JPH10289319A - Method and device for checking stripe-shaped defect - Google Patents

Abstract

PROBLEM TO BE SOLVED: To exactly detect stripe-shaped defects on the surface of object by adding a specified point to a concerned pixel when the angle difference of maximum and next lightness points on a contour estimated for the concerned pixel is a specified value, binarizing the pixel based on this point, removing isolate points and deleting a small area. SOLUTION: An image on the surface of object is optically inputted as a digital image, any arbitrary pixel is defined as the concerned pixel, and the set of pixels arranged in a square away from this concerned pixel for (k) pieces is defined as contour pixels. When a difference Θ between the angles of semistraight lines drawn to a maximum lightness point M1 and a next lightness point M2 among these contour pixels is close to π, that pixel is judged as the candidate point of stripe-shaped defect, and the fixed point is added to the concerned image. Further, among the nearby pixels included in the square estimated around the concerned pixel, the point of central pixel having a stripe direction toward the concerned pixel is added to the point of concerned pixel. Then, the pixels are binarized based on these points, the isolate points are removed and the small areas are deleted so that the stripe-shaped defect area can be found.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for detecting streak-like defects existing on an object surface. Since it is a streak-like defect, it is an anisotropic defect that is longer on one side and shorter in a direction perpendicular to it. It can also be called a linear defect. It is a long and thin continuous defect. Not necessarily a straight line. It may be a curve. In this case, it also means that the radius of curvature is large to some extent. The number of defects is not limited to one on the screen, and there may be a plurality of defects. In addition, there may be intersecting linear defects. Here, these continuous linear defects are called streak defects.

[0002] Localized round, square and wide defects are not included in streak defects. When a soft material is rubbed in contact with an uneven object, a streak-like defect occurs. For example, linear scratches on the end face of the fiber and linear scratches on the plastic plate are the streak-like defects referred to herein. The detection of a continuous linear flaw on a surface to be smooth, which is an issue here, is of interest here.

[0003]

2. Description of the Related Art JP-A-8-247957, "Defect detection method and defect detection apparatus for visual inspection of liquid crystal substrate or IC wafer"
The present invention relates to an apparatus for optically detecting various defects on a wiring pattern or a Si wafer or a glass substrate provided in a device lattice shape. This means that the wafer or substrate on which the device was made has various defects. Defects are detected optically, but the way of irradiating light is different for detecting defects of different types. It is said that it is necessary to irradiate a diffraction light obliquely to detect a shot shift defect. It states that in the case of a gap, it is necessary to apply oblique scattered light. In the case of defects such as unevenness in film thickness or scattering, it is necessary to irradiate interference light from above at right angles. It is better to perform image processing by irradiating light from three directions for the detection of three types of defects, but this takes too much time.

Therefore, three types of light are simultaneously applied to a wafer and a substrate to obtain all of diffracted light, scattered light, and interference light as an image, and various defects are to be obtained. The purpose is to reduce the inspection time. The details of the image processing are not described. It merely describes the ingenuity that three types of light beams are required and applied simultaneously. It states that the directions of the vertical and horizontal grid wirings are determined, and that the diffracted light is irradiated in a direction perpendicular to the directions. It is not intended to detect accidental scratches on a flat mirror surface. I don't know anything like streak defects.

[0005] Sumika Kawashima, "Series: Topic Products and Technologies About General-purpose Image Processing Devices," Image Lab, June 1995 (P75) states that general-purpose image processing devices have various functions, Edge extraction, center line detection,
It states that there are an expansion function, a contraction function, filling in holes, removal of isolated points, thinning, contour point extraction, feature point extraction function, area measurement function, centroid measurement function, and the like. This states that the binarization function is the most important among them, and states the problem of the current binarization function. Binarization removes many features,
The information you want disappears. He says this is the problem. He said that simply enhancing the contrast and binarizing it had no effect. The author states that by performing binarization using a different threshold value for each pixel, image recognition similar to that seen by human eyes can be achieved. However, there is no description about the recognition of a continuous thin line defect such as a line defect.

Japanese Patent Application Laid-Open No. 6-105157 discloses a “binary image generating method and apparatus” which obtains a difference image between an original image and an image obtained by shifting the original image by one pixel, and sums the plus side and the minus side of the difference image. Are determined as edge values, edge values smaller than a certain threshold value are dropped, the remaining edge values are integrated, and further binarized by a certain threshold value. The image shifted by one pixel and the image of the difference of the original image increase have a value where there is a change in shading. When the sum of them is obtained, the continuous change of the shading is changed to a steep discontinuous step change. Discard small change (differential). With this, light gray-scale changes such as noise and shading can be neglected. After integration, an image resembling a rectangular original image is reproduced. This not only reduces noise, but also functions as a waveform shaping circuit. Further, since the image is binarized in comparison with the threshold value, a binarized image free of faint noise or the like can be generated. However, this cannot be used to detect flaws when there are spots or dark noise. After all, it can be said that it is based on the premise that the noise density change is faint. In the case of a dark spot, it cannot be removed by such an operation. The light scratches disappear.

[0007] Japanese Patent Application Laid-Open No. 7-30753 discloses a "binarization processing method and apparatus" that obtains a difference image between an original image and an image obtained by shifting the original image by one or more pixels, and binarizes the density of the difference image by a certain threshold value. However, the threshold value is sequentially changed. It says that noise can be cut and binarized. Differentiating and binarizing is based on the same idea.
This also removes noise and the like assuming that shading and noise change in noise is small. There is no mention of flaw detection. Furthermore, when there is no difference between light and dark changes in the scratch and other defects, the two cannot be distinguished from each other. No streak defect (linear scratch) can be detected by any of the above.

[0008]

For example, a problem arises with a valve seat made of resin. Hot water and cold water come out of the same faucet, and the faucet internally combines water from the hot and cold water pipes and sends it out to the outlet pipe. A rotating valve body (valve seat) is used to adjust the inflow amounts H and C from the hot water pipe and the cold water pipe. This uses a resin sheet having excellent sliding characteristics. For example, a material in which a soft resin layer is coated on the front and back with a hard coating and cut into an appropriate shape is defined as a valve seat. Since the valve seat slides over the three openings of the hot water inlet, the cold water inlet, and the outlet provided in the circumference, it must have good wear resistance. It is required to have a longer life. Scratches on the surface can cause breakage and breakage.

Therefore, the resinous valve seat must be flat, smooth and intact. In the manufacturing process, linear scratches may be formed on the surface. Damaged items need to be inspected and eliminated. Conventionally, visual inspection was used. Each inspector visually inspects for scratches. However, manual visual inspection varies from person to person. The same person's pass / fail judgment changes according to the physical condition. Further, there is a problem that the inspection cost increases. There is a need for a technique that can automatically perform a flaw inspection without relying on visual observation.

