Chip defect visual inspection method based on clustering algorithm

Abstract

Provided is a chip defect visual inspection method, comprising: acquiring an image data set of a group of chips, the image data set comprising image data corresponding to an image of each chip in the group of chips, the image data comprising a size of the image and a size of a defect in the image; performing standardization processing on the size of the defect of the image of each chip in the image data set; based on the size of the defect subjected to the standardization processing, obtaining a first parameter which defines the maximum distance between two points in a neighborhood; and constructing a DBSCAN-based clustering model based on the first parameter and a second parameter representing the minimum point number in the neighborhood, and applying the clustering model to the image data set to obtain a clustering result of the defects of the group of chips, the clustering result comprising at least one cluster.

Description

Technical Field

[0001] The present disclosure generally relates to a chip defect visual inspection, and more particularly, to a chip defect visual inspection method based on a clustering algorithm. Background Art

[0002] In semiconductor manufacturing technology, it is inevitable that various defects will occur on the chip, such as foreign matter, contamination, scratches, and capacitor defects. Various defects on the chip surface will seriously affect product quality, so it is necessary to inspect the chip surface defects and perform cluster analysis on the type, location, size, etc. of the defects, so as to find the causes of recurring defects and facilitate management of various links such as production and transportation to avoid such defects.

[0003] Therefore, a solution is needed that can quickly, accurately and conveniently perform cluster analysis on chip surface defects. Summary of the invention

[0004] In view of the above problems, the present disclosure provides a chip defect visual inspection method, including: acquiring an image data set of a group of chips with the same defect category, the image data set including image data corresponding to the image of each chip in the group of chips, the image data including the size of the image and the size of the defect in the image; standardizing the size of the defect in the image of each chip in the image data set; obtaining a first parameter based on the size of the defect after the standardized processing, the first parameter limiting the maximum distance between two points in a neighborhood; and constructing a DBSCAN-based clustering model based on the first parameter and a second parameter representing the minimum number of points in the neighborhood, and applying the clustering model to the image data set to obtain a clustering result of the defects of the group of chips, the clustering result including at least one cluster. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate embodiments of the present disclosure and, together with the description, further serve to explain the principles of the present disclosure and enable those skilled in the relevant art to make and use the present disclosure.

[0006] Figure 1 A schematic diagram showing a defect visual inspection in one embodiment is shown.

[0007] Figure 2 A flow chart of a defect visual inspection method based on a clustering model according to an embodiment is shown.

[0008] Figure 3 A schematic diagram of chip defects represented by boxes according to an embodiment is shown.

[0009] Figure 4 A schematic representation of a clustering result according to an embodiment is shown.

[0010] Figure 5A and Figure 5B An exemplary visual GUI interface for clustering model evaluation according to an embodiment is shown.

[0011] Figure 6 The graph is shown when the clustering model according to the embodiment of the present disclosure is used to perform defect analysis in daily production.

[0012] Various embodiments will be described with reference to the accompanying drawings. DETAILED DESCRIPTION

[0013] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and is not a limitation of the scope of protection, applicability or examples set forth in the claims. The function and arrangement of the elements discussed can be changed without departing from the scope of protection of the present disclosure. Various examples can omit, replace or add various processes or components as needed. For example, the described method can be performed in an order different from the described order, and various steps can be added, omitted or combined. In addition, the features described relative to some examples can also be combined in other examples.

[0014] It should be noted that the references to "one embodiment", "embodiment", "some embodiments" and the like in the specification indicate that the described embodiments may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, such expressions do not necessarily refer to the same embodiment. In addition, when specific features, structures or characteristics are described in conjunction with an embodiment, it should be within the knowledge of technicians in the relevant field to implement such features, structures or characteristics in conjunction with other embodiments that are explicitly or not explicitly described.

[0015] As is known in the art, during the chip manufacturing process, the inspection of chip surface defects is a particularly important step. Usually, during the defect inspection process, each chip is first photographed by an optical inspection device, and the chip surface defects are inspected in the photographed image and marked with boxes to indicate the location and size of the defects, and the defects are preliminarily classified. This classification can help technicians count the number of chips with defects of each category, that is, the loss percentage. Figure 1As shown in the figure above, the types of losses can include capacitor damage, foreign matter, contamination, and scratches, but are not limited to these. The line chart reflects the trend of the proportion of chips in each loss category over time. As shown in the figure, the proportion of one loss category shows a clear upward trend, so technicians need to pay attention to and analyze the cause of this loss in order to improve the yield.

