KR20100121250A - Vision system for inspection of qfn semiconductor package and classifying method - Google Patents

Abstract

PURPOSE: A vision system for inspecting the external defects of a quad-flat-no-leads(QFN) semiconductor package and a method for classifying the defects are provided to efficiently manage a semiconductor manufacturing process by installing an optical system for inspecting the defects. CONSTITUTION: A charge coupled device(CCD) camera(120) is installed over the upper side of a semiconductor package and takes an image for reflected light which is reflected by the package and reflected light which is reflected by foreign materials and chip defects on the surface of the package. Illuminating units(140, 150) are located between the camera and the package. A half-mirror refracts light from the illuminating units and radiates the light to the package.

Description

Vision system for visual defect inspection of QFN semiconductor package and defect classification method using same {Vision System for inspectiON of QFN Semiconductor package and classifying method}

The present invention relates to an optical system configuration, an algorithm for classifying QFN package defects, and an optical system for extracting semiconductor package appearance defect feature information.

There are cracks, foreign materials, chip-outs, chips, voids, etc., in which the appearance defects of semiconductors are different, and the inspection system should be able to perform effective process management by performing defect classification and defect classification.

On the other hand, lead-free packages that are in the spotlight recently include Leadless Chip Carriers (LCC), Bumped Chip Carriers (BCC), and Quad Flat No Leads (QFN). These packages are attracting much attention because they are small in size, have no leads, reduce the area on the PCB by about 50%, and the production cost is similar to the existing small packages. Packages without leads have several advantages, but they are smaller in size and different from conventional semiconductor devices, and thus require inspection technology. Recently, advanced inspection targets such as an increase in automatic inspection items and a reduction in inspection time are required. Process control by furnace defect classification is becoming difficult.

1 is a cross-sectional view showing the foreign matter 21 and the chip defect 22 in the form of such a representative package defect.

Representative methods for solving such defects include object classification using image data, which is classified into binary image and light image. The classification methods in binarized images include textures, block heights and white pixel lengths, and chaincodes.

The image classification methods in the contrast image include artificial intelligence and statistical techniques, which include neural networks, fuzzy theories, and decision trees.

Among them, the neural network shows good classification performance by assigning different weights to the data without special regularity, but there is a problem that the calculation amount is increased and different weights are assigned to each variable.

Fuzzy theory has been successfully applied to pattern classification, but it has a disadvantage in that the calculation becomes very complicated when there are many classes to classify or various characteristics of objects.

The decision tree was used to classify variables that are not related to each other. Statistical techniques include minimum distance, ML, nearest neighbor, and parallel-piped. Among them, ML needs to assume that the data for the class has Gaussian distribution, and the classification performance depends on the quality of the learned class.

  2 is a configuration diagram of a conventional visual inspection optical system. As shown in the drawing, a CCD camera 12, a coaxial light 14, a side light 15, a half mirror 13, and a photographing controller 16 for inspecting the external appearance of the semiconductor package 11 to be inspected are configured. have. However, since the side lighting 15 is illuminated at a predetermined inclined angle on the package 11, there is a problem in extracting feature information about appearance defects, particularly chip defects, on the surface of the package 11.

  An object of the present invention for solving the above-described problems, the selection of feature information to have the highest classification performance using ML classification for foreign material and chip defects that are difficult to detect and classify among the appearance defects of the semiconductor package; An object of the present invention is to provide an optical system and a defect classification method for accurately acquiring feature information.

  In addition, another object of the present invention is to provide a location of appearance defects, a brightness characteristic information of a semiconductor package, a classification method using ML, and an optical system that is easy to acquire characteristic information.