[0010] Alternatively, a streak-like defect may also occur on the end face of the optical fiber. Since a long optical fiber is cut with a knife to a suitable length and polished, the end face may be damaged. Scratches on the end face may undesirably scatter light. Therefore, the inspector observes with a microscope and removes any scratches. This also varies from person to person. It is also desirable to automate the inspection of the optical fiber end face.

It is a first object of the present invention to provide a method capable of accurately detecting streak-like defects on the surface of an object in response to such a demand. It is a second object of the present invention to provide a method capable of detecting only streak-like defects ignoring spot-like isolated defects. It is a third object of the present invention to provide a method capable of detecting even a streak defect which is shallow and has a small difference in unevenness. A fourth object of the present invention is to evaluate the degree of defect of a product having streak defects so that defective products can be automatically eliminated.

[0012]

According to the method of detecting a streak defect of the present invention, an image of the surface of an object is optically input, converted into a digital image, brightness information of each pixel is stored in an image memory, and all pixels are detected. any pixel taken in the order as the pixel of interest D _j from
The set of pixels arranged in a square from the pixel of interest D _j k-pieces apart and contour pixels Nk, obtains pixel M1 with a maximum or minimum lightness of the outline pixels Nk from brightness information, except for the peripheral u pixels of the pixel M1 Out of the remaining contour pixels Nk, a pixel M2 having the maximum or minimum brightness is obtained, and M
1. The difference 角度 between the angles of the two half-lines drawn on M2 is π-α
≦ Θ ≦ π + α that pixel if it is (0 <α ≦ π / 3 ) is determined as a candidate point F _j of the line defect, the process of plus a constant scoring C1 to the pixel of interest D _j of the result image memory The contour pixel Nk is changed a plurality of times, and the streak direction L _{j of the} pixel of interest is determined as the direction of the half line drawn from the pixel of interest D _j to the pixel of interest M1.
Accommodated at the position of D _j Determination trace direction accommodating memory, do the same for all pixels of the original image {D _j}, determine the candidate score {F _j} and streak direction {L _j}, again any pixel taken sequentially from all the pixels to the target pixel D _j, W therearound
Among the neighboring pixels included in the × W square, a centripetal pixel D _k whose streak direction is directed to the target pixel is obtained, a candidate score F _k of the centripetal pixel D _k is added to the candidate score F _j of the target pixel, and the resultant image memory is stored in the result image memory. Then, the same operation is performed for all the pixels, an added candidate score {F _j } is obtained, the pixels are binarized by a certain threshold Q, and a streak defect which is a set of pixels satisfying F _j ≧ Q is obtained. The area Gm is obtained.

Further, isolated points are removed from the streak defect area, and a streak defect having no isolated points is obtained. Area S of streak defect
(M) is measured, and if it is larger than the allowable area S ₀ , it is determined that the product is defective.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention provides means for automatically detecting streak-like defects (scratches) on an inspection object by image processing. An image of the object is taken by a camera, the image is decomposed into pixels, brightness is given to the pixels as data, and an operation is performed for each pixel to determine whether or not it is a line defect candidate. Since the word streak defect is long, it is sometimes simply referred to as a scratch. The details of each procedure will be described below.

1. Optical image input: An image is taken with a TV camera by irradiating an object with light. The reflected light from the object becomes an image of the TV camera. The image is divided into, for example, 512 × 512 vertical and horizontal pixels. Pixels are hereinafter represented by D _j .
j is the number of the pixel. The two suffixes in the x and y directions are collectively written as j. Each pixel has brightness, that is, brightness gradation. The gradation is, for example, 256 steps. No color is required. Only the brightness information is required. Brightness is the brightness of a pixel. It is also called shading, density and gradation. Here we use the word brightness. Pixel data is stored in the image memory.

2. Target pixel setting: The same operation described below is performed for each pixel of the image. It will be referred to as pixels that are currently performing the operation as a pixel of interest D _j. The operation target pixel, that is, the target pixel is changed in the raster order. The processing of the pixels in one row is performed leftward from the pixels in the leftmost column of the uppermost row, and the process proceeds to the leftmost pixel in the second row to process the pixels in that row. In this manner, the same operation is performed for each row to reach the lower right pixel.

3. The pixel rows arranged on a square of the sides around the set ... pixel of interest D _j contour is referred to as a contour N. The size of the contour is represented by an integer k. A group consisting of squares including pixels that are vertically and horizontally diagonally separated by k from the target pixel is defined as Nk.
In other words, the outline of a square whose side is (2k + 1) is k
That. FIG. 6 shows the definition of the contour. N1 is a (3 × 3) square eight pixels immediately around the _target pixel Dj. N2 is the next (5 × 5) square 16 pixels. Generally, Nk is a set of 8k pixels. Nk is determined one in the pixel of interest D _j. If the pixel of interest D _j is different Nk also different. It should be written as Nk (D _j ), but it is written as Nk for simplicity.

In the ordinary image processing, the contour line refers to the outermost pixel row of the black pixels of the binary image. It is irregular and means the outer shell line of the image of the object. Such a contour line, also called an outline, is clearly different from the contour of the present invention. In the present invention, a square having a size of (2k + 1) is assumed around the pixel of interest, and this is referred to as a contour. Don't be confused with traditional contours.

The same processing is repeated for a plurality of continuous contours Nk (k = s to e). Ns is the minimum contour, N
e is the maximum contour. If there is a satin pattern, k
Pixels adjacent to the pixel of interest such as = 1 and 2 are not included in the processing. In order to remove small round spots, k is increased to some extent. For example, s = 6 and e = 9. If it is not a layer, one can start with s = 1.
S, e of the minimum (start) contour Ns and the maximum (end) contour Ne
Can be a parameter that can be set freely.

4. Obtain the pixel having the maximum (or minimum) brightness among the 8k pixels maximum (minimum) determining lightness point M1 ... contour of the pixel of interest _{D j Nk (s ≦ k ≦} e). Whether to determine the maximum brightness or the minimum brightness depends on the nature of the object and the streak defect. If the scratch (streak defect) is brighter than the target object, the pixel with the maximum brightness is selected, and if it is dark, the pixel with the minimum brightness is selected. Since the maximum (or minimum) is inconvenient, it is hereinafter referred to as the maximum brightness.

