Feedback-driven Chip Defect Classification Method and System Based on Visual Analysis

Abstract

The present invention discloses a feedback-driven chip defect classification method and system based on visual analysis, which relates to the technical field of image processing. It includes establishing a database, collecting historical defect pictures, and performing label classification according to different defects; obtaining the picture to be detected, converting it into a single-channel detection pixel map, performing region division, and calculating the block constant of the error-prone area; using blocks of different sizes, performing adaptive histogram equalization in the normal area, and performing enhanced histogram equalization according to the block constant in the error-prone area to obtain the detection map; judging the defect type according to the similarity and performing defect classification. The present invention performs region division on the single-channel detection pixel map, uses blocks of different sizes, performs adaptive histogram equalization in the normal area, and performs enhanced histogram equalization in the error-prone area to obtain a detection map that focuses on reflecting smaller local features inside, improves the detection effect, reduces the computational complexity, and decreases the amount of calculation.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and specifically to a feedback-driven chip defect classification method and system based on visual analysis. Background Art

[0002] A feedback-driven chip is an integrated circuit used to control and regulate signals or power in electronic devices. It adjusts the input signal according to the feedback information by real-time monitoring the output signal to ensure system stability and accuracy. Such chips are widely used in fields such as motor control, power management, and audio amplification. Defect classification of feedback-driven chips plays an important role in semiconductor manufacturing and quality control, which can effectively improve the quality and efficiency of chip production.

[0003] The defect classification of feedback-driven chips is generally carried out through visual inspection. For example, an intelligent chip surface defect detection method and system with the patent publication number CN116523869A, the method includes: extracting the first defect type in the preset chip defect types; forming a target chip set; extracting the first chip and obtaining the first chip features, where the first chip features have a corresponding relationship with the first defect type; constructing a chip defect database and storing it in the defect detection support vector machine; collecting the target chip video of the target chip to obtain a target chip feature set and inputting it into the defect detection support vector machine to obtain an output result; obtaining the target detection result. It solves the problems of long detection time, poor real-time performance, and low test accuracy existing in the existing chip surface defect detection by manual visual inspection, and achieves the effects of improving the real-time performance, efficiency, and accuracy of chip surface defect detection.

[0004] Chip defect detection generally collects images of chips by using an optical imaging system, and then analyzes the images through image processing algorithms to detect defects on the chip surface, such as missing pins, short circuits, bridging, etc. This method has a fast detection speed and can detect some tiny defects that are difficult to find by manual visual inspection. The above method realizes defect detection by preprocessing the image to be detected and performing feature recognition. However, in the detection of feedback-driven chips, the defect positions are generally small, and on the same production line, the positions where defects are likely to occur are relatively fixed. Preprocessing the entire image cannot focus on detecting the positions where defects are likely to occur, and the calculation amount is large. Summary of the Invention

[0005] The purpose of the present invention is to provide a feedback-driven chip defect classification method and system based on visual analysis to solve the problems raised in the above background art.

[0006] To achieve the above purpose, the present invention provides the following technical solution: A feedback-driven chip defect classification method based on visual analysis, including:

[0007] Build a database, collect historical defect pictures and record the generation time of the historical defect pictures, convert the historical defect pictures into single-channel historical pixel maps, and classify them by tags according to different defects;

[0008] Obtain the picture to be detected, convert the picture to be detected into a single-channel detection pixel map, and divide the single-channel detection pixel map into regions by the region prediction method according to the historical defect pictures. The single-channel detection pixel map is divided into a normal area and an error-prone area, and the block constant of the error-prone area is calculated by the block constant calculation method;

[0009] Adopt blocks of different sizes, perform adaptive histogram equalization in the normal area, and perform enhanced histogram equalization in the error-prone area according to the block constant to obtain a detection map, which can improve the contrast of the single-channel pixel map, and focus on reflecting the smaller local features inside the error-prone area, improve the detection effect and reduce the computational complexity;

[0010] Compare the detection map with the single-channel historical pixel map, judge the defect type according to the similarity, and classify the defects. When there are defects, save the detection map to the database to update the single-channel historical pixel map.

[0011] Preferably, the region prediction method includes:

[0012] Convert the historical defect map into a single-channel historical pixel map and mark the defect positions;

[0013] Reassign the channel values of the defect positions and change the channel values of the remaining positions to 0 to obtain an initial region map, and assign weights to the initial region map according to the generation time of the single-channel historical pixel map;

[0014] According to the weights of the initial region map, add the initial region maps in a matrix manner to obtain a predicted region map. Specifically:

[0015] , where is the channel value of the predicted region map at the position (x, y), is the channel value of the i-th initial region map at the position (x, y), is the weight of the i-th initial region map, which is set by the technician and its value is related to the generation time of the corresponding single-channel historical pixel map. The earlier the time, the smaller the value, is the total number of single-channel pixel maps;

[0016] In the predicted region map, ≥ the pixels are marked, and the region composed of all the marked pixel points is the error-prone area, where is a set regional division threshold, set by technicians, representing the area size of the divided region.

