JPH05264240A - Visual inspection device - Google Patents

Abstract

PURPOSE:To enhance the performance of a visual inspection device in detecting the defects of printed wiring patterns by detecting a category proposed as being deficient, and selecting an upper category code A corresponding to the relation between the proposed category and a specific category. CONSTITUTION:A radial measuring circuit 3 measures the length of a patterned portion according to a picture element signal of a printed wiring board 100. In accordance with a coded 4 measured value and a detected 5 center picture element a category changing circuit 6 selects and outputs a corresponding category code from stored categories showing the kinds of pattern configurations. The selected category code is compared with category codes proposed as being deficient, and a category proposed as being deficient is detected 7 and a distribution of the category codes is stored 81 and output. In accordance with the relation between the category proposed as being deficient and a specific target category in the distribution of the category codes around the proposed category, a CPU 80 selects a corresponding upper category code from upper category codes stored on registration rule memory 83. Whether or not the pattern of the wiring board 100 is good is determined 200 according to the selected code.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a visual inspection apparatus for visual inspection of a two-dimensional pattern such as a printed wiring pattern.

In recent years, printed wiring boards have been increasingly densified, and now an automatic visual inspection apparatus is indispensable for manufacturing. The technique for automatically inspecting the wiring pattern of the printed wiring board for defects is roughly divided into the following two techniques.

A detection technique for optically reading a wiring pattern and converting it into an electric signal. Generally, a binary image is obtained here. Inspection logic to detect defects from the obtained image.

In particular, there is a demand for improving the defect detection performance in the inspection logic.

[0005]

2. Description of the Related Art Conventionally, as this type of appearance inspection apparatus, there is one disclosed in Japanese Patent Application No. 61-107407. This conventional visual inspection device measures the pattern lengths in a plurality of directions with a radial length measuring sensor in the pattern center of the printed wiring board to be inspected, and creates a radial code by coding the measured values, This radial code and the dictionary data of the dictionary in which the quality of the radial code is stored in advance are compared to determine the quality of the pattern. The dictionary data stored in the dictionary can be created as a plurality of data corresponding to various patterns, and by storing the plurality of data by exchanging the plurality of data in a dictionary, it has versatility that can be adapted to various patterns. Becomes

This appearance inspection apparatus judges the quality of the pattern from one radial code. That is, in the inspection using the radial code in this appearance inspection apparatus, the center of the pattern is detected, and if it is the center, the length measurement value is coded (radial code), and the quality of the code is determined by referring to the dictionary.

The encoding is performed as follows.
After detecting the center of the pattern, encoding processing is performed in four directions of vertical, horizontal and diagonal. The code includes a length code (2 bits), a balance code (1 bit), and a direction code (1 bit) for each of the four sets of directions. 180 ° different directions (eg 0 ° and 1
The sum (diameter) of the measured values of the radius in the 80 ° direction is calculated.
This value is divided into four values. That is, when it is smaller than the threshold value (1), it is set to "S" (short). "C" when the threshold value (1) or more and the threshold value (2) or less
(Collect). When it is larger than the threshold value (2), it is set to "L" (long). Furthermore, when the length measurement value on one side is equal to or more than the “0V threshold value”, it is set to “0V” (overflow). When one or more of the two directions is "0V", the diameter is also "0V". A length code (2 bits) is created by the above processing.

Next, the difference between the above two length measurement values is obtained.
When the value is equal to or less than the "balance margin" (that is, it is almost in the center and balanced), the balance code is "1", and when not, it is "0". If either one is 0V, both are 0V, and one is 0V, it is unbalanced.

Further, when the balance is "0", which is longer is indicated by a direction code (1 bit).
It is "1" when the top is long and "0" when the bottom is long.
When the balance is "0", the direction code is also "0".

By the above processing, a 4-bit code is obtained for one set of directions. This is obtained in four sets of directions and is used as a 16-bit code. This is the radial code. Further, as another appearance inspection apparatus that is an improvement of the conventional appearance inspection apparatus, the quality is determined not by one radial code but by the distribution of the radial code. However, since there are many combinations in the radial code distribution, before that,
Convert a radial code into a category that represents a shape. One way to convert to categories is to extend the dictionary to one that changes into categories. Another method is to use a neural network.

[0011]

Since the above-described conventional visual inspection apparatus is constructed as described above, it has a problem of being excessively detected. That is, when there is a small copper residue in the vicinity of the wiring pattern, it is determined that the gap is insufficient, and as a result, this wiring pattern is determined to be a defect.
Such small copper residue should be judged as a good product by a human visual inspection unless the copper residue is small and it is not close to other conductors. As described above, the conventional device has a problem in that it is impossible to make a high-level judgment because other peripheral information cannot be used because the quality is judged from one radial code.