FIG. 3 illustrates a case where the maximum lightness point is obtained from the separation of N2. Since lightness L _j of all pixels already in the image memory, the maximum brightness point on the contour Nk is found immediately. This is called the maximum lightness point M1. This is a pixel that may be part of a streak defect. Maximum lightness point M
Because 1 is a function of only the target pixel D _j and the contour Nk. One of maximum brightness point M1 when determining the D _j and Nk is determined. In other words, it should be written as M1 (D _j , Nk). M1 does not mean that the point has the maximum brightness in the entire image. Don't get me wrong.

5. Decision ... same pixel of interest D _j next mark lightness point M2, the contour Nk, excluding the maximum brightness point M1 and its neighboring u pixel in the contour Nk, obtain the pixel having the maximum or minimum brightness. If the defect is bright, the one with high brightness is searched, and if the defect is dark, one with low brightness is searched. Since the distinction becomes complicated, the lightness (or darkness) following the maximum lightness point is referred to as a next-point lightness point M2. This is also D _j , Nk
Is determined only once. FIG. 4 shows the determination of the next lightness point in the case of the contour N2. The maximum lightness point M1 and the next lightness point M2 are a point having the maximum lightness on the same contour Nk and a point having the next lightness.

However, since there is no meaning when the pixel adjacent to M1 is the next point, it is necessary to search for the next lightness point distant from M1. For this purpose, M1 and its adjacent u pixels (for example,
From the (8k-1-2u) points excluding (2 pixels), the next lightness point is searched. The next lightness point M2 is searched from 8k-3 pixels when removing one pixel before and after, and from 8k-5 pixels when removing two pixels before and after. The next-point lightness point M2 is also a point that may become a part of the streak defect. Thus some point be part of the two wounds will have been extracted on the contour Nk of D _j.

6. Measurement of direction difference between M1 and M2: target pixel D
_It is checked whether or not the angle Θ formed by the two half lines drawn from _j to the maximum lightness point M1 and the next lightness point M2 is sufficiently close to π. This is a procedure that the pixel of interest D _j to assess whether a part of the wound (line defect). It is highly possible that the maximum lightness point and the next lightness point are part of the scratch. However, when setting the frame (contour) around the pixel of interest and checking the brightness,
This is not for checking whether the pixel on the frame is a part of the flaw but for checking whether the pixel of interest is a part of the flaw. If the scratch (streak defect) passes through the pixel of interest, the maximum lightness point, the pixel of interest, and the next lightness point should be substantially aligned. If it is on a straight line, the angle formed by a half line drawn from the target pixel to the maximum lightness point and the next lightness point is π.

It is checked whether 調 べ る is close to π for such a reason. FIG. 4 shows the procedure. If M1 and M2 are point-symmetric with respect to the pixel of interest, Θ is π. However, the pixel of interest may be a part of the flaw even if it is not completely point-symmetric but only slightly deviates therefrom. If the deviation margin is α and the angle difference 、 is π ± α, it is determined that the pixel of interest is likely to be a part of a flaw (streak defect). Points that may be part of the streak defect are referred to as “streak defect candidates” or simply “candidates”. That is, when π−α <Θ <π + α (1), the target pixel is set as a candidate for a line defect. α is π
/ 3 or less (0 <α ≦ π / 3).

Although a pixel is a point, it should be called a candidate point, but for simplicity, it is called a candidate. As shown in FIG. 5, a result storage image memory is created corresponding to the original image memory of the original image. This has the same dimensions as the original image memory and has a 1: 1 correspondence between pixels. The result-accommodated image is an image in which a streak score _Fj indicating that each pixel may be a part of a streak defect is to be accumulated. At first, all the pixels are reset to F _j = 0, and the possibility of being a line defect is added as a candidate score F _j . If this if the pixel of interest D _j satisfy (1) is a candidate. Add scores C1 at a position corresponding to the pixel of interest D _j in that case results accommodating image. If (1) is not satisfied, no points will be added.

7. Change contour and repeat ... one contour N
For s, the maximum lightness point, the next lightness point, and the angle difference as described above were obtained, and a score indicating that the target pixel was a flaw candidate was calculated. For the same pixel of interest D _j, it does the same for a plurality of contours. Why repeat the same thing for multiple contours? In FIG. 4, the maximum lightness point and the next lightness point are obtained for the contour N2, but the lightness is not checked for the pixel of interest. It has not been confirmed that the pixel of interest is directly a damaged pixel. Therefore, at this stage, only two pixels that may be flawed pixels are found around the pixel of interest. This is not enough. It is possible that an isolated wound happens to be on N2. If the maximum lightness point or the next lightness point is really the point above the wound, there should be a point on the wound on the extension of the straight line connecting the two points. Since the defect is a streak defect, the defect points must be aligned on a straight line.

Therefore, the maximum lightness point and the next lightness point must be obtained not only on one outline but also on a plurality of outlines surrounding the outline. Then, after Ns, the maximum brightness point M is similarly set for the contour of Ns + 1.
1. The next-point lightness point M2 and the angle difference Θ are obtained. This is, of course, different from the Ns M1 and M2 described above. However, the pixel of interest D _j is the same point. This is because the contour Nk is assumed to be concentric for the same target pixel. And Θ is (1)
When the expression is satisfied, the score C1 is added to the same D _j of the result accommodation image. If (1) is not satisfied, no points will be added.

As described above, the same operation is performed for the contour Nk (k = s,..., E). The same operation is performed (e−s + 1) times. The number of repetitions m = (e−s + 1) is 2 or more, but is appropriately determined according to the properties of the target image. Target pixel D _j by m single operation to obtain a score F _j of 0～MC. Which is stored in the column of D _j results accommodating image in FIG. The higher the score _Fj , the higher the possibility that the pixel of interest is a part of the scratch. It does not matter the brightness of the target pixel itself, and gives the probability that the target pixel at the center is flawed from the brightness of the contour. Also, the relationship between the contours is not considered at all.

The same is repeated for Ns to Ne, but how are the maximum lightness points and the next lightness points distributed in adjacent contours? That is not a problem. The maximum lightness point and the next lightness point in each contour should also be aligned on a straight line. However, the relationship between the arrangement of the maximum lightness point and the next lightness point between adjacent contours is not examined. Always angle of the pixel of interest D _j and these M1, M2 is the only problem whether near [pi. This is because the calculation is not complicated. If the correlation between the distribution of the maximum lightness point and the distribution of the next lightness point between adjacent contours is included, more accurate calculation will be possible.

8. Impart of muscle direction ... to define what that direction of the pixel of interest D _j to impart this to the D _j. A pixel is a single point, and the point is essentially non-directional. Therefore, it is strange that a pixel has a direction. This is strange, but nevertheless the present invention assumes a direction to the pixel of interest. This is one of the important features of the present invention. There are many other directions as well,
The direction attached to the pixel is referred to herein as the "streak direction" (L).
muscle direction of j pixel to be represented by L _j. This is defined to indicate the direction of the flaw assuming that the target pixel exists on the streak defect (flaw).