[0016] Next, select the chip images in this upward trend loss category for analysis. Figure 1 As shown in the figure below, it is a top view image of a chip taken by an optical inspection device, and 12 chip images are shown as examples. It can be seen that there is a box mark on each chip image to indicate the location and size of the defect on the corresponding chip. For example, in the upper two rows of chip images in the figure, the box marks are located at the lower left corner, while in the bottom row of chip images, the box is located at a position other than the lower left corner. Therefore, the upper two rows of chip images in the figure are considered to be a signal because of the great similarity of the location of their defects. The signal can reflect the corresponding problems in the production materials or equipment, thereby helping technicians to determine the cause of the defects and thus improve the yield. At the same time, the bottom row of chip images in the figure can be considered as random noise because the defect location distribution has no similar pattern compared to the upper two rows of images. The above analysis process is cluster analysis, and the resulting "cluster" refers to the process and result of grouping data points with similar characteristics in a data set. Specifically, clustering is an unsupervised learning method that aims to group objects in a dataset so that objects within the same group are similar to each other, while objects in different groups are dissimilar. These groups are called "clusters".

[0017] The above-mentioned method of clustering analysis of the similarity of the position and size of defects in chip images is usually completed manually. That is, the technician observes the characteristics of the defects on the image, distinguishes and records them one by one, and then finds the similarity and regularity of the defect characteristics in a large number of pictures, thereby clustering the defects. Obviously, this requires a lot of manual time, and the number of images that can be processed every day is very limited. In addition, this manual method is highly dependent on the experience of the technician and is highly subjective, resulting in different judgments for each technician. Therefore, for manual cluster analysis, there is a lack of a feasible similarity discrimination indicator, which leads to the occurrence of misjudgment. Therefore, an accurate, time-saving and efficient cluster analysis method is needed to perform cluster analysis on defects of different categories.

[0018] like Figure 2 As shown, the defect visual inspection method 200 based on the clustering model according to the embodiment of the present disclosure includes operations 201-205. Those skilled in the art can understand that Figure 2 The method shown may be executed by a processor or other device capable of processing data.

[0019] In operation 201, image data of the chip to be analyzed or inspected is first obtained. As mentioned above, in the defect inspection process, an optical inspection device is first used to take pictures of the front surface and the back surface of each chip, respectively, to identify and frame the surface defects in the image, and to preliminarily classify the defects and assign them loss codes. For example, different loss codes are assigned to scratch defects, foreign body defects, and contamination defects. Subsequently, the number of losses under the corresponding loss codes is monitored and recorded. Once it is found that the number of losses under the loss code of a certain defect category shows a clear upward trend, it is used as a sample set to be analyzed for cluster analysis. That is, a sample set to be analyzed, or a group of chips to be analyzed, belongs to the same defect category and has the same loss code. However, those skilled in the art will understand that although a group of chips to be analyzed have the same loss code, they can have different chip numbers or names. In an embodiment, in operation 201, image data of the sample set to be analyzed is obtained, and the image data includes the size of the image taken of each chip and the size of the defect on the image. In an embodiment, the size of the image includes the width of the image (for example, Figure 1 The x-dimensions shown in the top view of the chip on the bottom side) and the height (e.g. Figure 1 The size of the defect includes the lateral size and the longitudinal size of the defect. Figure 3 As shown in , each defect is represented by a box, which represents the area occupied by defects of different irregular shapes, wherein the x-direction dimension of the box is the lateral dimension of the defect, and the y-direction dimension of the box is the longitudinal dimension of the defect. It should be noted here that, depending on the defect category, the defect may have different shapes, such as linear scratches, or irregularly shaped contaminants / foreign matter. In one embodiment of the present disclosure, a box is used to frame the defect to calculate its size, but those skilled in the art should understand that in other embodiments, different methods may be used to calculate the size of the defect.

[0020] As mentioned above, usually a manual method is used to perform cluster analysis on the location and size of chip defects of each chip in the sample set to be analyzed. However, this method is time-consuming and the accuracy needs to be improved.

