CN117115148A - Chip surface defect intelligent identification method based on 5G technology - Google Patents

Abstract

The invention provides a chip surface defect intelligent identification method based on a 5G technology, which relates to the technical field of defect identification and comprises the following steps: performing chip defect association analysis to obtain association defect types, performing negative sampling to obtain a plurality of defect sample images, performing pixel feature analysis to obtain a plurality of defect region pixel gradients, representing pixel value deviation of a defect region and a normal region of a chip to which the defects belong, traversing, judging whether the pixel value deviation is larger than or equal to a pixel gradient threshold value, adding the association defect types into the pixel segmentation association defect types, performing negative sampling to obtain a plurality of defect sample image sets, performing defect comparison on the chip images to be detected, and obtaining a pixel segmentation association defect type identification result. The invention solves the technical problems that the generalization capability of a machine learning model is weak, the defect recognition accuracy cannot be ensured, the types of defects on the surface of a chip are various, the shape is complex, and the requirements on accuracy and robustness cannot be met in the prior art.

Description

Chip surface defect intelligent identification method based on 5G technology

Technical Field

The invention relates to the technical field of defect identification, in particular to a chip surface defect intelligent identification method based on a 5G technology.

Background

During the manufacturing and assembly of chips, many types of defects, such as impurities, cracks, scratches, bumps or depressions, may occur, which may be caused by material flaws, process problems, equipment failures or operation errors, etc., and if not found and handled in time, may affect the quality, reliability and performance of the chips, while as the chip manufacturing process is continuously developed, the chip size is continuously reduced, and the detection and classification of defects becomes increasingly difficult, so that accurate and automated chip surface defect identification is becoming an urgent requirement.

The conventional chip surface defect recognition method still has certain defects, a plurality of machine learning models are developed in the prior art to detect the surface defects of the chip, but the generalization capability of the machine learning model is weaker due to the problem of small data of production samples of different scenes of the chip, and the defect recognition accuracy cannot be ensured, so that a certain liftable space exists for the chip surface defect recognition.

Disclosure of Invention

The application provides a chip surface defect intelligent identification method based on a 5G technology, which aims to solve the technical problems that a machine learning model is weak in generalization capability, defect identification accuracy cannot be guaranteed, and the chip surface defects are various in types and complex in forms, so that accuracy and robustness requirements cannot be met in the prior art.

In view of the above problems, the application provides a chip surface defect intelligent identification method based on a 5G technology.

The application discloses a first aspect, which provides a chip surface defect intelligent identification method based on a 5G technology, comprising the following steps: transmitting the chip processing technological parameters and the chip model information to the industrial Internet for chip defect correlation analysis, and obtaining the correlation defect type; transmitting the associated defect type and the chip model information to the industrial Internet for negative sampling, and obtaining a plurality of defect sample images, wherein the plurality of defect sample images are sample images of detected defects, and any one defect sample image is uniquely associated with one associated defect type; traversing the plurality of defect sample images to perform pixel characteristic analysis to obtain a plurality of defect region pixel gradients, wherein the defect region pixel gradients represent pixel value deviations of a defect region and a normal region of a chip to which the defect belongs; traversing the pixel gradients of the defect areas, and judging whether the pixel gradients are larger than or equal to a pixel gradient threshold value; adding the associated defect types with the pixel gradients of the defect areas being greater than or equal to the pixel gradient threshold into a pixel segmentation associated defect type; transmitting the pixel segmentation associated defect type and the chip model information to the industrial Internet for negative sampling to obtain a plurality of defect sample image sets; and performing defect comparison on the chip images to be detected based on the plurality of defect sample image sets, and obtaining a pixel segmentation association defect type identification result.

In another aspect of the present disclosure, an intelligent recognition system for a chip surface defect based on 5G technology is provided, where the system is used in the above method, and the system includes: the correlation analysis module is used for transmitting the chip processing technological parameters and the chip model information to the industrial Internet to perform chip defect correlation analysis and obtain the correlation defect type; the negative sampling module is used for transmitting the associated defect type and the chip model information to the industrial Internet for negative sampling to obtain a plurality of defect sample images, wherein the plurality of defect sample images are sample images with detected defects, and any one defect sample image is uniquely associated with one associated defect type; the pixel characteristic analysis module is used for traversing the plurality of defect sample images to perform pixel characteristic analysis and obtaining a plurality of defect area pixel gradients, wherein the defect area pixel gradients represent pixel value deviations of a defect area and a chip normal area to which the defect belongs; the pixel gradient judging module is used for traversing the pixel gradients of the defect areas and judging whether the pixel gradient is larger than or equal to a pixel gradient threshold value or not; a correlation type adding module for adding the correlation defect types with the pixel gradients of the plurality of defect areas being greater than or equal to the pixel gradient threshold into a pixel segmentation correlation defect type; the sample image acquisition module is used for transmitting the pixel segmentation association defect type and the chip model information to the industrial Internet for negative sampling to acquire a plurality of defect sample image sets; and the defect comparison module is used for comparing the defects of the chip images to be detected based on the plurality of defect sample image sets and obtaining a pixel segmentation associated defect type identification result.

One or more technical schemes provided by the application have at least the following technical effects or advantages:

performing chip defect association analysis to obtain association defect types, performing negative sampling to obtain a plurality of defect sample images, performing pixel feature analysis to obtain a plurality of defect region pixel gradients, representing pixel value deviation of a defect region and a normal region of a chip to which the defects belong, traversing, judging whether the pixel value deviation is larger than or equal to a pixel gradient threshold value, adding the association defect types into the pixel segmentation association defect types, performing negative sampling to obtain a plurality of defect sample image sets, performing defect comparison on the chip images to be detected, and obtaining a pixel segmentation association defect type identification result. The method solves the technical problems that the generalization capability of a machine learning model is weak, the defect recognition accuracy cannot be guaranteed, the types of the defects on the surface of the chip are various, the form is complex, and the requirements on accuracy and robustness cannot be met in the prior art, realizes the calculation and communication capability by means of a 5G technology, performs chip defect recognition based on small sample data, realizes high-speed and accurate chip surface defect detection and classification, and achieves the technical effects of improving the efficiency, the accuracy and the degree of automation of chip surface defect recognition.