The optical system configuration for solving the technical problem of the present invention, the vision system for inspecting the appearance defects of the semiconductor package, the optical system is installed on the upper side of the semiconductor package, the reflected light that is supplied from the lighting means reflected from the package; A CCD camera for photographing reflected light reflected by foreign matter and chip defects present on the package surface; Illuminating means positioned between the lower portion of the camera and the upper portion of the package to emit light; A half mirror for refracting light supplied from the lighting means at a predetermined angle to irradiate the package; And a photographing controller unit for controlling the process of providing the light source, storing and transmitting photographed image data, wherein the lighting unit comprises a side illumination for irradiating light from the side of the package to be inspected and an upper side of the side illumination. Transmitting light through the coaxial light to drop the light source at a 90 ° angle to the package, characterized in that to form an optical path of the side lighting and coaxial light in parallel.

In this case, the light emitting source constituting the lighting means is characterized in that the LED.

The defect classification method for solving the above technical problem of the present invention uses a maximum likelihood ratio (ML) method of sequentially estimating data from a training pattern using a probability density function and extracting a feature pattern of a distribution from an input pattern. To extract feature information independent of size, position, and orientation for classification, to remove specific regions from the image and to maintain the original shape of defects and removal of unnecessary regions that appear after the thresholding process. It is characterized by using a defect classification algorithm using morphology technique.

Here, the algorithm, the step of inputting an image; Binarizing the input image using a minimum error threshold and naming it image 1; Performing a closing operation with a 3 × 3 box strutching element to remove a hole having a size of one or two pixels that may occur in the image 1; Performing an opening operation with a strutching element of an 11 X 11 box to minimize noise components distributed in a size of 5 to 10 pixels in the background and naming the resultant image as image 2; applying a conditional dilation technique using the image 1 and the image 2 to minimize the defect area that would have lost its original shape due to the opening; And performing blob analysis on the divided region for extracting feature information.

As described above, the optical system according to the present invention provides a very effective tool for selecting feature information and obtaining feature information on foreign matter and chip defects, which are difficult to detect and classify among appearance defects of a semiconductor package.

   In addition, it provides effective semiconductor package process management through effective semiconductor package defect identification and defect classification by providing the location of appearance defects of semiconductor package, classification method using brightness characteristic information and ML, and optical system that is easy to acquire characteristic information. do.

Hereinafter, with reference to the accompanying drawings, diagrams, graphs, equations described in detail with respect to the optical system configuration and feature information acquisition, the defect classification method through the same according to the present invention.

    First, the foreign material described below means foreign matter, and the residue of the molding component remains in the package due to a poor cleaning after the sawing process and appears as a defect. In this case, defects with a width of at least 100 μm or more are usually examined. Chips are caused by poor water-knife used during the sawing process, and the molded part of the package is broken, and the width of the chip is at least 100μm such as foreign matter. Since these foreign materials and chip defects are the same material and the structure of the surface is different, the illumination that can maximize the difference in the structure is required for effective detection.

   To this end, the configuration of the optical system proposed by the present invention will be described with reference to FIG. 3.

   As shown in FIG. 3, the optical system 100 includes a reflection light in which light supplied from the luminaires 140 and 150 is reflected on the package 110 and installed on the semiconductor package 110. 110 is located between the CCD camera 120 for capturing the reflected light reflected by the foreign matter 21 and the chip defect 22 existing on the surface, and located between the lower portion of the camera 120 and the upper portion of the package 110 Illumination means (140, 150) for emitting a circle, the half mirror 130 and the light source for refracting the light supplied from the illumination means (140, 150) at a predetermined angle to the package 110 provided , The shooting controller 160 to control the process of storing and transmitting the captured image data.

Here, the lighting means (140, 150) is configured on the side lighting 150 and the side lighting 150 for irradiating light from the side of the package 110 which is the inspection object and transmits the light through the lens The package 110 includes a coaxial light 140 for dropping the light source at a 90 ° angle. In this case, preferably, the optical paths of the side lighting 150 and the coaxial lighting 140 are formed in parallel. That is, the axis formed by the optical path of the side lighting 150 is preferably configured to be parallel to the horizontal axis of the package 110 serving as the inspection target sample.

However, when only the side lighting 150 is used, characteristics of the uneven surface of the semiconductor package 110 also appear simultaneously. It is desirable to use coaxial illumination 140 together to reduce this noise component.

