CN110992326B - QFN chip pin image rapid inclination correction method - Google Patents

Abstract

The invention discloses a quick tilt correction method for a pin image of a QFN chip, comprising the following steps: (1) collecting the pin image of the QFN chip, filtering and binarizing it; (2) extracting the center pad of the chip image by using a polygon approximation method contour; (3) propose an improved Harris corner detection algorithm to obtain contour vertices; (4) use the least squares method to fit a straight line to the farthest vertex, and use the straight line as the angle to identify the direction; (5) use the image centroid coordinates For the center of rotation, quickly correct the chip pin image and remove white edges. The correction method provides a certain theoretical basis for faster and more accurate correction of QFN chips, and improves the visual inspection efficiency of QFN package defects.

Description

A fast tilt correction method for QFN chip pin image

technical field

The invention belongs to the field of image processing algorithm design, and proposes an improved Harris corner point detection algorithm, combined with a polygon approximation method, to design a fast tilt correction method for QFN chip pin images.

Background technique

QFN (Quad Flat No-lead Package) is a leadless package, which is square or rectangular. It uses the middle pad at the bottom of the package to conduct heat, and there are conductive pads around the package periphery for electrical connection around the middle pad. There will be a certain size error in the production process of QFN chips, so there will be a margin when manufacturing the tray carrier port, but this will cause the chip to be tilted when it is placed, which will affect the visual inspection of the chip packaging quality.

At the same time, due to the small size of the chip, the manual detection and identification method with low accuracy is far from meeting the requirements of current chip production. It is urgently needed to develop an effective technology that can quickly and accurately correct the chip image. solved problem. The research on QFN chips in the prior art mainly focuses on the improvement of QFN chip structure, manufacturing process, appearance inspection and QFN chip defect detection. At present, there is no literature or patent related to the fast correction method of QFN chip image tilt. Therefore, Designing a fast tilt correction method for QFN chip pin images to fill the gap of existing research in this aspect is particularly important for improving the efficiency of visual inspection of QFN package defects.

SUMMARY OF THE INVENTION

In view of the problems existing in the prior art, the present invention proposes an improved Harris corner detection algorithm, combined with the polygon approximation method, to design a fast tilt correction method for the pin image of a QFN chip, which is faster and more efficient than the traditional algorithm. .

The technical scheme of the present invention is: a method for fast tilt correction of a pin image of a QFN chip, comprising the following steps:

1. Preprocess the chip pin image collected by the industrial computer:

1.1 Image filtering:

In order to remove noise and reduce image distortion, a 5×5 Gaussian filter is used to convolve the image to smooth the image and reduce the obvious noise effect on the edge detector. In image processing, a two-dimensional Gaussian function is often used to filter and calculate The formula is as follows:

Among them, G(x, y) is the two-dimensional Gaussian function, (x, y) is the point coordinate, σ is the standard deviation, and A is the normalization coefficient, so that the sum of different weights is one;

1.2. Binarization processing:

The fixed threshold method is used to binarize the image, and the calculation formula is as follows:

Among them, f(x, y) represents the distribution function of image pixel values, g(x, y) represents the pixel value distribution function after threshold segmentation, and the fixed threshold T=145;

2. Use the polygon approximation method to extract the target contour, including:

2.1 Edge detection on chip pin image:

Using Canny operator edge detection, the edge contour information of the chip pin image is obtained;

2.2 Use the polygon approximation method to extract the target contour, that is, the center pad contour of the chip pin image:

The contour of the center pad of the chip pin image is extracted by the polygon approximation method, and the remaining contour parts are filtered out, which simplifies the subsequent image processing. The polygon approximation method is to pick out the two farthest points from the target contour and connect them; Find a point on the target contour with the farthest distance from the off-line segment, add this point to the new contour after approximation, that is, the triangle formed by connecting the three points as the contour; finally select any side of the triangle to start, repeat the previous step, and add Add a new contour to the farthest point, and iterate continuously until the output accuracy requirements are met;