The foregoing description is only an overview of the present application, and is intended to be implemented in accordance with the teachings of the present application in order that the same may be more clearly understood and to make the same and other objects, features and advantages of the present application more readily apparent.

Drawings

FIG. 1 is a schematic flow chart of a chip surface defect intelligent identification method based on a 5G technology according to an embodiment of the application;

fig. 2 is a schematic flow chart of a possible process for obtaining a size division associated defect type identification result in a chip surface defect intelligent identification method based on a 5G technology according to an embodiment of the present application;

FIG. 3 is a schematic diagram of a possible flow for obtaining a recognition result of a manually-segmented associated defect type in a chip surface defect intelligent recognition method based on a 5G technology according to an embodiment of the present application;

fig. 4 is a schematic diagram of a possible structure of a chip surface defect intelligent recognition system based on a 5G technology according to an embodiment of the present application.

Reference numerals illustrate: the system comprises a relevance analysis module 10, a negative sampling module 20, a pixel characteristic analysis module 30, a pixel gradient judging module 40, a relevance type adding module 50, a sample image acquisition module 60 and a defect comparison module 70.

Detailed Description

The embodiment of the application solves the technical problems that the generalization capability of a machine learning model is weak, the defect recognition accuracy cannot be guaranteed, the types of the defects on the surface of the chip are various and the shapes are complex in the prior art, so that the requirements on accuracy and robustness cannot be met, the computing and communication capability by means of the 5G technology is realized, the chip defect recognition is carried out based on small sample data, the high-speed and accurate chip surface defect detection and classification are realized, and the technical effects of improving the efficiency, the accuracy and the degree of automation of the chip surface defect recognition are achieved.

Having described the basic principles of the present application, various non-limiting embodiments of the present application will now be described in detail with reference to the accompanying drawings.

Example 1

As shown in fig. 1, an embodiment of the present application provides a chip surface defect intelligent identification method based on a 5G technology, which is applied to a chip surface defect intelligent identification system, wherein the system is based on a 5G technology and an industrial internet communication connection, and the method includes:

step S100: transmitting the chip processing technological parameters and the chip model information to the industrial Internet for chip defect correlation analysis, and obtaining the correlation defect type;

Specifically, the chip surface defect intelligent recognition method based on the 5G technology is applied to a chip surface defect intelligent recognition system, and the chip surface defect intelligent recognition system is connected with the industrial Internet in a communication mode based on the 5G technology.

Further, the step S100 of the present application includes:

step S110: transmitting the chip model information and the chip processing process parameters to the industrial Internet to obtain chip surface defect detection record data, wherein the industrial Internet is a private network for realizing information sharing for a plurality of chip manufacturing enterprises;

step S120: acquiring a defect detection record type and a record type detection frequency according to the chip surface defect detection record data;

step S130: adding the defect detection record type into the associated defect type when the record type detection frequency is greater than or equal to a first detection frequency threshold;

specifically, processing technological parameters of a chip, including key parameters such as temperature, humidity and material concentration, are obtained from a production line, model information of the chip, such as production batch, specification and the like, are obtained, the obtained chip processing technological parameters and the obtained chip model information are transmitted to an industrial internet platform through 5G technology and industrial internet communication connection, the uploaded chip model information and the uploaded processing technological parameters are used for inquiring in a private network of the industrial internet, and surface defect detection record data related to the chip are obtained, wherein the record data comprise detection records of each defect and corresponding attribute information.

Classifying the detected records of each defect, for example, analyzing and comparing the image characteristics, the positions, the shapes and the like, so as to determine the detected records of different types of defects, counting the frequency of each record type in all the detected records of the defect, obtaining the detected frequency of each record type by counting the occurrence times of each record type or the proportion of the total record number, and determining the occurrence condition of the defects of different types in the chip manufacturing process by acquiring the detected record types and the detected frequency of the record types, thereby providing reference for quality control and improvement.

And setting a first detection frequency threshold based on experience, demand or previous data analysis results, taking the first detection frequency threshold as a judgment standard of detection frequency, judging whether the detection frequency of each record type is greater than or equal to the first detection frequency threshold, if so, adding the record type into an associated defect type set, traversing all record types, and adding all record types meeting the conditions into the associated defect type set.

Wherein the defect detection record type and the record type detection frequency are obtained,

Step S121: the chip surface defect detection record data comprises first enterprise defect detection record data and second enterprise defect detection record data up to Nth enterprise defect detection record data, wherein the number of record bars from the first enterprise defect detection record data to the second enterprise defect detection record data up to the Nth enterprise defect detection record data is the same;

step S122: acquiring a first detection frequency and a second detection frequency of the defect detection record type until an Nth detection frequency;

step S123: setting the defect detection record type to be an irrelevant detection defect type when the number of the first detection frequency, the second detection frequency and the second detection frequency up to the Nth detection frequency is smaller than or equal to a second detection frequency threshold value is larger than or equal to floor (0.7 x N), wherein floor () is a downward rounding function;

step S124: and when the number of the first detection frequency and the second detection frequency up to the Nth detection frequency which are smaller than or equal to a second detection frequency threshold is larger than or equal to floor (0.7 x N), acquiring the average value of the first detection frequency and the second detection frequency up to the Nth detection frequency, and setting the average value as the recording type detection frequency of the defect detection recording type.

Specifically, through reasonable sampling and unified data preparation processes, the defect detection record data of the first enterprise, the second enterprise and the Nth enterprise are collected and arranged, the record data are ensured to contain the same number of record pieces, and the defect detection record data of all enterprises are combined together to obtain a complete defect detection record data set.

And sequentially taking out the first to N record types according to the complete defect detection record data set, wherein N is the number of record types in the defect detection record data set, for example, 25 record types are used, N is 25, the occurrence times of each record type in all defect detection records are counted respectively, the ratio of each occurrence time to the total number of times is calculated, and the first detection frequency and the second detection frequency are obtained until the N detection frequency.

