Method for detecting small defects of high-density integrated circuit packaging substrate

Abstract

The invention discloses a method for detecting tiny defects of a high-density integrated circuit packaging substrate. The method comprises the following steps: acquiring image data of the surface of the high-density integrated circuit packaging substrate; processing the acquired images in sequence by adopting a mean filtering method, morphological dilation and a texture operator; processing the image by adopting an image binarization processing technology to obtain a binarized image; dividing the binary image into a plurality of energy regions, and calculating an entropy difference through the divided energy regions; performing processing operation on the image data through a discrete cosine transform technology to obtain a DCT coefficient, dividing the DCT coefficient into a high-frequency component and a low-frequency component, and performing calculation to obtain a ratio of a low-frequency coefficient mean value to a high-frequency coefficient mean value; the feature extraction model extracts defect shape features based on the entropy difference and the ratio; comparing a similarity measurement index calculated according to the defect shape characteristics of the packaging substrate with a preset threshold value, and judging whether the packaging substrate has a shape defect or not; and precise detection of small defects of the high-density integrated circuit packaging substrate is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a detection method for micro defects of high-density integrated circuit packaging substrates. BACKGROUND

[0002] In existing research, a high-density flexible packaging substrate defect detection method based on the CRS-YOLO algorithm optimizes the target detection algorithm to achieve fast and accurate detection of packaging substrate defects. However, due to the CRS-YOLO algorithm, some defects may be misjudged or missed when dealing with complex backgrounds and various defect types, reducing the accuracy of the detection results. In low-resolution ceramic substrate image defect detection based on knowledge distillation, knowledge distillation technology is used to detect low-resolution ceramic substrate image defects. However, knowledge distillation technology relies on high-quality training data and teacher models, and when the training data is insufficient or the teacher model performance is poor, the accuracy of the detection results may be affected. The PCBA on-board DDR chip solder joint defect detection research combines image processing technology and machine learning algorithms to accurately identify the defect types of solder joints. However, this method mainly targets specific types of defects, and the detection accuracy for other types of defects needs to be improved. In the ceramic substrate defect detection based on improved YOLOV4, the algorithm structure is improved and the parameter settings are optimized to improve the detection speed and accuracy. However, this method may be affected by factors such as light changes and noise interference, resulting in inaccurate detection results. SUMMARY

[0003] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a detection method for micro defects of high-density integrated circuit packaging substrates, which improves detection accuracy and sensitivity by in-depth analysis of micro defect characteristics and the use of advanced algorithms and technical means to ensure the accuracy and reliability of the results.

[0004] Technical scheme: To achieve the above purpose, the detection method for micro defects of high-density integrated circuit packaging substrates of the present application comprises the following steps:

[0005] Step one, obtain image data of the surface of the high-density integrated circuit packaging substrate;

[0006] Step two, sequentially process the obtained image using mean filtering, morphological dilation, and texture operators;

[0007] Step three, use image binarization processing technology to process the processed image again to obtain a binary image; divide the binary image into several energy regions, and calculate the entropy difference through the divided energy regions;

[0008] Step four, the image data is processed by discrete cosine transform technology to obtain DCT coefficients, the DCT coefficients are divided into high frequency components and low frequency components, and the ratio of the low frequency coefficient mean and the high frequency coefficient mean is calculated;

[0009] Step five, the feature extraction model extracts the shape features of the packaging substrate defects based on the difference between the feature entropy under different energy suppression ratios and the ratio of the low frequency coefficient mean and the high frequency coefficient mean;

[0010] Step six, according to the shape features of the packaging substrate defects, the similarity metric is calculated, the calculated similarity metric is compared with the preset threshold, and it is judged whether the packaging substrate has micro shape defects.

[0011] Further, in the step two, the mean filter method is used to smooth the image data, and the calculation process is as follows:

[0012]

[0013] In the formula, f'(x, y) represents the pixel value of the filtered image at the position (x, y), f(i, j) represents the pixel value of the original image at the point (i, j), W represents the window centered at (x, y), and M represents the total number of pixel points in the window.