[0017] Preferably, the region prediction method includes:

[0018] Convert the historical defect map into a single-channel historical pixel map, mark the defect positions, and obtain the central pixel coordinates of the marked regions;

[0019] Based on the central pixel coordinates of multiple defect positions, establish the coordinate relationship of the defect positions, specifically:

[0020] , where is the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, that is, the correlation coefficient of the x coordinate and the y coordinate, is the average value of the abscissas in each central pixel coordinate, is the abscissa value of the jth defect in the ith single-channel historical pixel map, is the number of intermediate pixel coordinates in the ith single-channel historical pixel map, is the ordinate value of the jth defect in the ith single-channel historical pixel map, is the average value of the ordinates in each central pixel coordinate, is the total number of single-channel pixel maps;

[0021] Based on the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, establish a distribution prediction model for the defect constant at each pixel position in the image to be detected, specifically:

[0022] , where is the defect constant of the pixel with coordinates (x, y) in the image to be detected, is the standard deviation of the abscissas in each central pixel coordinate, is the standard deviation of the ordinates in each central pixel coordinate, is the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, is the average value of the abscissas in each central pixel coordinate, is the average value of the ordinates in each central pixel coordinate, is the abbreviation of the exponential function, used to represent the exponential function with the natural constant e as the base, ;

[0023] Calculate each pixel coordinate in the image to be detected through the distribution prediction model, and mark the pixels where ≥ . All the marked pixel points form the error-prone area, where is a set regional division threshold, representing the area size of the divided region.

[0024] Preferably, the regional division threshold can be updated according to a threshold iteration method, and the threshold iteration method includes:

[0025] Detect whether the defect position of the image to be detected exceeds the error-prone area;

[0026] When the defect position exceeds the error-prone area, detect whether the defect position exceeds the error-prone area simultaneously in the positive or negative direction of the abscissa or the positive or negative direction of the ordinate. When it exceeds the error-prone area simultaneously, obtain the maximum exceeded distance and calculate the change amount of the regional division threshold, specifically:

[0027] , where is the set regional division threshold, is the regional division threshold calculated in the previous calculation, is the maximum distance exceeding the error-prone area, is the update coefficient, set by the technician, and its value represents the sensitivity of the update range;

[0028] When the defect position does not exceed the error-prone area, obtain the minimum distance between the edge of the defect position and the edge of the error-prone area, and calculate the change amount of the regional division, specifically:

[0029] , where represents the number of times of increasing the regional division threshold before, is the set regional division threshold, is the regional division threshold calculated in the previous calculation, is the minimum distance between the edge of the defect position and the edge of the error-prone area, is the update coefficient, set by the technician, and its value represents the sensitivity of the update range.

[0030] Preferably, the block constant calculation method includes:

[0031] Record the number of pixels whose calculated value exceeds the maximum number of channels in the previous enhanced histogram equalization calculation;

[0032] Obtain the total number of pixels in the error-prone area detected last time, and calculate the block constant, specifically:

[0033] , where is the block constant, is the block constant calculated last time, is the original channel value whose channel value is greater than the maximum channel value after the enhanced histogram equalization calculation in the previous calculation, is the upper limit of the channel value of the single-channel detection pixel map, is the number of pixels with channel value i in the error-prone area, is the block calculation coefficient, set by the technical staff, is the total number of pixels in the error-prone area detected last time.

[0034] Preferably, the adaptive histogram equalization includes:

[0035] Dividing the pixels in the normal area of the single-channel detection pixel map into several non-overlapping blocks;

[0036] Calculating each block separately:

[0037] , where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of the single-channel detection pixel image pixel i in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively;

[0038] Mapping all the channel values in the normal area of the single-channel detection pixel map can obtain the corresponding detection map of the normal area.

[0039] Preferably, the enhanced histogram equalization includes:

[0040] Dividing the pixels in the error-prone area of the single-channel detection pixel map into several non-overlapping blocks;

[0041] Calculating each block separately:

[0042] , where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of the single-channel detection pixel image pixel i in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively, is the block constant;

[0043] Mapping all the channel values in the error-prone area of the single-channel detection pixel map can obtain the detection map corresponding to the error-prone area.