[0021] Therefore, it is proposed to use a clustering model to perform cluster analysis on chip defects in the sample set to be analyzed. As known in the art, clustering models may include clustering based on the K-means algorithm, hierarchical clustering, and clustering based on the DBSCAN algorithm, but are not limited thereto. Among them, the clustering model based on the DBSCAN algorithm is density clustering, which defines clusters by finding areas with sufficiently high density, can handle noise and outliers, and is suitable for discovering clusters of arbitrary shapes, but it is sensitive to parameter settings.

[0022] The clustering model based on the DBSCAN algorithm includes two key parameters, namely, the neighborhood radius EPS and the minimum number of points. Specifically, EPS defines the maximum distance between two points in a neighborhood. For example, if the mutual distance between two points is less than or equal to the specified EPS, then they are in the same neighborhood. Therefore, EPS is the distance used by the DBSCAN algorithm to determine whether two points are similar and belong to the same class. A larger EPS will produce larger clusters (containing more points), and a smaller EPS will construct smaller clusters. In an embodiment, Figure 3 As shown in the figure, the defects of different sizes are represented by boxes, and the two intersecting boxes can be considered as one cluster, while the other adjacent but non-intersecting box on the left may be considered as the same cluster or another cluster, or may be considered as noise, depending on the parameter settings. On the other hand, the minimum number of points parameter defines the minimum number of points in a neighborhood. These points are considered as a cluster. A lower minimum number of points helps the algorithm build more clusters with more noise or outliers, while a higher minimum number of points will ensure more robust clusters, but if the cluster is too large, the smaller cluster will be merged into the larger cluster. Therefore, the selection of the two parameters EPS and minimum number of points is particularly important.

[0023] However, the existing clustering model can only perform cluster analysis on the location distribution of points, while the attributes of chip surface defects include not only the location distribution but also the size information of the defects, such as the diagonal size and diameter of the defects. Therefore, if the existing clustering model is directly used to perform cluster analysis on the image data of the sample set to be analyzed, the clustering results about the size cannot be obtained, and the obtained clusters may also be inaccurate. Therefore, in order to use the clustering model based on the DBSCAN algorithm to perform cluster analysis on the chip surface defects in the sample set to be analyzed in order to obtain accurate and fast clustering analysis results, it is necessary to improve and optimize the parameters of the algorithm model, and substitute the recalculated model parameters into the clustering model based on the DBSCAN algorithm to obtain an improved clustering model based on the DBSCAN algorithm suitable for cluster analysis of chip surface defects.

[0024] In operation 202, the image data of the sample set to be analyzed obtained in operation 201 are preprocessed. As mentioned above, the image data of the sample set to be analyzed includes the size of the image taken of each chip and the size of the defect on the image, wherein the size of the image includes the width of the image ( Figure 1 The x-dimensions shown in Figure 1 The defect size includes the y-direction dimension shown in , and the defect size includes the lateral dimension of the defect ( Figure 3 The x-direction dimension shown in Figure 3 In an embodiment, the preprocessing includes performing a standardization process on the image data of the group of chips to be analyzed.

[0025] In an embodiment, the normalization process may be calculated according to the following equation.

[0026]

[0027] The original defect size refers to the size of a defect on an image of a chip in a group of chips to be analyzed, for example, including the lateral size of the defect obtained from the original image ( Figure 3 The x-direction dimension shown in Figure 3 The new defect size represents the defect size obtained from the original image after the defect size is standardized or converted, for example, including the standardized or converted lateral size (x-direction size) and longitudinal size (y-direction size). Here, the standardized lateral size of the lateral size of the defect and the standardized longitudinal size of the longitudinal size of the defect can be calculated by the above equations. Those skilled in the art will appreciate that the diagonal size of the defect (e.g., Figure 3 shown).

[0028] In addition, the standard maximum size in the above equation represents the maximum image size among the image sizes of all chips in the group of chips to be analyzed, including the maximum image width and the maximum image height. The original size in the above equation represents the size of the image captured by the chip currently undergoing standardization, including the current image width and the current image height.

[0029] It can be understood that the standard maximum size and the original size in the above equation both include the width size and the height size. In the above-mentioned conversion process from the original defect size to the new defect size, according to The four known values ​​contained in (i.e., the maximum image width and the maximum image height, the current image width and the current image height) are scaled in the width and height directions respectively by bilinear interpolation, thereby realizing the standardized transformation of the defect size in the image.

