A dynamic positioning method for chip wire bonding

Abstract

The present invention discloses a chip wire bonding dynamic positioning method, which pre-processes the chip image and marks the chip bonding points, generates the center point coordinates and the radius of each bonding point, and generates a bonding point data set based on the center point coordinates and radius of each bonding point; generates a chip template image; determines the chip center point based on the centroid of the template image; calculates the bonding point feature descriptor using an improved ORB algorithm; and integrates the KCF algorithm and the improved Medianflow algorithm to achieve chip dynamic positioning. The method of the present invention first improves the way the ORB algorithm calculates feature descriptors to generate feature data that fits the chip wire bonding scenario, then applies the KCF and Medianflow positioning and tracking algorithms to bonding point positioning, and uses the ORB feature extraction algorithm for improvement, effectively improving the efficiency and accuracy of the chip bonding process, and has a broad application market space and economic value.

Description

Technical Field

[0001] The invention relates to a chip wire bonding dynamic positioning method, belonging to the field of chip packaging intelligent manufacturing. Background Art

[0002] With the continuous evolution of technology, chip integration is increasing exponentially, posing unprecedented challenges to chip packaging technology. Wire bonding is a crucial step in chip packaging equipment, directly impacting chip operation and performance. Traditionally, wire bonding equipment relies primarily on improving the mechanical precision of the motion platform to ensure bonding quality. However, the complexity of modern chip packaging processes presents fundamental limitations to this approach. First, the bonding process involves dozens of precision steps, each of which introduces random errors. Second, these errors have a cumulative effect, propagating and amplifying to downstream processes along the process chain. Against this backdrop, real-time error detection and compensation technology based on machine vision presents unique advantages. By establishing a closed-loop control system based on "perception-decision-compensation," vision systems can dynamically correct for deviations in each process step, thus breaking through the technical bottleneck of relying solely on mechanical precision. However, the increasing complexity of chips also poses significant challenges to vision systems' detection and positioning, resulting in poor positioning accuracy and low efficiency. Summary of the Invention

[0003] In response to the existing background technology problems, the present invention proposes a dynamic positioning method for chip wire bonding. The method first improves the way the ORB algorithm calculates feature descriptors to generate feature data that fits the chip wire bonding scenario, then applies the KCF and Medianflow positioning and tracking algorithms to the bonding point positioning, and uses the ORB feature extraction algorithm for improvement, effectively improving the efficiency and accuracy of the chip bonding process, and has a broad application market space and economic value.

[0004] The technical solutions adopted in the present invention are:

[0005] A chip wire bonding dynamic positioning method, comprising:

[0006] 1) Generation of chip template images and feature data:

[0007] Preprocessing the chip image and marking the bonding points of the chip, generating the center point coordinates and the radius of each bonding point, and generating a bonding point data set based on the center point coordinates and the radius of each bonding point;

[0008] Generate a chip template image based on the pre-processed chip image and bonding point data (the center point coordinates and the radius of each bonding point);

[0009] Determining the center point of the chip according to the centroid of the template image;

[0010] 2) Calculate the bond point feature descriptor using the improved ORB feature extraction algorithm:

[0011] Selecting a bonding point from the bonding point data set, defining a circular area with the center point coordinates of the bonding point as the center, calculating the coordinates of the center of the circular area, and calculating the azimuth angle from the center point coordinates of the bonding point to the center of the circular area;

[0012] According to the image features of the chip bonding point, point pairs are selected from eight directions of the bonding point to form multiple feature point pairs, and the feature descriptors of these feature point pairs are calculated;

[0013] 3) Fusion of KCF algorithm and improved Medianflow algorithm to achieve dynamic chip positioning:

[0014] Determine the position of the chip to be bonded based on the KCF algorithm;

[0015] In the chip to be bonded position, the improved MedianFlow algorithm is used to search for bonding points, and the improved ORB feature extraction algorithm is used to calculate the feature descriptors of the searched bonding points;

[0016] The Hamming distance between the feature descriptor of the searched bond point and the feature descriptor of the corresponding bond point in the template image is calculated. If the Hamming distance is less than a preset threshold, the bond point is determined to be the target bond point. Otherwise, the search is repeated and the calculation is repeated until the target bond point is found.

