancing the feature by enhancing the feature, but the number of input data to the neural network is large and the network becomes large in scale. It takes a lot of time to process in the network. Further, the Fourier transform processing also requires a lot of time.

2) Japanese Unexamined Patent Publication No. 3-91878 This technique uses the result of Fourier transform as a feature amount as in the technique of 1) above, but it can be executed at a higher speed than the technique of 1) because it uses optical calculation. In this technique, in addition to Fourier transform as a primary process for obtaining the feature amount, image processing such as edge extraction from a two-dimensional image, thinning, rotation, enlargement, reduction, etc. is performed, and this may be input as the feature amount. It is possible, but both are premised on processing by an optical neural network. Further, these feature amounts are different from the feature amounts handled in the present invention.

3) Study on application of neural network to steel sheet surface flaw inspection (Nippon Steel & Kimitsu Works) In this technique, the presence or absence of flaws is inspected by detecting the reflected waveform of the laser beam irradiated in the width direction of the steel sheet surface. The inspection is performed for each range A having a size that covers a single flaw in the width direction.

The inputs to the neural network are two items, the feature quantity of the waveform intensity and the feature quantity of the waveform shape.
The number of input data to the neural network is reduced. The feature amount of the waveform intensity and the feature amount of the waveform shape are obtained as follows. That is, assuming that the position in the width direction is i, the intensity of the reflected light at the position i is xi, the maximum intensity within the range A is X _max , and the minimum intensity is X _min , the standard deviation σ within the range A according to the following equation (1). The feature quantity of the intensity of the waveform normalized by is calculated. Further, the quadratic average y _A of the range A is normalized by the quadratic average y _B of the range B wider than the range A by the following equation (2) to obtain the feature amount of the waveform shape.

In the expressions (3) to (4), N is the number of data in the range A, and in the expression (5), N is the number of data in the range A or the range B for calculating the quadratic average. .

Waveform intensity feature amount = (X _max −X _min ) / σ (1) Waveform shape feature amount = y _A / y _B (2)

[0009]

[Equation 1]

In this method, since the feature quantity of the waveform intensity and the feature quantity of the waveform shape, which are the statistical information independent of the position, are used as the feature quantity, the detection capability is deteriorated due to the position shift of the signal input to the neural network. But not
Since only the one-dimensional information in the width direction is input, the flaw in the length direction cannot be detected. Further, these characteristic amounts are different from the characteristic amounts of the present invention.

The present invention has been made in order to solve the above-mentioned problems. With respect to an image of an inspection target expressed in light and shade (shades), a neural network is used to determine the relation between a simple feature amount of the image and the inspection result. It is an object of the present invention to provide a surface defect detection apparatus that automatically generates a defect inspection algorithm by using less learning data by learning by using it and automatically determines a defect of an inspection target.

[0012]

In order to achieve the above object, the present invention provides a feature amount for extracting a feature amount of an image of an inspection region where an image of the inspection object is picked up so that a flaw of the inspection object appears as a contrast of the image. Defect determining means including an extracting means and a neural network, and automatically determining a defect of an inspection target by inputting a feature amount of the extracted image to the neural network and learning an inspection result of an inspector as teacher data, It is configured to include.

[0013]

The imaged characteristic amount extraction means of the present invention extracts the characteristic amount of the image of the inspection region in which the inspection object is imaged so that the flaw of the inspection object appears as the light and darkness of the image, that is, the light and shade. As the feature amount, either or both of the statistical information on the brightness of the image and the information related to the size of the flaw can be used. The defect determining means is equipped with a neural network, and the feature amount of the extracted image is input to the neural network to learn the inspection result of the inspector as teacher data to automatically generate the defect inspection algorithm and inspect the inspection target. Defects are automatically determined. As described above, according to the present invention, since the defect inspection algorithm can be automatically generated, the defect to be inspected can be determined even when it is difficult to create the algorithm.

[0014]

Embodiments of the present invention will now be described in detail with reference to the drawings. As shown in FIG. 1, the surface defect detection apparatus according to the present embodiment includes an image pickup unit 10 that picks up an image of an inspection target so that a defect of the inspection target appears as the darkness or darkness of the screen. A TV camera, a CCD camera, or the like can be used as the image pickup means 10. The image pickup means 10 is connected to the image feature quantity extraction means 14 for extracting, as a simple feature quantity of the image, 10 items of statistical information of brightness and darkness of the image, 16 items of information related to the size of the flaw, or both as a simple feature amount of the image by a method described later. Has been done.