[0044] Preferably, the enhanced histogram equalization further includes an optimization method for the block edge, specifically:

[0045] Mark all the block edge pixels, calculate the mapped channel values of the target pixel in the adjacent blocks, and perform weighted summation on the mapped channel values according to the number of adjacent pixels in the corresponding adjacent blocks, specifically:

[0046] , where is the mapped channel value of the target pixel by the adjacent block z (including the block where the target pixel is located), is the edge pixel optimization value, is the number of adjacent pixels of the target pixel in the adjacent block z (including the block where the target pixel is located), is the total number of adjacent pixels of the target pixel;

[0047] After calculating the edge pixel optimization values of all the block edge pixels, use the edge pixel optimization values to reassign the channel values of the target pixels to obtain the optimized detection map.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] Dividing the single-channel detection pixel map into a normal area and an error-prone area, calculating the block constant of the error-prone area through the block constant calculation method, using blocks of different sizes, performing adaptive histogram equalization in the normal area, and performing enhanced histogram equalization according to the block constant in the error-prone area to obtain the detection map, which can improve the contrast of the single-channel pixel map, highlight the smaller local features inside the error-prone area, improve the detection effect and reduce the computational complexity, and reduce the amount of calculation.

[0050] At the same time, the regional prediction method is used to calculate the boundaries of the normal area and the error-prone area in real time, update the division of the error-prone area according to the results of each detection, and update the block constant of the error-prone area through the block constant calculation method to ensure the accuracy of the division of the error-prone area and improve the pixel contrast of the error-prone area in the detection map, thereby improving the accuracy of the detection result.

[0051] In the enhanced histogram equalization, by optimizing the block edge, the sense of fragmentation between each block is reduced, the smoothness of the transition between blocks is improved, the details of the image are better retained and the edge processing effect is better, and the original features of the image can be better maintained during the image processing process, further improving the accuracy of the detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic flow chart of the feedback-driven chip defect classification method of the present invention;

[0053] Figure 2 It is a schematic flow chart of the region prediction method (Example 1) of the present invention;

[0054] Figure 3 It is a schematic flow chart of the region prediction method (Example 2) of the present invention;

[0055] Figure 4 It is a schematic flow chart of the threshold iteration method of the present invention;

[0056] Figure 5 It is a schematic flow chart of the adaptive histogram equalization of the present invention;

[0057] Figure 6 It is a schematic flow chart of the enhanced histogram equalization of the present invention. Detailed implementation manners

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] In this application, the method steps used for easy understanding do not need to be executed in the order of the steps in this embodiment during actual operation. In some other embodiments, these steps can be performed synchronously or in a changed order.

[0060] Example 1:

[0061] In the detection of feedback-driven chips, the defect positions are generally small, and on the same production line, the positions where defects are likely to occur are relatively fixed. When classifying defects of feedback-driven chips through visual detection, preprocessing the entire picture cannot focus on detecting the positions where defects are likely to occur, and the calculation amount is relatively large. This embodiment can perform partition detection on the detected pictures, reduce the calculation amount, refine the detection of error-prone areas, focus on reflecting the smaller local features inside, and improve the detection effect.

[0062] As Figures 1 - 6 shown, the present invention provides a technical solution: a feedback-driven chip defect classification method based on visual analysis, including:

[0063] Establish a database, collect historical defect pictures and record the generation time of the historical defect pictures, convert the historical defect pictures into single-channel historical pixel maps, and perform label classification according to different defects;

[0064] Obtain the image to be detected, convert the image to be detected into a single-channel detection pixel map, and divide the single-channel detection pixel map into regions according to historical defect images through the region prediction method. The single-channel detection pixel map is divided into a normal region and an error-prone region, and the block constant of the error-prone region is calculated by the block constant calculation method;

[0065] It should be noted that there are various conversion methods for converting the image to be detected into a single-channel detection pixel map, such as the maximum channel number method, the average channel number method, and the weighted channel number method, etc. These are all existing technologies and will not be elaborated here. Different conversion methods can be adopted according to different acquisition devices and production environments.

[0066] Adopt blocks of different sizes, perform adaptive histogram equalization in the normal region, and perform enhanced histogram equalization in the error-prone region according to the block constant to obtain the detection map, which can improve the contrast of the single-channel pixel map, and focus on reflecting the smaller local features inside the error-prone region, improve the detection effect and reduce the computational complexity;

[0067] It should be noted that when performing adaptive histogram equalization in the normal region, larger blocks are used, and when performing enhanced histogram equalization in the error-prone region, smaller blocks are used, which can reduce the computational amount in the normal region and improve the detection effect on the error-prone region.

[0068] Compare the detection map with the single-channel historical pixel map, judge the defect type according to the similarity, and perform defect classification. When there is a defect, save the detection map to the database and update the single-channel historical pixel map.