[0014] Further, in the step two, the image after mean filtering is processed by morphological dilation, and the morphological dilation operation increases the image brightness by selecting the maximum value in the field defined by the structure element; the calculation process is as follows:

[0015]

[0016] In the formula, represents the pixel value of the morphologically dilated image at the position (x, y), f'(x+s, y+t) represents the pixel value of the filtered image at the point (x+s, y+t), b represents the structure element, and max represents the maximum value operation, represents the addition operation.

[0017] Further, in the step two, a texture operator is designed, which connects the center pixel and the neighborhood pixel in the image by a specific encoding method; the calculation process is as follows:

[0018]

[0019] In the formula, T(x, y) represents the output of the texture operator at the point (x, y), N represents the field centered at (x, y), represents a specific encoding method, and s(·) represents a sign function.

[0020] Further, in the step three, the image processed by the mean filter method, morphological dilation and texture operator is processed by the binarization processing technology to separate the foreground and background of the image; the pixel gray value of each pixel in the image is converted to obtain a binary image, and the pixel gray value conversion calculation process of the image is as follows:

[0021]

[0022] In the formula, g(x, y) represents the pixel gray value after the binarization processing, and t represents the gray threshold value.

[0023] Further, the binary image is divided into a plurality of energy regions, the energy of any one energy region is defined as Ei, and the total energy of the binary image is Esum; the feature entropy is calculated based on the energy value of the energy region to calculate the entropy difference, and the entropy difference reflects the change of the feature entropy under different energy suppression ratios; the calculation process of the entropy difference AH is as follows:

[0024]

[0025] In the formula, Q represents the number of energy regions of the binary image.

[0026] Further, in the step four, the discrete cosine transform (DCT) technology is introduced to perform the blocking, DCT transform, quantization and encoding operation on the image processed by the mean filter method, morphological dilation and texture operator to obtain the DCT coefficient of the image; the DCT coefficient is divided into low-frequency coefficient and high-frequency coefficient; the low-frequency coefficient mean value and the high-frequency coefficient mean value are calculated respectively, and the ratio of the low-frequency coefficient mean value and the high-frequency coefficient mean value is calculated; the calculation process is as follows:

[0027]

[0028] In the formula, H and L are the number of rows and columns of the rectangular block respectively, and C(u, v) represents the coefficient after the DCT transform.

[0029] Further, in the step six, a similarity measure index is constructed, the dot product sum between the shape feature of the template image and the direction vector of each corresponding shape feature point in the to-be-detected image is calculated, and the direction vector is normalized; the calculation process of the normalized similarity measure index is as follows:

[0030]

[0031] In the formula, R is the normalized similarity measure index, D is the total number of corresponding shape feature points, and respectively represent the direction vector of the template image shape feature and the direction vector of the i-th shape feature corresponding point in the to-be-detected image, and · is the dot product operation. and respectively represent the length of the corresponding vector.

[0032] Further, the calculated similarity metric indicator is compared with the preset threshold RT, to obtain the comparison result of the similarity metric indicator R and the preset threshold RT; when R>RT, it is determined that the packaging substrate has a micro shape defect; when R<RT, it is determined that the packaging substrate does not have a micro shape defect.

[0033] Further, when R=RT, the image obtained in step two needs to be divided into several sub-images, and steps three to six are repeated; when the comparison result obtained after the operation of steps three to six on any one sub-image or multiple sub-images in the several sub-images is R>RT, it is determined that the packaging substrate has a micro shape defect at this time.

[0034] Further, when the calculated similarity metric indicator R=RT, the image processed by the mean filtering method, morphological dilation and texture operator needs to be uniformly divided into four first sub-images, and steps three to six are performed on the four first sub-images respectively to obtain the comparison result of the second similarity metric indicator R and the preset threshold RT; if there is still R=RT in the second obtained comparison result, the first sub-image corresponding to this comparison result is again uniformly divided into four second sub-images, and steps three to six are performed on the four second sub-images respectively to obtain the comparison result of the third similarity metric indicator R and the preset threshold RT; in this way, until there is no R=RT in the comparison result of the similarity metric indicator R and the preset threshold RT.