[0017] Furthermore, the bonding points of the chip are marked, and the center coordinates and radius of each bonding point are generated. The data of each bonding point is expressed as:

[0018] ;

[0019] Where i is the index of the bonding point, 、 are the relative coordinates of each bonding point, is the radius of the bonding point;

[0020] generating a chip template image based on the preprocessed chip image and bonding point data;

[0021] The chip center point is determined by the centroid of the template image, which is calculated as follows:

[0022] ;

[0023] in, 、 are the horizontal and vertical coordinates of the centroid, 、 is the coordinate of the pixel point, is the grayscale value of the pixel, is the index of each pixel;

[0024] Based on the template image, chip center point and bonding point data, the position feature data of the bonding point is generated. The number of bonding points is set to , the generated location feature dataset is ,in is the center point of the template image.

[0025] Furthermore, the improved ORB feature extraction algorithm has the following improvements:

[0026] (1) Optimize the point pair selection method of the traditional ORB feature extraction algorithm:

[0027] According to the characteristics of the chip bonding point, point pairs are evenly selected in the eight directions of the bonding point, with 4 points selected in each direction, forming 4 layers of feature points in 8 directions;

[0028] (2) Optimize the feature descriptor generation process of the traditional ORB feature extraction algorithm:

[0029] When generating feature descriptors, a screening step for the average value of point pairs is added. After removing the 16 point pairs with the smallest average value, 256 point pairs with relatively obvious features are obtained for generating feature descriptors.

[0030] Furthermore, the improved ORB feature extraction algorithm calculates the bond point feature descriptor, and the process is as follows:

[0031] Selecting a bonding point from the bonding point data set, defining a circular area with the center point coordinates of the bonding point as the center, calculating the coordinates of the center of the circular area, and calculating the azimuth angle from the center point coordinates of the bonding point to the center of the circular area;

[0032] According to the image features of the chip bonding points, point pairs are selected from eight directions of the bonding points to form multiple feature point pairs, and the feature descriptors of these feature point pairs are calculated.

[0033] Further, a bonding point is selected from the bonding point data set , with the bonding point as the center and the radius as Delineate a circular area;

[0034] Calculate the image moment m within the circular area, and the calculation formula is:

[0035] ;

[0036] Among them, x and y are the coordinates of the pixel point, p is the gray value of the pixel point, ;

[0037] Calculate the centroid coordinates C of the circular area using the following formula:

[0038] ;

[0039] Calculate bonding points The azimuth angle θ to the center of mass is calculated as:

[0040] ;

[0041] According to the image features of the chip bonding point, point pairs are selected from eight directions of the bonding point, with 4 points selected in each direction, forming 4 layers of feature points in 8 directions;

[0042] Calculate the pixel average for each point pair, remove the 16 point pairs with the smallest average values, and obtain 256 point pairs with obvious features;

[0043] The coordinates of each point are rotated by an angle of θ to correct the direction. The coordinate transformation formula is:

[0044] ;

[0045] Compare the pixel intensity of each pair of points from top to bottom in a clockwise direction, starting from just above the bonding point. The comparison function is used to calculate the comparison result ψ, and its calculation formula is:

[0046] ;

[0047] Where I represents the gray value of the pixel;

[0048] The comparison results of all point pairs are concatenated to obtain the feature descriptors of the bonding points.

[0049] Furthermore, the improved Medianflow algorithm has the following improvements:

[0050] (1) The feature descriptor generated by the improved ORB feature extraction algorithm is used to replace the feature extractor in the traditional Medianflow algorithm;

[0051] (2) Before precise positioning, the KCF algorithm is used to determine the position of the chip to be bonded.