The streak direction has a geometrical meaning to indicate the direction of the flaw. Originally, a pixel is just a point. However, the idea is that if it is over the flaw, the pixel itself is also oriented according to the direction of the flaw. The line direction L is also defined when the pixel of interest is not over the scratch. At this stage, it is not known whether or not the pixel of interest is on the scratch, so the scratch (streak defect)
The direction L is also defined for a pixel of interest not above. In that case, the direction does not have a definite geometric meaning that it is the direction of the wound.

Since it is difficult to handle the direction with a continuous amount,
Classify only into 8 vertical and horizontal directions. That is, the angle between the space around the pixel of interest and the x-axis is 337.5 °, 22.
5 ゜, 67.5 ゜, 112.5 ゜, 157.5 ゜, 20
Divided by a half line of 2.5 °, 247.5 °, 292.5 °, and divided into eight directions a, b, c, d, −a, −b, −c,
-D. FIG. 19 shows the classification in eight directions. Actually, it does not matter whether the sign is positive or negative. FIG. 20 shows classification in four directions in which the distinction between positive and negative is omitted. The muscle direction L is a variable that takes four values.

[0034] So how can give the most accurate if the muscle direction L _j of the pixel D _j? This is the problem. (I) to the m maximum brightness point M1 for the m contour from the pixel of interest D _j catching half line. Its half-line is the direction that matches when matching the range of the m least four directions classification and streak direction of D _j. A half line drawn from the target pixel to the maximum lightness point M1 will be simply called an M1 half line. The above rules when entering one of the m M1 half lines are both 4 classification is that it its direction and streak direction of D _j.

[0035] (ii) if the m M1 half-line does not match There are several methods for determining the muscle direction D _j. (Ii) Maximum method: A method in which the most frequent direction in the four directions of m M1 half-lines is adopted as the muscle direction L. Even if the directions of the m M1 half-lines do not all match, the majority will fall into one category. (Iib) Contrast method: A method in which the direction of M1 that maximizes the difference in brightness between the direction orthogonal to the M1 half-line and the direction of the M1 half-line is adopted as the streak direction L. This is because the brightness should be clearly different between the direction in which the scratch grows and the direction at right angles. (Iic) Average method: A method in which the average value of the angles of m M1 half-lines is determined, and the direction of the average value is adopted as the muscle direction L.

When all M1 semi-linear directions do not match
By any of the above three methods, D _j Determine the muscle direction
Is also good. However, if they do not match, D_j Is on the wound
Most likely not a point. Muscle direction L_j Is D_j Is on the wound
It only means to suggest the direction of the wound when it is done. D
_j If there is no over the wound, line direction L_j In the first place it makes no sense
No. It doesn't make sense, so the decision may contain some errors
That's why. (Iia)-(iic) throat
There is no big difference even if you decide by the method of.

9. Repeat the operations 2 to 7 with the target pixel changed ... By the above steps, one target pixel D _j
Maximum brightness point M1, the next mark lightness point M2, the extraction of directional difference Θ for the calculation of the score F _j as a candidate, and confer muscle direction L _j could be performed. As data on D _j , a candidate score F _j and a muscle direction L _j are obtained. Such a process is sequentially performed for all the pixels in the image. For example, 1. Repeat the same operation in raster order as described in.

It is assumed that the number of pixels is R, and 0, 1, 2,.
.., R-1. R is the total number of pixels and is the product of the number of horizontal pixels and the number of vertical pixels. j = 0,1, ..., will promote the R-1, the score for all pixels _{{D j} {F j}} ,
The muscle direction {L _j } is determined. These are stored in the result storage image memory. Pixel score F _j is large is a high probability that the pixel on the line defect. That is, the score can be considered as an index indicating the probability of being a pixel on the defect. However, it is premature to judge that the pixel is part of the scratch just because it is high. It is necessary to confirm that pixels with high scores are continuously arranged in a line. The following steps are for examining continuity as an attribute of streak defects.

10. Added the number of centered pixels (consolidation processing)
... again assumed to be that the pixel of interest D _j. Previous F _j
In the same way, which is to repeat each and every similar operation in raster order and when to seek and L _j. Consider a square Tj having a size of W × W centering on a _target pixel Dj. W is an odd number such as 3, 5, 7, or the like. It contains W ² -1 pixels. These pixel groups Tj are not the contours described above. The square to distinguish it from contour will be called "near" Tj of the pixel of interest D _j. Let that element be {D _k }. The neighboring pixel D _k is already provided with the line direction L _k and the score F _k .

The sum ΔF _k of the scores F _k of the target pixel D _j centered on the muscle direction L _k among the neighboring pixels D _k is added to the score of the target pixel. This is the operation of the steps here. Since the operation is difficult to understand, it will be described in detail in order.

[0041] Among the neighboring pixels W ² -1, pick only the muscle direction L _k faces the pixel of interest D _j. FIG. 8 illustrates this.
This is the case in the vicinity of the minimum of W = 3, but the same is true in the vicinity of a larger dimension. Since there is no distinction of the muscle direction between positive and negative, there are only four directions. The direction of the arrow may be either. The eight neighboring pixels are 80 to 87. Muscle direction of neighboring pixels 80 is not fit to D _j of the center is y (vertical) direction. Muscle direction of the pixel 81 is not fit to be D _j is x (horizontal) direction. 83 also does not face the D _j. 8
4 does not face D _j either. 86 is also not suitable for the D _j faces the diagonal. 87 is also not suitable for the D _j is facing sideways. Those neighboring pixels not facing D _j are discarded. Such pixels are termed scattered pixels.

[0042] 82 is oriented in the direction of D _j. 85 is also D
It is pointing in the direction of _j . Pixels having a streak direction oriented at D _j is called a centripetal pixel. The fact that the streak direction such as 82 or 85 has a pixel (centripetal pixel) that points to the target pixel means that the target pixel is likely to be a part of the streak defect. Since the streak direction is the direction of M1 and M1 may be a part of the streak defect, if the neighboring pixel has a streak direction toward the target pixel, it is highly likely that the target pixel is the streak defect. Will be. As described above, the present invention employs a perspective method of clarifying the properties of the target pixel by the properties of surrounding pixels. From such a standpoint, the pixels ₈₂ and ₈₅ having the streak direction facing the target pixel are selected, and the sum of the scores C ₈₂ and C ₈₅ of the centered pixels is added to the score F _j of the target pixel. That is, a value _obtained by adding the sum ΣF _k of the scores of the centered pixels to F _j of the target pixel obtained last time is newly set as the score of the target pixel. F _j + ΣF _k → F _j .