[0035] Further, the loss function is integrated with the Focal Loss strategy, and by dynamically adjusting the sample weight, the model is more focused on difficult classification samples, and the calculation process is as follows:

[0036]

[0037] In the formula, FL adj represents the adjusted Focal Loss function; β t is a coefficient for balancing the weights of positive and negative samples, and δ represents an adjustment factor, and q j represents the prediction probability of the image sample.

[0038] Beneficial effects: the detection method of the micro defect of the high-density integrated circuit packaging substrate of the application, including mean filtering to suppress Gaussian noise, morphological dilation to highlight micro defects, and enhanced image adaptability by designing texture operators. In the feature extraction stage, the foreground and background are separated by binary processing, the shape feature is extracted by energy zone division and feature entropy calculation, and the detail feature is extracted by DCT, which provides key information for defect detection. In the defect detection link, the standard template area is constructed, the shape template matching method is adopted, the external interference is effectively dealt with, the Focal Loss strategy is integrated to improve the model performance, and the accurate detection of the micro defect of the high-density integrated circuit packaging substrate is realized. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flow chart of the detection method of the micro defect of the high-density integrated circuit packaging substrate is shown in the figure.

[0040] Figure 2 The comparison result graph of the positioning accuracy of the micro defect of the packaging substrate by four methods is shown in the figure.

[0041] Figure 3 The comparison result graph of the detection sensitivity of the micro defect of the packaging substrate by four methods is shown in the figure. DETAILED DESCRIPTION

[0042] The application will be further described below in combination with the drawings.

[0043] As shown in the figure, a detection method of the micro defect of the high-density integrated circuit packaging substrate comprises the following steps: Figure 1

[0044] Step one, in order to better detect the micro defect of the high-density integrated circuit packaging substrate, the image data of the surface of the high-density integrated circuit packaging substrate is obtained by using a precise appearance intelligent detector experimental platform;

[0045] Step two, since the collected image data of the high-density integrated circuit packaging substrate contains a large amount of noise and redundant information, the collected image data is first subjected to a series of systematic processing; that is, the mean filtering method, morphological dilation and texture operator are sequentially used to process the obtained image;

[0046] Step three, the processed image is processed again by using the image binarization processing technology to obtain a binary image; the binary image is divided into a plurality of energy zones, and the entropy difference is calculated through the divided energy zones;

[0047] Step four, the image data is processed by using the discrete cosine transform technology to obtain DCT coefficients, the DCT coefficients are divided into high-frequency components and low-frequency components, and the ratio of the low-frequency coefficient mean value and the high-frequency coefficient mean value is calculated; ​

[0048] Step five, the feature extraction model extracts the shape feature of the defect of the packaging substrate based on the difference of the feature entropy under different energy suppression ratios and the ratio of the mean of the low frequency coefficient and the mean of the high frequency coefficient;

[0049] Step six, calculating the similarity metric index according to the shape feature of the defect of the packaging substrate, comparing the calculated similarity metric index with the preset threshold, and determining whether the packaging substrate has a micro shape defect.

[0050] In the step two, when processing the image, in order to eliminate the possible Gaussian noise interference, the mean filtering method is used to smooth the image data, and the calculation process is as follows:

[0051]

[0052] In the formula, f'(x, y) represents the pixel value of the filtered image at the position (x, y), f(i, j) represents the pixel value of the original image at the point (i, j), W represents the window centered at (x, y), and M represents the total number of pixel points in the window. Through mean filtering, the Gaussian noise in the high-density integrated circuit packaging substrate image can be effectively suppressed.