[0069] It should be noted that when performing similarity judgment, methods such as local binary pattern and histogram of oriented gradients can be used. These are all existing technologies and will not be elaborated here.

[0070] As Figure 2 shown, the region prediction method includes:

[0071] Convert the historical defect map into a single-channel historical pixel map and mark the defect positions;

[0072] Reassign the channel values at the defect positions and change the channel values at the remaining positions to 0 to obtain the initial region map, and assign weights to the initial region map according to the generation time of the single-channel historical pixel map;

[0073] According to the weights of the initial region map, add the initial region maps in matrix form to obtain the predicted region map. Specifically:

[0074] , where To predict the channel value of the prediction region map at the position (x, y), is the channel value of the i-th initial region map at the position (x, y), is the weight of the i-th initial region map, which is set by the technician, and its value is related to the generation time of the corresponding single-channel historical pixel map. The earlier the time, the smaller the value, is the total number of single-channel pixel maps;

[0075] In the prediction region map, ≥ Pixels are marked, and the area composed of all marked pixel points is the error-prone area, where is the set region division threshold, which is set by the technician and represents the area size of the divided region.

[0076] It should be noted that for the convenience of understanding, the simulated data after assignment is shown in Table 1 and Table 2 below (the positions in the tables represent pixel positions, and the numbers represent pixel channel values). Table 1 and Table 2 represent different initial region maps respectively:

[0077] Table 1: Initial Region Figure 1

[0078]

[0079] Table 2: Initial Region Figure 2

[0080]

[0081] For the convenience of calculation, all are set to 1, is 0.5.

[0082] Calculate the of each pixel according to the above data, as shown in Table 3 below:

[0083] Table 3: Prediction Region Figure 1

[0084]

[0085] Compare the in Table 3 with . The area where the numbers in the figure are greater than 0.5 is the error-prone area, and vice versa is the normal area. Mapping the pixel positions to the single-channel detection pixel map can achieve the division of the error-prone area and the normal area.

[0086] As Figure 5 shown, adaptive histogram equalization includes:

[0087] Divide the pixels in the normal area of the single-channel detection pixel map into several non-overlapping blocks;

[0088] Calculate each block separately:

[0089] , where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of pixel i in the single-channel detection pixel image in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively;

[0090] Map all the channel values in the normal area of the single-channel detection pixel map to obtain the corresponding detection map of the normal area.

[0091] It should be noted that for the convenience of calculation, it is set that is 3, and the normal area in the single-channel detection pixel map is as shown in Table 4 below (the positions in the table represent pixel positions, and the numbers represent the channel values of the pixels):

[0092] Table 4: Pixels in the normal area Figure 1

[0093]

[0094] Divide the pixels in the normal area of the above single-channel detection pixel map into two non-overlapping blocks with a block size of 2x3, and calculate the upper block according to the formula (the values of i are 0, 1, 2, 3 respectively):

[0095] are 2, 1, 3, 0 respectively; are 2, 3, 6, 6 respectively; are 0, 1, 3, 3 (after rounding); since the two blocks in the simulation data are exactly the same, the values of the remaining blocks are the same as those of the above calculated block, and the normal area in the obtained detection map is as shown in Table 5 below:

[0096] Table 5: Pixels in the normal area Figure 2

[0097]

[0098] Compared with the single-channel detection pixel map, the contrast is improved, and it is easier to identify defect features. Moreover, in actual use, there are more pixels, and larger blocks can be used for calculation, reducing the amount of calculation.

[0099] As Figure 6 shown, the enhanced histogram equalization includes:

[0100] Dividing the pixels in the error-prone area of the single-channel detection pixel map into several non-overlapping blocks;

[0101] Calculating each block separately:

[0102] ,

[0103] where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of pixel i in the single-channel detection pixel image in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively, is the block constant;

[0104] Mapping all the channel values in the error-prone area of the single-channel detection pixel map can obtain the corresponding detection map of the error-prone area.

[0105] It should be noted that, for the convenience of calculation, it is set that is 3, is 2, and the error-prone area in the single-channel detection pixel map is as shown in Table 6 below (the positions in the table represent pixel positions, and the numbers represent the channel values of the pixels):

[0106] Table 6: Error-prone area pixels Figure 1

[0107]

[0108] Dividing the pixels in the error-prone area of the above single-channel detection pixel map into four non-overlapping blocks with a block size of 2x2, and calculating the upper left block according to the formula (the values of i are 0, 1, 2, 3 respectively):

[0109] are 1, 1, 2, 0 respectively; are 1, 2, 4, 4 respectively; They are 0, 2, 3, 3 respectively (when the calculation result exceeds the upper limit of the channel value, the maximum channel value is taken); since the four blocks in the simulated data are exactly the same, the values of the remaining blocks are the same as those of the above calculated blocks. The error-prone areas in the detection map are shown in Table 7 below:

[0110] Table 7: Pixels in the error-prone area Figure 2

[0111]

[0112] Compared with the single-channel detection pixel map, the contrast is improved, and it is easier to identify defect features. Moreover, compared with the normal area of the detection map, smaller blocks are used for calculation, and the channel value is further enlarged by the block constant, which can reflect smaller local features and improve the accuracy of detection.