Further, since the other conventional appearance inspection apparatus uses the dictionary memory to judge the quality of the category distribution, the dictionary memory capacity must be enormous as the number of categories increases. Had a problem. For example, a category is represented by 5 bits and the number is 4
If the number of categories is 20, the memory (1 M words) having an address of 20 bits can be used, but if the number of categories increases to 9, the dictionary memory capacity needs to have 45 bits of addresses. It may be possible. Further, although the category distribution can be processed only with about 3 × 3 pixels, in reality, category information with a target range of about 30 × 30 pixels is necessary for correct recognition. There was a problem that correct defect detection could not be performed. Furthermore, it is difficult to create dictionary data to be stored in the dictionary memory.

Furthermore, in the case of detecting a defect using a neural network, the input data is 3
Although it is possible to perform arithmetic processing in the case of a category or the like targeting a region of × 3 pixels, there is a problem that accurate defect determination cannot be performed in such a narrow region. In particular, when the input data is classified into a category based on a relatively wide area in order to accurately determine the result, the neural network has a problem that arithmetic processing cannot be performed.

The present invention has been made to solve the above problems, and proposes a visual inspection apparatus capable of improving the defect detection performance of a printed wiring pattern and correctly determining the type of the detected defect. With the goal.

[0015]

FIG. 1 shows the principle of the present invention. In the figure, the appearance inspection apparatus according to the present invention is
Image capturing means 1 for capturing the pattern formed on the sample 100, and image binarizing means 2 for binarizing the captured image.
A radial length measuring means 3 for measuring a pattern portion length of a pattern portion in each pixel row extending in a plurality of directions from an arbitrary central pixel on the binarized image; A center detecting means 4 for detecting a substantially center of a pattern, an encoding means 5 for converting each measured value in the plurality of directions into a radial code, and a category code indicating a type of a pattern shape corresponding to various radial codes are stored in advance. And a category conversion means 6 for selecting a stored category based on the radial code converted by the encoding means 5 and converting it into a category code, and an appearance for detecting a pattern defect based on the category code. The detection device stores a previously defined defect candidate among the category codes indicating the type of the stored pattern shape, and stores the category change. The defect candidate detecting means 7 for detecting the defect candidate category by comparing the category code output from the means 6 with the category code of the defect candidate, and the relationship between the defect candidate category and the target category are classified and classified into the classification. Stores various upper category codes indicating the types of corresponding pattern shapes, and relates to the relationship between the detected defect candidate category and a specific category that is a target in the category distribution in a predetermined area centered on the defect candidate category. Based on the upper category code, the upper category conversion means 8 for selecting and outputting the upper category code corresponding to the relationship is detected, and the pattern defect is detected based on the upper category code.

FIG. 2 illustrates another principle of the present invention. In the figure, another appearance inspection apparatus according to the present invention is a sample 10
0, an image pickup unit 1 for picking up a pattern formed on 0, an image binarization unit 2 for binarizing the picked up image, and a plurality of directions extending from arbitrary central pixels on the binarized image. Radial length measuring means 3 for measuring the pattern portion length of the pattern portion in each pixel row, center detecting means 4 for detecting the approximate center of the pattern from the length measuring values in the plurality of directions, and length measuring in each of the plurality of directions. Encoding means 5 for converting a value into a radial code, and category codes indicating types of pattern shapes corresponding to various radial codes are stored in advance, and categories are stored based on the radial codes converted by the encoding means 5. In the appearance detecting apparatus for detecting a pattern defect based on the category code. The radial category detection means 9 for detecting a plurality of categories radially distributed from the center pixel around an arbitrary center pixel in the plurality of categories sequentially output from the means 6, and the detected plurality of categories are related to each other. The neural network 10 outputs a higher category by additionally processing it as data of a hierarchical network structure, and detects a pattern defect based on the higher category.

[0017]

In the present invention, the defect candidate detecting means detects the defect candidate category based on the distribution of the category of the pattern of the sample sequentially outputted from the category converting means, and the defect candidate category is centered on the defect candidate category. The upper category is selected from the relationship between the target specific category and the defect candidate category, and the defect of the sample pattern is detected based on the upper category, so it is judged whether the pattern is good or bad by using the code in a wide area of the sample pattern. Therefore, it is possible to reduce the detection error due to excessive detection as much as possible, and improve the defect detection performance. Also,
The detected defect type can be more accurately discriminated.

According to another aspect of the present invention, a plurality of categories are radially detected by the radial category detection means based on the distribution of the category of the sample pattern sequentially output from the category conversion means, with the central pixel of this distribution as the center, The detected multiple categories are related to each other and processed as data of a hierarchical network structure by a neural network, the upper category is calculated and output, and the pattern defect of the sample is detected based on the upper category. The code can be used to make a pass / fail judgment, and the defect detection performance is improved. In particular, since the inspection can be performed by learning the neural network, it becomes easy to create a dictionary, which has been difficult in the past. Further, the defect type can be accurately discriminated.