[0030] Thereafter, the method 200 proceeds to operation 203 , in which the EPS parameters of the DBSCAN-based clustering model are calculated using the standardized new defect size obtained in operation 202 , and then the DBSCAN-based clustering model is established.

[0031] In an embodiment, the EPS parameter is calculated according to the following equation:

[0032]

[0033] Wherein, k represents a proportional coefficient, group represents the above-mentioned sample set to be analyzed or a group of chips to be analyzed, new defect x size and new defect y size represent the x-direction size and y-direction size of the new defect size obtained in operation 202, min_dis represents a minimum distance parameter, and max_dis represents a maximum distance parameter.

[0034] In an embodiment, the proportionality coefficient k is a statistical value, which is an empirical value obtained based on long-term observation by technicians. In other embodiments, the proportionality coefficient k can also be adjusted based on the clustering results. In one embodiment, the k value can be 2. In other embodiments, other k values ​​can be adopted.

[0035] In an embodiment, the minimum distance parameter min_dis and the maximum distance parameter max_dis are selected based on the image size of all chips in the group of chips to be analyzed and the new defect size after the standardized conversion. Therefore, based on the above calculation formula, those skilled in the art can understand that the diagonal length of the defect is first calculated based on the new defect x size and the new defect y size (the diagonal length is the standardized diagonal length), and then the median of the diagonal lengths of all defects in the group of chips to be analyzed is taken, and then multiplied by the proportional coefficient k. Then, the value calculated by multiplying the diagonal length by the proportional coefficient k is compared with the minimum distance parameter min_dis and the maximum distance parameter max_dis. For example, if the calculated value is larger than the maximum distance parameter max_dis, the EPS takes the maximum distance parameter max_dis, and if the calculated value is smaller than the minimum distance parameter min_dis, the EPS takes the minimum distance parameter min_dis. If the calculated value is between the minimum distance parameter min_dis and the maximum distance parameter max_dis, the EPS takes the calculated value. Here, it should be noted that the maximum distance parameter max_dis is selected in combination with the image size of each chip in the group of chips to be analyzed, and usually does not exceed half of the image size of the chip currently being detected, and the minimum distance parameter min_dis is selected based on the specified size that meets the minimum scrap standard. For example, if a 1mm scratch is set as the scrap standard, 0.8mm may be set as the minimum distance parameter min_dis.

[0036] Subsequently, based on the EPS parameters calculated above, combined with the selection of an appropriate minimum number of points parameter, a clustering model based on DBSCAN in the present disclosure can be established. In an embodiment, the minimum number of points parameter can be an integer greater than or equal to 2. In one embodiment, the minimum number of points parameter can be selected as 3. In other embodiments, the minimum number of points parameter can be selected as 4. It can be understood that the smaller the minimum number of points parameter, the more clusters and more noise or outliers can be established, and the larger the minimum number of points parameter, the more robust clusters can be established, that is, more points are contained in a cluster.

[0037] In an embodiment, after establishing a clustering model based on DBSCAN, the aforementioned set of image data of the chips to be analyzed is input into the clustering model based on DBSCAN to obtain a clustering result. In an embodiment, the clustering result includes a plurality of clusters divided by density distribution, wherein the cluster division not only takes into account the defect location distribution but also the defect size, and wherein the points in each cluster have great similarity, and there is no or almost no similarity between the points in different clusters.

[0038] like Figure 4As shown, it shows an example of the clustering results output by the clustering model based on DBSCAN. In the embodiment, the ranges framed by circles of different colors represent different clusters. For example, in Figure 4 In the figure, two clusters are indicated by arrows, namely Cluster1 and Cluster2. Figure 4 As can be seen, each cluster encircled by each circle contains several points, which is a group of objects with similarity.

[0039] As mentioned above, the choice of EPS parameters and minimum number of points parameters may affect the clustering results and lead to different cluster divisions. Therefore, in order to evaluate the quality of clustering results, Figure 2 In the method 200 shown, at operation 204, an evaluation of the model is performed. In an embodiment, the silhouette coefficient and the intersection-over-union ratio may be used to evaluate the clustering results.