[0053] In the step two, after the filtering process, in order to further highlight the micro defect, the image is subjected to morphological processing; the morphological dilation operation increases the image brightness by selecting the maximum value in the field defined by the structure element, and the image after mean filtering is subjected to morphological dilation processing; the morphological dilation operation increases the image brightness by selecting the maximum value in the field defined by the structure element, and reduces the black details; the calculation process is as follows:

[0054]

[0055] In the formula, represents the pixel value of the morphologically dilated image at the position (x, y), f'(x+s, y+t) represents the pixel value of the filtered image at the point (x+s, y+t), b represents the structure element, and max represents the maximum value operation, represents the addition operation. Through morphological dilation, the micro defect in the high-density integrated circuit packaging substrate image is further highlighted.

[0056] In the step two, a texture operator is designed, which connects the bit positions of the center pixel and the neighborhood pixels in the image through a specific coding method; the calculation process is as follows:

[0057]

[0058] In the formula, T(x, y) represents the output of the texture operator at point (x, y), N represents the field centered at (x, y), represents a specific encoding method, s(·) represents a sign function, is used to construct a ternary pattern, and a tolerance factor is introduced to enhance the robustness of the operator to light and dark changes. This series of processing steps constitutes a systematic strategy for processing high-density integrated circuit packaging substrate data, providing a clear and accurate image basis for subsequent effective extraction of defect features.

[0059] In step three, the image processed by the mean filtering method, morphological dilation and texture operator is processed by the binarization processing technology to separate the foreground and background of the image; the pixel gray value of the image is converted to obtain a binary image, and the pixel gray value conversion calculation process of the image is as follows:

[0060]

[0061] In the formula, g(x, y) represents the pixel gray value after binarization processing, and t represents the gray threshold.

[0062] In order to extract the shape features of high-density integrated circuit packaging substrates, the energy region division and feature entropy calculation method is introduced; the binary image is divided into several energy regions, the energy of any energy region is defined as Ei, and the total energy of the binary image is Esum, which can be obtained by summing the energy of all energy regions; the feature entropy is calculated based on the energy value of the energy region to calculate the entropy difference, and the feature entropy H1 and H2 of the image under two energy suppression ratios can be calculated based on these energy values, and the entropy difference reflects the change of the feature entropy under different energy suppression ratios; the calculation process of the entropy difference ΔH is as follows:

[0063]

[0064] In the formula, Q represents the number of energy regions of the binary image, y1 and y2 represent the images under two energy suppression ratios, respectively.

[0065] In step four, in order to further extract the detail features of the defects from the frequency domain, the discrete cosine transform (DCT) technology is introduced to perform block, DCT transform, quantization and encoding operations on the image processed by the mean filtering method, morphological dilation and texture operator to obtain the DCT coefficients of the image; the DCT coefficients are divided into low frequency coefficients and high frequency coefficients, which respectively reflect the general outline and detail gray change of the high-density integrated circuit packaging substrate image; the low frequency coefficient mean value and the high frequency coefficient mean value are calculated respectively, and the ratio of the low frequency coefficient mean value and the high frequency coefficient mean value, which can more comprehensively extract the detail features of the defects; the calculation process is as follows:

[0066]

[0067] In the formula, H and L are the number of rows and columns of the rectangular block respectively, and C(u, v) represents the coefficient after DCT transformation.

[0068] The feature extraction model is based on the difference between the feature entropy under two energy suppression ratios and the ratio of the calculated average of the frequency coefficients and the average of the high frequency coefficients; thus, clear foreground and background boundaries of the high-density integrated circuit packaging substrate and the details of the defect and other key features are extracted, and the shape features of the packaging substrate defects are mainly extracted.