[0019]

【Example】

a) One Embodiment of the Present Invention One embodiment of the present invention will be described below with reference to FIGS. FIG. 3 is an overall block diagram of the apparatus of this embodiment, and FIGS. 4 to 8 are detailed explanatory diagrams of circuits constituting each block in FIG.

In the appearance inspection apparatus according to the present embodiment shown in FIG. 3, the sample printed wiring board 100 placed on the scanning XY stage 101 is captured by the image pickup section 1 as a pixel signal of image data. The pixel signal is binarized by the binarizing circuit 2 and stored in the image memory 21, the noise component is removed by the preprocessing circuit 22, the pattern length of the printed wiring board 100 is measured by the radial length measuring circuit 3, and the code The digitizing circuit 4 encodes the length measurement values of the pattern portion length in a plurality of directions, and the center detection circuit 5 detects the center pixel corresponding to the approximate center of the printed wiring board 100 from the length measurement values in the plurality of directions and encodes the center pixel. The category conversion circuit 6 selects the corresponding category code from a plurality of categories indicating the type of the pattern shape stored in advance based on the length measurement value and the center pixel and the defect is caused. And outputs to the auxiliary detection circuit 7 and the category distribution memory 81.

The defect candidate detection circuit 7 stores a previously defined defect candidate among the stored category codes indicating the type of pattern shape, and the category code output from the category conversion circuit 6 is the category code of the defect candidate. The defect candidate category is detected by comparing with, and this defect candidate category is output to the CPU 80. The category distribution memory 81 stores the distribution of category codes output from the category conversion circuit 6 on condition that a defect candidate category is detected by the defect candidate detection circuit 7, and stores the stored category code distribution. Output to the CPU 80.

The CPU 80 mainly includes a category search circuit 82, an upper category output circuit 84, and a registration rule reference circuit 85. Based on the defect candidate category and the distribution of the categories, various upper classes stored in advance in the registration rule memory 83. The corresponding upper category code is selected from the category codes and output to the defect determination circuit 200.

That is, the category search circuit 82 searches for a specific target category from the distribution of category codes stored in the category distribution memory 81, and the relationship between the searched category and the detected defect candidate category. The upper category code stored in the registration rule memory 83 corresponding to is read out under the control of the registration rule reference circuit 85 and output by the upper category output circuit 84. The defect determination circuit 200 determines the quality of the pattern on the printed wiring board 100 based on the upper category code.

Next, the defect detecting operation of this embodiment based on the above-mentioned structure will be described with reference to FIGS. 9 to 11.
First, the sample printed wiring board 1 is mounted on the XY stage 101.
00 is placed, the line illuminating device 14 irradiates the surface of the printed wiring board 100 with a linear light flux from obliquely above, and the linear light flux reflected on the printed wiring board 100 forms an imaging lens. Line sensor 16 through 12
The image is formed and received on the light receiving surface of. The line sensor 16 is electrically scanned by a driver (not shown), and each pixel signal is sequentially output to the binarization circuit 2. The pixel signal binarized by the binarization circuit 2 is the XY stage 101.
And a preprocessing circuit for a binary image of the wiring pattern (hereinafter, pattern binary image) stored in the image memory 21 at an address based on the operation position of the line sensor 13 and based on the stored binary image signal. 22 performs processing for removing noise components, and various processing such as edge smoothing. The pattern binary image thus variously processed by the preprocessing circuit 22 is sequentially output line by line to the subsequent radial length measuring circuit 3.

The radial length measuring circuit 3 is shown in FIG.
As shown in (A), it has one shift register constituted by connecting flip-flops for seven lines in cascade. The pattern binary image is scanned on the radial length measuring circuit 3. The radial length measuring circuit 3 is provided with a radial length measuring sensor area 3a of 7 bits × 7 bits.
As shown in FIG. 4B, the radial measurement sensor area 3a
Is provided with a radial length measuring sensor 3b having pixels radially extending in eight directions from the bit at the center thereof.
The radial length measurement sensor 3b is for detecting the number of images having "1" in each of the eight directions, that is, the length of the wiring pattern portion. In FIG. 4B, numbers 0 to 3 indicating the position from the center are entered along each direction.

FIGS. 4A and 4B show a simplified structure for the sake of explanation. Actually, for example, as shown in FIG. 5, the radial length measuring sensor 3c is constructed in an area of 65 bits × 65 bits. , Has length measuring arms A1 to A16 extending in the direction of the line sensor 16 from the center, and each length measuring arm Ai (i =
1 to 16) have a length of 33 bits (including the central bit). With this radial length measurement sensor 3c, wiring pattern portion length measurement values r1 to r16 in each direction from the central pixel of the radial length measurement sensor 3c can be obtained.