[0113] The calculation method of the block constant includes:

[0114] Record the number of pixels whose calculated values exceed the maximum channel number in the previous enhanced histogram equalization calculation;

[0115] Obtain the total number of pixels in the error-prone area detected last time, and calculate the block constant, specifically:

[0116] , where is the block constant, is the block constant calculated last time, is the original channel value whose channel value is greater than the maximum channel value after the enhanced histogram equalization calculation in the previous calculation, is the upper limit of the channel value of the single-channel detection pixel map, is the number of pixels with channel value i in the error-prone area, is the block calculation coefficient, set by the technical staff, is the total number of pixels in the error-prone area detected last time.

[0117] It should be noted that for the convenience of calculation, the calculation data of the above enhanced histogram equalization is used. In the above data is 2, which means that in the previous calculation, the channel values corresponding to the pixels with channel values of 2 and above in the single-channel detection pixel map exceeded 3 after calculation. The number of pixels with channel value 2 in the error-prone area is 8, the number of pixels with channel value 3 is 0, and the total number of pixels in the error-prone area is 16, is 2, set is 1, so according to the formula , it can be calculated that is 1.5.

[0118] It should be noted that in the enhanced histogram equalization calculation, when the calculated channel value exceeds the maximum channel value, the maximum channel value is taken. From the above data, when a part of the calculated channel values exceed the maximum channel value, the block constant for the next calculation can be reduced, thereby reducing the degree of expansion of the channel value and preventing excessive assimilation of the channel value, improving the detection accuracy.

[0119] Embodiment 2:

[0120] In Embodiment 1, the calculation amount of the region prediction method is small, but it completely depends on the positions of the previous error-prone regions and can only divide the error-prone regions based on the positions of the defects that have already occurred. When the position of the defect deviates greatly, it is easy to divide the defect position outside the error-prone region, and the adaptability is poor. The difference between this embodiment and Embodiment 1 is that another region prediction method is provided to reduce the dependence on the positions of the previous error-prone regions and further improve the accuracy of region division.

[0121] The region prediction method includes:

[0122] Convert the historical defect map into a single-channel historical pixel map, mark the defect positions, and obtain the central pixel coordinates of the marked regions;

[0123] According to the central pixel coordinates of multiple defect positions, establish the coordinate relationship of the defect positions, specifically:

[0124] , where is the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, that is, the correlation coefficient of the x coordinate and the y coordinate, is the average value of the abscissas in each central pixel coordinate, is the abscissa value of the jth defect in the ith single-channel historical pixel map, is the number of intermediate pixel coordinates in the ith single-channel historical pixel map, is the ordinate value of the jth defect in the ith single-channel historical pixel map, is the average value of the ordinates in each central pixel coordinate, is the total number of single-channel pixel maps;

[0125] According to the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, establish a distribution prediction model for the defect constant at each pixel position in the image to be detected, specifically:

[0126] , where is the defect constant of the pixel with coordinates (x, y) in the image to be detected, is the standard deviation of the abscissas in each central pixel coordinate, is the standard deviation of the ordinates in each central pixel coordinate, is the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, is the average value of the abscissas in each central pixel coordinate, is the average value of the ordinates in each central pixel coordinate, is the abbreviation of the exponential function, used to represent the exponential function with the natural constant e as the base, ;

[0127] Each pixel coordinate in the image to be detected is calculated through the distribution prediction model, and ≥ The pixels are marked, and the area composed of all the marked pixel points is the error-prone area, where is the set area division threshold, representing the area size of the divided area.

[0128] It should be noted that for the convenience of calculation, the simulated data is as shown in Table 8 below:

[0129] Table 8: Defect Location Table

[0130]

[0131] Calculate is approximately 141, is approximately 101, is approximately 47, is approximately 28, is approximately -0.11. Select any coordinate in the single-channel detection pixel map to calculate the defect constant and compare it with When the defect constant at this pixel position is greater than , the pixel position can be marked as the error-prone area, otherwise it is marked as the normal area. After calculating all the pixel positions, the normal area and the error-prone area can be divided. For example, (160, 110) and (100, 160), f(100, 160) is approximately 1.04x10 -5 , calculate f(160, 110) is approximately 1.036x10 -4 , set to be 1x10 -4 , then f(100, 160) is less than , indicating that the probability of a defect occurring is small, and the pixel position is marked as the normal area, while f(160, 110) is greater than , indicating that the probability of a defect occurring is large, and the pixel position is marked as the error-prone area.