And sequencing the acquired first detection frequency and the second detection frequency to the Nth detection frequency from high to low, setting a second detection frequency threshold according to requirements, screening out all detection frequencies smaller than the second detection frequency threshold in an sequencing table, and carrying out digital statistics.

floor (0.7×n) means that 0.7×n is rounded downward, for example, 25 is taken for N, floor (0.7×n) =floor (0.7×25) =floor (17.5), that is, 17.5 is rounded downward, an integer of 17 or less is expressed, the count result is compared with floor (0.7×n), if the number is greater than or equal to floor (0.7×n), it is indicated that the number satisfying the condition reaches the specified threshold, and the defect detection record type corresponding to the number satisfying the condition is set as the irrelevant defect detection type.

For the list of the first to nth detection frequencies satisfying the condition, the average value thereof is calculated, and the obtained average value is set as the recording type detection frequency of the defect detection recording type.

By acquiring the frequency of defect type detection and record type detection which are not related to each other, the defect detection records which have low frequency of occurrence in the overall data and have no importance or relevance are identified, and more accurate frequency information is provided for the defect detection records which have low frequency of occurrence in the overall data but still have certain relevance.

Step S200: transmitting the associated defect type and the chip model information to the industrial Internet for negative sampling, and obtaining a plurality of defect sample images, wherein the plurality of defect sample images are sample images of detected defects, and any one defect sample image is uniquely associated with one associated defect type;

specifically, through the 5G technology and the industrial Internet communication connection, data are transmitted to an industrial Internet platform, a negative sampling operation is performed, a proper image is selected from an existing defect sample library to serve as a defect sample, and the negative sampling refers to that a part of samples are randomly selected from non-target class samples to serve as training samples so as to train a classification model better. By recording the associated defect type information of each defect sample image, it is ensured that any given defect sample image is uniquely associated with an associated defect type, and the use of the sample images helps to train and improve the defect recognition algorithm, and improves the performance of the chip surface defect intelligent recognition system.

Step S300: traversing the plurality of defect sample images to perform pixel characteristic analysis to obtain a plurality of defect region pixel gradients, wherein the defect region pixel gradients represent pixel value deviations of a defect region and a normal region of a chip to which the defect belongs;

specifically, for each defect sample image, the image is segmented into a defective area and a normal area according to known defect and non-defect labels, for example, using image processing and segmentation algorithms, such as color, shape, or deep learning based methods, to automatically segment the defects. For each divided defect area, calculating the difference between each pixel point and the pixel value of the normal area of the corresponding chip by calculating the Euclidean distance between the pixel values, obtaining the pixel gradient of the pixel point, wherein the gradient represents the change rate of the pixel value and is used for representing the pixel value deviation of the defect area and the normal area of the chip to which the defect belongs, and repeating the steps until all the defect sample images are traversed. The gradients obtained by this step can be used as inputs to defect detection and classification algorithms to help accurately distinguish defective areas from normal areas.

Step S400: traversing the pixel gradients of the defect areas, and judging whether the pixel gradients are larger than or equal to a pixel gradient threshold value;

Specifically, a pixel gradient threshold is defined for determining whether the pixel gradient is large enough to be considered a defect, comparing the current pixel gradient to the pixel gradient threshold for each defective region, and if the current pixel gradient is greater than or equal to the pixel gradient threshold, then treating as a defect; otherwise, the method is regarded as normal, the pixel gradients of the defect areas are recorded to exceed the threshold value according to the judging result, and the steps are repeated until the pixel gradients of all the defect areas are traversed.

Step S500: adding the associated defect types with the pixel gradients of the defect areas being greater than or equal to the pixel gradient threshold into a pixel segmentation associated defect type;

specifically, a set of pixel segment associated defect types is created for storing defect types associated with defect areas having pixel gradients exceeding a threshold, for which, if the pixel gradient of the current defect area satisfies a condition and is associated with a certain associated defect type, the associated defect type is added to the set of pixel segment associated defect types, and the steps are repeated until the pixel gradients of all defect areas are traversed, thereby identifying the defect type associated with the potential defect area and providing useful information for the next step of defect identification and classification.

Step S600: transmitting the pixel segmentation associated defect type and the chip model information to the industrial Internet for negative sampling to obtain a plurality of defect sample image sets;

specifically, through the 5G technology and the industrial Internet communication connection, the pixel segmentation associated defect type and the chip model information are transmitted to an industrial Internet platform, a negative sampling operation is performed, an appropriate image is selected as a defect sample according to the pixel segmentation associated defect type and the chip model information, and a plurality of selected defect sample images are returned to the chip surface defect intelligent recognition system to form a defect sample image set, wherein the images can be used for training, verifying or testing a defect recognition algorithm.

Step S700: and performing defect comparison on the chip images to be detected based on the plurality of defect sample image sets, and obtaining a pixel segmentation association defect type identification result.

Further, step S700 of the present application includes:

step S710: acquiring a chip concept image, wherein the chip concept image is a standard chip image, the chip concept image is provided with a plurality of identification areas and a plurality of characteristic pixels, and the deviation of the characteristic pixels of any two identification areas is larger than or equal to the pixel gradient threshold;

Step S720: based on the plurality of identification areas, performing image shearing on any one defect sample image set of the plurality of defect sample image sets to obtain a plurality of groups of local sample images, wherein the plurality of groups of local sample images are in one-to-one correspondence with the plurality of identification areas;

step S730: based on the plurality of identification areas, carrying out area division on the chip image to be detected to obtain a plurality of area images to be detected;

specifically, a representative chip image is selected as a standard chip image, the chip image has a plurality of known defects and normal areas, a plurality of identification areas are selected according to priori knowledge or experience of field experts, the areas correspond to the known defects or other specific targets, feature pixels are extracted from any two identification areas, the feature pixels are median filtering pixel values of the corresponding areas, the feature pixel deviation is larger than or equal to the pixel gradient threshold, that is, the two identification areas have enough difference, and therefore the plurality of identification areas are combined into the standard chip image to form a chip concept image to be used as a reference.

And traversing a plurality of identification areas of each defect sample image set, and performing cutting operation on images in the current defect sample image set by using the position and size information of the corresponding identification areas so as to reserve local sample images corresponding to the identification areas until all the identification areas are traversed, so as to generate a plurality of local sample image groups corresponding to the identification areas one by one, namely a plurality of groups of local sample images corresponding to each identification area.