The measured values r1 to r16 are supplied to the center detection circuit 5 and the coding circuit 4 shown in FIG. The center detection circuit 5 determines the pattern binary image 1 from the measured values r1 to r16.
The approximate center of 00a is detected. In addition, the coding circuit 4
The shape of the pattern binary image 100a at this length measurement position is coded from the length measurement value ri (i is an odd number of 15 or less).

As shown in FIG. 6, the center detection circuit 5 includes balance determination circuits 51a to 51h having the same structure.
For example, the balance determination circuit 51a uses the length measurement value r1 of the length measurement arm A1 and the length measurement arm A in the direction opposite to the length measurement arm A1.
The difference from the length measurement value r9 of 9 is calculated by the difference calculator 51a ₁ , and the result is absoluteized and supplied to the comparator 51a ₂ for comparison with a predetermined threshold value of the balance margin BM. This balance margin BM is, for example, 2 and | r1-r9 |
When <2, the output of the balance determination circuit 51a becomes "1". The balance determination circuits 51b to 51h are similar to the balance determination circuit 51a. The balance condition determination circuit 52 is similar to each of the balance determination circuits 51a to 51a. The balance condition determination circuit 52 is for each balance determination circuit 5.
The number which is "1" among the outputs 1a to 51h is obtained, this value is compared with a preset reference value, and if this value is equal to or larger than the preset reference value, the center detection signal is output. .. Further, the length measuring sensor 3c of the radial length measuring circuit 3
From the measurement values r1 and r9 are input to the unidirectional coding circuit 41a of the coding circuit 4, and the difference calculator 41a _{1 having} the same circuit configuration as the balance determination circuit 51a of the center detection circuit 5 and the comparison The device 41a ₂ determines whether or not the balance is achieved. The difference calculator 41a ₁ and the comparator 41a ₂ are the balance determination circuit 51a of the center detection circuit 5.
It may be configured to be commonly used with. When the circuit determines that the balance is not achieved, the direction bit indicating which is longer is output, and the 2-bit code for the balance and the direction is output.

The outputs r1 and r9 of the length measurement values are also input to the length coding circuit 42a of the coding circuit 4 at the same time, and the adder 42a ₁ of the length coding circuit 42a calculates the sum of both outputs. , The sum (r1 + r9) of both outputs is compared with the coding thresholds V _th1 and V _th2 by the comparators 42a ₂ and 42a ₃ to output a comparison value indicating a magnitude relationship. Further, the outputs r1 and r9 are compared with the 0V threshold value by the comparators 43a and 44a, and the comparison value indicating whether or not the voltage is 0V is output.

The data conversion circuit 45a encodes a 4-bit data value as a unidirectional code based on the 2-bit code output and the comparison value and outputs the data value to the gate circuit 40. Similarly to the outputs r1 and r9, the unidirectional coding circuits 41b to 41h also output the unidirectional codes to the gate circuit 40 for the respective length measurement values of the outputs r2 and r10, ..., R8 and r16.

The gate circuit 40 opens its gate on condition that the center detection signal output from the center detection circuit 5 is input, and outputs a 16-bit radial code to the category conversion circuit 6 of the subsequent circuit. As shown in FIG. 8, the category conversion circuit 6 is formed by a dictionary of ROM-structured categories, converts the categories into a category code stored at an address specified by a radial code, and stores them in the defect candidate detection circuit 7 and the category distribution memory 81. Output each. The radial code and the category code are associated with each other in advance and stored in the ROM of the category conversion circuit 6 as a category dictionary. As shown in FIGS. 9A to 9J, the contents of the category corresponding to this category code are a straight lead, a bent lead, a land, a land with a lead, a thin lead, a wire break, a protrusion, a T branch, a land protrusion, and copper. The rest, etc.
It is defined to include both normal and defective parts. Also,
This category has both a conductor portion and an air space portion, and generally includes several dozen types including the example shown in FIG. further,
In the case of a category having a directivity such as a straight lead in the example of FIG. 9, it is possible to classify each direction by using this directivity as a parameter.

As shown in FIG. 8, the defect candidate detection circuit 7 is composed of a defect candidate reading section 71 and a dictionary 72, and the defect candidates stored in the dictionary 72 corresponding to the category code input from the category conversion circuit 6. The defect candidate reading unit 71
Read out and compare the read defect candidate with the category code, and output it to the CPU 80 as a result candidate category. Further, the category code from the category conversion circuit 6 is also input to the category distribution memory 82, and the category distribution memory 82 receives the detection signal indicating that the defect candidate category is detected by the defect candidate detection circuit 7. The category code is stored as a condition code distribution as a condition. The reason why the category code is stored under the condition that the detection signal is input is to minimize the storage capacity of the category distribution memory 82.