The above example is a case where only the neighborhood of 8 where W = 3 is considered. However, W = 4 near 15 and W = 5 2
The same can be done in the vicinity of 4. However, in this case, there occurs a pixel in which the streak direction quantized in four directions does not always face the target pixel. These neighboring pixels are scattered pixels from the beginning.

The pixel of interest having a centripetal pixel is likely to be a part of a streak defect, and in that case, a certain point may be added for each centripetal pixel. However, here, the score of the centripetal pixel is added. That is also significant. Assuming that the target pixel has a centripetal pixel, the larger the value of the centripetal pixel itself is, the higher the probability that the target pixel is a line defect is. Thus, the score of the centered pixel is added to the candidate score of the target pixel. Such an operation is added for all the pixels {D _j }. The pixels on the streak defect add scores to each other, so that a flaw (streak defect) can be strongly highlighted. This is called connection processing.

Pixels with many candidate scores are likely to be points above streak defects. Then, pixels having a score Q (threshold) or more are selected. Only those that satisfy F _j ≧ Q are considered to form flaws. However, continuous pixels are not always selected as shown in FIG. Since pixels having a high score added each other as neighboring pixels to obtain a high score, it is highly probable that those pixels are connected vertically and horizontally and diagonally. This is also a point where the present invention has been elaborated. The operation is difficult to understand, so I described it in detail.

11. Binarization processing of concatenated processing image _Fj
Pixels satisfying ≧ Q are extracted as candidates for forming a streak defect, but are extracted in white, and in other cases (F _j <Q)
Is binarized as black. Until now, 256 gradations have been used, but here binarization is performed to obtain 2 gradations. In many cases, a region having an area is selected. This region is referred to as a line-like defect candidate region Gm. m is an area number. It is clearly divided into streak defect candidates and those that are not. Streaky defects first appear as linear continuum. This means that the streak defect has been detected by the binarization. This is the end of the above steps in the case of flaw detection. The following describes an inspection process in which a defective product is inspected using the result of the line defect detection and the defective product is removed.

12. Isolated point removal: Although the processing in the preceding stage is a concatenation process, since the continuity itself is not extracted, isolated points may be included. Therefore, an isolated point is searched for from the candidate area and eliminated. An isolated point means that one white pixel exists in isolation. This is done by examining the 8-neighborhood pixels D _j. Excluding the isolated points means that at least white pixels form a group of two or more pixels.

13. The line defect of the area measurement ... line defect is too large, the defective, even if a line defect and a non-defective if somewhat S ₀ or less. The entire area of the streak defect is S (m) and the area is measured. Therefore, if S (m) <S ₀ , this is determined to be good. If S (m) ≧ S _{0, an} article having such a wide streak defect is determined to be defective. Then, defective products are automatically removed. This enables the quality inspection to be automated.

[0049]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The procedure for detecting streak defects according to the present invention will be described again with reference to FIG. Subsequent to the start 20 of the program, an image input 21 is performed optically. Since the image is a digital image, lightness (darkness) information for each pixel can be obtained. Next, a line defect candidate is detected 22 for each pixel. As described above, this defines an outline Nk around the pixel of interest, finds the maximum lightness point M1 and the next lightness point M2 on the outline, and forms an angle formed by a half line drawn from the pixel of interest to M1 and M2. It is checked whether Θ is close to π. If Θ is close to π, + C is added to the processing result image.

The details of that part are depicted in FIG. The starting contour Ns is appropriately selected, and the process is started. The choice of the processing contour depends on the nature of the spots, uneven lighting and other noise. In order to remove speckles, s is set to 5 or 6, and the target pixel and the contour are separated. The same operation 34-40 from the initial contour Ns to the final contour Ne
repeat. Initially, the process is started from the initial contour with N = Ns.

The outline Nk (k = s,..., E) will be described in general. The pixel M1 having the maximum value of lightness on the contour
Ask for. This is the case for a light wound. In the case of a dark flaw, the pixel having the minimum value of lightness is M1. Since it is complicated to refer to each of the two cases, the brightest scratch is targeted, and the maximum lightness point is M1. The area around M1 is masked, and the next brighter pixel M2 is searched. Masking the surrounding u pixels is because the probability that the next pixel exists near M1 is high, but it is only the brightness that has been caught by M1 and it is not possible to extract it as two points on the line. Because it is appropriate. Since the next lightness point M2 distant from M1 has been obtained, a half line is drawn from the target pixel to M1 and M2, and the included angle is defined as Θ. This is expressed as a direction difference.

It is checked whether or not the direction difference Θ is π−α <π <π + α. That is, it is checked whether M1 and M2 are in opposite directions with respect to the target pixel. M
Since both 1 and M2 have high lightness, there is a high probability that they are points on streak defects. If this is on the opposite side of the target pixel, it means that there is a high probability that the target pixel itself is above the streak defect.
So we check if it is on the other side by the above inequality. α is a margin, but a positive value of π / 3 or less. If Θ is approximately π, the pixel of interest is a streak defect candidate point.
Add C1. C1 is a predetermined constant. If Θ is not π, C1 is not added. The value F _j of the processing result image is initially reset to zero.

After doing this for a certain contour, N = N + 1, and similar steps 34 to 39 are performed for the next contour. The same is repeated until N becomes Ne. The target pixel is given a score of F = C (e−s + 1) at the maximum. At the same time, the direction of M1 from the target pixel is given to the target pixel as the streak direction L.

Returning to the step "complete all pixels" 24 in FIG. The calculation of the score F _{j and} the calculation of the line direction L _j as described above are performed for all the target pixels. The end of all pixels 24 means that the same calculation is performed for all pixels (j = 0, 1, 2,..., R−
It means that you went about 1). Up to this point, for all the pixels {D _j }, the candidate score {F _j }, the line direction {L}
_jデ_ータ data is obtained.

Next, a step called a linking process 25 of streak defect candidates is performed. Although this has been often described, the pixel groups are not directly connected. This operation is to add the score of a pixel whose stripe direction is toward the center among the W × W neighboring pixels to the score of the central target pixel and to increase the score of a pixel whose stripe direction is opposite as a stripe defect candidate. The detailed steps are shown in FIG. Begins at start 43. The same processing is performed on all pixels one by one in raster order. The pixel to be processed is also referred to as a pixel of interest here. Around the pixel of interest D _j assume something called near a square W × W. Check streak direction L _j neighboring pixels D _k. Muscle direction L _k neighboring pixels D _k is, those facing the pixel of interest D _j in the center (the centripetal pixel) (positive and negative do not think) Select. This is step 45.