[0132] Compared with Embodiment 1, the calculation and division result of Embodiment 2 is more accurate, and it can allow a larger fluctuation in the defect position, and the calculation result is more accurate.

[0133] Embodiment 3:

[0134] In the second embodiment, the area is divided according to the central coordinates of the defect, and the area division threshold represents the divided area of the defect. However, in actual production, the area of the defect may often change. Therefore, the division method in the second embodiment may result in an error-prone area that is too large or too small, which is not conducive to detection. Based on the second embodiment, this embodiment provides a threshold updating method to further improve the accuracy of the area of the error-prone area.

[0135] The threshold updating method includes:

[0136] Detect whether the defect position of the image to be detected exceeds the error-prone area;

[0137] When the defect position exceeds the error-prone area, detect whether the defect position exceeds the error-prone area in both the positive and negative directions of the abscissa or the positive and negative directions of the ordinate. When it exceeds the error-prone area simultaneously, obtain the maximum distance of the excess, and calculate the change amount of the area division threshold. Specifically:

[0138] , where is the set area division threshold, is the area division threshold calculated last time, is the maximum distance exceeding the error-prone area, is the update coefficient, set by the technical personnel, and its value represents the sensitivity of the update range;

[0139] When the defect position does not exceed the error-prone area, obtain the minimum distance between the edge of the defect position and the edge of the error-prone area, and calculate the change amount of the area division, specifically:

[0140] , where represents the number of times the area division threshold was increased before, is the set area division threshold, is the area division threshold calculated last time, is the minimum distance between the edge of the defect position and the edge of the error-prone area, is the update coefficient, set by the technical personnel, and its value represents the sensitivity of the update range.

[0141] It should be noted that for the convenience of calculation, the following simulated data is set:

[0142] The defect position does not exceed the error-prone area, and the minimum distance between the edge of the defect position and the edge of the error-prone area is 2000 pixels, is 2, is 1x10 -4 , Adopt the method in the second embodiment The data is 1x10 -4 ;

[0143] It can be calculated that is 1.2x10 -4 .

[0144] Thus, when the defect area is small, the region division threshold can be increased, so as to reduce the number of those greater than the region division threshold in the defect constant, thereby narrowing the range of the error-prone area, further improving the accuracy of the division of the error-prone area, reducing the calculation amount, and improving the detection efficiency. Similarly, when the defect area is large, the region division threshold can be decreased to expand the range of the error-prone area and improve the detection accuracy.

[0145] Example 4:

[0146] In Example 1, when calculating the error-prone area through enhanced histogram equalization, the blocks are calculated separately, which may cause insufficient smoothness between the blocks and affect the detection effect. This example provides an optimization method to further optimize the mapping channel values at the block edges to improve the smoothness between the blocks.

[0147] The optimization method includes:

[0148] Mark all the block edge pixels, calculate the mapping channel values of the target pixel in the adjacent blocks, and perform weighted summation on the mapping channel values according to the number of adjacent pixels in the corresponding adjacent blocks. Specifically:

[0149] , where is the mapping channel value of the adjacent block z (including the block where the target pixel is located) for the target pixel, is the edge pixel optimization value, is the number of adjacent pixels of the target pixel in the adjacent block z (including the block where the target pixel is located), is the total number of adjacent pixels of the target pixel;

[0150] After calculating the edge pixel optimization values of all the block edge pixels, use the edge pixel optimization values to reassign the channel values of the target pixels to obtain the optimized detection map.

[0151] It should be noted that, for the convenience of calculation, the data of enhanced histogram equalization in Example 1 is used:

[0152] Taking the bottom-right pixel of the upper-left block in Table 7 as an example, its channel value in Table 6 is 2, it has two adjacent pixels in the upper-left block ( is 2), and the mapping channel value of the upper-left block for the pixel with channel value 2 is 3 ( is 3); it has one adjacent pixel in the upper-right block ( is 1), and the mapping channel value of the upper right block for the pixel with channel value 2 is 3 ( is 3), and it has an adjacent pixel in the lower left block ( is 1), and the mapping channel value of the lower left block for the pixel with channel value 2 is 3 ( is 3), according to the formula , it can be calculated that the edge pixel optimization value of this pixel is 3 (since the mapping channel values of the four blocks for the pixel with channel value 2 are all 3, so the edge pixel optimization value is also 3. In actual use, when the mapping channel values of adjacent blocks are different, the edge pixel optimization value will also change). Recalculating the mapping channel values located at the block edges can improve the smoothness between blocks, thereby improving the detection accuracy.