[0052] Furthermore, the KCF algorithm and the improved Medianflow algorithm are integrated to realize dynamic chip positioning. The process is as follows:

[0053] Determine the position of the chip to be bonded based on the KCF algorithm;

[0054] Within the determined position of the chip to be bonded, the improved MedianFlow algorithm is used to roughly locate the bonding point, and the improved ORB feature extraction algorithm is used to calculate the feature descriptor of the roughly located bonding point;

[0055] The Hamming distance between the calculated feature descriptor of the bonding point and the feature descriptor of the corresponding bonding point in the template image,

[0056] If the Hamming distance is less than the average Hamming distance of all points, the currently located bond point is considered to be the target bond point; otherwise, a point is randomly selected from the located bond point feature point pair as the new bond point to be located, and the feature descriptor is recalculated until the Hamming distance between the new bond point descriptor and the template bond point descriptor is less than the average Hamming distance.

[0057] Furthermore, let the descriptors of the bonding points in the template image and the coarsely located bonding points be F1 and F2 respectively, both of which are binary strings. The Hamming distance D is calculated as:

[0058] .

[0059] The present invention has the following beneficial effects:

[0060] The method of the present invention first improves the way in which the ORB algorithm calculates feature descriptors to generate feature data that fits the chip wire bonding scenario, then applies the KCF and Medianflow positioning and tracking algorithms to bond point positioning, and uses the ORB feature extraction algorithm for improvement, effectively improving the efficiency and accuracy of the chip bonding process, and has a broad application market space and economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of chip template image and bonding point marking.

[0062] Figure 2 Schematic diagram for selecting bonding point neighborhood pairs.

[0063] Figure 3 Schematic diagram of the order of comparing pixel intensities for point pairs.

[0064] Figure 4 Flowchart for locating chip bonding points. DETAILED DESCRIPTION

[0065] The present invention will be further described below with reference to the accompanying drawings.

[0066] Chip positioning accuracy has a crucial impact on wire bonding quality. In traditional wire bonding vision systems, chip positioning involves image matching between a pre-stored template and the chip to be positioned to determine the relative position between the bonding mechanism and the chip. However, this method only considers the position of the chip; the positioning of the bond points is performed based on alignment rules, which still results in a certain degree of error. Due to the complex structure of the chip, with its numerous circuits, pins, and other components, it is easy for the bond points to be confused with other components, thus affecting the positioning of the chip and the bond points.

[0067] The present invention improves the chip positioning method based on the ORB feature detection algorithm. It first completes the local positioning of the chip and then generates a feature description of the chip bonding point, providing a basis for the subsequent precise positioning of the bonding point. The specific steps include the following.

[0068] Step 1: Generation of chip bonding point template image and feature data.

[0069] First, the image is preprocessed. The image preprocessing mainly includes Gaussian filtering, binarization, and edge extraction operations. At the same time, the image preprocessing in the present invention also includes a bonding point positioning operation. By marking the bonding points of the chip, the center point coordinates and the radius of each bonding point are generated. Figure 1 As shown, the data of each point is represented by : ;

[0070] in is the index of each bond point, 、 are the relative coordinates of each bonding point, is the radius of the bonding point.

[0071] A chip template image is generated based on the pre-processed chip image and the bonding point data.

[0072] The coordinates of the bonding point in the previous step are relative to the center of the chip. The center of the chip is determined by the centroid of the template image and is calculated as follows:

[0073] ;

[0074] in 、 are the horizontal and vertical coordinates of the centroid, 、 is the coordinate of the pixel point, is the pixel gray value of the point, The index of each pixel.

[0075] Through the above steps, we can finally get the position characteristic data of the bonding points. Let the number of bonding points be , the generated location feature dataset is ,but The CCP has The data of the points are as follows, where is the center point of the chip template image.

[0076] ;

[0077] Step 2: Calculate the bond point feature descriptor using the improved ORB feature extraction algorithm.

[0078] The improvements of the improved ORB feature extraction algorithm are mainly reflected in the following aspects:

[0079] Improvement 1: Optimization of point selection method

[0080] Traditional ORB algorithm: When generating feature descriptors, point pairs are randomly selected within a certain range of neighborhood around the bonding point.