[0069] In step six, in order to perform defect detection on the shape features, a set of standard template images is constructed, which are standard template images of the high-density integrated circuit packaging substrate without defects; on this basis, a shape-based template matching method is adopted; the core is to construct a similarity measure index, which can effectively cope with the interference of external factors such as occlusion, noise and illumination changes; specifically, the dot product sum between the shape feature of the template image and the direction vector of each corresponding shape feature point in the image to be detected is calculated, and the direction vector is normalized, and the index is thereby reduced. The adverse effects of illumination changes on matching accuracy; the calculation process of the normalized similarity measure index is as follows:

[0070]

[0071] In the formula, R is the normalized similarity measure index, D is the total number of corresponding shape feature points, and respectively represent the direction vector of the shape feature of the template image and the corresponding point of the i-th shape feature in the image to be detected, and · is the dot product operation, and respectively represent the length of the corresponding vector.

[0072] The calculated similarity measure index is compared with the preset threshold RT to obtain the comparison result of the similarity measure index R and the preset threshold RT; when R>RT, it is determined that the packaging substrate has a small shape defect; when R<RT, it is determined that the packaging substrate does not have a small shape defect.

[0073] When R = RT, the image obtained in step two needs to be divided into several sub-images, and steps three to six are repeated; the image processed by the mean filtering method, morphological dilation and texture operator is uniformly divided into several sub-images, and when the comparison result obtained after any one or multiple sub-images are operated by steps three to six is R > RT, it is determined that the packaging substrate has a micro shape defect. That is, when the sub-image obtained by dividing the image processed by the mean filtering method, morphological dilation and texture operator has a similarity measure index R greater than the threshold value RT, it is determined that the surface of the packaging substrate has a micro shape defect.

[0074] When the calculated similarity measure index R = RT, the image processed by the mean filtering method, morphological dilation and texture operator is uniformly divided into several sub-images, and steps two to six are performed on the sub-images respectively to obtain the comparison result of the similarity measure index R and the preset threshold value RT. Whether the packaging substrate has a micro shape defect is obtained by comparing the comparison result. When the calculated similarity measure index R is equal to the preset threshold value RT, it is because the shape features of the template image and the shape features of the sum calculated by the packaging substrate image to be detected are too much, resulting in that the calculated similarity measure index is equal to the threshold value. After further dividing the image and performing steps three to six, the comparison result obtained may appear that the similarity measure index R calculated by a part of the sub-images is less than the threshold value RT, and the similarity measure index R of another part of the sub-images is greater than the threshold value RT. Therefore, the image is further divided and steps three to six are performed again to perform more precise shape defect detection on the surface of the original packaging substrate.

[0075] At the same time, dividing the image of the packaging substrate into several images can further accurately detect whether the packaging substrate has a micro shape defect, and can determine the position of the micro shape defect. Uniformly dividing the sub-image again can more accurately determine the specific position of the micro shape defect of the packaging substrate. Of course, to more accurately obtain the position of the micro shape defect in the image of the packaging substrate, the image with the similarity measure index R greater than the preset threshold value RT can also be uniformly divided into several sub-images, and then steps three to six are performed. The position of the sub-image with the comparison result R > RT is selected, that is, the specific position of the micro shape defect.

[0076] When the calculated similarity metric R = RT, the image processed by the mean filtering method, morphological dilation and texture operator is uniformly divided into four first sub-images, the four first sub-images are respectively operated in steps three to six to obtain the comparison result of the second similarity metric R and the preset threshold RT. If there is still R = RT in the second comparison result, the first sub-image corresponding to the comparison result is again uniformly divided into four second sub-images, the four second sub-images are respectively operated in steps three to six to obtain the comparison result of the third similarity metric R and the preset threshold RT. Similarly, until there is no R = RT in the comparison result of the similarity metric R and the preset threshold RT. If there is still R = RT in the third comparison result, the second sub-image corresponding to the R = RT comparison result is again uniformly divided into four third sub-images, the four third sub-images are respectively operated in steps three to six to obtain the comparison result of the fourth similarity metric R and the preset threshold RT. Similarly, if there is still R = RT in the n-1th comparison result, the n-2th sub-image corresponding to the R = RT comparison result is again uniformly divided into four n-1th sub-images, the four n-1th sub-images are respectively operated in steps three to six to obtain the comparison result of the nth similarity metric R and the preset threshold RT. Until there is no R = RT in the comparison result of the similarity metric R and the preset threshold RT. By dividing the image into four only once, the calculation is more rapid to obtain the comparison result, the calculation amount of the redundant sub-image is reduced, and the calculation amount is further avoided due to too many divided sub-images.

