CN108596875B - Image segmentation algorithm-based semiconductor chip flash rapid detection method - Google Patents

Abstract

The invention discloses a rapid detection method for semiconductor chip overflow based on an image segmentation algorithm, which aims to identify the excess material that overflows onto pins and exposed carriers during the packaging process of integrated circuit EMC (Epoxy Molding Comppond), so as to provide EMC packaging. The process provides timely feedback. First, preprocess the original image of the EMC package collected by the industrial camera, and then use the Selective search algorithm to segment the image to obtain the target candidate area, and filter out the area of the chip to be detected. The area of is subjected to threshold segmentation to obtain the overflow area. The method can give the flash detection result of each chip in a very short time, which not only optimizes the flash detection method in the EMC packaging process, but also provides real-time feedback information for the EMC packaging process, thereby improving the EMC packaging process. performance.

Description

A fast detection method for semiconductor chip flashing based on image segmentation algorithm

technical field

The invention belongs to the field of integrated circuit EMC package overflow detection, and in particular relates to a semiconductor chip overflow rapid detection method based on an image segmentation algorithm.

Background technique

EMC is an important microelectronic packaging material. The semiconductor chip is covered by the packaging process to form protection to avoid damage from the external environment. At the same time, EMC also has a certain heat dissipation effect. EMC has the characteristics of large-scale production and reasonable reliability. It is one of the common packaging materials for semiconductor packaging. It occupies more than 95% of the packaging market with its unique advantages. However, in the semiconductor plastic packaging process, due to the gap between the lead frame and the plastic packaging film, the glue-like plastic packaging material will leak out from the gap, solidify on the surface of the lead body, and form black overflow, which is the integrated circuit pin and exposed excess on the carrier. The overflow itself will not adversely affect the performance of the plastic-encapsulated product, but because the overflowing plastic-encapsulated chatter covers the lead pins, if it is not treated, an insulating area will be formed, which will affect the subsequent plating quality and product reliability, resulting in product disconnection and virtual welding. And other issues. Therefore, how to reduce the occurrence of flash and how to remove the flash has always been an important problem for researchers and manufacturers.

The reasons for EMC packaging overflow are usually: 1) The plastic sealing compound has good fluidity and low viscosity, and the plastic sealing compound/epoxy resin may overflow from the parting surface and cover the lead pins if the mold is not tightly closed; 2) After long-term use of the plastic packaging mold, the surface is worn or the base is not flat, which causes flashing; 3) During the semiconductor packaging process, the frame accuracy does not match the mold, and the lack of meshing causes the problem of edge blanking. After the overflow is formed, it has no effect on the performance of the plastic encapsulated product, but because the overflowed plastic encapsulates the lead pins, it may form plating defects and affect the reliability of the product. At present, manufacturers generally use a file, sandpaper grinding, and weak acid cleaning to remove residual flash after the flash removal process. However, this method is easy to scratch the plastic body and cause the surface of the lead pins to be rough, which affects the appearance quality. and electroplating quality, but also affect the production progress. Therefore, how to avoid and reduce the generation of flash is particularly important. In conclusion, it is of great practical significance to develop a visual green and pollution-free overflow detection method, which enables rapid detection of chip overflow to meet the needs of timely provision of reference for the EMC packaging process.

Thanks to the development of computer technology and image processing technology, a new type of visual flash detection technology has emerged. The detection result is a grayscale image. It has the characteristics of high detection accuracy, flexible use and large amount of information, which is very suitable for chip packaging production. Process overflow detection.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a fast flash detection strategy for the flash problem existing in the existing EMC packaging process. This strategy can accurately assess the chip overflow situation on site, and then provide feedback information for the adjustment and maintenance of in-service equipment, which is an important reference for the EMC packaging process.

The object of the present invention is achieved through the following technical solutions: a method for fast detection of semiconductor chip flashing based on an image segmentation algorithm, the method comprises the following steps:

A method for rapid detection of semiconductor chip flashing based on an image segmentation algorithm, characterized in that the method comprises the following steps:

(1) Using a semiconductor packaging press (FSAM120-1US) and a QFNQFNWB7X7-48L (T0.75) mold to produce a packaged original, and use an industrial camera to obtain multi-frame grayscale images of the packaged original;

(2) Gaussian blur preprocessing is performed on each frame of image;

(3) Image segmentation based on selective search, to obtain hundreds of candidate regions and screen out the chip region to be analyzed, this step is realized by the following sub-steps:

(3.1) Over-segment a single frame image, and obtain _n initial regions r ₁ , r ₂ ,..., rn , and form the initial region set R={r ₁ , _{r 2} ,..., rn } ; At the same time, construct an initialization similarity set S,

(3.2) Calculate the similarity s(r _i , r _j ) of each two adjacent regions, the similarity index adopts color similarity, size similarity, texture similarity, fit similarity, and four kinds of complementary similarity, respectively. Consolidate regions from different aspects;

(a) The calculation method of color similarity is as follows:

其中，

表示第i个初始区域中第k个像素值区间的归一化频数；Divide each initial area into q intervals according to the pixel value to obtain a normalized color histogram

in,

Indicates the normalized frequency of the kth pixel value interval in the ith initial region;

The color similarity of two adjacent regions r _i , r _j is:

The new region r _ij obtained by merging r _i , r _j , its color histogram is:

size(r _ij )=size(r _i )+size(r _j ) (19)

where size( _ri ) is the size of the initial region _ri ;

(b) The calculation method of texture similarity is:

得到作为纹理特征。For each initial region, calculate its SIFT (scale-invariant feature transform) feature, where the scale factor is 1 in this patent, and the histogram interval width is 10, obtaining

obtained as texture features.

表示第i个初始区域中第k个纹理特征；in,

Represents the kth texture feature in the ith initial region;

The texture similarity of two adjacent regions r _i , r _j is:

The new region r _ij obtained by merging r _i , r _j has a texture histogram of

(C) The calculation method of size similarity is:

where size(im) represents the size of the entire grayscale image.

(d) The calculation method of coincidence similarity is:

where BB _ij is a bounding box _{that only encloses regions ri and r j .}

Finally, the four similarities are combined to obtain the final similarity:

s(r _i ,r _j )=a ₁ Scolor(r _i ,r _j )+a ₂ Stexture(r _i ,r _j )+a ₃ Ssize(r _i ,r _j )+a ₄ Sfill(r _i ,r _j )(24)

where a _i ∈ {0,1}, indicating whether the corresponding similarity is adopted.

(3.3) All the similarities constitute the similarity set S, the two adjacent regions with the largest similarity in the similarity set s(r _i , r _j )=max(S), merge these two regions r _ij =r _i ∪r _j , calculate the similarity between the new region r _ij and adjacent regions, update the similarity set: delete all the similarities s(r _{i ,} r _* ), and All the similarities s(r _j , r _# ), which are related to the merged region r _j , are added to the updated similarity set as: adding the similarity between the new region r _ij and the adjacent region;

(3.4) Loop steps 3.2-3.3, when the similarity set S is an empty set, the loop ends. Complete the merging of the initial regions to obtain the candidate region set R;

和

这4个关键几何点，封装原片的分辨率为a*b。(3.5) Screen out the upper and lower chip regions to be analyzed from the candidate region set R, so that the size of the region to be analyzed is in the range of one-third to one-half of the size of the original package, and the two regions to be analyzed need to be separately Include

and

For these four key geometric points, the resolution of the original package is a*b.

(4) image segmentation based on the threshold is performed on the screened area to be analyzed, and this step is realized by the following sub-steps:

(4.1) Calculate the normalized histogram of the area to be analyzed. Let {0,1,...,L-1} denote L different gray levels in an area image of size M*N pixels, and n _z denote the number of pixels with gray level z. The total number of pixels in the area image is MN=n ₀ +n ₁ +n ₂ +...+n _L-1 , p _z =n _z /MN,z=0,1,2,...,L-1 Represents the normalized frequency of the gray level z in this histogram.

(4.2) For u=0,1,2,...,L-1, calculate the cumulative sum P ₁ (u), the calculation formula is as follows:

(4.3) For u=0,1,2,...,L-1, calculate the cumulative mean value m(u), the calculation formula is as follows:

(4.4) Calculate the global grayscale mean value m _G , the calculation formula is as follows:

计算公式如下：(4.5) For u=0,1,2,...,L-1, calculate the inter-class variance

Calculated as follows:

(4.6) Otsu threshold value u ^* is obtained by making the objective function (13) obtain the maximum value, if the maximum value is not unique, u ^* is obtained by the average of the respective detected maximum values u;

(4.7) Segment the regional image based on the optimal threshold u ^* using the following formula to obtain the segmented image:

f(x,y) is the pixel of the pixel point (x,y) in the image of the area to be analyzed. After segmentation, the pixel point of 1 represents the detected flashing part in the production process.