The feature quantity extracting means 14 is connected to the neural network 16. The neural network 16 is composed of an input layer, one layer or a plurality of intermediate layers, and an output layer, and inputs the simple feature amount of the image extracted by the feature amount extraction means 14 to each of the input elements of the input layer as input data. Then, the inspection result of the inspector obtained in advance is learned as the desired output of the teacher data. This learning uses a generally known backpropagation method. The number of input elements in the input layer of the neural network 16 is the same as the number of input items, the number of intermediate elements in the intermediate layer is arbitrary, and the number of output elements in the output layer is one or more according to the form to be output. The output form is how to assign the presence or absence, degree or type of defect to the output element. For example, when assigning the presence or absence, degree or type of defect to the output value of the output element, one is assigned to the output element. When allocating, it becomes plural. Note that FIG. 1 shows a three-layer structure with one output element.

The operation of this embodiment will be described below. The flaw of the inspection target is picked up by the image pickup means 10 as the brightness of the image. The feature amount extraction means 14 extracts statistical information of light and darkness of the image and information related to the size of the flaw as a simple feature amount of the image by a method described later, and normalizes it to a value in the range of 0.0 to 1.0. Convert and output. The neural network 16 outputs the inspection result of the inspector as a desired output using the simple feature amount of the image extracted and normalized by the feature amount extracting unit 14 and the information related to the size of the flaw, or both as input data. Is learned as teacher data. The output value of the neural network 16 after learning is output according to the set output form, and the presence or absence or the degree of the defect is determined.

After the learning of the neural network 16 is completed, if necessary, the input values other than the focused input item are fixed, and the influence on the judgment result when the focused input value is changed is investigated, You may re-learn by removing the input items that have little influence on the determination. By removing the input items in this way, the number of input data can be further reduced, and the processing time can be shortened.

According to this embodiment, since the defect detection algorithm can be automatically generated from the simple feature amount of the image,
It is not necessary to develop an algorithm, so that the development period of the surface defect detection device can be shortened. In particular, it is effective when it is difficult to generate an algorithm because the inspection standard of the inspector is not digitized. Further, since it is not necessary to extract a feature amount that undergoes complicated processing from the image, it is possible to shorten the image processing time. Also, since the feature amount obtained by compressing the image information into a small amount is input to the neural network, the neural network can be made small,
The processing time can be shortened and the data required for learning can be reduced.

Next, a simple method for extracting a feature amount of an image will be described. From the cut-out image of the inspection region, a simple feature amount that is input data to the neural network is calculated by the following method. This feature amount is obtained by compressing information excluding the influence of the defect occurrence position. Note that the addresses of the pixels forming the image are i in the X direction and j in the Y direction.
, And the pixel brightness is px [i] [j].

1) In the case of statistical information on the brightness of the image In order to emphasize the brightness of the image, the difference dx [i] [j] of the brightness between the pixel of interest and the eight pixels (points) around the pixel of interest is calculated by the following equation. . However, only the image range is handled, and for example dx [0] [0]
Is the difference of 3 pixels in the image range.

Dx [i] [j] = 8 × px [i] [j] −px [i−1] [j−1] −px [i−1] [j] −px [i−1] [ j + 1] -px [i] [j-1] -px [i] [j + 1] -px [i + 1] [j-1] -px [i + 1] [j] -px [i + 1] [j + 1] ( 6) From the brightness px [i] [j] and the brightness difference dx [i] [j] between the pixel of interest and the surrounding pixels, which is obtained from the equation (6), the following 10 items of statistical information about the brightness of the image are calculated. Get the feature quantity.

Average value, maximum value, minimum value of brightness px [i] [j],
Difference between maximum value and minimum value, standard deviation Mean value, maximum value, minimum value, difference between maximum value and minimum value, standard deviation of the difference dx [i] [j], standard deviation 2) No. related to the size of the flaw In the case of 1 information (when a dark pixel corresponds to a defect) When the brightness px [i] [j] of the pixel of interest is darker than the brightness of surrounding pixels, the brightness of 8 pixels around the pixel of interest (whether there is a defect) Then, the total value bx [i] [j] of dark pixels is obtained as a feature amount related to the size of the flaw by the following equation.