[0040] Specifically, the silhouette coefficient is an indicator used to evaluate the effectiveness of clustering, and its value ranges from -1 to 1. The silhouette coefficient combines the compactness within the cluster and the separation between clusters, and measures the similarity between each data point and its own cluster and the nearest neighbor cluster by calculating a silhouette coefficient for each data point. The larger the silhouette coefficient, the better the clustering effect; close to 1 means that the data point matches its own cluster well and is very different from the adjacent cluster; close to -1 means that the point is more suitable for the adjacent cluster rather than the current cluster. In an embodiment, the silhouette coefficient is calculated according to the following equation:

[0041] Silhouette coefficient = (ba) / max(a,b)

[0042] where a represents the average intra-cluster distance between points in a cluster (e.g., Figure 4 ), and b represents the inter-cluster distance between individual clusters (e.g., Figure 4 The distance between Cluster1 and Cluster2 shown in ).

[0043] On the other hand, the present disclosure uses the intersection-over-union ratio to evaluate the clustering effect. Specifically, the intersection-over-union ratio (IOU) of individual defects is a cluster (e.g., Figure 4 The area of ​​intersection between a point in Cluster1 and its neighboring points (e.g., Figure 3 The schematic diagram shown shows the intersection) and the ratio of the total area of ​​the point and the neighboring points. The average intersection-over-union (IOU) within a cluster is the average of the individual defect IOUs of each defect within the cluster. The overall intersection-over-union (IOU) is the average of all the average IOUs within the cluster obtained for a set of chips to be analyzed.

[0044] In an embodiment, the higher the average IOU value within the same cluster, the better the clustering effect. In an embodiment, the average IOU is greater than or equal to 60%. In another embodiment, the average IOU ranges from 60% to 70%.

[0045] According to the model evaluation at operation 204, if the evaluation result does not meet the requirements, then Figure 2 The method 200 shown may proceed to operation 205, where the model parameters are optimized according to the evaluation results. Otherwise, if the evaluation results have reached the requirements, the method 200 may proceed to operation 206, where the final clustering model based on the DBSCAN algorithm is output.

[0046] In operation 205, according to an embodiment of the present disclosure, the clustering results are evaluated by the silhouette coefficient and the average intersection-over-union ratio, and the technician can adjust the model parameters, such as the EPS parameter and the minimum number of points parameter, in real time according to the evaluation results. For example, if the average intersection-over-union ratio of the cluster is low, the EPS parameter and / or the minimum number of points can be appropriately reduced so that the cluster contains fewer points, that is, more clusters are divided to improve the similarity within the cluster.

[0047] As an example, Figure 5A and Figure 5B A schematic visualization GUI interface for clustering model evaluation is shown. Figure 5A In the table interface at the top, each row represents the corresponding parameters of each chip in the current sample set, and each column shows the various parameters of the chip, such as the image width and height, the width and height of the box used to mark defects, the k value used to calculate the EPS parameter, the value of the EPS parameter, the value of the minimum point parameter, the maximum and minimum values ​​of the EPS parameter, and the number of chips included in the current sample set, etc. In an example, the silhouette coefficient and intersection ratio used for model evaluation can also be displayed in Figure 5A For example, in Figure 5A In the table shown at the bottom of , you can adjust the parameters of the clustering model by entering different K values, EPS parameters, and minimum number of points to obtain new clustering results. Figure 5A In the example shown, the K value used to calculate the EPS parameter is 1.5, the value of the EPS parameter is 100, the minimum number of points is 2, and Figure 5A The chips in the interface are grouped into the second cluster (Cluster 2), where the second cluster contains a total of 500 chips and a total of 527 images are generated.

[0048] exist Figure 5BIn the figure, the eight small figures on the left show the clustering results obtained by using different parameter value combinations for different groups (different sample sets, including different numbers of chip images), where different colors are used in each small figure to represent different clusters. Figure 5B In the eighth figure (i.e., the figure at the bottom of the second column), the number of chips included in the eighth group is 1352, the K value used to calculate the EPS parameter is 1.5, the value of the EPS parameter is 100, the minimum number of points is 2, and the clustering result contains many clusters, and a cluster at the lower left corner (for example, Figure 5A Cluster 2) in the Figure 5B The upper right corner of the Figure 5B In the lower right corner of each chip image, the position of the defect represented by the cluster is indicated by an arrow.