3. An improved Harris corner detection algorithm is proposed to obtain the target contour vertices, including:

3.1. Determine the selection threshold and obtain the corner points of the target contour:

Under normal circumstances, when the difference between the pixel gray levels of the two black and white images is less than 10% to 15% of the maximum pixel gray value, it is difficult for the human eye to distinguish, so the threshold N is selected to extract the corners of the target contour of the image, and the formula is as follows:

N=255×12%≈30

3.2. Extract the corner points of the target contour:

Extract and save all the corner points, read the three points _Man , M _a , and M _a+n in the corner points in sequence, the three points constitute a template consisting of three elements, among which the three points take M _a as the template center , the subscript of the point represents the sequence number of the corner point in all the corner points, traverse all the corner points, starting from the initial value, M _a is determined as a real-time operation point, take the two points whose front and rear numbers are separated by n, and M _a are respectively the same as The other two points in the template form two sides, the angle formed by the two sides is used as the corner response value of point Ma, and the distance between point _Ma and point _Man determines the edge L ₁ , point _Ma and point M _{a +} The distance of _n determines the side L ₂ , the distance between the point _Man and the point M _{a +n} determines the side L ₃ , and the angle of the point Ma can be obtained from the three sides according to the cosine law. The calculation formula is as follows:

The angle formed by the two sides, that is, the response value of the corner point, is used to determine whether to retain the corner point. Let g be the response value of the corner point. If the point M _a ≥ g, save it as the required corner point; otherwise, remove it;

3.3. Eliminate adjacent corners and retain contour vertices:

After the contour corner points are extracted, there may be other corner points around. To eliminate this phenomenon, the adjacent corner points are removed, and the remaining points are taken as the contour corner vertices. Set the image height as H and the image width as W, Corner(x, y) indicates whether there is a corner point at the image (x, y), when Corner(x, y)=1, there is a corner point at (x, y), m×m (m>1 ) is the size of the matrix centered at (x, y), then:

Among them, m≤x≤H, m≤y≤W, count refers to the corner points of the matrix range with (x, y) as the center, and all the adjacent corner points of the matrix range with (x, y) as the center are removed and reserved The remaining corner points at the corners of the target contour are used as vertices to facilitate subsequent image correction processing;

4. Least squares fit straight line:

Using the least squares method, the two vertices of the longest side of the target contour are fitted with a straight line as the angle identification direction of the chip pin image;

5. Quickly correct the chip and remove the white edge according to the image centroid, including:

5.1. Taking the image centroid as the rotation center:

In order to correct the image accurately, the centroid method is used to determine the center position of the image, and the centroid coordinate O point is used as the rotation center of the image;

5.2. Obtain the image tilt angle and correct the chip pin image.

Further, in the step 4, set the coordinates of the two vertices to be p(x, y) and q(x, y) respectively, and the least squares fitting straight line calculation formula is as follows:

y=ax+b

a and b are the slope and intercept of the straight line equation, respectively, then:

分别为顶点p、q的横坐标与纵坐标的均值，

in

are the mean values of the abscissa and ordinate of vertices p and q, respectively,

Further, in the step 5.1, the upper left corner of the image is set as the starting point coordinates (0, 0), the lower right corner is set as the end point coordinates (m, n), and the image centroid formula is as follows:

Among them, (x ₀ , y ₀ ) are the coordinates of the centroid, m and n are the number of rows and columns of the image respectively (m, n are both integers greater than or equal to 2), f(x, y) is the image at the point ( gray value at x, y).

Further, in the step 5.2, the chip image tilt angle α can be calculated by the following calculation formula:

The offset value x of the pin image in the horizontal direction can be obtained:

x=|p _x -h _x |

Then the pin image tilt angle α is:

Among them, l is the vertical distance from point p to point O, OA is the extension of the centroid coordinate to the positive direction of the x-axis, and h(x, y) is the coordinate of the intersection point of pq and OA.

The beneficial effects of the present invention are:

The invention discloses a quick tilt correction method for a pin image of a QFN chip, which provides a certain theoretical basis for correcting the QFN chip more quickly and accurately, and improves the visual inspection efficiency of QFN package defects.

