JP3635795B2 - Method and apparatus for detecting fine line defects - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a method and apparatus for detecting a fine linear defect from an image obtained by imaging a detection object.
[0002]
[Prior art]
Attempts have been made to automatically detect fine linear defects such as welds, metal hair cracks, or cracks in the concrete surface using image processing techniques.
[0003]
Conventionally, in order to detect a defect from an image obtained by imaging an object to be detected, a two-step threshold method is used when the defect portion is relatively large and has a large shading difference.
[0004]
In the two-step threshold method, first, binarization is performed at a low threshold value, so that only a very dark portion 11a which is a part but surely a defect is extracted as shown in FIG. Then, the image 10a including no noise is created. Next, by binarizing with a high threshold value, as shown in FIG. 10B, an image 10b in which all the defective areas 11b including the noise 12b are extracted is created. Finally, as shown in FIG. 10C, in an image 10c obtained by superimposing both images 10a and 10b, a binarized image region with a low threshold value in a binarized image with a high threshold value. The region 13 including 11c is left, and the region 12c not including the binarized image region by the low threshold is removed as noise. As a result, a region having a low density value is extracted as a defect region 11d among regions binarized with a high threshold as shown in an image 10d in FIG.
[0005]
[Problems to be solved by the invention]
However, if the area to be extracted, such as hair cracks or cracks, is thin and the contrast with the surroundings is small, lowering the binarization threshold results in discontinuous breaks in the area line, and higher noise also extracts more noise. Therefore, even if the two-stage threshold method is used, it does not operate effectively.
[0006]
This will be described with reference to FIG. 9. FIG. 9A shows a binarized image 15a with a low threshold value that does not contain noise, and FIG. 9B shows an image binarized with a threshold value that can extract all lines. 15b and FIG. 9 (c) show an image 15c binarized with a threshold value higher than the threshold value of FIG. 9 (a).
[0007]
The image 15a shown in FIG. 9A and the image 15b shown in FIG. 9B are overlapped to form an image 15d shown in FIG. 9D. From this image 15d, an image including the linear defect area 16a included in the image 15a. Although the noise 17b can be removed as shown in the image 15f of FIG. 9F even if the region 16b of 15b is extracted, the region to be extracted is discontinuously broken like the linear defect region 16b of the image 15b. Therefore, some linear defects cannot be extracted with the extraction result 15f.
[0008]
Further, in order to extract all the linear defects, the image 15b in FIG. 9B and the image 15c in FIG. 9C are overlapped to form an image 15e shown in FIG. 9E. If the region including the linear region 16c of the image 15c is extracted, the problem that the noise 17g is also included in the defect region 16g as shown in the image 16g of FIG.
[0009]
As described above, when extracting with the two-step threshold value, the area to be extracted cannot be extracted or the area to be extracted can be extracted depending on the threshold setting, but a lot of noise is extracted. Therefore, it is difficult to set optimal parameters.
[0010]
Accordingly, an object of the present invention is to solve the above-described problems and provide a fine linear defect detection method and apparatus capable of accurately extracting a linear defect without changing parameters every time an object is extracted. It is in.
[0011]
[Means for Solving the Problems]
In order to achieve the above object, according to a first aspect of the present invention, in a method for detecting a fine line defect from an image obtained by imaging a detection object, after performing normalization processing on the image, shading unevenness is detected from the normalized image. For the image that has been removed and the shading unevenness has been removed, the binarized area with a low density threshold that can remove noise is used as the initial nucleus, and a new binary is added in a density threshold range slightly higher than this area. If the converted region is close to the region obtained so far, the method of connecting to the obtained region is repeated in multiple steps from the low concentration threshold range to the high concentration threshold range. This is a method for detecting a fine linear defect that is detected as a candidate for a linear defect by repeating the process. In the invention of claim 2, after extracting the linear defect candidates, each linear defect candidate is described in a straight line segment, and the segments in which the continuity of the direction of the straight line segment is locally maintained are integrated, It is determined whether two or more segments generally form a straight line or a quadratic curve, and when forming a quadratic curve, the constituent segments are combined, and local defect regions are combined. The fine linear defect detection method according to claim 1, wherein the fine linear defect is extracted by combining the global defect regions.