When the defect candidate categories and the category code distributions are respectively input to the CPU 80, the category search circuit 82 predetermines a particular category of interest in the category code distribution centered on the defect candidate categories. Search based on the rules. The upper category output circuit calculates the relationship between the searched specific category and the defect candidate category, and based on this relationship, the corresponding upper category is read from the plurality of upper categories stored in the registration rule memory 83. The upper category output circuit 85 outputs the defect determination circuit 200.

The upper categories corresponding to the relationship between the specific category (denoted by symbol B) and the defect candidate category (denoted by symbol A) are shown in FIGS. 10A to 10G as an example. In FIG. 7A, defect candidate category A
If the target specific category B exists within a certain distance from, it is set as the higher category N ₁ (rule). In the same figure (B), when the target specific category B exists within a certain distance where the pattern binarized image 100a is continuous from the defect candidate category A, the higher category N ₂ is set (rule). .. In the same figure (C), when the target specific category B exists in a specific direction (within the range of the elevation angle) and within a certain distance with respect to the defect candidate category A, the higher category N ₃ (Rule) In FIG. 5D, if the target category B is within a certain distance from the defect candidate category A, if the target category B is in the specific direction, it is set as the upper category N ₄ (rule). In FIG. 6E, when the target specific category B and the other specific category C exist at a distance within a certain range of n pixels from the defect candidate category A, the higher category N ₅ is set (rule). In FIG. 6F, when the target category B exists within a certain distance from the defect candidate category A, it is regarded as the upper category N ₆ . However, when the defect candidate category A exists further in the direction of the category B with respect to the defect candidate category A, the higher category N ₇
(Rule) Further, in FIG. 6G, when there are a predetermined number or more of defect candidate categories A within a certain distance, the defect candidate category A is set as a higher category N ₈ (rule).

The rules defined and described in FIGS. 10A to 10G are based on the rules shown in Table 1 below.
It is configured by adding each element as shown by. Further, in addition to the configuration shown in Table 1, each element may be combined to create more rules. For example, N _m is set when a target specific category exists in a specific direction and within a continuous constant distance with respect to the defect candidate category by combining rules.

[0036]

[Table 1] In Table 1, the continuity among the respective elements indicates that the pattern binary image 100a continuously extends, the reference direction is the direction of the target specific category with respect to the defect candidate category, and the code direction is the reference code. Specific direction of multiple, multiple category references have multiple target specific categories, extension has more defect candidate categories before the specific target category, multiple have multiple defect candidate categories within a certain range Indicates each element that does.

The category search circuit 82 is shown in FIG.
Within a certain distance, for example, the center, by radial scanning in 16 directions (FIG. 7A), square area scanning (FIG. 3B), circular area scanning (FIG. 3C), etc. From 20
It is searched whether or not there is a specific category of interest within a range of about pixels. The method of each search is shown in FIG.
There are merits and demerits as shown in, and it is necessary to use them properly according to each condition.

[0038]

[Table 2] That is, the radial scan is the fastest in terms of search speed, but has the drawback that omissions occur in the search results. The square area assessment is easy to process without omission of search, but has drawbacks such as slow search speed and inaccurate distance. The circular area scanning has no omission in the search and the distance is accurate, but has a drawback that the search speed is slow and the processing is difficult.

Further, as shown in FIG. 12 in order to eliminate the defect of the rectangular area scanning, a rectangular area (shown by a chain line) is formed on the line in the direction determined by the kind of the defect candidate category A at the center. It can also be configured to scan and search.

B) Another Embodiment of the Present Invention Another embodiment of the present invention will be described with reference to FIG. In FIG. 1, the appearance inspection apparatus according to the present embodiment is the same as that shown in FIG.
2 has a circuit configuration from the image pickup device 1 to the defect candidate detection circuit 7 and the category distribution memory 91 similarly to the appearance inspection apparatus shown in FIG. 2, and in addition to this configuration, the defects detected by the defect candidate detection circuit 7 A plurality of categories stored in the candidate category and category distribution memory 91 are respectively input,
A category search unit 92 that radially searches for a specific target category centered on the defect candidate category from among the plurality of input categories, and rotates / reverses the front and back of the category distribution that includes the plurality of retrieved categories. A normalization circuit 11 for normalizing the input data and a plurality of the normalized categories are input, the plurality of categories are associated with each other, processed as a hierarchical network structure, and an upper category is calculated and output. Neural network 1
0, and a defect determination circuit 200 for determining the quality of the wiring pattern on the printed wiring board 100 of the sample based on the higher category.