And obtaining a chip image to be detected from a data source or other sources as input, for each identification area, carrying out area division in the chip image to be detected according to the position and size information of the current identification area, for example, determining the position and the range of the area to be detected by using a geometric shape or a boundary frame, extracting the area corresponding to the current identification area from the chip image to be detected to form an area image to be detected, repeating the step until all the identification areas are traversed, and obtaining a plurality of area images to be detected, wherein the area images to be detected are used for subsequent area growth and defect comparison operation so as to further analyze and identify defects in the chip image to be detected.

Step S740: performing region growing on the plurality of region images to be detected according to the plurality of characteristic pixels to obtain a plurality of region growing results, wherein the plurality of region growing results comprise a plurality of growth covered regions and a plurality of growth uncovered regions;

further, step S740 of the present application includes:

step S741: acquiring a pixel growth similarity threshold;

step S742: taking a kth pixel point of an ith to-be-detected area of the plurality of to-be-detected area images as an initial growing pixel point, and taking the boundary of the ith to-be-detected area as a growth constraint boundary to acquire adjacent pixel points of the initial growing pixel point;

Step S743: traversing the adjacent pixel points and the initial growing pixel points to evaluate the color similarity, and obtaining the color similarity;

step S744: setting the adjacent pixel points with the color similarity larger than or equal to the pixel growth similarity threshold value as first generation growth pixel points;

step S745: adding the adjacent pixel points with the color similarity smaller than the pixel growth similarity threshold value into an ith growth uncovered area;

step S746: repeating growth based on the first generation of growth pixels until the pixel point of the ith area to be detected is traversed, and acquiring an ith growth coverage area;

step S747: adding the i-th growth coverage area to the plurality of growth coverage areas, and adding the i-th growth uncovering area to the plurality of growth uncovering areas.

Specifically, based on actual requirements and application scenarios, a range of pixel growth similarity thresholds is determined that are used to evaluate pixel color similarity, sufficient to enable a growing region to be partitioned into contiguous and close color regions while avoiding classifying dissimilar color pixels into the same region.

Selecting a kth pixel point in an ith to-be-detected area of an image of the to-be-detected area as an initial growing pixel point according to the ith to-be-detected area, wherein i and k respectively represent any one of the to-be-detected area and a similar point, one representative pixel point can be selected according to a specific strategy or rule, for example, a center point or random selection is performed, the boundary of the ith to-be-detected area is used as a growth constraint boundary, the boundary can be extracted by using an edge detection method of the area, such as a Canny operator, for the selected initial growing pixel point, adjacent pixel points are determined, wherein the adjacent pixel points refer to pixel points which have a direct connection relation with the pixel points in space, and the adjacent pixel points can be defined in different manners, for example, 8 communication or 4 communication, and in the 8 communication case, the adjacent pixel points comprise 8 pixels around the original pixel point; in the 4-way case, the adjacent pixel points only include pixels in four directions of up, down, left and right of the original pixel point.

And traversing the selected adjacent pixel points, comparing and evaluating the pixel points with the initial growing pixel points one by one, for example, adopting a correlation coefficient method to calculate the color similarity between the pixel points, wherein the similarity values are used as the basis for the growth of the subsequent region.

Judging whether the color similarity is larger than or equal to a pixel growth similarity threshold value, if the color similarity meets the condition, setting the adjacent pixel points as first generation growth pixel points which serve as starting points of subsequent region growth, providing a basis for defect comparison and pixel segmentation association defect type identification, and repeating the steps to check all the adjacent pixel points until all the adjacent pixel points are completed.

If the color similarity is less than the pixel growth similarity threshold, indicating that the color similarity does not meet the condition, adding the adjacent pixels to an ith growth uncovered area, which may contain pixels that are not sufficiently similar to other grown areas, may be potential defects or new growth portions, further processing and analysis are required, and repeating the steps to process all the adjacent pixels until the traversal is completed.

And using the first generation of growing pixel points as starting points, selecting adjacent pixel points of the pixel points to expand according to various growth criteria and conditions such as color similarity, connectivity and the like, calculating and evaluating the color similarity between the adjacent pixel points and a current area, ensuring that the adjacent pixel points meet the growth criteria to be added into the current area, adding the adjacent pixel points meeting the conditions into the current area, marking the adjacent pixel points as accessed, repeating the steps until all pixel points of an ith area to be detected are traversed, and finally generating an area covering all relevant pixel points as an ith growth coverage area, wherein all pixel points which are connected with the initial growing pixel points and have similar colors are included.

Adding the ith growth coverage area to a set of multiple growth coverage areas to store all generated growth coverage areas, so that all growth coverage area results for each area to be inspected can be recorded; the i-th grown uncovered area is added to a collection of grown uncovered areas to store all of the grown uncovered areas, so that information about the grown uncovered areas, such as areas that may be potential defects or portions that require further processing and analysis, may be retained.

Step S750: and performing defect comparison on the plurality of growth uncovered areas based on the plurality of groups of local sample images to obtain the identification result of the pixel segmentation associated defect types.

Further, step S750 of the present application includes:

step S751: obtaining comparison characteristic pixels of an ith growth uncovered area;

step S752: obtaining M defect characteristic pixel sets of an ith group of local sample images corresponding to an ith region to be detected;

step S753: obtaining a pixel value deviation maximum value and a pixel value deviation minimum value of the M defect characteristic pixel sets, and constructing a pixel deviation interval;

step S754: traversing the M defect characteristic pixel sets to calculate pixel value deviations based on the comparison characteristic pixels, and obtaining M pixel value deviations;

step S755: and acquiring the proportionality coefficient of the M pixel value deviations belonging to the pixel deviation interval, setting the ith growing uncovered area as the associated defect type corresponding to the plurality of groups of local sample images when the proportionality coefficient is larger than or equal to a proportionality coefficient threshold value, and adding the associated defect type identification result of the pixel segmentation.

Specifically, appropriate locations are selected in the i-th growth non-covered region to extract alignment feature pixels, which may be center points of the region, key pixel points on the boundary, or other important locations selected based on a priori knowledge and experience, and corresponding pixel values or feature descriptors, including gray scale pixel values, color components, texture features, edge features, etc., are extracted from the i-th growth non-covered region based on the selected locations for subsequent defect alignment.