[0012]
According to a third aspect of the present invention, there is provided an apparatus for detecting a fine line defect from an image obtained by imaging a detection object, an image input device for imaging the detection object, and a normalized image after normalizing the image. In the image from which the shading unevenness is removed from the image, the binarized area with a low density threshold that can remove noise is used as an initial nucleus, and the density threshold range slightly higher than this area is used. If the newly binarized area is close to the area obtained so far, the method of connecting to the obtained area is multi-staged from the low density threshold range to the high density threshold range. The image processing apparatus that detects the line defect candidates and the line defect candidates are described in the straight line segments, and the continuity in the direction of the straight line segments is locally maintained. Segment In the case where two or more segments generally form a straight line or a quadratic curve, and a quadratic curve is formed, the constituent segments are combined, This is a fine line defect device including a determination device that extracts fine line defects by combining defect regions and global defect regions.
[0013]
According to the above configuration, a good linear defect candidate is detected and distinguished from noise only from the image grayscale information, and the obtained linear defect candidate is combined with the defect portion by local / global search. Further, it is possible to detect a linear defect more satisfactorily.
[0014]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings.
[0015]
First, the device configuration of the fine line defect detection device will be described with reference to FIG.
[0016]
The object 20 to be inspected is imaged by the image input device 21, the captured image is subjected to image processing by the image processing device 22, and the fine linear defect extracted by the determination device 23 is determined.
[0017]
The outline of the image processing and determination will be described with reference to FIG. 1. After the processing is started 25 and the image input 26 is made, the image normalization processing 27 is performed to convert the average value and variance value of the density histogram into constant values. After that, the line candidate area extraction 28 is performed by the multi-stage slice method, and the extracted line candidates are determined to be continuous lines by the area combination 29 by the global search and the area combination by the local search. After performing the evaluation 31 on both sides and performing the extraction 31 of the fine linear defect region, the process is ended 32.
[0018]
Hereinafter, these processes will be described in order.
[0019]
(1) Image normalization processing Since the input image varies in density depending on the image pickup situation, density normalization processing is performed.
[0020]
The normalization process refers to a process of converting an input image having a density average value m and a standard deviation σ into an image having a density average m _N and a standard deviation σ _N according to the equation (1). In general, normalization processing is effective when standardizing images of the same type captured by sensors having different sensitivities.
[0021]
I _N (x, y) = (σ _N / σ) (I (x, y) −m) + m _N (1)
However,
I (x, y): Gray value at the point (x, y) of the input original image I _N (x, y): Gray value at the point (x, y) of the normalized image.
[0022]
FIG. 3A shows a density histogram of an image that is dark as a whole and has low contrast. By normalizing such an image, the brightness is balanced as a whole as shown in FIG. 3B, and the image is converted into an image with good contrast. That is, even if an image that is too dark, too bright, or has a low contrast is input depending on the illumination conditions or the input device setting conditions, the density range of the defect area is made substantially the same range by normalization. Converted. As a result, parameters in subsequent processing can be substantially fixed, and automation can be achieved.
[0023]
(2) Extraction of Linear Candidate Region by Multiple Multistep Slice Method The multiple multistep slice method extracts hair cracks that repeat local shading changes as shown in FIG. 4A and includes noise. This is how to avoid it.
[0024]
As shown in FIG. 4A, it is assumed that a linear region 41 and noise 42 are present in a linear density distribution 40 in an image. In this case, the concentration of the extracted linear region 41 is distributed in the concentration range T _{0 to} T ₄ , the concentration of the linear region 41 is distributed in T _{1 to} T ₄ , and T ₄ or more is the target region. The surface.
[0025]
First, the shading unevenness at the time of image shooting is removed from the normalized image by shading morphology processing. That is, by performing density morphology processing on the normalized image with a predetermined filter size and obtaining the difference between the density values of the image and the normalized image, an image from which density unevenness has been removed can be obtained. By binarizing this image with a very low density threshold value T ₁ , as shown in (1) of FIG. A part determined to be present is extracted. As shown in FIG. 5A, the extracted image 50 shows a region of the linear defect 51a that does not include noise.
[0026]
Next, as shown by (2) in FIG. 4B, the portions 41b and 42b of the density threshold value ranges T _{1 to} T ₂ that are slightly higher than the threshold value T ₁ are extracted by binarization. It is checked whether or not it is close to the portion 41a extracted in 1 ▼, and if it is close, it is combined as shown in (3) in FIG. In this case, the noise portion 42b is not connected because it is not close to the linear defect portion 41a. Similarly, at this stage, the central portion 41b0 of (2) in FIG. 4B is not connected.