The beneficial effects of the present invention are as follows: the present invention is a rapid flash detection method oriented to the EMC packaging process, and the image is divided into hundreds to thousands of candidate regions through a selective search algorithm, and the chip region to be analyzed is screened from the candidate regions. . Further, the optimal threshold is used to segment the image to be analyzed to obtain the overflow image, so that the seriousness of the overflow can be measured. For each frame of image, the present invention can provide the detection result within 20 seconds, can quickly feed back the flashing information for the EMC packaging process, and provide a reference basis, thereby avoiding the occurrence of more flashing situations.

Description of drawings

Fig. 1(a) is a flow chart of the semiconductor packaging of the present invention, Fig. 1(b) is the parameter setting in the actual production operation, and Fig. 1(c) is the final package original sheet;

Fig. 2 is the semiconductor package overflow detection system of the present invention, the original grayscale image is obtained by the industrial camera, and transmitted to the computer for the next step analysis;

Fig. 3 is the encapsulated original sheet grayscale image obtained by the present invention;

4 is a reduced resolution image obtained by the method of the present invention utilizing Gaussian blurring preprocessing, which retains the complete information of the image while reducing the image resolution;

5 is a partial result display obtained after the method of the present invention is selectively searched to segment the image;

Fig. 6 is the to-be-analyzed area screened out by the method of the present invention;

FIG. 7 is the final flash image obtained by the method of the present invention after threshold segmentation.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific examples.

(1) EMC (Epoxy molding compound) package: As shown in Figure 1(a), place the lead frame in the mold so that each chip is located in the cavity, and place the bulk EMC in the mold air after the mold is closed. Set the parameters of the packaging press as shown in Figure 1(b). At high temperature, the EMC begins to melt, flows to the acupoints along the track, and gradually covers the chip from the bottom until it is completely covered and wrapped, forming and solidifying, as shown in Figure 1(c) shown;

(2) Image acquisition and preprocessing: The image-based flash detection system is shown in Figure 2. The industrial camera is used to obtain the grayscale image of the packaged original film as shown in Figure 3, and its resolution is 3264*4896. The blurring process reduces its resolution to 408*612, which retains the integrity of the image information, removes unnecessary noise, reduces the image resolution, and speeds up the flash analysis and detection, as shown in Figure 4.

(3.1) Over-segment a single frame image, and obtain _n initial regions r ₁ , r ₂ ,..., rn , and form the initial region set R={r ₁ , _{r 2} ,..., rn } ; At the same time, construct an initialization similarity set S,

(3.2) Calculate the similarity s(r _i , r _j ) of each two adjacent regions, the similarity index adopts color similarity, size similarity, texture similarity, fit similarity, and four kinds of complementary similarity, respectively. Consolidate regions from different aspects;

(a) The calculation method of color similarity is as follows:

其中，

表示第i个初始区域中第k个像素值区间的归一化频数；Divide each initial area into q intervals according to the pixel value to obtain a normalized color histogram

in,

Indicates the normalized frequency of the kth pixel value interval in the ith initial region;

The color similarity of two adjacent regions r _i , r _j is:

The new region r _ij obtained by merging r _i , r _j , its color histogram is:

size(r _ij )=size(r _i )+size(r _j ) (33)

where size( _ri ) is the size of the initial region _ri ;

(b) The calculation method of texture similarity is:

得到作为纹理特征。For each initial area, calculate its SIFT (scale-invariant feature transform) feature, in the present invention, the scale factor is 1 in this patent, and the histogram interval width is 10, obtaining

obtained as texture features.

表示第i个初始区域中第k个纹理特征；in,

Represents the kth texture feature in the ith initial region;

The texture similarity of two adjacent regions r _i , r _j is:

The new region r _ij obtained by merging r _i , r _j has a texture histogram of

(C) The calculation method of size similarity is:

where size(im) represents the size of the entire grayscale image.

(d) The calculation method of coincidence similarity is:

where BB _ij is a bounding box _{that only encloses regions ri and r j .}

Finally, the four similarities are combined to obtain the final similarity:

s(r _i ,r _j )=a ₁ Scolor(r _i ,r _j )+a ₂ Stexture(r _i ,r _j )+a ₃ Ssize(r _i ,r _j )+a ₄ Sfill(r _i ,r _j )(38)

where a _i ∈ {0,1}, indicating whether the corresponding similarity is adopted.