Description of drawings

Figure 1a is the original image of the QFN chip, Figure 1b is the Gaussian filtered image, and Figure 1c is the binarized image;

Figure 2a is a Canny edge detection image, and Figure 2b is a polygon approximation contour image;

Fig. 3 is the corner image obtained according to the threshold;

Fig. 4 is to extract the image target contour vertex;

Fig. 5 is to select two vertices of the longest contour, and utilize the least squares method to carry out straight line fitting;

Fig. 6 is the schematic diagram of angle deviation;

Fig. 7a is the image after rotation correction, Fig. 7b is the image after removing the white edge;

FIG. 8 is a schematic diagram of the correction labeling after the Hough transform algorithm detects the angle;

FIG. 9 is a schematic diagram of calibration and labeling after the minimum circumscribed moment method detects the angle;

FIG. 10 is a schematic diagram of the correction annotation after the angle is detected based on the improved Harris corner detection algorithm;

Figure 11 is the angle deviation data statistics table of the three algorithms;

Figure 12 is the operation time data statistics table of the three algorithms;

FIG. 13 is a flow chart of a method for fast tilt correction of a pin image of a QFN chip.

Detailed ways

The following examples further illustrate the content of the present invention, but should not be construed as limiting the present invention. Modifications and substitutions made to the methods, steps or conditions of the present invention without departing from the essence of the present invention all belong to the scope of the present invention.

In order to improve the visual detection efficiency of QFN package defects, this embodiment discloses a method for fast tilt correction of QFN chip pin images. Fig. 1a is the original image, and Fig. 1a is used as the explanatory image of this embodiment. The specific correction process includes the following steps :

(1) Preprocess the chip pin image collected by the industrial computer:

(1.1) Image filtering:

In order to remove noise and reduce image distortion, a 5 × 5 Gaussian filter is used to convolve the image (as shown in Figure 1b) to smooth the image and reduce the obvious noise effect on the edge detector. In image processing, two-dimensional Gaussian function is often used for filtering, and the calculation formula is as follows:

Where G(x, y) is a two-dimensional Gaussian function, (x, y) is the point coordinate, σ is the standard deviation, and A is the normalization coefficient, so that the sum of different weights is one.

(1.2) Binarization processing:

The fixed threshold method is used to binarize the image (as shown in Figure 1c). The calculation formula is as follows:

Where f(x, y) represents the distribution function of image pixel values, g(x, y) represents the pixel value distribution function after threshold segmentation, and the fixed threshold T=145;

(2) Extract the target contour by using the polygon approximation method, which specifically includes:

(2.1) Edge detection on the chip pin image:

Using Canny operator edge detection, the edge contour information of the chip pin image is obtained, as shown in Figure 2a;

(2.2) Use the polygon approximation method to extract the target contour, that is, the center pad contour of the chip pin image:

The contour of the center pad of the chip pin image is extracted by the polygon approximation method (as shown in Figure 2b), and the remaining contour parts are filtered out, which simplifies the subsequent image processing. The polygon approximation method is to pick out the two farthest points from the target contour and connect them; then find a point with the furthest distance from the off-line segment from the target contour, and add this point to the new contour after approximation, that is, connect the three The triangle formed by the points is used as the contour; finally, select any side of the triangle to start, repeat the previous step, add the farthest point to the new contour, and iterate continuously until the output accuracy requirements are met.

(3) An improved Harris corner detection algorithm is proposed to obtain the target contour vertices, including:

(3.1) Determine the selection threshold and obtain the corner points of the target contour:

Under normal circumstances, when the difference between the pixel gray levels of two black and white images is less than 10% to 15% of the maximum pixel gray value, it is difficult for the human eye to distinguish, so the threshold N is selected to extract the corners of the target contour of the image (as shown in Figure 3). ), the formula is as follows:

N=255×12%≈30

(3.2) Extract the corner points of the target contour:

Extract and save all the corner points, read the three points _Man , M _a , and M _a+n in the corner points in sequence, the three points constitute a template consisting of three elements, among which the three points take M _a as the template center , the subscript of the point represents the sequence number of the corner point among all the corner points. Traverse all the corner points, starting from the initial value, M _a is determined as a real-time operation point, take the two points whose front and rear serial numbers are separated by n, M _a and the other two points in the template form two sides respectively, and the two sides constituted by The included angle is taken as the corner response value of point _Ma . The side L ₁ is determined by the distance between the point Ma and the point Man, the side L ₂ is determined by the distance between the point _Ma and the point M _{a +n ,} and the side L ₃ is determined by the distance between the point _Man and the point M _a +n. The law of cosines obtains the angle of the point M _a , and the calculation formula is as follows:

The angle formed by the two sides, that is, the response value of the corner point, is used to determine whether to retain the corner point. Let g be the response value of the corner point, if the point Ma ≥ _g , save it as the required corner point; otherwise, remove it.

(3.3) Eliminate adjacent corners and retain contour vertices:

After the contour corners are extracted, there may be other corners around. To eliminate this phenomenon, the adjacent corners are removed, and the remaining points are taken as contour corner vertices. Let the image height be H, the image width be W, Corner(x, y) indicates whether there is a corner at the image (x, y), and when Corner(x, y) = 1, there is a corner at (x, y) point, m×m (m>1) is the size of the matrix centered at (x, y), then:

Where m≤x≤H, m≤y≤W, count refers to the corner points of the matrix range centered on (x, y), remove all adjacent corner points of the matrix range centered on (x, y), and keep the target The remaining corner points at the corners of the contour are used as vertices (as shown in Figure 4) to facilitate subsequent image correction processing.

(4) Fitting a straight line by the least squares method:

Using the least squares method, the two vertices of the longest side of the target contour are fitted with a straight line, as the angle recognition direction of the chip pin image, as shown in Figure 5.

(5) Quickly correct the chip and remove the white edge according to the image centroid, including:

(5.1) Taking the image centroid as the rotation center:

In order to correct the image accurately, the centroid method is used to determine the center position of the image, and the centroid coordinate O point is used as the rotation center of the image.

(5.2) Obtain the image tilt angle and correct the chip pin image:

When the QFN chip is packaged into the carrier tape, there is an angle deviation that is difficult to distinguish with the naked eye. In order to calculate the deviation angle, an improved Harris corner detection algorithm is proposed. Combined with the polygon approximation outline, the least squares method is used to fit the two vertices of the longest side. , at this time, the chip tilt angle α and the chip offset value x in the horizontal direction have a right triangle relationship. As shown in Figure 6, the chip is corrected according to the tilt angle α (as shown in Figure 7a), and the rotated corrected chip pin image exists The white edge is removed (as shown in Figure 7b).

In the step (4), set the coordinates of the two vertices to be p(x, y) and q(x, y) respectively, and the least squares fitting straight line calculation formula is as follows:

y=ax+b

a and b are the slope and intercept of the straight line equation, respectively, then:

分别为顶点p、q的横坐标与纵坐标的均值，

in

are the mean values of the abscissa and ordinate of vertices p and q, respectively,

In the step (5.1), the upper left corner of the image is set as the starting point coordinates (0, 0), the lower right corner is set as the end point coordinates (m, n), and the image centroid formula is as follows:

Among them, (x ₀ , y ₀ ) are the coordinates of the centroid, m and n are the number of rows and columns of the image respectively (m, n are both integers greater than or equal to 2), f(x, y) is the image at the point ( gray value at x, y).

In the step (5.2), the tilt angle α of the chip image can be calculated by the following formula:

The offset value x of the pin image in the horizontal direction can be obtained:

x=|p _x -h _x |

Then the pin image tilt angle α is:

Among them, l is the vertical distance from point p to point O, OA is the extension of the centroid coordinate to the positive direction of the x-axis, and h(x, y) is the coordinate of the intersection point of pq and OA.