[0027]
In the same manner, a portion of {T ₁ + (i−1) α <T <T ₁ + iα i = 1,... N} is taken out and it is checked whether or not it is close to the portion extracted so far. proximity if you are in extended registered as a line defect, ▲ 3 ▼ the connecting portion 41c of the, by connecting the portion 41d of the ▲ 4 ▼ of T ₂ ~T ₃ ▲ 5 ▼ a connection portion 41e of further to this connection to the portion 41e ▲ 6 ▼ by connecting portion 41f of the T ₃ through T ₄ of ▲ 7 ▼ connecting portion 41g of the.
[0028]
This operation is repeated N times (three times in the example of FIG. 4), and after the connection to the concentration range T _{1 to} T ₄ is completed, the above is repeated N times from the low threshold value T _1. The part which totaled is made into a linear area | region. Thus, by repeating the same processing loop N times (three times in the figure), the portion 41d0 of (4) is continuous with the connecting portion 41g of (7), and therefore becomes the connecting portion 41h of (8). Further, since the central portion 41b0 of (2) is continuous with the connecting portion 41h of (8), all the linear regions are extracted as the connecting portion 41i of (9), and noises 42b and 42d. , 42f are not extracted because they are not continuous with the portion 41a of (1).
[0029]
As described above, the connection region that is not included in the initial region of the linear defect 51a shown in FIG. 5A is the growth of the first loop, the linear defect 51g as shown in FIG. , 3 loops, it is possible to extract the linear defects 51h and 51i that have grown sequentially, as shown in FIGS. 5 (c) and 5 (d).
[0030]
By extracting the line candidates by the multiple multi-stage slice method in this way, it becomes possible to suppress the noise and extract the line defects with high accuracy than the two-stage threshold method.
[0031]
Next, the present invention will be further described.
[0032]
In the connection processing by the multiple multi-step slicing method, a noise component far away from the linear region is basically not detected, but the noises 53 and 54 close to the linear defect 52 are shown in FIG. c), as shown in FIG. 5 (d) is detected.
[0033]
Therefore, noises close to the linear defect are removed by the following process.
[0034]
(3) Even if the combined multiple multi-step slicing method of defect areas by local and global search is used, only noises close to the linear defect area cannot be removed. Such noise cannot be removed even if a simple image processing technique is combined because the linear defect region to be extracted has the same density range, region shape, region area, and the like.
[0035]
Therefore, the linear regions obtained in a jumping manner are described in the straight line segments, and the straight line segments are combined as the same defect portion based on geometric continuity and fit. As a result, one continuous linear defect can be extracted and noise close to the linear defect can be distinguished. Two approaches are used to combine straight line segments: local search and global search.
[0036]
a. Combining defect areas by local search As shown in FIG. 6, this processing is close to the segment group S described in a straight line, and the continuity of the direction is maintained. Things are integrated. That is, first, a relatively long straight line segment S0 is named a base segment, and a ribbon-like region 80 is generated by thickening the base segment S0 by a width W. Next, the segment S1 having the end point within the distance r from the end point of the base segment S0 and having the same direction among the straight segments included in the ribbon region 80 is integrated, and this operation is repeated until the growth stops. . At this time, the segment s indicated by a dotted line not included in the ribbon region 80 is not included in the connection candidate. Thus, by sequentially searching for the continuous segment S in the ribbon-shaped region 80, a noise component having a direction different from that of the segment S can be removed.
[0037]
b. Combining defect areas by global search In the local search shown in FIG. 6, defect candidates that are far apart cannot be connected to each other because adjacent defect candidates are combined while looking around. Therefore, it is determined by hypothesis verification whether or not two or more straight line segments form a straight line or a quadratic curve, and processing for integration is performed even if they are far apart. As a result, it is possible to detect a group of straight lines composed of straight lines or quadratic curves from a macro perspective.
[0038]
FIG. 7 (a) determines, for each base segment S0, whether or not a straight line l can be formed with a pair with another candidate segment S'0, and if possible, a straight line is generated by the least square method. To do.