This is an illumination that can maximize the structural difference for the effective detection of the foreign matter 21 and chip defects 22 on the surface of the package 110, and it can be seen that this shows a clear effect in the optical system 100 image of FIG. have.

In this case, the light emitting source constituting the lighting means is preferably composed of LED.

Since the photographing controller 160 has the same configuration as that used in the optical system for inspecting a general package defect, a detailed description thereof will be omitted.

The image taken by the proposed optical system and the image taken by the existing optical system are shown in FIG. 4. FIG. 4B is an image taken by an existing optical system, and FIG. 4A is an image obtained by improving contrast by a histogram-stretch method. FIG. 4C is an image taken by the proposed optical system. 4B and 4C show the results of applying the region division method to FIGS. 4D and 4E. Here, it can be seen that the image captured by the conventional optical system does not detect chip defects even after improving the contrast. Detecting defects is essential to accurate classification of defects. However, it can be seen that both foreign bodies and chip defects are detected in the results of the images taken by the proposed optical system.

In addition, as another method for comparing FIG. 4B and FIG. 4C, intra-class variance and inter-class variance in each image histogram were used. Table 1 shows the values obtained by dividing the variance in the class by the variance between the classes. The smaller the value, the easier the image can be distinguished from the background.

S _W / S _B Conventional optical system 13270.12 Existing Optical System (Improved Image) 5425.24 Proposed optical system 153.91

: 클래스 내 분산,

: 클래스 간 분산*

: Distribution within a class,

: Distributed between classes

(Table 1)

As shown in Table 1, it can be seen that the image of the proposed optical system is clearly distinguished from the defect and the background, rather than the improved image of the existing optical system, so that the imaging of the defect can be easily taken.

The choice of lighting is very important in optical systems. Considerations for lighting selection include the shape of the light, the light source, and the color of the light. In the present invention, by changing the color of the side lighting 150 in the optical system 100 introduced above to check the effect of the effect on the classification of defects and extracted the color of the illumination most suitable for classification. The colors used are IR, red, blue, and white light sources, each of which has a different wavelength. Since the shape of the object varies depending on the wavelength, the selection of lighting is an important factor for defect classification.

Figure 5 shows the wavelength band according to the color of the LED light.

Hereinafter, a defect classification method according to the present invention using extraction of feature information of defects 21 and 22 using the optical system 100 will be described.

First, a classification method using a maximum likelihood ratio (hereinafter referred to as ML) is adopted, which will be described in detail as follows.

ML is a parametric model that sequentially estimates data from training patterns using probability density function as a method of extracting feature patterns of distribution from input patterns. The probability density function of Equation 1 is established under the assumption that generation of input patterns is independent of each other.

클래스

의 평균벡터(mean vector),

는 클래스

의 공분산행렬(covariance matrix)을 나타낸다. 평균벡터

와 공분산행렬

는 아래의 수학식 (2)와 (3)에 의해 훈련패턴으로 추정된다.

class

Mean vector of,

Class

Denotes the covariance matrix of. Mean vector

And covariance matrix

Is estimated as a training pattern by the following equations (2) and (3).

의 훈련패턴 수, 수학식 2의

는

의

번째 값을 나타내며, 수학식 3의

는

의

성분,

는 클래스

의

성분을 의미한다. 또한 위 수학식을 근거로 수학식 (4)와 같이 likelihood ratio를 나타낼 수 있다. 이때

또는

의 경우에 따라 ML 규준을 적용하여 feature vector

가 소속하는 클래스를 추정할 수 있게 된다.class

Number of training patterns,

Is

of

Th value,

Is

of

ingredient,

Class

of

Means an ingredient. Also, based on the above equation, it can represent likelihood ratio as shown in equation (4). At this time

or

In some cases, we apply ML criteria to feature vectors

It is possible to estimate the class to which the belongs.