Next, the operation of this embodiment based on the above configuration will be described with reference to FIG. The image pickup device 1 outputs the pattern image of the printed wiring board 100 as pixel signals,
Categories are sequentially output through the binarization circuit 2, the image memory 21, the preprocessing circuit 22, the radial length measurement circuit 3, the coding circuit 4, the center detection circuit 5, and the category conversion circuit 6, and defects are generated based on this category. While the candidate detection circuit 7 detects a defect candidate category and outputs it to the category search unit 92,
The categories are sequentially stored in the category distribution memory.

The category searching unit 92 searches for a category on a radial line in eight directions as shown in FIG. 14A from a plurality of categories stored in the category distribution memory centering on the detected defect candidate category. To do. This category search is performed when a specific category that is a target different from the central defect candidate category is detected from the category distribution. Further, when a pixel whose category is unknown is detected, the category detected first is used thereafter. In this way, the categories of to in FIG. 14 (A) can be detected. This central defect candidate category is shown as, and a total of 9 categories can be searched.

The search for the category is shown in FIG.
As shown in, it is possible to adopt a configuration in which the defect candidate category is divided into eight areas radially and the search is sequentially performed from the area closer to the defect candidate in each area.

Each category searched in this way is detected, and each category is a rotation of the distribution. The normalization circuit 11 arranges and adjusts the data processing direction for front-back inversion and outputs the normalized category to the neural network 10.
The neural network 10 processes each category that is sequentially input as a hierarchical network structure to calculate a higher category and outputs it to the defect determination circuit 200. The operation of this upper category is processed in a hierarchical network structure of a three-layer back-propagation type. In particular, it is desirable to use a dedicated LSI or DSP for the neural network in order to execute high speed processing. If the upper category that is output by calculation includes, for example, “poor spacing” and “fine copper residue”,
It can also be defined as "proximity minute copper residue". It is possible for the user to determine whether each of these upper categories is normal or defective, and only the defect is output together with the type code.

The result judging circuit 200 stores each upper category in which a plurality of types are classified in advance and defects are defined,
The input upper category is compared with each of the stored upper categories, and the quality of the wiring pattern is determined based on the comparison result.

It should be noted that it is also possible to configure as another embodiment as shown in FIGS. 15 (A) and 15 (B). FIG. 15A shows a configuration in which a category is searched from a plurality of category distributions centering on each of defect candidate categories ₁ to _9, and a neural network calculates and outputs a higher category based on each of the searched categories. You can also

Further, in FIG. 15B, a plurality of neural networks 10 ₁ to 10 _n are provided in parallel, and a higher category is calculated based on a plurality of categories to which the respective neural networks 10 ₁ to 10 _n are input. Then, the upper neural network may be input to the upper neural network, and the upper neural network may calculate the highest category. Based on this highest category, the defect judgment circuit 200 executes the defect selection judgment and can judge the quality of the wiring pattern. Since a plurality of categories can be input to each of the plurality of neural networks 10 ₁ to 10 _n in this manner, a wide range of pattern images corresponding to the plurality of categories can be targeted, and the inspection accuracy can be further improved. Become. Further, in the above-described embodiment, in the category search in the category search unit 92, the category is radially searched from the category distribution with the defect candidate category as the central pixel, but the category is radially searched from any central pixel in the category distribution. It is also possible to adopt a configuration in which the search is executed. In this case, it is not necessary to provide the defect candidate detection circuit 7 for detecting the defect candidate category in the preceding stage. Further, the category search unit 9 is provided without providing the category distribution memory 91.
It is also possible to adopt a configuration in which the category is directly searched for with 1. Although the distribution of the categories output from the category searching unit 91 is normalized by the normalization circuit 11, the distribution of the categories may be directly input to the neural network 10 to calculate the upper category.

[0048]

As described above, according to the present invention, the defect candidate detecting means detects the defect candidate category based on the distribution of the category of the pattern of the sample sequentially output from the category converting means, and the defect candidate category is centered on the defect candidate category. The upper category is selected from the relationship between the target specific category and the defect candidate category in the category distribution, and the defect in the sample pattern is detected based on the upper category. By using this, it is possible to judge the quality of the pattern, and it is possible to reduce the detection error based on the excessive detection as much as possible and improve the defect detection performance. Further, there is an effect that the detected defect type can be more accurately discriminated.

In another aspect of the present invention, a plurality of categories are radially detected by the radial category detecting means based on the distribution of the category of the sample pattern sequentially output from the category converting means, with the central pixel of this distribution as the center. , Correlating the detected plurality of categories with each other and processing them as data of a hierarchical network structure by a neural network to arithmetically output the upper category, and detecting the pattern defect of the sample based on the upper category, Since it is possible to make a pass / fail judgment using a code in a wide area, the defect detection performance is improved. In particular, since the inspection can be performed by learning the neural network, it has an effect of facilitating the creation of a dictionary, which has been difficult in the past. In addition, the defect type can be accurately determined.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is another principle explanatory diagram of the present invention.