[0039]
Next, as shown in FIG. 7B, all candidate segments S existing on the straight line l are detected. If the length of all straight lines formed by the segment S group is equal to or greater than a certain threshold value and the ratio of the segment S to the total length is high, a linear defect 71 as shown in FIG. Is determined. At that time, if the segment s is too far apart, as in the segment s indicated by the dotted line, the segmentation is performed and the hypothesis is verified for each straight line group.
[0040]
FIG. 8 shows a schematic diagram for extracting linear defects by local and global search.
[0041]
First, FIG. 8A shows an image 80a in which linear defect candidates extracted by the above-described multi-stage slice method are described in a straight line segment 81. FIG. As shown in FIG. 8 (b), a relatively long base segment S0 is extracted from the straight line segment 81. In FIG. 8 (c), a local search method is used, and the proximity and directivity are maintained. A segment group that sags is extracted to form a linear defect 81c. However, segments that are far apart like 81d cannot be extracted. In FIG. 8 (d), the linear defect 81d is extracted by using a global search method with segments that are far apart as straight lines. By synthesizing the two images 80c and 80d, an image 80e of only the linear defect 81e can be obtained as shown in FIG.
[0042]
【The invention's effect】
In short, according to the present invention, good linear defect candidates are detected and distinguished from noise only from image gray level information, and the obtained linear defect candidates are combined with defect portions by local / global search. Thus, a linear defect can be detected more satisfactorily.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating a processing flow for detecting fine linear defects according to the present invention.
FIG. 2 is a diagram showing an apparatus configuration for detecting fine linear defects according to the present invention.
FIG. 3 is a diagram illustrating normalization processing in the present invention.
FIG. 4 is an explanatory diagram for extracting a linear defect by a multiple multi-stage slice method in the present invention.
5 is a diagram showing an image of a process in which linear regions are connected and grown by the multiple multi-stage slice method of FIG. 4;
FIG. 6 is a diagram illustrating region combination by local search in the present invention.
FIG. 7 is a diagram illustrating region combination by global search in the present invention.
FIG. 8 is a diagram showing a schematic diagram until a linear defect is extracted from extracted linear defect candidates by a multiple multi-stage slicing method in the present invention.
FIG. 9 is a diagram for explaining extraction of a linear defect by a conventional two-step threshold method.
FIG. 10 is a diagram for explaining extraction of linear defects by the conventional two-step threshold method.
[Explanation of symbols]
41a, b, d, f linear defect 42 Noise T ₁ through T ₄ multi-step concentration range 41i linear defect candidate

Claims (3)

In a method for detecting a fine line defect from an image obtained by imaging an object to be detected, after normalizing the image, shading unevenness is removed from the normalized image, and noise is removed from the image from which shading unevenness has been removed. The region binarized with a low concentration threshold that can be removed is used as the initial nucleus, and the newly binarized region with a concentration threshold range slightly higher than this region is close to the region obtained so far. If this is the case, the method of connecting to the obtained region is repeatedly detected in multiple steps from the low density threshold range to the high density threshold range to detect as a linear defect candidate. A method for detecting fine line defects characterized by the above.   After extracting the linear defect candidates, each linear defect candidate is described in a straight line segment, and the segments in which the continuity of the direction of the straight line segments is locally maintained are integrated, and two or more segments are roughly divided. If a linear curve or a quadratic curve is determined and a quadratic curve is formed, the segments as constituent elements are combined, and the combination of local defect areas and global defect areas is combined. The method for detecting fine linear defects according to claim 1, wherein fine linear defects are extracted. In an apparatus for detecting a fine line defect from an image obtained by imaging a detection object, an image input apparatus for imaging the detection object, and after normalizing the image, removing shading unevenness from the normalized image, For an image from which shading unevenness has been removed, an area obtained by binarizing with a low density threshold that can remove noise is used as an initial nucleus, and an area newly binarized with a density threshold range slightly higher than this area. However, if it is close to the area obtained so far, the method of connecting to the obtained area is repeatedly performed in multiple steps from the low density threshold range to the high density threshold range. by an image processing apparatus for detecting a linear defect candidate, along with the respective linear defect candidates described straight segment, the direction of the continuity of these straight line segments is to integrate a segment which is locally maintained It is determined whether two or more segments generally form a straight line or a quadratic curve, and in the case of forming a quadratic curve, the constituent segments are combined, and local defect regions are combined. A fine linear defect device comprising: a determination device that extracts a fine linear defect from a combination of global defect regions.

Images