[0081] The improved ORB algorithm selects evenly spaced point pairs along eight directions of the bond point based on the characteristics of the chip bond point. Four points are selected in each direction, forming four layers of feature points in eight directions. The angle between two adjacent directions is 45°, which improves the stability and robustness of the feature descriptor at different rotation angles.

[0082] Improvement 2: Optimization of feature descriptor generation process

[0083] Traditional ORB algorithm: randomly select a certain number of point pairs to generate feature descriptors.

[0084] Improved ORB algorithm: When generating feature descriptors, a step is added to filter the pixel averages of each point pair. Specifically, the pixel average of each point pair is calculated, and the 16 pairs with the smallest averages are removed. The remaining 256 pairs with distinct features are used to generate feature descriptors. This makes the feature descriptors more discriminative and can more accurately describe the characteristics of the chip bond points.

[0085] The process of calculating the bond point feature descriptor using the improved ORB feature extraction algorithm is as follows:

[0086] First, select a bonding point from the bonding point coordinate set in step 1 (except ), set as , with the bonding point as the center and the radius as Delimit a circular area and first calculate the image moment in the area , the calculation method is as follows:

[0087] ;

[0088] in, 、 is the coordinate of the pixel point, is the pixel gray value of the point, , then calculate the coordinates of the region centroid .

[0089] ;

[0090] In order to ensure the rotation invariance of the bond point feature description, it is necessary to calculate the bond point Azimuth to the center of mass , which is calculated as follows:

[0091] .

[0092] To generate bond point descriptors, the traditional ORB algorithm randomly selects a certain number of point pairs (e.g., 128, 256) within a certain range of neighborhood around the bond point (e.g., 31×31).

[0093] In the present invention, based on the characteristics of the chip bonding point, since the bonding point is usually the pin area of ​​the chip, the pixel grayscale values ​​in this area are usually similar within a certain range. Therefore, the point selection method is changed. In a larger area, 4 points are selected from each of the eight directions of the bonding point to form a total of 4 layers of feature points in 8 directions, such as Figure 2 As shown in the figure, 4 points in each direction form pairs, totaling 6 pairs. A total of 48 pairs are generated across all 8 directions. Furthermore, feature points in each layer form pairs across all 8 directions, totaling 224 pairs. The pixel average is calculated for each point pair. After removing the 16 pairs with the smallest averages, 256 pairs of points with distinct features are obtained, and their feature descriptors are calculated.

[0094] Then rotate the coordinates of each point The angle is used to correct the direction. The calculation of the rotation angle is shown in the following formula, where 、 are the coordinates of the points after selection.

[0095] ;

[0096] Compare the pixel intensities of each pair of points from top to bottom in a clockwise direction, starting from just above the bonding point, such as Figure 3 shown.

[0097] Let a point pair be , use the function shown in the following formula to compare, and the comparison result is set as ,in Represents the pixel grayscale value of the bonding point.

[0098] ;

[0099] The comparison results of all point pairs are concatenated to obtain the bond point descriptor.

[0100] Step 3: Integrate the KCF algorithm and the improved Medianflow algorithm to achieve dynamic chip positioning.

[0101] During the chip bonding process, the bonding mechanism needs to constantly change position. Due to the complexity of various circuits and components on the chip, it is easy to lose the target. The present invention combines the MedianFlow tracking algorithm and the ORB feature detection algorithm to achieve precise positioning during the chip bonding process.

[0102] 31: Chip positioning based on KCF algorithm.

[0103] The core principle of the KCF algorithm is to solve a target template with the maximum response value. It introduces a circulant matrix and uses the data obtained by cyclic shift of the target area to approximate the simulation of the training sample set.

[0104] One-dimensional vector For example, suppose , the circulant matrix As shown in the following formula.

[0105] ;

[0106] right The cyclically shifted vector sequence can be obtained by left-multiplying the cyclic shift matrix in sequence.