[0077] For the micro defect samples in high-density integrated circuit packaging substrates that are difficult to accurately detect, the Focal Loss strategy is integrated into the loss function, and by dynamically adjusting the sample weight, the model focuses more on difficult classification samples. The calculation process is as follows:

[0078]

[0079] In the formula, FL adj is the adjusted Focal Loss function; β t is the coefficient for balancing the weights of positive and negative samples, δ is the adjustment factor, and q j is the prediction probability of the image sample. Through this strategy, the model can focus on detecting difficult defects, thereby improving the detection accuracy of micro defects in high-density integrated circuit packaging substrates.

[0080] Embodiment one

[0081] When the calculated similarity metric R = RT, the image obtained in step one can be directly divided into several sub-images, and the sub-images are respectively operated in steps two to six to obtain the comparison result of the similarity metric R and the preset threshold RT, and whether the packaging substrate has a micro shape defect is obtained through the comparison result.

[0082] Similarly, when the calculated similarity metric R = RT, the image obtained in step one is not divided uniformly after the mean filtering method, morphological dilation and texture operator processing, but is divided into four first sub-images, and the four first sub-images are respectively operated in steps two to six to obtain the comparison result of the second similarity metric R and the preset threshold RT. If the comparison result obtained in the second time still has R = RT, the first sub-image corresponding to the comparison result when R = RT is again divided into four second sub-images, and the four second sub-images are respectively operated in steps two to six to obtain the comparison result of the third similarity metric R and the preset threshold RT. Until there is no similarity metric R = RT in the obtained comparison result. Compared with the previous division operation, the direct division of the original image in this division operation can make the accuracy of the image in judging whether the packaging substrate has a micro shape defect higher.

[0083] Example two

[0084] In order to better detect the micro defects of high-density integrated circuit packaging substrates, first, the experimental environment is built, and the key parameter configuration is shown in Table 1; during the experimental preparation stage, the experimental environment will be built according to the configuration shown in Table 1. The devices or software of the experimental hardware and software environment configuration include microscope, scanning electron microscope, data storage device, power supply, operating system, image processing software, deep learning framework and data management software; the model of the microscope is Olympus BX53M, the magnification range is 50x-1000x, and it is equipped with a high-resolution digital camera; the model of the scanning electron microscope is Zeiss Merlin, the resolution is 0.8nm@15kV, and the acceleration voltage range is 0.2kV-30kV; the data storage device, the solid state disk is Samsung 970EVO Plus 1TB, and the external hard disk is Western Digital My Book 4TB; the power supply, the UPS uninterruptible power supply is APC Smart-UPS1500VA, which ensures the stability of the power supply during the experiment; the operating system is Windows 10Pro, 64-bit version; the image processing software MATLAB R2023a contains Image Processing Toolbox; the deep learning framework is TensorFlow 2.7.0, and the Python 3.8 environment; the data management software, FileZilla (for file transfer) and MySQL (for data storage and query).

[0085] After the experimental environment is built, a plurality of types of micro defect samples on the high-density integrated circuit packaging substrate as shown in Table 1 are selected; the present application prepares the common micro defect type samples in the high-density integrated circuit packaging substrate, and strictly controls the number and size range of each defect. The sample processing follows the experimental requirements, including cleaning, drying and marking, to ensure cleanliness and consistency.

[0086] Table 1 Information table of micro defect samples of high-density integrated circuit packaging substrate

[0087]

[0088] As Figure 2As shown, in order to verify the superiority of the research method, the present application designs the following comparative experiments. The detection method of micro defects of high-density integrated circuit packaging substrate proposed by the present application is taken as method 1, the high-density flexible packaging substrate defect detection method based on CRS-YOLO algorithm is taken as method 2, the low-resolution ceramic substrate image defect detection based on knowledge distillation is taken as method 3, and the ceramic substrate defect detection based on improved YOLOV4 is taken as method 4. The above three methods are used to detect the micro defects of the high-density integrated circuit packaging substrate respectively. After the experiment, the performances of the four methods in defect positioning accuracy are compared, and the results are shown in FIG. 4. Figure 2 As shown in FIG. 4.