(3.3) All the similarities constitute the similarity set S, the two adjacent regions with the largest similarity in the similarity set s(r _i , r _j )=max(S), merge these two regions r _ij =r _i ∪r _j , calculate the similarity between the new region r _ij and adjacent regions, update the similarity set: delete all the similarities s(r _{i ,} r _* ), and All the similarities s(r _j , r _# ), which are related to the merged region r _j , are added to the updated similarity set as: adding the similarity between the new region r _ij and the adjacent region;

(3.4) Loop steps 3.2-3.3, when the similarity set S is an empty set, the loop ends. Complete the merging of the initial regions to obtain the candidate region set R;

和

这4个关键几何点，封装原片的分辨率为a*b。(3.5) Screen out the upper and lower chip regions to be analyzed from the candidate region set R, as shown in Figure 6, so that the size of the region to be analyzed is within the range of one-third to one-half of the size of the original package, while the two Each area to be analyzed needs to contain

and

For these four key geometric points, the resolution of the original package is a*b.

(3.4) At the end of the cycle, a candidate region R is obtained; the above-mentioned selective search algorithm steps are performed on the grayscale image shown in Figure 4, and 214 candidate regions will be obtained, some of which are shown in Figure 5.

(3.5) Region screening: set the screening conditions, and select the chip region to be analyzed from the candidate region R. Screening the grayscale image shown in Fig. 4 will obtain the area to be analyzed as shown in the images on the right side of Fig. 6 .

(4) image segmentation based on the threshold is performed on the screened area to be analyzed, and this step is realized by the following sub-steps:

(4) image segmentation based on the threshold is performed on the screened area to be analyzed, and this step is realized by the following sub-steps:

(4.1) Calculate the normalized histogram of the area to be analyzed. Let {0,1,...,L-1} denote L different gray levels in an area image of size M*N pixels, and n _z denote the number of pixels with gray level z. The total number of pixels in the area image is MN=n ₀ +n ₁ +n ₂ +...+n _L-1 , p _z =n _z /MN represents the normalized frequency of the gray level z in the histogram , z=0,1,2,...,L-1.

(4.2) For u=0,1,2,...,L-1, calculate the cumulative sum P ₁ (u), the calculation formula is as follows:

(4.3) For u=0,1,2,...,L-1, calculate the cumulative mean value m(u), the calculation formula is as follows:

(4.4) Calculate the global grayscale mean value m _G , the calculation formula is as follows:

计算公式如下：(4.5) For u=0,1,2,...,L-1, calculate the inter-class variance

Calculated as follows:

(4.6) Otsu threshold value u ^* is obtained by making the objective function (13) obtain the maximum value, if the maximum value is not unique, u ^* is obtained by the average of the respective detected maximum values u;

(4.7) Segment the regional image based on the optimal threshold u ^* using the following formula to obtain the segmented image:

f(x,y) is the pixel of the pixel point (x,y) in the image of the area to be analyzed. After segmentation, the pixel point of 1 represents the detected flashing part in the production process. The black part in the dashed box in Figure 7 is the flash component in the chip, and the flashing situation can be seen intuitively. If the flashing is serious, the packaging process should be stopped immediately, and the mold parameters should be checked to avoid more flashing.

It should be understood that the present invention, as a detection algorithm, is not limited to the specific experimental facilities and experimental conditions of the above-mentioned specific embodiments, and those skilled in the art can also make equivalent deformations or substitutions without departing from the spirit of the present invention. , these equivalent modifications or substitutions are all included within the scope defined by the claims of the present application.

Claims (4)

1. A semiconductor chip flash rapid detection method based on an image segmentation algorithm is characterized by comprising the following steps:

(1) acquiring a multi-frame gray image of a packaging original sheet by using an industrial camera;

(2) performing Gaussian blur preprocessing on each frame of image;

(3) based on selective search image segmentation, hundreds of candidate regions are obtained and chip regions to be analyzed are screened out, and the method is realized by the following substeps:

(3.1) over-dividing the single-frame image to obtain n initial regions r₁,r₂,...,r_nForming an initial region set R ═ { R ═ R₁,r₂,...,r_n}; meanwhile, an initialization similarity set S is constructed,

(3.2) calculating the similarity s (r) of every two adjacent areas_i,r_j) The similarity index adopts color similarity, size similarity, texture similarity, coincidence similarity and four complementary similarities, and the four complementary similarities are respectively combined for the areas from different aspects;

(a) the color similarity is calculated as follows:

dividing each initial region into e intervals according to pixel values to obtain a normalized color histogram

Wherein,

the normalization frequency number of the ith pixel value interval in the ith initial area is represented;

two adjacent regions r_i,r_jThe color similarity of (a) is:

merge r_i,r_jThe new region r obtained_ijThe color histogram is:

size(r_ij)＝size(r_i)+size(r_j) (4)

wherein size (r)_i) Is an initial region r_iThe size of (d);