[0153] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended embodiments and their equivalents.

Claims

1. A feedback-driven chip defect classification method based on visual analysis, characterized in that: Including: Establish a database, collect historical defect pictures and record the generation time of the historical defect pictures, convert the historical defect pictures into single-channel historical pixel maps, and perform label classification according to different defects; Obtain the picture to be detected, convert the picture to be detected into a single-channel detection pixel map, and according to the historical defect pictures, divide the single-channel detection pixel map into regions by the region prediction method, divide the single-channel detection pixel map into a normal region and an error-prone region, and calculate the block constant of the error-prone region through the block constant calculation method; Adopt blocks of different sizes, perform adaptive histogram equalization in the normal region, and perform enhanced histogram equalization in the error-prone region according to the block constant to obtain a detection map, which can improve the contrast of the single-channel pixel map, and can highlight the smaller local features inside the error-prone region, improve the detection effect and reduce the computational complexity; Compare the detection map with the single-channel historical pixel map, judge the defect type according to the similarity, and perform defect classification. When there is a defect, save the detection map to the database to update the single-channel historical pixel map; The block constant calculation method includes: Record the number of pixels whose calculated value exceeds the maximum number of channels in the previous enhanced histogram equalization calculation; Obtain the total number of pixels in the error-prone area detected last time, and calculate the block constant, specifically: where is the block constant, is the block constant calculated last time, is the original channel value whose channel value is greater than the maximum channel value after enhanced histogram equalization calculation in the last calculation, is the upper limit of the channel value of the single-channel detection pixel map, is the number of pixels with channel value i in the error-prone area, is the block calculation coefficient, set by the technical staff, is the total number of pixels in the error-prone area detected last time; The enhanced histogram equalization includes: Divide the pixels in the error-prone region of the single-channel detection pixel map into several non-overlapping blocks; Calculate each block separately: , where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of pixel i in the single-channel detection pixel image in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively, is the block constant; Map all the channel values in the error-prone region of the single-channel detection pixel map to obtain the detection map corresponding to the error-prone region.

2. The feedback-driven chip defect classification method based on visual analysis according to claim 1, wherein: The region prediction method includes: Convert the historical defect map into a single-channel historical pixel map and mark the defect positions; Reassign the channel values of the defect positions and change the channel values of the remaining positions to 0 to obtain an initial region map, and perform weight assignment on the initial region map according to the generation time of the single-channel historical pixel map; According to the weights of the initial region maps, the initial region maps are added in a matrix manner to obtain the predicted region map, specifically as follows: where is the channel value of the predicted region map at the position (x, y), is the channel value of the i-th initial region map at the position (x, y), is the weight of the i-th initial region map, which is set by the technician and its value is related to the generation time of its corresponding single-channel historical pixel map. The earlier the time, the smaller the value, is the total number of single-channel pixel maps; In the prediction area map, ≥ Pixels are marked, and the area composed of all the marked pixel points is the error-prone area, where is the set area division threshold, set by the technician, representing the area size of the divided area.

3. The feedback-driven chip defect classification method based on visual analysis according to claim 1, wherein: The region prediction method includes: Convert the historical defect map into a single-channel historical pixel map, mark the defect positions, and obtain the central pixel coordinates of the marked regions; Based on the central pixel coordinates of multiple defect positions, establish the coordinate relationship of the defect positions, specifically as follows: Among them is the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, that is, the correlation coefficient of the x coordinate and the y coordinate, is the average value of the abscissas in each central pixel coordinate, is the abscissa value of the j-th defect in the i-th single-channel historical pixel map, is the number of intermediate pixel coordinates in the i-th single-channel historical pixel map, is the ordinate value of the j-th defect in the i-th single-channel historical pixel map, is the average value of the ordinates in each central pixel coordinate, is the total number of single-channel pixel maps; According to the correlation coefficients of the abscissa and ordinate of the central pixel coordinates, establish a distribution prediction model for the defect constant at each pixel position in the picture to be detected, specifically: wherein is the defect constant of the pixel at coordinates (x, y) in the image to be detected, is the standard deviation of the abscissas in the respective central pixel coordinates, is the standard deviation of the ordinates in the respective central pixel coordinates, is the correlation coefficient of the abscissa and ordinate of the central pixel coordinates, is the average value of the abscissas in the respective central pixel coordinates, is the average value of the ordinates in the respective central pixel coordinates, is the abbreviation of the exponential function, used to represent the exponential function with the natural constant e as the base, ; Calculate each pixel coordinate in the image to be detected through the distribution prediction model, and ≥ Pixels are marked, and the area composed of all marked pixel points is the error-prone area, where Is the set area division threshold, representing the area size of the divided area.