[0107] ;

[0108] Finally, we get:

[0109] ;

[0110] The KCF algorithm uses the least squares method to train the classifier. The original problem can be expressed as minimizing the sum of squared errors (ridge regression):

[0111] ;

[0112] In the above formula, is the training sample, for The corresponding target value, is the target template, for With target template The dot product in the frequency domain, is the regularization coefficient to prevent the target template from overfitting, is the index of the training data. Overall use To express, Overall use To express it, the above formula can be expressed as:

[0113] ;

[0114] right Taking partial derivatives we get:

[0115] ;

[0116] in is the identity matrix. Transforming the above formula to the complex domain yields:

[0117] ;

[0118] in for The complex conjugate of .

[0119] Calculate according to the corresponding formula , but the amount of direct calculation is large. KCF can greatly speed up the calculation by introducing the circulant matrix. Diagonalization:

[0120] ;

[0121] in is the initial sample The value obtained after Fourier transformation is is the constant discrete Fourier matrix, yes The complex conjugate matrix of is a constant, is the natural logarithm constant, Is an imaginary unit.

[0122] ;

[0123] The product of the circulant matrix and the vector is equivalent to the convolution of the inverse of the generating vector and the vector, which can be further transformed into the Fourier space multiplication, and finally The calculation is transformed into the following formula, where Represents dot product.

[0124] ;

[0125] Solve at this time The algorithm complexity is ,The introduction of circulant matrix not only expands the number of samples and improves ,the accuracy, but also simplifies the calculation and increases the speed of ,chip dynamic positioning.

[0126] Step 22: Chip bonding point positioning based on Medianflow.

[0127] The MedianFlow target positioning algorithm is based on the optical flow tracking algorithm and uses forward and backward evaluation point tracking, NCC cross-validation feedback, and other methods to achieve high accuracy in tracking and positioning dynamic targets. Traditional MedianFlow algorithms first detect the target area, extract feature points, and then forward track these feature points. Backward tracking is also used to verify the accuracy of tracking and positioning, thereby adjusting the positioning accuracy.

[0128] In the present invention, the feature extractor in the traditional Medianflow algorithm is improved. First, the rough positioning of the bonding point is performed. According to the chip bonding point set obtained in step 1, the target area is searched to complete the positioning of the bonding point. The target area is located by the KCF algorithm.

[0129] After completing the coarse positioning, the feature point selection method in the improved ORB feature extraction algorithm of the present invention is used to calculate the descriptor of the coarsely positioned bond points using the same method. Let the descriptors of the bond points in the template image and the coarsely positioned bond points be and , both descriptors are binary strings, so the Hamming distance between them is calculated using the following formula, where Indicates bitwise comparison, It represents the Hamming distance, and the formula is:

[0130] .

[0131] When the Hamming distance When the distance between the bond point and the template bond point descriptor is less than the average Hamming distance of all points, the currently located bond point is considered to be the target bond point. Otherwise, a point is randomly selected from the located bond point feature point pair as the new bond point to be located, and the feature descriptor is recalculated until the Hamming distance between the new bond point descriptor and the template bond point descriptor is less than the average Hamming distance. The flowchart of chip bond point positioning is as follows: Figure 4 shown.

[0132] The above description is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be regarded as the scope of protection of the present invention.