Next, the sum 候補 of the candidate score F _k of the centered pixel D _k
Calculate F _k . This is the meaning of the “calculation of the data sum of the result storage image of the corresponding pixel” 46. The sum of the scores of the centered pixels ΣF _k is added to the score F _j of the central target pixel, and this is set as a new score F _j . F _j + ΣF _k → F _j
That is. This is step 46 "sum total data addition". The connection process does not directly perform the connection operation, but adds the score to the connected pixels. The same addition is performed for all the pixels by changing the _target pixel Dj. Here, the process returns to "all pixels completed" in step 26 of FIG.

The pixel of the line defect candidate has already obtained a high score. Other pixels have only a low score at or near zero. Here, the threshold value Ft and the score _Fj are compared to binarize all pixels. If F _j ≧ Ft, a line defect pixel is determined. If F _j <Ft, it is determined that there is no streak defect. This is the binarization 27 of the connection processing image.

Next, isolated point removal 28 is performed. An isolated point having a small area that is meaningless as a streak defect is eliminated. What remains is a streak defect (scratch). Things like spots and shading have already fallen. The detection of the streak defect is completed by this point. The detection is completely performed by the previous operation.

The following is a procedure for automatically sorting out defective products. The area S (m) of the line defect Gm is measured. This is step 29. Comparing their total area to a constant value S _0. If the total area S (m) is equal to or less than S ₀ , it is determined that the product is good even if there is such a defect. Line defect total area S (m) is S ₀ or more _{(S (m) ≧ S 0} ) as long as those having such a long line defect is judged to be defective.
Items determined to be defective are excluded. This is step 3
0. At this point, only good products remain.

Example 1 (Scratch on Acrylic Plate) A scratch on a white acrylic plate was detected by the method of the present invention.
The original image is shown in FIG. When a streak is formed on a white acrylic plate, it becomes a darker line than the surrounding area. In addition, illumination unevenness (shading) 91 exists. Although this is shown as a dark line, in reality the brightness is continuously changing, and it is bright near the center of the image and darkens gradually toward the outer periphery of the image. Since the scratch is darker than the background (background), the darkness should be measured. M1 should be the minimum lightness point, and M2 should be the next point lightness point. However, since M1 has been referred to as the maximum lightness point and M2 has been referred to as the next lightness point, the same expression should be used in place of replacing lightness with darkness. M1 should be the maximum darkness point, M2 should be the next darkness point, but since we are used to terms so far, we will use the word lightness here, M1 is the maximum lightness point, and M2 is the next lightness point. I say that.

FIG. 10 shows the configuration of the system used for the test. The inspection light from the light source 2 is applied to the acrylic plate 1 which is the inspection object. The inspection light is reflected by the half mirror 3 and impinges on the inspection object 1. The reflected light is emitted from the inspection object (acrylic plate) 1 and enters the CCD camera 4. This forms an image of the inspection object with a lens, and the image information is decomposed into two-dimensional pixels by a two-dimensional CCD. This is A /
The signal is converted into a digital signal by the D converter 5. A lightness (darkness) signal is stored in the original image memory 6 as a digital signal for each pixel. The arithmetic processing unit 7 executes the above-described operation for all pixels.

The parameters are C1 = 4, Ns = 1, Ne =
6, W = 13 and α = π / 4. Start the contour
= 1. This means starting from the contour adjacent to the pixel of interest. There is no such thing as spots so there is no need to drop such things.

[0063] The first data of each pixel of the original image memory 6 based on, go take a pixel of interest D _j in raster order. In for a certain pixel of interest D _j to the contour Nk, maximum brightness point M1, the next mark lightness point M2, determining the angle difference theta. 3π / 4 ≦
If Θ ≦ 5π / 4, four candidate scores are added to the target pixel. 24 if ６ is in that range for all six contours
Get points score. Thus for all the pixels, it calculates the candidate point F _j as line defect, to accommodate this result image memory 8 (memory results stored image in FIG. 5). This is performed for all the pixels {D _j } in order. Scores {F _j } for all pixels are stored in the result image memory 8.

[0064] This is also all the pixels from the muscle direction L _j is also determined simultaneously for each pixel D _j from the direction of the maximum brightness point M1 {D
_j }, all muscle directions {L _j } are obtained and a muscle direction memory 9
Housed in

[0065] from all the pixels, again take any of the pixel of interest D _j. 13 × 13 neighboring pixels (168) D around D _j
Assume _k . The candidate scores of the afferent pixels whose muscle direction L _k is directed to the target pixel D _j are summed up (ΔF _k ) and added to the target pixel score F _j (F _j + ΔF _k → F _j ). The operation is performed for all the pixels {D _j }. An added score {F _j } is obtained. The score F _j is compared with a threshold value Q, and is binarized into a score higher than the score Q and a score lower than the score. Those with higher scores are connected streaky defect areas. The binarized data is stored in the binarized image memory 10.

The image of the streak defect obtained as a result is D / A converted and the result of the image processing is displayed on the monitor 13. FIG. 12 shows the result. The shading 91 is removed, and only the flaw (streak defect) 90 is extracted. This shows the advantages of the present invention. This is the end of the line defect detection.

Inspection of defective products having streak defects can be automated based on the result of extracting only scratches. An isolated point is removed from the streak defect area generated by binarization, the area is further calculated, the defect area S (m) is compared with the allowable area S _0, and if S (m) ≧ S _0, this is Since the product has an excessively large streak defect, it is excluded as a defective product.
The result determined by the area measurement is a digital input / output 11
(FIG. 10) is output to the outside as a signal.
Defective products can be automatically removed using an external control device.

FIGS. 13 and 14 show the results of processing the same inspection object by the conventional simple binarization method. The one in FIG. 13 is successful in eliminating shading. This is because the threshold value was increased. If the threshold is high, the shading 91 which is a weak dark area disappears. However, some of the scratches have disappeared, resulting in discontinuous scratches 92. There is a strong possibility that some streak defects are overlooked. In FIG. 14, the threshold is lowered so that the scratch 92 does not disappear.
However, instead, the shading 91 is entirely shaded, causing a remarkable malfunction. In both of FIGS. 13 and 14 based on the simple binarization method, only the flaw cannot be extracted cleanly.

Example 2 (Scratch on Sintered Parts) Scratch on the surface of the alumina sintered member was detected by the method of the present invention.
FIG. 15 shows an original image of the sintered member taken by a CCD camera. There are long scratches 94. Since it is a sintered member, even a good product has fine spots 95 on the surface. This is scratch 94
Instead, it is inherent to the surface of the sintered body. The spot 95 is not a mark indicating a defective product. Therefore, it is necessary to prevent spots from being extracted in the wound extraction. In the conventional binarization method, spots remain in the final result anyway, and the flaw extraction cannot be performed. The present invention has a remarkable effect in such a case. Since the flaw is bright, a point having a large brightness is detected as a flaw candidate in this example. The maximum brightness point and the next brightness point are literally the maximum brightness and the next brightness point.