FIG. 3 is an overall block configuration diagram of a visual inspection apparatus according to an embodiment of the present invention.

4 is a principle configuration diagram of a radial length measuring circuit in the embodiment apparatus shown in FIG.

FIG. 5 is an explanatory view of a radial length measuring sensor which constitutes a radial length measuring circuit in the apparatus of the embodiment shown in FIG.

FIG. 6 is a block configuration diagram of a center detection circuit in the embodiment apparatus shown in FIG.

FIG. 7 is a circuit configuration diagram of an encoding circuit in the embodiment apparatus shown in FIG.

8 is a schematic configuration diagram of a category conversion circuit and a defect candidate detection circuit in the apparatus of the embodiment shown in FIG.

FIG. 9 is a diagram illustrating a specific example of various category definitions used in the apparatus of the embodiment shown in FIG.

10 is a definition explanatory diagram of a higher category corresponding to the relationship between a specific category and a defect candidate category used in the apparatus of the embodiment shown in FIG.

11 is a scanning mode diagram of a category search circuit in the apparatus of the embodiment shown in FIG.

FIG. 12 is another scanning mode diagram of the category search circuit.

FIG. 13 is a block diagram of a main part of a visual inspection apparatus according to another embodiment of the present invention.

FIG. 14 is a search explanatory diagram of a category search unit in the apparatus of the embodiment described in FIG.

FIG. 15 is a search explanatory view of a category search unit and a block configuration diagram of a neural network in another embodiment apparatus of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Imaging part (imaging means) 2 ... Binarization circuit (binarization means) 3 ... Radial length measurement circuit (radial length measurement means) 4 ... Coding circuit (coding means) 5 ... Center detection circuit (center detection) Means) 6 ... Category conversion circuit (category conversion means) 7 ... Defect candidate detection circuit (defect candidate detection means) 8 ... Upper category conversion means 9 ... Radial category detection means 10 ... Neural network 11 ... Normalization circuit (normalization means) 21 ... Image memory 22 ... Preprocessing circuit 81 ... Category distribution memory 82 ... Category search circuit 83 ... Registration rule memory 84 ... Upper category output circuit 85 ... Registration rule reference circuit 91 ... Category distribution storage unit 92 ... Category search unit 100 ... Print Wiring board 101 ... XY stage 200 ... Defect determination circuit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification number Internal reference number FI Technical display location G06F 15/70 460 Z 9071-5L H05K 3/00 Q 6921-4E

Claims (18)