For the ith to-be-detected area, selecting a corresponding ith group of local sample images, determining the number M of defect characteristic pixel sets to be extracted according to actual requirements and application scenes for the local sample images, wherein the M can be adjusted according to the complexity degree of defects, detection requirements, algorithm performances and the like. In the i-th group of partial sample images, a suitable position is selected to extract M defect feature pixel sets.

For each defect characteristic pixel set, calculating the pixel value deviation between the defect characteristic pixel set and the characteristic pixels at the corresponding positions in the region to be detected, traversing M defect characteristic pixel sets, respectively recording the maximum value of the pixel value deviation, and setting the maximum value asAnd minimum value of ∈>Constructing a pixel deviation interval according to the recorded maximum value and minimum value, wherein the range of the interval is +.>。

For each defect characteristic pixel set, calculating pixel value deviation with the comparison characteristic pixels one by one, for example, calculating by using a Euclidean distance method, obtaining the pixel value deviation, traversing M defect characteristic pixel sets, and recording the obtained pixel value deviation to form a set containing M pixel value deviations.

Counting how many pixel value deviations in M pixel value deviations fall in a pixel deviation interval In the method, the pixel deviation number of the coincidence interval is recorded as W, a scaling factor is calculated, namely the scaling factor is W/M, the scaling factor is compared with a scaling factor threshold value, if the scaling factor is larger than or equal to the scaling factor threshold value, the growth uncovered area corresponds to a plurality of groups of local sample images and has associated defect type characteristics, the ith growth uncovered area is set as an associated defect type corresponding to a plurality of groups of local sample images, and the growth uncovered area is added into a pixel segmentation associated defect type identification result set.

Further, as shown in fig. 2, the present application further includes:

step S810: traversing the plurality of defect sample images for size analysis to obtain size deviations of a plurality of defect areas;

step S820: when the pixel gradient of the plurality of defect areas is smaller than the pixel gradient threshold, judging whether the pixel gradient of the plurality of defect areas is larger than or equal to a size deviation threshold or not based on the size deviation of the plurality of defect areas, wherein the size deviation threshold is the minimum size deviation which can be identified by a size classifier;

step S830: adding a size division associated defect type to the associated defect type for which the pixel gradient of the plurality of defect areas is less than the pixel gradient threshold and the size deviation of the plurality of defect areas is greater than or equal to the size deviation threshold;

Specifically, for each defect sample image, a size analysis is performed, for example, measuring the width, height, area or other relevant size parameters of the defect region, for each defect sample image, a size deviation from a reference size is calculated, the reference size may be a predefined standard size, average size or other reliable reference value, and a plurality of defect region size deviations are obtained according to the deviation calculation result.

Checking whether pixel gradients of a plurality of defect regions are smaller than a pixel gradient threshold value, if the pixel gradients are smaller than the threshold value, setting a size deviation threshold value as a minimum size deviation which can be identified as a defect according to the performance and accuracy of the size classifier, checking whether the size deviation of each defect region is larger than or equal to the size deviation threshold value, and if so, indicating that the defect region has significant size deviation and can be identified as a defective region.

Checking whether pixel gradients of the plurality of defect regions are smaller than a pixel gradient threshold, if a condition is met, indicating that pixel gradient differences are small, checking whether size deviations of the defect regions meeting the pixel gradients are larger than or equal to a size deviation threshold, and adding corresponding associated defect types to the size division associated defect type set when the pixel gradients meet the threshold and the size deviations are larger than or equal to the threshold, so as to identify the defect regions with small pixel gradients but significant size deviations.

Step S840: based on the size division associated defect type and the chip model information, training a size division model to perform defect identification on the chip image to be detected, and obtaining a size division associated defect type identification result, wherein the method comprises the following steps:

step S841: based on the size division associated defect type and the chip model information, acquiring a positive sample image of the defect to be detected and a negative sample image of the defect to be detected, wherein the negative sample image of the defect to be detected has defect size identification information;

step S842: and taking the positive sample image of the defect to be detected and the negative sample image of the defect to be detected as input data, taking the defect size identification information as supervision data, and carrying out size defect classification training based on a convolutional neural network to obtain the size segmentation model.

Specifically, according to the size division associated defect type and the chip model information, determining the defect type to be detected and classified, collecting positive sample images of the defect to be detected, wherein the images contain target defects and are marked in the form of images, and the images can be generated by using an image processing algorithm; negative sample images of defects to be inspected are acquired, which images have defect size identification information, wherein the negative sample images do not contain features similar to those of the target defects, but carry identifications representing the sizes of the defects so that the training model can learn the correct size classification.

A Convolutional Neural Network (CNN) model is constructed that contains a convolutional layer, a pooling layer, and a fully-connected layer to extract and learn features. And training on a convolutional neural network by using the positive sample image and the negative sample image as input data and using defect size identification information as supervision data, optimizing parameters of a model, and performing iterative training until the model can accurately distinguish and classify defect samples with different sizes, thereby obtaining the size segmentation model.

After model training is completed, defect identification is carried out on the chip image to be detected by utilizing the trained size segmentation model, a segmentation algorithm is applied to the image, a defect area in the image is marked, the corresponding size segmentation associated defect type is predicted, the identification result of the size segmentation associated defect type is extracted from the size segmentation model output, and the chip model information is checked and verified.

Further, as shown in fig. 3, the present application further includes:

step S910: adding the associated defect types of which the pixel gradients of the defect areas are smaller than the pixel gradient threshold and the size deviation of the defect areas is smaller than the size deviation threshold into the manually-segmented associated defect types;

Step S920: when the defect detection quantity of the pixel segmentation association defect type identification result and the size segmentation association defect type identification result is zero, transmitting the chip image to be detected to a manual quality inspection channel to identify the manual segmentation association defect type, and acquiring a manual segmentation association defect type identification result.

Specifically, it is checked whether the pixel gradient of the plurality of defect areas is smaller than a pixel gradient threshold value, and whether their size deviation is smaller than a size deviation threshold value, and when the pixel gradient of the plurality of defect areas is smaller than the threshold value and the size deviation is smaller than the threshold value, the defect areas and the associated defect types are added to the manually divided associated defect types, which require further manual processing and judgment.