4. The feedback-driven chip defect classification method based on visual analysis according to claim 3, wherein: The region division threshold can be updated according to the threshold iteration method, and the threshold iteration method includes: Detect whether the defect position of the picture to be detected exceeds the error-prone region; When the defect position exceeds the error-prone area, it is detected whether the defect position exceeds the error-prone area simultaneously in the positive and negative directions of the abscissa or the positive and negative directions of the ordinate. When it exceeds the error-prone area simultaneously, the maximum distance of the excess is obtained, and the change amount of the region division threshold is calculated. Specifically: Where is the set region division threshold, is the region division threshold calculated in the previous time, is the maximum distance exceeding the error-prone area, is the update coefficient, which is set by the technical staff, and its value represents the sensitivity of the update range; When the defect location does not exceed the error-prone area, obtain the minimum distance between the edge of the defect location and the edge of the error-prone area, and calculate the change in area division, specifically: Among them represents the number of times of increasing the area division threshold before, is the set area division threshold, is the area division threshold calculated in the previous time, is the minimum distance between the edge of the defect location and the edge of the error-prone area, is the update coefficient, set by the technical staff, and its value represents the sensitivity of the update range.

5. The feedback-driven chip defect classification method based on visual analysis according to claim 1, characterized in that: The adaptive histogram equalization includes: Divide the pixels in the normal region of the single-channel detection pixel map into several non-overlapping blocks; Calculate each block separately: , where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of pixel i in the single-channel detection pixel image in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively; Map all the channel values in the normal region of the single-channel detection pixel map to obtain the detection map corresponding to the normal region.

6. The feedback-driven chip defect classification method based on visual analysis according to claim 1, characterized in that: The enhanced histogram equalization also includes an optimization method for the block edges, specifically: Mark all the edge pixels of the blocks, calculate the mapped channel value of the target pixel in the adjacent blocks, and perform a weighted sum of the mapped channel values according to the number of adjacent pixels in the corresponding adjacent blocks. Specifically: Where is the mapped channel value of the adjacent block z for the target pixel, is the edge pixel optimization value, is the number of adjacent pixels of the target pixel in the adjacent block z, is the total number of adjacent pixels of the target pixel; After calculating the edge pixel optimization values of all the block edge pixels, use the edge pixel optimization values to reassign the channel values of the target pixels to obtain an optimized detection map.

7. A feedback-driven chip defect classification system based on visual analysis, characterized in that: Including: Data storage module: used to establish a database, collect historical defect pictures and record the generation time of the historical defect pictures, convert the historical defect pictures into single-channel historical pixel maps, and perform label classification according to different defects; Data sorting module: Obtain the image to be detected, convert the image to be detected into a single-channel detection pixel map, and divide the single-channel detection pixel map into regions by the region prediction method according to historical defect images. The single-channel detection pixel map is divided into a normal region and an error-prone region, and the block constant of the error-prone region is calculated by the block constant calculation method. The block constant calculation method includes: Record the number of pixels whose calculated values exceed the maximum number of channels in the previous enhanced histogram equalization calculation; Obtain the total number of pixels in the error-prone area detected last time, and calculate the block constant, specifically as follows: where is the block constant, is the block constant calculated last time, is the original channel value whose channel value is greater than the maximum channel value after enhanced histogram equalization calculation in the last calculation, is the upper limit of the channel value of the single-channel detection pixel map, is the number of pixels with channel value i in the error-prone area, is the block calculation coefficient, set by the technical staff, is the total number of pixels in the error-prone area detected last time; Data processing module: Use blocks of different sizes to perform adaptive histogram equalization in the normal region and perform enhanced histogram equalization in the error-prone region according to the block constant to obtain a detection map, which can improve the contrast of the single-channel pixel map and highlight the smaller local features inside the error-prone region, improve the detection effect and reduce the computational complexity. The enhanced histogram equalization includes: Divide the pixels in the error-prone region of the single-channel detection pixel map into several non-overlapping blocks; Calculate each block separately: , where is the number of pixels with channel value i in the block, is the cumulative function, that is, the number of pixels with channel value i and below in the block, is the upper limit of the channel value in the single-channel detection pixel map, is the mapped channel value of pixel i in the single-channel detection pixel image in the detection map, is the upper limit of the channel value of the single-channel detection pixel map, and are the minimum and maximum values in the cumulative function respectively, is the block constant; Map all the channel values in the error-prone region of the single-channel detection pixel map to obtain the detection map corresponding to the error-prone region; Data output module: Compare the detection map with the single-channel historical pixel map, judge the defect type according to the similarity, classify the defects, and when there are defects, save the detection map to the database to update the single-channel historical pixel map.

Images