Figure 13 is a flow chart of the method for fast tilt correction of the pin image of the QFN chip. The experimental verification and comparison of the above method of fast tilt correction of the pin image of the QFN chip are as follows:

(1) The correction method disclosed in this embodiment is compared with the correction accuracy of the traditional Hough transform and the least second moment method:

In this implementation manner, the memory is 4GB, the processor is AMD A10-7300RadeonR6, the operating system is 10ComputeCores4C+6G@1.9GHz, and the Visual Studio version is 2013. Select 10 different QFN chip pin images as experimental objects, run the program in the same environment, and compare the correction method disclosed in this embodiment with the traditional Hough transform and the minimum external moment method. FIG. 8 , FIG. 9 , and FIG. 10 are schematic diagrams of correction labels after the angles are detected by the above three algorithms, respectively. The corrected center pad of the chip is used for angle detection and verification. The gray frame line is the minimum circumscribed rectangle of the pentagon part of the chip obtained by different algorithms, and the black frame line is the ideal circumscribed rectangle of the artificially marked pentagon part of the chip. It is not difficult to find that there is a certain angular deviation between FIG. 8 and FIG. 9 . The schematic diagram of the circumscribed rectangle marked by the correction method disclosed in this embodiment is shown in FIG. 10 . The marked gray frame line and the manually marked black frame line are nearly coincident. Therefore, the correction method disclosed in this embodiment can obtain a more accurate inclination angle. . Figure 11 lists the tilt angles of the chip detected by the three algorithms.

(2) The correction method disclosed in this embodiment is compared with the correction time of the traditional Hough transform and the least second moment method:

In order to detect the difference of the running time between the calibration method disclosed in this embodiment and the traditional Hough transform and the least external moment method, the running time of 10 different QFN chip pin images are experimentally compared. Figure 12 lists the running times of the three algorithms. Taking figure number 5 as an example, the traditional Hough transform correction time is 357ms, and the minimum circumscribed moment method correction time is 116ms, while the correction method disclosed in this embodiment only takes 18ms to complete the chip image correction process. The average time for Hough transform to correct 10 chip pin images is 412.6ms, and the average time for minimum external moment method to correct 10 chip pin images is 125.8ms. Compared with traditional Hough transform, the running time of the correction method disclosed in this embodiment is only Its 1/34, compared to the minimum external moment method, the running time of the calibration method disclosed in this embodiment is only 1/10 of that. Therefore, the fast tilt correction method for the QFN chip image proposed in this embodiment not only has high accuracy, but also greatly reduces the running time and has higher computational efficiency.

To sum up, a method for fast tilt correction of QFN chip pin images proposed in this embodiment is a faster and more efficient method than traditional algorithms, and can be used in the production of QFN chip correction links, which is the basis for the tilt of the chip. Correcting and detecting chip appearance defects provides clear and accurate images and improves the visual inspection efficiency of QFN package defects.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. However, the above descriptions are only specific embodiments of the present invention, and the technical features of the present invention are not limited thereto. within the scope of the invention patent.

Claims (4)

1. a method for quick tilt correction for QFN chip pin image, is characterized in that, comprises the following steps: (1) Preprocess the chip pin image collected by the industrial computer: 1.1) Image filtering: In order to remove noise and reduce image distortion, a 5×5 Gaussian filter is used to convolve the image. In image processing, a two-dimensional Gaussian function is used for filtering. The calculation formula is as follows:

Among them, G(x, y) is the two-dimensional Gaussian function, (x, y) is the point coordinate, σ is the standard deviation, and A is the normalization coefficient, so that the sum of different weights is one; 1.2) Binarization processing: The fixed threshold method is used to binarize the image, and the calculation formula is as follows:

Among them, f(x, y) represents the distribution function of image pixel values, g(x, y) represents the pixel value distribution function after threshold segmentation, and the fixed threshold T=145; (2) Extract the target contour by using the polygon approximation method, which specifically includes: 2.1) Edge detection on the chip pin image: Using Canny operator edge detection, the edge contour information of the chip pin image is obtained; 2.2) Use the polygon approximation method to extract the target contour: The contour of the center pad of the chip pin image is extracted by the polygon approximation method, and the remaining contour parts are filtered out. The polygon approximation method is to pick out the two farthest points from the target contour and connect them; then find an offline segment from the target contour. The farthest point is added to the new contour after approximation, that is, the triangle formed by connecting the three points is used as the contour; finally, any side of the triangle is selected to start, and the previous step is repeated to add the farthest point to the new contour. , and iterate continuously until the accuracy requirements of the output are met; (3) An improved Harris corner detection algorithm is proposed to obtain the target contour vertices, including: 3.1) Determine the selection threshold and obtain the corner points of the target contour: Under normal circumstances, when the difference between the pixel gray levels of two black and white images is less than 10% to 15% of the maximum pixel gray value, it is difficult for the human eye to distinguish. Therefore, the threshold N is selected to extract the corner points of the target contour of the image. The formula is as follows: N=255×12%≈30 3.2) Extract the corner points of the target contour: Extract and save all the corner points, read the three points _Man , M _a , and M _a+n in the corner points in sequence, the three points constitute a template consisting of three elements, among which the three points take M _a as the template center , the subscript of the point represents the sequence number of the corner point in all the corner points, traverse all the corner points, starting from the initial value, M _a is determined as a real-time operation point, take the two points whose front and rear numbers are separated by n, and M _a are respectively the same as The other two points in the template form two sides, the angle formed by the two sides is used as the corner response value of point Ma, and the distance between point _Ma and point _Man determines the edge L ₁ , point _Ma and point M _{a +} The distance of _n determines the side L ₂ , the distance between the point _Man and the point M _{a +n} determines the side L ₃ , and the angle of the point Ma can be obtained from the three sides according to the cosine law. The calculation formula is as follows:

The angle formed by the two sides, that is, the response value of the corner point, is used to determine whether to retain the corner point. Let g be the response value of the corner point. If the point M _a ≥ g, save it as the required corner point; otherwise, remove it; 3.3) Eliminate adjacent corners and retain contour vertices: After the contour corner points are extracted, there may be other corner points around. To eliminate this phenomenon, the adjacent corner points are removed, and the remaining points are taken as the contour corner vertices. Set the image height as H and the image width as W, Corner(x, y) indicates whether there is a corner point at the image (x, y), when Corner(x, y) = 1, there is a corner point at (x, y), m×m is based on (x, y) , y) is the size of the centered matrix, where m>1, then:

Among them, m≤x≤H, m≤y≤W, count refers to the corner points of the matrix range with (x, y) as the center, and all the adjacent corner points of the matrix range with (x, y) as the center are removed and reserved The remaining corner points at the corners of the target contour are used as vertices to facilitate subsequent image correction processing; (4) Fitting a straight line by the least squares method: Using the least squares method, the two vertices of the longest side of the target contour are fitted with a straight line as the angle identification direction of the chip pin image; (5) Quickly correct the chip and remove the white edge according to the image centroid, including: 5.1) Take the image centroid as the rotation center: In order to correct the image accurately, the centroid method is used to determine the center position of the image, and the centroid coordinate O point is used as the rotation center of the image; 5.2) Obtain the image tilt angle and correct the chip pin image. 2. a kind of method for quick tilt correction for QFN chip pin image as claimed in claim 1, is characterized in that, in step 4, set two vertex coordinates to be respectively p(x, y), q(x, y), the least squares fitting straight line calculation formula is as follows: y=ax+b a and b are the slope and intercept of the straight line equation, respectively, then:

in,

are the mean values of the abscissa and ordinate of vertices p and q, respectively,

3. a kind of method for fast tilt correction of QFN chip pin image as claimed in claim 2, is characterized in that, in step 5.1, the upper left corner of the image is set as the starting point coordinates (0, 0), and the lower right corner is set as The coordinates of the end point (m, n), the image centroid formula is as follows:

Among them, (x ₀ , y ₀ ) are the centroid coordinates, m and n are the number of rows and columns of the image, respectively, m and n are integers greater than or equal to 2, and f(x, y) is the image at point (x) , the gray value at y). 4. a kind of method for quick tilt correction for QFN chip pin image as claimed in claim 3, is characterized in that, in step 5.2, chip image tilt angle α can be obtained by following calculation formula:

The offset value x of the pin image in the horizontal direction can be obtained: x=|p _x -h _x | Then the pin image tilt angle α is:

Among them, l is the vertical distance from point p to point O, OA is the extension of the centroid coordinate to the positive direction of the x-axis, and h(x, y) is the coordinate of the intersection of pq and OA.

Images