와

항에 의한 수학식 5와 같이 정의 할 수 있다.Next, Fisher's linear discriminant will be described. Fisher's linear discriminant has been proposed to find the slope of the line that best divides the two classes when the two-dimensional D-dimensional points are mapped to a one-dimensional straight line. In this case, in order to find a straight line dividing the two classes, it is necessary to determine the degree of division of the two classes mapped to the one-dimensional straight line. The function J (W), which determines the degree of division of the class mapped to a one-dimensional straight line where the D-dimensional point is the gradient vector W,

Wow

It can be defined as Equation 5 by the term.

,

만 을 이용해 2차 성능 검증을 하였다. Here, we need to find the transformation matrix W. In this paper, W is not found.

,

Secondary performance verification was performed using.

는 값이 클수록 좋고,

는 값이 작을수록 좋다. 2차 검증 방법에서는

/

을 사용하여 가장 큰 값을 가지는 feature를 선택 한다.In conclusion

Is higher value is better,

The smaller the value, the better. In the second verification method

Of

Use to select the feature with the largest value.

Next, a feature selection method will be described.

The foreign matter and chip defects, which are defects to be classified in the present invention, are not constant in size, position, and orientation, so that the characteristic information for classification should be independent of size, position, and orientation. In this experiment, we applied both feature information to compare the proposed feature information with the commonly used feature information. First, compactness, eccentricity, and roundness were selected as candidates for general feature information, and connected boundary ratio (CB) and difference of mean (DM) were selected as proposed feature information. These are described below.

일 때 수학식 9와 같이 나타낼 수 있고 그 값이 ‘1’이면 직선, ‘0’이면 원을 나타낸다. Roundness는 해당 영역의 단축길이(minor axis length)와 장축길이의 비율로서 ‘1’이면 원, ‘0’이면 직선을 나타낸다.Compactness is the ratio of the area to the perimeter. When the area is a and the perimeter is l, it can be expressed as Equation 8. Eccentricity is the ratio of the focal and major axis lengths of the ellipse that contains the area, and f is the distance from the center of the ellipse to the focal point.

In the case of Equation 9, a value of '1' indicates a straight line and a value of '0' indicates a circle. Roundness is the ratio between the minor axis length and the major axis length of the area.

, 결함의 평균밝기를

라 하면 수학식 11과 같이 나타낼 수 있다. The CB proposed in the present invention may be expressed as Equation 10 when a part of the periphery of the region is connected to the outside of the semiconductor package and the number of pixels connected to the outside is c and the perimeter is l. DM is the ratio of the average brightness of the package face to the average brightness of the defect area.

Average brightness of the defect

If it can be expressed as Equation (11).

Here, the feature information considers characteristics in which the pixel value of the defect is different from that of the package due to the intrinsic characteristics of the chip defects generated from the outside of the semiconductor package and the illumination.

Hereinafter, a defect classification algorithm according to the present invention will be described. 6 is a flowchart illustrating a defect area classification according to the present invention.

The threshold method is the most widely used method for the separation of specific regions in an image. The threshold technique used in the present invention is Kittler & Illingworth's minimum error threshold method. In addition, the morphology technique was used to remove unnecessary areas and maintain the original shape of defects after the threshold treatment.

3 box의 structuring element로 closing 연산(S230)을 하였다. 또한 배경에서 5~10개 화소 크기로 분포하는 노이즈 성분을 최소화하기 위해 11

11 box의 structuring element로 opening 연산을 하고 결과영상을 Image2라 하였다(S240). Opening으로 인하여 본래 모양을 잃었을 결함 영역을 최소화하기 위해 Image1과 Image2를 이용하여 conditional dilation(조건확대)기법을 적용(S250)하였다. 마지막으로 특징정보 추출을 위해 분할 된 영역에 blob analysis를 수행(S260)하였다.As shown in FIG. 6, first, the input image S210 is binarized using a minimum error threshold, and this is called Image1 (S220). And, to remove 1 ~ 2 pixel size hole that can occur in Image1, 3

The closing operation (S230) was performed using the three-box structuring element. Also, to minimize the noise component distributed in the size of 5-10 pixels in the background,

The opening operation was performed using the 11-box structuring element, and the resultant image was called Image2 (S240). In order to minimize the area of defects due to the opening, the conditional dilation technique was applied using Image1 and Image2 (S250). Finally, blob analysis was performed on the divided region for feature information extraction (S260).