[0049] pass Figure 5A and Figure 5B This visual interface allows technicians to see the parameters of each chip in detail and to Figure 5A The parameters can be adjusted conveniently by modifying the corresponding model parameters on the interface, and the clustering results after parameter adjustment can be intuitively seen, which greatly improves the efficiency of clustering work.

[0050] After adjusting the model parameters, method 200 may return to operation 203, wherein the above-mentioned set of image data of the chip to be analyzed is input into the adjusted DBSCAN-based clustering model again to obtain an updated clustering result. Thereafter, the silhouette coefficient method and intersection-over-union ratio may be applied again in operation 204 to evaluate the updated clustering result, and the model parameters may be optimized again in operation 205. In an embodiment, the model parameters may be iteratively optimized by repeating the above-mentioned evaluation and parameter adjustment process until the obtained clustering result meets the requirements in the evaluation at operation 204. Thus, the finally established improved DBSCAN-based clustering model according to the present disclosure may be output at operation 206.

[0051] Here, those skilled in the art can understand that a DBSCAN-based clustering model for one defect category can be established based on the process shown in the above method 200. Subsequently, the above method 200 can be used to establish corresponding DBSCAN-based clustering models for various other defect categories, so as to facilitate defect analysis of different categories.

[0052] In an embodiment, the optimized final DBSCAN-based clustering model can be applied in daily production to inspect and analyze chip surface defects. In an embodiment, another group of chips with the same defect category can be selected as samples and input into the corresponding DBSCAN-based clustering model. The model will output the results of several clusters to guide technicians to focus on the cluster with the highest density, that is, the location where the defects are most densely distributed.

[0053] As an example, Figure 6 The curve graph obtained when the DBSCAN-based clustering model according to the present disclosure is used for defect analysis in daily production is shown. As can be seen from the figure, for a chip of a certain code, its image data within 30 days is collected by optical inspection equipment, and the DBSCAN-based clustering model according to the present disclosure is used for defect analysis. Figure 6 As shown, the vertical axis of the left figure represents the number of chips with the same chip number (VID), and the horizontal axis represents each cluster. The vertical axis of the right figure represents the number of chips with different chip numbers (VID), and the horizontal axis also represents each cluster. The blue column represents the clustering result of the first category of defects, and the orange column represents the clustering result of the second category of defects. It can be seen that the 118th cluster, the 6th cluster, and the 124th cluster in the figure are the clusters with the most points. Therefore, technicians can observe the defects on the chip images in these clusters and trace the chips through VID, and then analyze the causes of the corresponding defects from the perspectives of production process, transportation process, etc., so as to avoid such defects in subsequent production.

[0054] For example, in one example, assuming that the defect category represented by the orange column is surface scratches, then according to Figure 6 From the clustering results shown, it can be seen that the scratches represented by cluster 6 appear the most frequently. The technician can then check the corresponding chip images in the cluster to learn the location and size of the scratches in the cluster, and further analyze the cause of the scratches.

[0055] To the extent that various operations or functions are described herein, they may be described or defined as software code, instructions, configurations, and / or data. Content may be directly executable ("object" or "executable" form), source code, or differential code ("incremental" or "patch" code). The software content of the embodiments described herein may be provided via an article having content stored thereon, or via a method of operating a communication interface to send data via a communication interface. A machine-readable storage medium may enable a machine to perform the functions or operations described, and includes any mechanism for storing information in a form accessible by a machine (e.g., a computing device, an electronic system, etc.), such as a recordable / non-recordable medium (e.g., a read-only memory (ROM), a random access memory (RAM), a disk storage medium, an optical storage medium, a flash memory device, etc.). A communication interface includes any mechanism for communicating with another device through any hardwired, wireless, optical, or other media interface, such as a memory bus interface, a processor bus interface, an Internet connection, a disk controller, etc. The communication interface may be configured by providing configuration parameters and / or sending signals to prepare the communication interface for providing a data signal describing the software content. The communication interface may be accessed via one or more commands or signals sent to the communication interface.

[0056] The various components described herein may be units for performing the described operations or functions. Each component described herein includes software, hardware, or a combination of these. These components may be implemented as software modules, hardware modules, dedicated hardware (e.g., application-specific hardware, application-specific integrated circuits (ASICs), digital signal processors (DSPs), etc.), embedded controllers, hard-wired circuits, etc.