Checking whether the defect detection quantity of the pixel segmentation associated defect type identification result and the size segmentation associated defect type identification result is zero, if the defect detection quantity of the two results is zero, indicating that the automatic identification cannot accurately identify the defects, transmitting the chip image to be detected to a manual quality inspection channel, and carrying out defect identification and classification by a manual operator, wherein the manual operator adopts a manual segmentation technology, manually draws defect areas and the like, and obtains the identification result of the manual segmentation associated defect type through manual operation.

Through the above steps, when the pixel difference and the size difference are smaller, the automated pixel segmentation and size segmentation may not accurately detect the defects, then the defect identification is performed by using the manual segmentation channel to ensure a higher quality standard, in fact, the pixel values of most defects and the regions to which the pixel values belong have larger differences, the rest of the defects belong to the size difference, and the smallest part of the defects belong to the type, so that the workload of the manual quality inspection is less or no, and only auxiliary detection is performed.

In summary, the chip surface defect intelligent identification method based on the 5G technology provided by the embodiment of the application has the following technical effects:

performing chip defect association analysis to obtain association defect types, performing negative sampling to obtain a plurality of defect sample images, performing pixel feature analysis to obtain a plurality of defect region pixel gradients, representing pixel value deviation of a defect region and a normal region of a chip to which the defects belong, traversing, judging whether the pixel value deviation is larger than or equal to a pixel gradient threshold value, adding the association defect types into the pixel segmentation association defect types, performing negative sampling to obtain a plurality of defect sample image sets, performing defect comparison on the chip images to be detected, and obtaining a pixel segmentation association defect type identification result.

The method solves the technical problems that the generalization capability of a machine learning model is weak, the defect recognition accuracy cannot be guaranteed, the types of the defects on the surface of the chip are various, the form is complex, and the requirements on accuracy and robustness cannot be met in the prior art, realizes the calculation and communication capability by means of a 5G technology, performs chip defect recognition based on small sample data, realizes high-speed and accurate chip surface defect detection and classification, and achieves the technical effects of improving the efficiency, the accuracy and the degree of automation of chip surface defect recognition.

Example two

Based on the same inventive concept as the chip surface defect intelligent recognition method based on the 5G technology in the foregoing embodiment, as shown in fig. 4, the present application provides a chip surface defect intelligent recognition system based on the 5G technology, the system is based on the 5G technology and industrial internet communication connection, and the system includes:

the correlation analysis module 10 is used for transmitting the chip processing technological parameters and the chip model information to the industrial internet to perform chip defect correlation analysis, and acquiring the correlation defect type;

the negative sampling module 20 is configured to transmit the associated defect type and the chip model information to the industrial internet for negative sampling, and obtain a plurality of defect sample images, where the plurality of defect sample images are sample images of detected defects, and any one defect sample image is uniquely associated with one associated defect type;

The pixel characteristic analysis module 30 is configured to traverse the plurality of defect sample images to perform pixel characteristic analysis, and obtain a plurality of defect region pixel gradients, where the plurality of defect region pixel gradients characterize pixel value deviations of a defect region and a normal region of a chip to which the defect belongs;

a pixel gradient judging module 40, wherein the pixel gradient judging module 40 is configured to traverse the pixel gradients of the plurality of defect areas and judge whether the pixel gradient is greater than or equal to a pixel gradient threshold;

a correlation type adding module 50, wherein the correlation type adding module 50 is configured to add the correlation defect types with the pixel gradients of the plurality of defect areas being greater than or equal to the pixel gradient threshold value into a pixel division correlation defect type;

the sample image acquisition module 60 is configured to transmit the pixel segmentation associated defect type and the chip model information to the industrial internet for negative sampling, so as to acquire a plurality of defect sample image sets;

and a defect comparison module 70, wherein the defect comparison module 70 is configured to perform defect comparison on the chip image to be inspected based on the plurality of defect sample image sets, and obtain a recognition result of the pixel segmentation associated defect type.

Further, the system further comprises:

the traversing module is used for traversing the plurality of defect sample images to carry out size analysis and obtain a plurality of defect area size deviations;

the deviation judging module is used for judging whether the size deviation of the plurality of defect areas is larger than or equal to a size deviation threshold value or not based on the size deviation of the plurality of defect areas when the pixel gradient of the plurality of defect areas is smaller than the pixel gradient threshold value, wherein the size deviation threshold value is the minimum size deviation which can be identified by a size classifier;

a size division association type adding module, configured to add the association defect types, in which the pixel gradient of the plurality of defect areas is smaller than the pixel gradient threshold, and the size deviation of the plurality of defect areas is greater than or equal to the size deviation threshold;

the model training module is used for training a size segmentation model to carry out defect recognition on the chip image to be detected based on the size segmentation associated defect type and the chip model information, and obtaining a size segmentation associated defect type recognition result, and comprises the following steps:

the sample image acquisition module is used for acquiring a positive sample image of the defect to be detected and a negative sample image of the defect to be detected based on the size segmentation associated defect type and the chip model information, wherein the negative sample image of the defect to be detected has defect size identification information;

The classification training module is used for performing size defect classification training based on a convolutional neural network by taking the positive sample image of the defect to be detected and the negative sample image of the defect to be detected as input data and the defect size identification information as supervision data, so as to obtain the size segmentation model.

Further, the system further comprises:

the manual segmentation associated type adding module is used for adding the associated defect types with the pixel gradients of the defect areas smaller than the pixel gradient threshold and the size deviation of the defect areas smaller than the size deviation threshold into the manual segmentation associated defect types;

and the manual segmentation association type recognition module is used for transmitting the chip image to be detected to a manual quality inspection channel to recognize the manual segmentation association defect type when the defect detection quantity of the pixel segmentation association defect type recognition result and the size segmentation association defect type recognition result is zero, and acquiring the manual segmentation association defect type recognition result.