[Claims] 1. An image pickup means (1) for picking up a pattern formed on a sample (100), an image binarization means (2) for binarizing the picked-up image, and an image on the binarized image. Radial measuring means (3) for measuring the pattern portion length of the pattern portion in each pixel row extending in a plurality of directions from any of the central pixels, and detecting the approximate center of the pattern from the measured values in the plurality of directions. A center detecting means (4), a coding means (5) for converting each measured value in the plurality of directions into a radial code, and a category code indicating a type of pattern shape corresponding to each radial code, A category conversion means (6) for selecting a stored category based on the radial code converted by the encoding means (5) and converting it into a category code; and a pattern based on the category code In the appearance detecting device for detecting defects, a pre-defined defect candidate among the stored category codes indicating the type of pattern shape is stored, and the category code output from the category conversion means (6) is used as the defect candidate. Defect candidate detection means (7) that detects a defect candidate category by comparing the defect candidate category with the category code of No. 1, and various types that indicate the type of pattern shape corresponding to the classification by classifying the relationship between the defect candidate category and the target category. An upper category code that stores a higher category code and corresponds to the detected defect candidate category and a specific category that is a target in a category distribution in a predetermined area centered on the defect candidate category Upper category conversion means (8) for selecting and outputting a category code, and based on the upper category code Appearance inspection apparatus and carrying out the detection of a pattern defect. 2. The appearance inspection apparatus according to claim 1, wherein the higher category conversion means (8) detects a defect of a category code distribution output from the category conversion means (6) by a defect candidate detection means (7). A category distribution storage unit (81) that stores a condition that a candidate category is detected, a category search unit (82) that searches for a specific target category in the stored distribution of category codes, and the defect. A registration rule storage unit (8) that classifies relationships between candidate categories and target categories, and stores various upper category codes that indicate types of pattern shapes corresponding to the categories.
3) and a registration rule storage unit (83) corresponding to the relationship between the searched category and the detected defect candidate category.
And an upper category output section (84) for selecting and outputting the upper category code stored in the appearance inspection apparatus. 3. The appearance inspection apparatus according to claim 1, wherein the upper category conversion means (8) determines whether a specific category of interest is present within a certain distance from a defect candidate category. An appearance inspection device characterized by determining a distribution and outputting a higher category based on this determination. 4. The appearance inspection apparatus according to claim 1, wherein the upper category conversion unit (8) is a specific category to be a target within a certain distance from a defect candidate category and a position where patterns are continuous. The appearance inspection apparatus is characterized in that the distribution of categories is judged depending on whether or not there exists, and a higher category is output based on this judgment. 5. The appearance inspection apparatus according to claim 1, wherein the upper category conversion means (8) is within a certain distance from a defect candidate category, and is there a specific category of interest in a specific direction? An appearance inspection apparatus characterized in that the distribution of categories is judged depending on whether or not the higher category is output. 6. The appearance inspection apparatus according to claim 1, wherein the upper category conversion unit (8) determines that a specific category that is a target within a certain distance from a defect candidate category and that has a direction in a specific direction is a target category. An appearance inspection apparatus characterized in that a category distribution is judged depending on whether or not it exists, and a higher category is output based on this judgment. 7. The appearance inspection apparatus according to claim 1, wherein the upper category conversion means (8) is within a certain distance from the defect candidate category, and the target specific category and other specific categories are simultaneously selected. An appearance inspection apparatus characterized in that a category distribution is judged depending on whether or not it exists, and a higher category is output based on this judgment. 8. The appearance inspection apparatus according to claim 1, wherein the upper category conversion unit (8) has a specific category within a predetermined distance from a defect candidate category and in a specific direction, An appearance inspection apparatus characterized in that the distribution of categories is judged depending on whether or not other specific categories exist in a specific direction, and the higher category is output based on this judgment. 9. The appearance inspection apparatus according to claim 1, wherein the upper category conversion means (8) is within a certain distance from the defect candidate category, and whether or not there is a specific number or more of the target specific categories. An appearance inspection apparatus characterized in that the distribution of categories is judged depending on whether or not the higher category is output. 10. The appearance inspection apparatus according to claim 1, wherein the higher category conversion means 8 searches for a specific target category based on radial measurement within a predetermined area. A visual inspection device characterized by outputting a higher category. 11. The appearance inspection apparatus according to claim 1, wherein the upper category conversion unit 8 searches for a specific target category in a predetermined rectangular area and outputs the upper category. A visual inspection device characterized in that 12. The appearance inspection apparatus according to claim 1, wherein the upper category conversion unit 8 searches for a specific category of interest in a predetermined circular area and outputs the upper category. A visual inspection device characterized in that 13. An image pickup means (1) for picking up a pattern formed on a sample (100), an image binarization means (2) for binarizing the picked-up image, and on the binarized image. Radial measuring means (3) for measuring the pattern portion length of the pattern portion in each pixel row extending in a plurality of directions from any of the central pixels, and detecting the approximate center of the pattern from the measured values in the plurality of directions. A center detecting means (4), a coding means (5) for converting each measured value in the plurality of directions into a radial code, and a category code indicating a type of pattern shape corresponding to each radial code, A category conversion means (6) for selecting a stored category based on the radial code converted by the encoding means (5) and converting the stored category into a category code; In the appearance detecting apparatus for detecting defects in the image, radial categories for detecting a plurality of categories radially distributed from the center pixel around an arbitrary center pixel in the plurality of categories sequentially output from the category conversion means (6). A detection means (9) and a neural network (1) which outputs a higher category by processing the detected plurality of categories in relation to each other and processing them as data of a hierarchical network structure.
0) and the pattern inspection is carried out based on the higher category. 14. The appearance inspection apparatus according to claim 13, wherein a defect candidate defined in advance among the stored category codes indicating the types of pattern shapes is stored and output from the category conversion means (6). Defect candidate detecting means (7) for detecting a defect candidate category by comparing the category code of the defect candidate with the category code of the defect candidate, and performing the detection of the category by the radial category detecting means (9) mainly on the defect candidate category. Appearance inspection device. 15. The appearance inspection apparatus according to claim 13 or 14, wherein the normalization means (11) normalizes the distribution of a plurality of categories output from the radial category detection means (9) with respect to rotation and front / back inversion. ), And outputs a plurality of normalized categories to the neural network. 16. The appearance inspection apparatus according to any one of claims 13 to 15, wherein the radial category detection means (9) detects each category in a radially divided division area in a category distribution. Appearance inspection device characterized by: 17. The appearance inspection apparatus according to claim 13, wherein the radial category detection means (9) uses a distribution of categories and distance information from a substantially central pixel of the distribution as a target. The visual inspection apparatus is characterized by detecting a specific category as follows. 18. The appearance inspection apparatus according to claim 13, wherein the neural network (10) has a plurality of lower neural networks connected in parallel at a first stage, and the plurality of lower neural networks are connected to each other. A visual inspection apparatus, wherein a plurality of output upper categories are input, and an upper neural network that calculates the highest category based on the plurality of upper categories is connected in stages.