(b) the texture similarity calculation method comprises the following steps:

calculating the SIFT (scale-invariant feature transform) feature of each initial region to obtain

Obtaining the number of the texture features as v; wherein,

representing the ith texture feature in the ith initial region;

two adjacent regions r_i,r_jThe texture similarity of (a) is:

merge r_i,r_jThe new region r obtained_ijThe histogram of the texture is

(C) The calculation method of the size similarity comprises the following steps:

wherein size (im) represents the size of the entire grayscale image;

(d) the calculation method of the coincidence similarity comprises the following steps:

wherein BB_ijTo surround only the region r_iAnd r_jThe bounding box of (1);

and finally, integrating the four similarities to obtain the final similarity:

s(r_i,r_j)＝a₁Scolor(r_i,r_j)+a₂Stexture(r_i,r_j)+a₃Ssize(r_i,r_j)+a₄Sfill(r_i,r_j) (9)

wherein a is_m∈{0,1}，m＝1,2,3,4, indicating whether the corresponding similarity is adopted;

(3.3) all the similarities form a similarity set S, and the two adjacent regions S (r) with the highest similarity in the similarity set_i,r_j) Max(s), merge the two regions r_ij＝r_i∪r_jCalculating a new region r_ijAnd (3) updating the similarity set with the similarity of adjacent regions: deleting the region r with which the data is merged_iAll similarities s (r) of the correlation_i,r_*) And with the region r being merged_jAll similarities s (r) of the correlation_j,r_#) And adding the updated similarity set as follows: adding a new region r_ijSimilarity to adjacent regions;

(3.4) circulating the steps (3.2) - (3.3), and ending circulation when the similarity set S is an empty set; completing the combination of the initial regions to obtain a candidate region set R;

(3.5) screening an upper chip area and a lower chip area to be analyzed from the candidate area set R, so that the size of the area to be analyzed is in the range of one third to one half of the size of the original packaging piece, and the two areas to be analyzed need to be respectively contained

And

the resolution of the packaging original sheet is a x b at the 4 key geometric points;

(4) performing threshold-based image segmentation on the screened region to be analyzed, wherein the step is realized by the following substeps:

(4.1) calculating a normalized histogram of the region to be analyzed; let {0, 1.,. L-1} denote L different gray levels in a region image of size M N pixels, N_zThe number of pixels having a gray level z is represented; the total number of pixels in the area image is MN ═ n₀+n₁+n₂+...+n_L-1，p_z＝n_zthe/MN represents a normalization frequency with z as a gray level in the histogram, where z is 0,1, 2.

(4.2) for u-0, 1,2₁(u), the calculation formula is as follows:

(4.3) for u 0,1, 2.., L-1, calculate the cumulative mean m (u) as follows:

(4.4) calculating the Global Gray mean m_GThe calculation formula is as follows:

(4.5) for u-0, 1,2

The calculation formula is as follows:

(4.6) obtaining the Otsu threshold u by maximizing the objective function (13)^*If the maximum values are not unique, u is obtained by averaging the respective detected maximum values u^*；

(4.7) based on the optimal threshold u^*The region image is segmented by the following formula to obtain a segmented image:

f (x, y) is the pixel of the pixel point (x, y) in the image of the area to be analyzed; and the segmented pixels are 1 and represent the detected flash part in the production process.

2. The method of claim 1, wherein the packaging raw sheet in step (1) is a packaging raw sheet produced by using a semiconductor packaging press FSAM120-1US and qfnqqnwb 7X7-48L T0.75.75 dies.

3. The method of claim 1, wherein the gaussian blurring step of step (2) is as follows:

(2.1) generating a gaussian operand of size (2k +1) × (2k +1) according to equation (1):

wherein k is an integer, w_pqIs the weight of the (p, q) coordinate position in the Gaussian operand, σ is the standard deviation of the Gaussian distribution;

(2.2) calculating the reciprocal of the sum of Gaussian operand coefficients

Multiplying each element in the Gaussian operand by the element to obtain a new Gaussian operand;

and (2.3) reading pixels from the image and performing convolution operation on the pixels and the Gaussian operand of the second step, thereby completing the Gaussian blur preprocessing.

4. The method according to claim 3, wherein in step (2.1), the larger σ is, the wider the frequency band of Gaussian blur is, the better the smoothing degree of the image is, and the higher the detail loss degree is; by adjusting the sigma parameter, the cancellation of image noise and preservation of integrity can be balanced.

Images