[0089] As can be seen from Figure 2 , the method of the present application exhibits excellent performance in detecting micro defects on high-density integrated circuit packaging substrates. The method not only can accurately capture five types of micro defects such as cracks, scratches, pinholes, delamination and bubbles, but also reaches the highest level in positioning accuracy. In contrast, other detection methods have misjudgment or omission when dealing with these small and complex defects on high-density integrated circuit packaging substrates, thereby affecting the accuracy and integrity of the detection. This significant advantage is mainly because the method of the present application successfully improves the detection capability of micro defects by deeply analyzing and mining the unique features of micro defects, combined with advanced algorithms and technical means, thereby ensuring the accuracy and reliability of the detection results.

[0090] As shown in Figure 3 , after comparing the positioning accuracy of the four methods for micro defects of packaging substrates, the sensitivity of the detection method is further evaluated, and the results are shown in FIG. 5. Figure 3 As can be seen from Figure 3 , the method proposed in this paper shows obvious advantages in sensitivity. In the detection of different defect sizes, method 1 shows higher detection sensitivity, indicating that the method has higher sensitivity and accuracy in detecting micro defects. Specifically, the method in this paper successfully improves the detection sensitivity through fine feature extraction and optimized algorithm design, so that even in the case of weak or complex background of micro defects, the defects can be accurately and quickly identified.

[0091] The above is only a description of the preferred embodiments of the present application, and those skilled in the art can make some modifications and optimizations based on the above disclosure without departing from the above basic principles. These improvements and optimizations should be considered as the scope of protection of the present application.

Claims

1. A method for detecting minute defects in a high-density integrated circuit package substrate, characterized by: The following steps are involved: Step 1: Acquire image data of the surface of a high-density integrated circuit package substrate; Step 2: using mean filtering, morphological dilation and texture operators to process the acquired image in sequence; Step 3: Use image binarization processing technology to process the processed image again to obtain a binary image; divide the binary image into several energy zones, and calculate the entropy difference based on the divided energy zones; Step 4: Process the image data using discrete cosine transform technology to obtain DCT coefficients, divide the DCT coefficients into high-frequency components and low-frequency components, and calculate the ratio of the mean of the low-frequency coefficients to the mean of the high-frequency coefficients; Step 5: The feature extraction model extracts the shape features of the defects of the package substrate based on the difference in feature entropy under different energy suppression ratios and the ratio of the mean of the low-frequency coefficient to the mean of the high-frequency coefficient; Step 6: Calculate a similarity metric index based on the shape characteristics of the defect of the packaging substrate, compare the calculated similarity metric index with a preset threshold, and determine whether there is a tiny shape defect on the packaging substrate.

2. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 1, wherein: In the second step, the image data is smoothed using the mean filter method, and the calculation process is as follows: Where f′(x, y) represents the pixel value of the filtered image at position (x, y), f(i, j) represents the pixel value of the original image at point (i, j), W represents the window centered at (x, y), and M represents the total number of pixels in the window.

3. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 2, wherein: In step 2, the image after mean filtering is subjected to morphological dilation. The morphological dilation operation increases the image brightness by selecting the maximum value within the defined area of ​​the structure element. The calculation process is as follows: Where, It represents the pixel value of the image at the position (x, y) after morphological expansion, f′(x+s, y+t) represents the pixel value of the filtered image at the point (x+s, y+t), b represents the structural element, and max represents the maximum value operation. Represents an addition operation.

4. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 3, wherein: In step 2, a texture operator is designed to connect the bits of the central pixel in the image with those of the neighboring pixels through a specific encoding method. The calculation process is as follows: Where T(x,y) represents the output of the texture operator at point (x,y), N represents the area centered at (x,y), It is represented as a specific encoding method, and s(·) is represented as a sign function.

5. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 1, wherein: In step 3, the image processed by the mean filter method, morphological dilation and texture operator is processed by binarization technology to separate the foreground and background of the image; and the pixel grayscale values ​​of several pixels in the image are converted to obtain a binary image. The pixel grayscale value conversion calculation process of the image is as follows: Where g(x,y) represents the grayscale value of the pixel after binarization, and t represents the grayscale threshold.

6. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 5, wherein: The binary image is divided into several energy regions, and the energy of any energy region is defined as Ei, and the total energy of the binary image is Esum. The entropy difference is calculated based on the energy value of the energy region to obtain the characteristic entropy. The entropy difference reflects the change of characteristic entropy under different energy suppression ratios. The calculation process of the entropy difference ΔH is as follows: Where Q represents the number of energy regions in the binary image.

7. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 1, wherein: In the step 4, the discrete cosine transform (DCT) technology is introduced to perform block division, DCT transformation, quantization and encoding operations on the image after the mean filter method, morphological dilation and texture operator processing to obtain the DCT coefficients of the image; the DCT coefficients are divided into low-frequency coefficients and high-frequency coefficients; the mean of the low-frequency coefficients and the mean of the high-frequency coefficients are calculated respectively, and the ratio of the mean of the low-frequency coefficients to the mean of the high-frequency coefficients is calculated; the calculation process is as follows: Wherein, H and L are the number of rows and columns of the rectangular block respectively, and C(u, v) represents the coefficient after DCT transformation.

8. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 1, wherein: In step six, a similarity metric index is constructed. By calculating the sum of the dot products between the shape features of the template image and the direction vectors of each corresponding shape feature point in the image to be detected, and normalizing the direction vectors; the calculation process of the normalized similarity metric index is as follows: In the formula, R is the normalized similarity metric, D is the total number of corresponding shape feature points, and They are respectively represented as the shape feature of the template image and the direction vector of the corresponding point of the i-th shape feature in the image to be detected, · is the dot product operation, and They are respectively represented as the modulus of the corresponding vectors.

9. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 1, wherein: The calculated similarity metric index is compared with a preset threshold RT to obtain the comparison result between the similarity metric index R and the preset threshold RT; when R > RT, it is determined that there are minor shape defects in the packaging substrate; when R < RT, it is determined that there are no minor shape defects in the packaging substrate.

10. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 9, wherein: When R = RT, the image obtained in step two needs to be divided into several sub-images, and steps three to six are repeated; when the comparison result obtained after performing steps three to six on any one or multiple of the several sub-images is R > RT, it is determined that there are minor shape defects in the packaging substrate at this time.

11. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 10, wherein: When the calculated similarity metric index R = RT, the image processed by the mean filtering method, morphological dilation, and texture operator needs to be evenly divided into four first sub-images, and steps three to six are performed on the four first sub-images respectively to obtain the comparison result between the second similarity metric index R and the preset threshold RT; If there is still R = RT in the comparison result obtained for the second time, the first sub-image corresponding to this comparison result is further evenly divided into four second sub-images, and steps three to six are performed on the four second sub-images respectively to obtain the comparison result between the third similarity metric index R and the preset threshold RT; and so on until there is no R = RT in the comparison result between the similarity metric index R and the preset threshold RT.

12. The method for detecting minute defects in a high-density integrated circuit package substrate according to claim 1, wherein: The Focal Loss strategy is incorporated into the loss function, and by dynamically adjusting the sample weights, the model focuses more on difficult-to-classify samples. The calculation process is as follows: FL adj =-β t (1-q j ) δ log(q j ) Where FL adj Represented as the adjusted Focal Loss function; β t is the coefficient for balancing the weights of positive and negative samples, δ is the adjustment factor, and q j is expressed as the predicted probability of an image sample.

Images