[0057] Through the above-mentioned clustering model based on the DBSCAN algorithm disclosed in the present invention, it is possible to cluster the defects of a group of chips with the same defect category (for example, scratch defects), find clusters with similarities in the location distribution and size of the defects, and then analyze the causes of the defects, thereby improving the yield. By using the improved clustering model based on the DBSCAN algorithm disclosed in the present invention, it is not only possible to provide accurate and fast clustering analysis, but also possible to save a lot of time for manual clustering analysis, thereby improving work efficiency.

[0058] The above description of the present disclosure is provided to enable any person of ordinary skill in the art to implement or use the present disclosure. Various modifications to the present disclosure will be apparent to those of ordinary skill in the art, and the general principles defined herein may be applied to other variations without departing from the scope of protection of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but is consistent with the widest range of principles and novel features disclosed herein.

Claims

1. A chip defect visual inspection method, comprising: Acquire an image data set of a group of chips, the image data set comprising image data corresponding to an image of each chip in the group of chips, the image data comprising a size of the image and a size of a defect in the image; normalizing the size of the defect in the image of each chip in the image data set; Based on the size of the defect after the normalization process, obtaining a first parameter, wherein the first parameter defines a maximum distance between two points in a neighborhood; as well as A DBSCAN-based clustering model is constructed based on the first parameter and a second parameter representing the minimum number of points in a neighborhood, and the clustering model is applied to the image data set to obtain a clustering result of defects of the group of chips, wherein the clustering result includes at least one cluster.

2. The method according to claim 1, further comprising: After obtaining the clustering result, evaluating the clustering result, and adjusting the first parameter and the second parameter based on the evaluation, so as to obtain an optimized clustering model and an optimized clustering result; as well as The optimized clustering model is evaluated again and the first parameter and the second parameter are adjusted again based on the evaluation, and a final clustering model is obtained by repeating the evaluation and adjustment process.

3. The method according to claim 2, wherein: The evaluation of the clustering result and the adjustment of the first parameter and the second parameter are performed through a visual interface.

4. The method according to claim 1, wherein: The normalization process is performed according to the following equation: wherein the original defect size represents the size of the defect on the image of each chip in the group of chips, and the new defect size represents the standardized size of the defect; and The standard maximum size is the maximum image size among the images of all chips in the group of chips, and the original size is the size of the image currently being processed.

5. The method according to claim 4, wherein: The standard maximum size includes a standard maximum width and a standard maximum height, the original size includes an original width and an original height, and wherein the conversion from the original defect size to the new defect size is obtained by bilinear interpolation based on the standard maximum width and the standard maximum height and the original width and the original height.

6. The method according to claim 4, wherein: The original defect size includes the original defect transverse size and the original defect longitudinal size, and the new defect size includes the new defect transverse size and the new defect longitudinal size, wherein the step of obtaining the first parameter is performed according to the following equation: Among them, k represents the scale factor, group represents the group of chips, min_dis represents the minimum distance parameter, and max_dis represents the maximum distance parameter.

7. The method according to claim 6, wherein: The minimum distance parameter and the maximum distance parameter are selected based on the sizes of images of all chips in the group of chips and the normalized new defect size.

8. The method according to claim 2 or 3, wherein: The evaluation criteria include silhouette coefficient and average intersection-over-union ratio.

9. The method according to claim 8, wherein: The silhouette coefficient is calculated according to the following equation: Silhouette coefficient = (ba) / max(a,b) Wherein, a represents the intra-cluster average distance between points within the cluster, and b represents the inter-cluster distance between the clusters.

10. The method according to claim 8, wherein: The average intersection-and-union ratio represents the average of the individual defect intersection-and-union ratios of each defect within the cluster, wherein the individual defect intersection-and-union ratio represents the ratio of the area of ​​intersection between a point within the cluster and its neighboring points to the total area of ​​the point and its neighboring points, and wherein the average intersection-and-union ratio is greater than or equal to 60%.

11. The method according to claim 1, wherein: The value of the second parameter is at least 2.

12. The method according to claim 3, further comprising: Acquire another set of chip image data sets; as well as The image data set of the other group of chips is input into the final clustering model to obtain a clustering result of chip defects.

13. A computer-readable storage medium for visual inspection of chip defects, having instructions stored thereon, which, when executed by a computing device, cause the computing device to execute the method according to any one of claims 1 to 12.

Images