Hereinafter, embodiments of a feature information extraction and classification method using the optical system according to the present invention will be described in detail.

768의 mono CCD 카메라를 사용하였고 검사할 반도체 패키지는 QFN 4mm

4mm패키지를 사용하였다. First, in the present invention, the camera used for image acquisition has a resolution of 1024.

A 768 mono CCD camera was used and the semiconductor package to be tested was QFN 4mm

4 mm package was used.

      (Table 2. Optical System Specifications)

Working Distance (mm) 216 Focal Length (mm) 134 Object size (QFN) (mm) 4 X 4 CCD Size (mm) 4.76 X 3.57 CCD Size (mm) 7.06 X 5.3 Pixel Resolution (μm / pixel) 6.9

In addition, the color of the illumination is evaluated to evaluate the classification performance and the accuracy of the characteristic data according to the color of the illumination when the data of the general characteristic information on the defect area and the proposed characteristic information CB and DM are applied to the ML method. About 800 foreign material defect images and chip defect images were taken for each of them by changing to IR, red, blue, and white, and classification for ML classification using two feature information among five feature data obtained from each image Models were created. In addition, two verification methods were used to evaluate and verify the ML classification performance of the generated classification models. As the first verification method, a total of 280 new defect images by color and defects of lighting were applied to the classification model and photographed. Table 3 shows the results of the verification method.

When the red side lighting was used in Table 3, the classification performance by the combination of selected feature information was the best with an average of 98.88%. It can be seen that in the blue light, the most inaccurate feature information data is obtained. In addition, the DM vs CB method showed 99.9% classification performance in all four colors, unlike other feature information whose performance changes according to the illumination color change.

   (Table 3. Classification and Evaluation Results by Lighting Type and Feature Information)

Illumination Color Features White (%) Blue (%) Red (%) IR (%) Comp vs Ecc 96.5 96.5 98.6 96.5 Ecc vs DM 97.2 99.3 99.9 99.3 DM vs CB 99.9 99.9 99.9 99.9 Round vs Comp 96.5 96.5 97.9 96.5 Round vs Ecc 94.4 30.1 97.9 94.4 Average performance 96.9 84.46 98.84 97.32 Standard deviation 1.98 30.42 1.91 2.26

    * Comp: compactness, Ecc: eccentricity, Round: roundness

As the second verification method, the variance between class distribution and class variance was obtained by calculating the variance between each class and variance in class in the feature domain. Table 4 below shows the results of the second verification.

    (Table 4. Comparison of variance between classes and variance in classes in feature domain)

Data characteristic information SB S _W S _B / S _W Comp vs Ecc 4.81 e-04 4.03 e-04 1.195 Ecc vs DM 5.20 e-04 9.24 e-05 5.629 DM vs CB 3.94 e-04 2.74 e-05 14.347 Round vs Comp 9.30 e-04 7.34 e-04 1.266 Round vs Ecc 3.98 e-04 5.53 e-04 0.721

: 클래스 내 분산,

: 클래스 간 분산*

: Distribution within a class,

: Distributed between classes

The above results also showed the best results when using the DM vs CB method. This result means that the two classes have high independence in the feature domain, which is directly related to the classification performance.

7 is a graph showing chip defects and foreign material defect classification models in the feature information domain of a red illumination image. Here, the characteristic information domain is shown in a graph, where the results of Table 4 can be easily confirmed.

In this case, in detail, FIG. 7A shows Compactness vs eccentricity, FIG. 7B shows Eccentricity vs DM, FIG. 7C shows DM vs CB, FIG. 7D shows Roundness vs compactness, and FIG. 7E shows Roundness vs eccentricity. .