If F (px [i] [j]) is 1, then equation (7) is executed. bx [i] [j] = F (px [i-1] [j-1] + F (px [i-1] [j]) + F (px [i-1] [j + 1]) + F (px [i] [j-1] + F (px [i] [j + 1]) + F (px [i + 1] [j-1] + F (px [i + 1] [j] ) + F (px [i + 1] [j + 1] (7) where F (x) is 1 when x is dark and 0 when x is bright.

From the total value bx [i] [j] of dark pixels, the first information relating to the size of the surrounding flaws with the number n of surrounding flaws of 1 to 8 is obtained by the following equation, and 16 items are obtained. Get the feature quantity of.

Accumulation of feature quantity of flaw size = Σ F1 (n, bx [i] [j]) (8) Average of feature quantity of flaw size = Σ F1 (n, bx [i ] [j]) / Σ F2 (n, bx [i] [j]) (9) where F1 (n, x) is 0 when x ≧ n and 0 F2 (when x <n n, x) is 1 when x ≧ n and 0 when x <n.

3) Second information related to the size of a flaw (when a bright pixel corresponds to a flaw) When the brightness px [i] [j] of the pixel of interest is brighter than the surroundings, The brightness (darkness) of the surrounding 8 pixels is checked, and the total value wx [i] [j] of bright pixels is obtained as the feature amount related to the size of the darkness by the following equation.

If F (px [i] [j]) is 1, then equation (10) is executed. wx [i] [j] = F (px [i-1] [j-1] + F (px [i-1] [j]) + F (px [i-1] [j + 1]) + F (px [i] [j-1] + F (px [i] [j + 1]) + F (px [i + 1] [j-1] + F (px [i + 1] [j] ) + F (px [i + 1] [j + 1] (10) where F (x) is 1 when x is bright and 0 when x is dark.

From the total value wx ij of the bright pixels, the second information relating to the size of the surrounding flaws in which the number n of the surrounding flaws is 1 to 8 is obtained by the following formula, and the feature quantity of 16 items is obtained. .

Accumulation of feature size of flaw size = Σ F1 (n, wx [i] [j]) (11) Average of feature size of flaw size = Σ F1 (n, wx [i ] [j]) / Σ F2 (n, wx [i] [j]) (12) where F1 (n, x) is 0 when x ≧ n and 0 F2 (when x <n n, x) is 1 when x ≧ n and 0 when x <n.

Next, a description will be given of the result of examining the optimum image feature amount when the defect detection algorithm is automatically generated and the defect is automatically detected using the surface defect detection apparatus of the present embodiment.

The surface of the cylinder head that contacts the valve seat is inspected, and the above feature amount extraction methods 1), 2), 3) are used.
Defects are obtained by performing learning with a neural network that uses any of the input information as input information and the inspection result of the inspector as a teacher, and inputting multiple inspection images that were not used for learning to the neural network after learning. The correct answer rates for the presence and absence are 100% and 99.4, respectively.
〔%〕Met.

In this embodiment, the valve seat contact surface of the cylinder head was inspected, but the present invention is not limited to this. For example, surface defects due to material processing such as cutting or drilling, casting or die casting. It can be applied to the detection of surface defects such as surface defects of products such as defects on the inspection site that can be proposed as the contrast of the image.

[0033]

As described above, according to the present invention, since the image feature amount is extracted and the extracted image feature amount is input to the neural network for learning, the brightness of the pixels forming the image is input. Compared to the above, the neural network can be made smaller, the processing time is shorter, the amount of data to be learned is less, and the surface defect can be automatically judged by automatically generating the algorithm for judging the surface defect. The effect of, is obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

[Explanation of symbols]

 10 Imaging Unit 14 Feature Extraction Unit 16 Neural Network

Claims (1)

[Claims] 1. A feature amount extraction means for extracting a feature amount of an image of an inspection region imaged of an inspection object such that a flaw of the inspection region appears as a contrast of an image, and a feature of the extracted image, which is equipped with a neural network. A surface defect detection device comprising: a defect determination means for automatically determining a defect to be inspected by inputting an amount into a neural network and learning an inspection result of an inspector as teacher data.