Claims

1. A chip wire bonding dynamic positioning method, characterized by: include: Preprocess the chip image and mark the bonding points of the chip, generate the center point coordinates and the radius of each bonding point, and generate a bonding point data set based on the center point coordinates and radius of each bonding point; Generate chip template image; Determining the center point of the chip according to the centroid of the template image; The improved ORB algorithm is used to calculate the bond point feature descriptor; Fusion of KCF algorithm and improved Medianflow algorithm to achieve dynamic chip positioning; The improvements of the improved ORB feature extraction algorithm are: (1) Optimize the point pair selection method of the traditional ORB feature extraction algorithm: According to the characteristics of the chip bonding point, point pairs are evenly selected in the eight directions of the bonding point, with 4 points selected in each direction, forming 4 layers of feature points in 8 directions; (2) Optimize the feature descriptor generation process of the traditional ORB feature extraction algorithm: When generating feature descriptors, a step of filtering the average value of point pairs is added. After removing the 16 point pairs with the smallest average value, 256 point pairs with relatively obvious features are obtained for generating feature descriptors. The improvements of the improved Medianflow algorithm are: (1) The feature descriptor generated by the improved ORB feature extraction algorithm is used to replace the feature extractor in the traditional Medianflow algorithm; (2) Before precise positioning, the KCF algorithm is used to determine the position of the chip to be bonded; The KCF algorithm and the improved Medianflow algorithm are integrated to realize dynamic chip positioning. The process is as follows: Determine the position of the chip to be bonded based on the KCF algorithm; Within the determined position of the chip to be bonded, the improved MedianFlow algorithm is used to roughly locate the bonding point, and the improved ORB feature extraction algorithm is used to calculate the feature descriptor of the roughly located bonding point; The Hamming distance between the calculated feature descriptor of the bonding point and the feature descriptor of the corresponding bonding point in the template image, If the Hamming distance is less than the average Hamming distance of all points, the currently located bond point is considered to be the target bond point; otherwise, a point is randomly selected from the located bond point feature point pair as the new bond point to be located, and the feature descriptor is recalculated until the Hamming distance between the new bond point descriptor and the template bond point descriptor is less than the average Hamming distance; The improved ORB feature extraction algorithm calculates the bond point feature descriptor. The process is as follows: Selecting a bonding point from the bonding point data set, defining a circular area with the center point coordinates of the bonding point as the center, calculating the coordinates of the center of the circular area, and calculating the azimuth angle from the center point coordinates of the bonding point to the center of the circular area; According to the image features of the chip bonding point, point pairs are selected from eight directions of the bonding point to form multiple feature point pairs, and the feature descriptors of these feature point pairs are calculated; Select a bond point from the bond point data set , with the bonding point as the center and the radius as Delineate a circular area; Calculate the image moment m within the circular area, and the calculation formula is: ； Among them, x and y are the coordinates of the pixel point, p is the gray value of the pixel point, ; Calculate the centroid coordinates C of the circular area using the following formula: ； Calculate bonding points The azimuth angle θ to the center of mass is calculated as: ； According to the image features of the chip bonding point, point pairs are selected from eight directions of the bonding point, with 4 points selected in each direction, forming 4 layers of feature points in 8 directions; Calculate the pixel average for each point pair, remove the 16 point pairs with the smallest average values, and obtain 256 point pairs with obvious features; The coordinates of each point are rotated by an angle of θ to correct the direction. The coordinate transformation formula is: ； Compare the pixel intensity of each pair of points from top to bottom in a clockwise direction, starting from just above the bonding point. The comparison function is used to calculate the comparison result ψ, and its calculation formula is: ； Where I represents the gray value of the pixel; The comparison results of all point pairs are concatenated to obtain the feature descriptors of the bonding points.

2. The chip wire bonding dynamic positioning method according to claim 1, wherein: Mark the bonding points of the chip and generate the center coordinates and radius of each bonding point. The data of each bonding point is expressed as: ； Where i is the index of the bonding point, 、 are the relative coordinates of each bonding point, is the radius of the bonding point; generating a chip template image based on the preprocessed chip image and bonding point data; The chip center point is determined by the centroid of the template image, which is calculated as follows: ； in, 、 are the horizontal and vertical coordinates of the centroid, 、 is the coordinate of the pixel point, is the grayscale value of the pixel, is the index of each pixel; Based on the template image, chip center point and bonding point data, the position feature data of the bonding point is generated. The number of bonding points is set to , the generated location feature dataset is ,in is the center point of the template image.

3. The chip wire bonding dynamic positioning method according to claim 1, wherein: Assume that the descriptors of the bonding points in the template image and the coarsely located bonding points are F1 and F2 respectively, both of which are binary strings. The Hamming distance D is calculated as: 。

Images