The classification models generated by the general feature information partially overlap the feature distribution of defects, but the classification model generated by the proposed feature information shows that the feature distribution of defects is almost independent. Referring to Figure 7 (c), which shows the DM vs CB feature domain, the proposed feature information CB is the feature information proposed by using the characteristics of the chip and the foreign material that occurs independently of the location. It is determined that the feature information is sufficient to improve, and the DM using the brightness difference information also generates an effective classification model to improve the classification performance.

As described above, in the optical system 100 according to the present invention, the side lighting 150 and the coaxial illumination (the same material) are used to maximize the difference between the defect surfaces and to minimize the noise due to the characteristics of the surface of the semiconductor package 110. 140) was used at the same time, it was confirmed through the proposed optical system it is easy to detect the chip defect 22, which is difficult to detect in the existing optical system. In addition, the feature information CB and DM using the position and brightness characteristics of the defects 21 and 22 are used rather than the general feature information, and for the verification, the experimental verification method using the bad image and the dispersion of the class generated by the obtained feature information The method of analysis was used. The experimental results show that the classification performance is significantly improved when the proposed feature information is used, compared with the general feature information (eccentricity, roundness, compactness). It can be seen in the above example that the classification performance.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto and is intended by those skilled in the art to which the present invention pertains. Of course, various modifications and variations are possible within the scope of equivalents of the claims to be described.

1 is a cross-sectional view showing a foreign matter and chip defects in the form of a defect

2 is a block diagram showing a conventional optical system configuration

3 is a block diagram showing the configuration of an optical system according to the present invention

4 is a photograph for comparing the optical system image

5 is a graph of wavelength characteristics of illumination used in another optical system according to the present invention.

6 is a flowchart illustrating a defect area classification according to the present invention.

6 is a graph of chip defects and foreign material defect classification models in the feature information domain of a red illumination image.

<Description of Symbols for Main Parts of Drawings>

100: optical system 110: semiconductor package

120: CCD camera 130: half mirror

140: coaxial light 150: side lighting

160: controller

Claims (4)

In the vision system for inspecting the external defect of the semiconductor package, A CCD camera installed at the upper side of the semiconductor package and configured to photograph the reflected light reflected by the light from the lighting means and the reflected light reflected by the foreign matter and chip defects present on the surface of the package; Illuminating means positioned between the lower portion of the camera and the upper portion of the package to emit light; A half mirror for refracting light supplied from the lighting means at a predetermined angle to irradiate the package; And And a shooting controller unit for controlling the providing of the light source and storing and transmitting the captured image data. The lighting means comprises a side illumination for irradiating light from the side of the package to be inspected and coaxial illumination configured to drop the light source at a 90 ° angle to the package by transmitting light through the lens and the upper side of the side illumination. Vision system for inspection of external defects in QFN semiconductor packages, characterized by parallel formation of side and coaxial light paths The method of claim 1, Vision system for inspection of external defects of the QFN semiconductor package, characterized in that the light emitting source constituting the lighting means is an LED A semiconductor package appearance defect classification method using the vision system according to claim 1 or 2, Using the probability density function, we estimate the data sequentially from the training pattern and use the maximum likelihood ratio (ML) method to extract the feature pattern of the distribution from the input pattern, and extract feature information independent of size, position, and direction for classification. QFN semiconductors, which use a threshold technique for separation of specific regions from an image, and a defect classification algorithm using a morphology technique to remove unnecessary regions appearing after the threshold processing and maintain the original shape of the defects. Package appearance defect classification method The method of claim 3, wherein The algorithm is Inputting an image (S210); Binarizing the input image using a minimum error threshold and naming it image 1 (S220); Performing a closing operation with a strutching element of a 3 × 3 box to remove holes having one or two pixel sizes that may occur in the image 1 (S230); In step S240, performing an opening operation with a strutching element of an 11 × 11 box and minimizing a noise component distributed in a size of 5-10 pixels in the background (S240); applying a conditional dilation technique using the image 1 and the image 2 in order to minimize the defect area which will lose its original shape due to the opening (S250); And Performing blob analysis on the divided region for feature information extraction (S260); QFN semiconductor package appearance defect classification method comprising a

Images