The parameters are C1 = 4, Ns = 6, Ne
= 9, W = 9, α = π / 4. The reason why N1 to N5 close to the pixel of interest are omitted from the contour is to eliminate short round defects such as spots. Since the diameter of the largest spot is smaller than 5 pixels, the spot can be prevented from being detected if a spot farther than N6 is used. Then, the maximum lightness point M1, the next lightness point M2, and the angle difference Θ are obtained for the quadruple contours N6 to N9. A maximum lightness point M1 is found on a certain contour Nk.

The next u lightness point M2 is obtained by masking the u pixels in the vicinity. Further, the angle difference 計算 is calculated, and this is 3
π / 4~5π / if 4 of the pixel of interest candidate score F _j to 4
Add. If Θ is within the range of all four contours, score 16 will be obtained. Also, from the direction of M1 to the muscle direction _Lj
Ask for. Perform such an operation on all pixels,
A set of candidate scores {F _j } and muscle directions {L _j } is obtained.

Further, a new pixel of interest is newly set, and among the 9 × 9 neighboring pixels (80 pixels), a streak direction L toward the pixel of interest is set.
the score F _k of centripetal pixels having a _k, added to the score of the pixel of interest. Thus, the connection operation is performed for all the pixels. Then, the image is binarized and divided into an area which is a set of points on the wound and a background which is not a scratch. FIG. 16 shows the result. The wound image 96 is correctly obtained. The speckles disappear and cannot be distinguished from the background 97. The reason why the speckles are erased is that the contours Ns to Ne are set at positions away from the pixel of interest.

In the conventional method of simple binarization, FIGS.
become that way. FIG. 17 is an example in which the threshold value for binarization is set high in order to reduce the spots. The image 99 of the spot 95 is faint but appears. An attempt to eliminate speckles by increasing the threshold has failed. Not only that but also wound 9
The image 98 of No. 4 has been interrupted. Since the threshold value is high, a thin portion of the line defect is cut off. This should not be. In the embodiment of FIG. 18, the threshold value is set low so that the image of the scratch is not interrupted. The wound image 100 is not interrupted,
The speckle image 101 is rather emphasized. I can't do this either. With simple binarization, it is not possible to successfully extract only a flaw. This is an aspect of the present invention.

The line defect is extracted by the above operation. Further, an isolated point is removed from the streak defect, the area S (m) of the streak defect is measured, compared with the allowable area S _0, and S
If (m) ≧ S ₀ , this is determined to be defective.
A streak defect that is too large is a defective product. However, according to the present invention, the defective product can be automatically discriminated and discriminated and eliminated.

[0075]

According to the present invention, assuming that the pixel of interest is a contour, the maximum brightness point M1 and the next brightness point M2 are searched on the contour, and when the angle difference Θ is close to π, the score C1 is assigned to the pixel of interest.
Is added, and the score of a centripetal pixel having a streak direction facing the pixel of interest is assumed in the vicinity of the pixel of interest, and the score of the centripetal pixel is added to the score of the pixel of interest. Then, an isolated point is removed, an area is measured, and a small area is deleted to obtain an image of a streak defect. Unlike simple binarization, since a measure is taken to increase the score of a pixel located above a streak defect, spots, shading, and the like can be removed. Streak defects can be accurately and automatically detected. It can be used for detecting flaws on valve seats, flaws on optical fiber end faces, and the like.

For example, in the case of a 30 mmφ valve seat,
Assuming that image processing is performed on a screen of 512 × 512 pixels, the object length per pixel is 30 mm / 512.
= 60 μm. Although the width of the flaw is larger or smaller than 60 μm, the method of the present invention can extract streak defects without error. For example, in the case of an end face of an optical fiber having a diameter of 250 μm, if image processing is performed on a screen of 512 × 512 pixels, the length per pixel is 250 μm.
μm / 512 = 0.5 μm.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a procedure of a main part of a streak defect detection method according to the present invention.

FIG. 2 shows M1 and M2 among the streak defect detection methods of the present invention.
5 is a flowchart of a part for evaluating whether or not the target pixel is a part of the streak defect.

FIG. 3 shows a square outline N1, N2,
.. Are defined, and outlines for explaining an operation of detecting a pixel M1 having the maximum or minimum brightness in the definition.

FIG. 4 is a diagram showing the outline of pixels N1, N2,... Defined around a pixel of interest, by masking the pixel M1 having the maximum or minimum brightness and u pixels in the vicinity thereof, and having the highest brightness or the lowest brightness from the remaining pixels. FIG. 9 is a contour diagram for explaining an operation of detecting a pixel M2 (next brightness point).

FIG. 5 is a photograph of an object taken by a camera, decomposed into information for each pixel, and a digital original image storing brightness information for each pixel, and the angle Θ between M1 and M2 on the contour of an arbitrary pixel of interest is π.
FIG. 9 is a diagram illustrating a result accommodation image for adding C1 of a target pixel and storing the result when C1 is close to the target pixel.

FIG. 6 is a diagram illustrating an example of an outline defined around a target pixel. The square containing k pixels away from the target pixel is the contour N
k.

FIG. 7 is a flowchart illustrating an operation of adding a score of a pixel having a streak direction toward a target pixel among neighboring pixels of W × W to the target pixel score in the streak defect detection method of the present invention.

FIG. 8 defines a neighborhood of a W × W square around a pixel of interest, sets a pixel having a streak direction facing the pixel of interest as a centered pixel among neighboring pixels, and diffuses a pixel having a stripe direction not facing the pixel of interest. FIG. 13 is a diagram illustrating a case where W = 3 for describing an operation of adding the number of centered pixels as a pixel to the candidate number of a pixel of interest.

FIG. 9 is a diagram for explaining a pixel whose streak direction is directed to a target pixel, that is, a pixel selected as a centripetal pixel among neighboring pixels.

FIG. 10 is a schematic configuration diagram of an apparatus for executing the streak defect detection method of the present invention.

FIG. 11 is an original image when a streak defect of the present invention is detected using a scratched acrylic plate as an inspection object.

FIG. 12 is an image showing the result of detecting streak defects by the method of the present invention using the scratched acrylic plate of FIG. 11 as an inspection object.

FIG. 13 is an image showing a result of defect detection by a conventional simple binarization method in which a threshold value is increased to remove shading. The wound disappears halfway.

FIG. 14 is an image showing a result of defect detection by a conventional simple binarization method in which a threshold value is reduced so that a scratch is not broken. The scratches are preserved, but the shading is enlarged, resulting in a large and wide defect image.