Further, the system further comprises:

the recording data acquisition module is used for transmitting the chip model information and the chip processing technological parameters to the industrial Internet to acquire chip surface defect detection recording data, wherein the industrial Internet is a private network for realizing information sharing for a plurality of chip manufacturing enterprises;

The detection frequency acquisition module is used for acquiring a defect detection record type and a record type detection frequency according to the chip surface defect detection record data;

a record type adding module, configured to add the defect detection record type to the associated defect type when the record type detection frequency is greater than or equal to a first detection frequency threshold;

wherein the defect detection record type and the record type detection frequency are obtained,

the recorded data description module is used for enabling the chip surface defect detection recorded data to comprise first enterprise defect detection recorded data and second enterprise defect detection recorded data to be the nth enterprise defect detection recorded data, wherein the number of the recorded strips of the first enterprise defect detection recorded data and the second enterprise defect detection recorded data to be the nth enterprise defect detection recorded data is the same;

the Nth detection frequency acquisition module is used for acquiring the first detection frequency and the second detection frequency of the defect detection record type until the Nth detection frequency;

an irrelevant detection type obtaining module, configured to set the defect detection record type to an irrelevant detection defect type when the number of the first detection frequency, the second detection frequency, and up to the nth detection frequency is greater than or equal to a floor (0.7×n) that is a downward rounding function;

And the average value calculation module is used for acquiring an average value of the first detection frequency and the second detection frequency until the Nth detection frequency when the number of the first detection frequency and the second detection frequency until the Nth detection frequency is smaller than or equal to a second detection frequency threshold value is larger than or equal to floor (0.7 x N), and setting the average value as the record type detection frequency of the defect detection record type.

Further, the system further comprises:

the system comprises a concept image acquisition module, a pixel gradient threshold value acquisition module and a pixel gradient threshold value acquisition module, wherein the concept image acquisition module is used for acquiring a chip concept image, the chip concept image is a standard chip image, the chip concept image is provided with a plurality of identification areas and a plurality of characteristic pixels, and the deviation of the characteristic pixels of any two identification areas is larger than or equal to the pixel gradient threshold value;

the image cutting module is used for cutting any one of the defect sample image sets based on the plurality of identification areas to obtain a plurality of groups of local sample images, wherein the plurality of groups of local sample images are in one-to-one correspondence with the plurality of identification areas;

the region dividing module is used for dividing the region of the chip image to be detected based on the plurality of identification regions to obtain a plurality of region images to be detected;

The region growing module is used for carrying out region growing on the plurality of region images to be detected according to the plurality of characteristic pixels to obtain a plurality of region growing results, wherein the plurality of region growing results comprise a plurality of growth covered regions and a plurality of growth uncovered regions;

and the defect comparison module is used for comparing the defects of the plurality of growth uncovered areas based on the plurality of groups of local sample images and obtaining the identification result of the pixel segmentation associated defect types.

Further, the system further comprises:

the threshold value acquisition module is used for acquiring a pixel growth similarity threshold value;

the adjacent pixel point acquisition module is used for taking the kth pixel point of the ith to-be-detected area of the plurality of to-be-detected area images as an initial growth pixel point, and acquiring the adjacent pixel point of the initial growth pixel point by taking the boundary of the ith to-be-detected area as a growth constraint boundary;

the traversing module is used for traversing the adjacent pixel points and the initial growing pixel points to evaluate the color similarity and acquire the color similarity;

a first generation growing pixel point obtaining module, configured to set the adjacent pixel points with the color similarity greater than or equal to the pixel growing similarity threshold as first generation growing pixel points;

The adjacent pixel point adding module is used for adding the adjacent pixel points with the color similarity smaller than the pixel growth similarity threshold value into an ith growth uncovered area;

the repeated growth module is used for carrying out repeated growth based on the first generation of growth pixel points, stopping until the pixel point of the ith area to be detected is traversed, and acquiring an ith growth coverage area;

a coverage area adding module, configured to add the ith growth coverage area to the plurality of growth coverage areas, and add the ith growth uncovered area to the plurality of growth uncovered areas.

Further, the system further comprises:

the comparison characteristic pixel acquisition module is used for acquiring comparison characteristic pixels of the ith growing uncovered area;

the defect characteristic pixel acquisition module is used for acquiring M defect characteristic pixel sets of the ith group of local sample images corresponding to the ith to-be-detected area;

the deviation interval construction module is used for acquiring the maximum pixel value deviation value and the minimum pixel value deviation value of the M defect characteristic pixel sets and constructing a pixel deviation interval;

the deviation calculation module is used for traversing the M defect characteristic pixel sets to calculate pixel value deviations based on the comparison characteristic pixels and obtaining M pixel value deviations;

And the proportion coefficient acquisition module is used for acquiring proportion coefficients of the M pixel value deviations belonging to the pixel deviation interval, setting the i growth uncovered area as the associated defect type corresponding to the plurality of groups of local sample images when the proportion coefficients are larger than or equal to a proportion coefficient threshold value, and adding the associated defect type identification result of the pixel segmentation.

The foregoing detailed description of the method for intelligently identifying the surface defects of the chip based on the 5G technology will clearly be known to those skilled in the art, and the device disclosed in the embodiment corresponds to the method disclosed in the embodiment, so that the description is relatively simple, and relevant places refer to the description of the method section.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (8)

1. The intelligent chip surface defect identification method based on the 5G technology is characterized by being applied to an intelligent chip surface defect identification system, wherein the system is connected with the industrial Internet in a communication way based on the 5G technology, and the method comprises the following steps:

transmitting the chip processing technological parameters and the chip model information to the industrial Internet for chip defect correlation analysis, and obtaining the correlation defect type;

transmitting the associated defect type and the chip model information to the industrial Internet for negative sampling, and obtaining a plurality of defect sample images, wherein the plurality of defect sample images are sample images of detected defects, and any one defect sample image is uniquely associated with one associated defect type;

traversing the plurality of defect sample images to perform pixel characteristic analysis to obtain a plurality of defect region pixel gradients, wherein the defect region pixel gradients represent pixel value deviations of a defect region and a normal region of a chip to which the defect belongs;

traversing the pixel gradients of the defect areas, and judging whether the pixel gradients are larger than or equal to a pixel gradient threshold value;

adding the associated defect types with the pixel gradients of the defect areas being greater than or equal to the pixel gradient threshold into a pixel segmentation associated defect type;

Transmitting the pixel segmentation associated defect type and the chip model information to the industrial Internet for negative sampling to obtain a plurality of defect sample image sets;

and performing defect comparison on the chip images to be detected based on the plurality of defect sample image sets, and obtaining a pixel segmentation association defect type identification result.