FIG. 15 is an original image when the method of the present invention is applied to the detection of a scratch on a sintered member.

FIG. 16 is a processed image showing a result of performing the method of the present invention on the original image of FIG. 15;

17 is an image showing a result of defect detection by a conventional simple binarization method in which a threshold value is increased in order to remove speckles from the original image of FIG. The spots have decreased, but the wound has been interrupted.

FIG. 18 is an image showing a result of a conventional method in which a threshold value is reduced so that a scratch is not interrupted on the original image of FIG. The wound is not interrupted, but the spots are emphasized and enlarged.

FIG. 19 is a view showing eight directions a, b, c, d, -a, -b, and-in all directions of 45 degrees in order to quantize the muscle direction into eight directions.
The figure explaining that it divided | segmented to c and -d.

FIG. 20 does not consider the direction of the streaks, so that all directions are divided into eight directions a, b, c, and d by 45 degrees in order to quantize the directions from the eight directions into the four directions. FIG. 9 is a diagram for explaining that the directions are given the same symbols and regarded as the same direction.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Acrylic board 2 Light source 3 Half mirror 4 CCD camera 5 A / D converter 6 Original image memory 7 Arithmetic processing part 8 Result image memory 9 Strike direction memory 10 Binary image memory 11 Digital output 12 D / A converter 13 Monitor

Claims (5)

[Claims] 1. A type the image of the object surface optically, and the digital image, the brightness information of each pixel is stored in the image memory, an arbitrary pixel taken sequentially from all the pixels as a pixel of interest D _j, attention pixels from D _j and k pieces apart contour pixels Nk the set of pixels arranged in a square, seeking pixel M1 with a maximum or minimum lightness of the outline pixels Nk lightness information, excluding the peripheral u pixels of the pixel M1 A pixel M2 having the maximum or minimum brightness is obtained from the remaining contour pixels Nk, and the angle difference Θ between two half lines drawn from the target pixel to M1 and M2 is π
Determines that -α ≦ Θ ≦ π + α ( 0 <α ≦ π / 3) in which the candidate point F _j of the pixel is the line defect if, by adding the target pixel D constant scoring C1 to _j of the result image memory If Θ is not the case, the processing of not adding the contour pixel Nk from Ns to Ne
By changing to do more than once, the attention from the pixel of interest D _j pixel M1
Accommodated at the position of D _j muscle direction L _j a determined muscle direction accommodating the memory of the target pixel as the direction of the half line drawn to perform the same operation for all the pixels of the original image {D _j}, candidate score {F _j } and streak direction {L _j } are obtained, and an arbitrary pixel taken again from all the pixels is set as a target pixel D _j, and W ×
W a candidate score F _k centripetal pixels D _k determined centripetal pixels D _k that among trace direction is oriented target pixel neighboring pixels included in a square applied to a candidate score F _j of the pixel of interest, the results stored in the image memory Then, the same operation is performed on all pixels to obtain an added candidate score {F _j }, binarize the pixels by a certain threshold Q, and obtain a streak defect area which is a set of pixels where F _j ≧ Q. An inspection method for streak defects, wherein Gm is obtained. 2. Assuming that the value of s of the start contour Ns of the contour Nk is an integer of 2 or more, the maximum lightness point or the minimum lightness point M1 and the next lightness point M2 are obtained for the contour Nk separated from the pixel of interest. 2. The inspection method for streak defects according to claim 1, wherein an image having a short length is removed. 3. The isolated point is removed from the line defect region, measuring the area of the line defect S (m), the greater than compared to the allowable area S ₀ is the product is defective 2. The inspection method for streak defects according to claim 1, wherein the judgment is made. 4. A camera for optically inputting an image of an object surface and decomposing an image having a brightness distribution into pixels, an A / D converter for converting brightness information of each pixel from an analog value to a digital value, An original image memory that stores the brightness value of each pixel of the original image as a digital value, a result image memory that has the same dimensions as the original image memory, and stores a score indicating that each pixel may be a streak defect, Attention is paid to a streak direction memory which has the same size as the memory and stores the direction of a half line drawn to the maximum brightness (or minimum pixel) M1 in each pixel as a streak direction of the pixel, and an arbitrary pixel taken in order from all pixels As the pixel Dj, a pixel set arranged in a square separated by k pieces from the pixel of interest Dj is defined as a contour pixel Nk, and a pixel M1 having the maximum or minimum brightness among the contour pixels Nk is obtained from the brightness information, and u pixels around the pixel M1 are determined. Minute Pixel M2 having the maximum or minimum brightness among the remaining contour pixels Nk excluding the above is obtained, and the difference 角度 between the angles of the two half lines drawn from the target pixel to M1 and M2 is π
When −α ≦ Θ ≦ π + α (0 <α ≦ π / 3), the pixel is determined to be a candidate point Fj of a streak defect, and a certain score C1 is added to the target pixel Dj of the result image memory to obtain Θ. Otherwise, the process of not adding is performed by changing the contour pixel Nk from Ns to Ne.
To the target pixel M1
The line direction Lj of the pixel of interest is determined as the direction of the drawn half-line, and stored at the position of Dj in the line direction storage memory, and the same operation is performed on all the pixels {Dj} of the original image to obtain the candidate score {Fj} And a line direction {Lj}, an arbitrary pixel taken in order from all the pixels again is set as a target pixel Dj, and W ×
Among the neighboring pixels included in the square of W, a centripetal pixel Dk whose streak direction is directed to the target pixel is obtained, a candidate score Fk of the centripetal pixel Dk is added to the candidate score Fj of the target pixel, the result is stored in the result image memory, and the same operation is performed. Is performed on all the pixels, an added candidate score {Fj} is obtained, the pixels are binarized by a certain threshold value Q, and a line defect area Gm which is a set of pixels satisfying Fj ≧ Q is obtained. And a binary image memory for storing an image composed of pixels satisfying Fj ≧ Q, and information of pixel brightness and direction stored in the original image memory, the result image memory, the streak direction storage memory, and the binary image memory. A digital-to-analog (D / A) converter for converting an image into an analog value, a monitor for displaying the contents of an original image memory, a result image memory, a streak direction storage memory, and a binary image memory using the digital / analog converted values; Size of Recessed, position, the inspection apparatus of the line defect, characterized in that it comprises an output device and for outputting the like. 5. The arithmetic processing unit further removes an isolated point from the streak defect area Gm, measures an area S (m) of the streak defect,
If compared to the allowable area S ₀ greater than it determines the product as defective, streaky claim 4, characterized in that so as to output to the outside to be a defective Defect inspection equipment.