2. The method as recited in claim 1, further comprising:

traversing the plurality of defect sample images for size analysis to obtain size deviations of a plurality of defect areas;

when the pixel gradient of the plurality of defect areas is smaller than the pixel gradient threshold, judging whether the pixel gradient of the plurality of defect areas is larger than or equal to a size deviation threshold or not based on the size deviation of the plurality of defect areas, wherein the size deviation threshold is the minimum size deviation which can be identified by a size classifier;

adding a size division associated defect type to the associated defect type for which the pixel gradient of the plurality of defect areas is less than the pixel gradient threshold and the size deviation of the plurality of defect areas is greater than or equal to the size deviation threshold;

based on the size division associated defect type and the chip model information, training a size division model to perform defect identification on the chip image to be detected, and obtaining a size division associated defect type identification result, wherein the method comprises the following steps:

Based on the size division associated defect type and the chip model information, acquiring a positive sample image of the defect to be detected and a negative sample image of the defect to be detected, wherein the negative sample image of the defect to be detected has defect size identification information;

and taking the positive sample image of the defect to be detected and the negative sample image of the defect to be detected as input data, taking the defect size identification information as supervision data, and carrying out size defect classification training based on a convolutional neural network to obtain the size segmentation model.

3. The method as recited in claim 2, further comprising:

adding the associated defect types of which the pixel gradients of the defect areas are smaller than the pixel gradient threshold and the size deviation of the defect areas is smaller than the size deviation threshold into the manually-segmented associated defect types;

when the defect detection quantity of the pixel segmentation association defect type identification result and the size segmentation association defect type identification result is zero, transmitting the chip image to be detected to a manual quality inspection channel to identify the manual segmentation association defect type, and acquiring a manual segmentation association defect type identification result.

4. The method of claim 1, wherein transmitting the chip processing process parameters and the chip model information to the industrial internet for chip defect correlation analysis, obtaining the correlation defect type, comprises:

Transmitting the chip model information and the chip processing process parameters to the industrial Internet to obtain chip surface defect detection record data, wherein the industrial Internet is a private network for realizing information sharing for a plurality of chip manufacturing enterprises;

acquiring a defect detection record type and a record type detection frequency according to the chip surface defect detection record data;

adding the defect detection record type into the associated defect type when the record type detection frequency is greater than or equal to a first detection frequency threshold;

wherein the defect detection record type and the record type detection frequency are obtained,

the chip surface defect detection record data comprises first enterprise defect detection record data and second enterprise defect detection record data up to Nth enterprise defect detection record data, wherein the number of record bars from the first enterprise defect detection record data to the second enterprise defect detection record data up to the Nth enterprise defect detection record data is the same;

acquiring a first detection frequency and a second detection frequency of the defect detection record type until an Nth detection frequency;

setting the defect detection record type to be an irrelevant detection defect type when the number of the first detection frequency, the second detection frequency and the second detection frequency up to the Nth detection frequency is smaller than or equal to a second detection frequency threshold value is larger than or equal to floor (0.7 x N), wherein floor () is a downward rounding function;

And when the number of the first detection frequency and the second detection frequency up to the Nth detection frequency which are smaller than or equal to a second detection frequency threshold is larger than or equal to floor (0.7 x N), acquiring the average value of the first detection frequency and the second detection frequency up to the Nth detection frequency, and setting the average value as the recording type detection frequency of the defect detection recording type.

5. The method of claim 1, wherein obtaining a pixel segmentation associated defect type recognition result based on defect comparison of the chip image to be inspected based on the plurality of defect sample image sets, comprises:

acquiring a chip concept image, wherein the chip concept image is a standard chip image, the chip concept image is provided with a plurality of identification areas and a plurality of characteristic pixels, and the deviation of the characteristic pixels of any two identification areas is larger than or equal to the pixel gradient threshold;

based on the plurality of identification areas, performing image shearing on any one defect sample image set of the plurality of defect sample image sets to obtain a plurality of groups of local sample images, wherein the plurality of groups of local sample images are in one-to-one correspondence with the plurality of identification areas;

Based on the plurality of identification areas, carrying out area division on the chip image to be detected to obtain a plurality of area images to be detected;

performing region growing on the plurality of region images to be detected according to the plurality of characteristic pixels to obtain a plurality of region growing results, wherein the plurality of region growing results comprise a plurality of growth covered regions and a plurality of growth uncovered regions;

and performing defect comparison on the plurality of growth uncovered areas based on the plurality of groups of local sample images to obtain the identification result of the pixel segmentation associated defect types.

6. The method of claim 5, wherein performing region growing on the plurality of region images to be inspected according to the plurality of feature pixels, and obtaining a plurality of region growing results, wherein the plurality of region growing results include a plurality of growth covered regions and a plurality of growth uncovered regions, comprising:

acquiring a pixel growth similarity threshold;

taking a kth pixel point of an ith to-be-detected area of the plurality of to-be-detected area images as an initial growing pixel point, and taking the boundary of the ith to-be-detected area as a growth constraint boundary to acquire adjacent pixel points of the initial growing pixel point;

traversing the adjacent pixel points and the initial growing pixel points to evaluate the color similarity, and obtaining the color similarity;

Setting the adjacent pixel points with the color similarity larger than or equal to the pixel growth similarity threshold value as first generation growth pixel points;

adding the adjacent pixel points with the color similarity smaller than the pixel growth similarity threshold value into an ith growth uncovered area;

repeating growth based on the first generation of growth pixels until the pixel point of the ith area to be detected is traversed, and acquiring an ith growth coverage area;

adding the i-th growth coverage area to the plurality of growth coverage areas, and adding the i-th growth uncovering area to the plurality of growth uncovering areas.

7. The method of claim 5, wherein performing defect comparison on the plurality of grown uncovered areas based on the plurality of sets of local sample images to obtain the pixel segmentation associated defect type identification result comprises:

obtaining comparison characteristic pixels of an ith growth uncovered area;

obtaining M defect characteristic pixel sets of an ith group of local sample images corresponding to an ith region to be detected;

obtaining a pixel value deviation maximum value and a pixel value deviation minimum value of the M defect characteristic pixel sets, and constructing a pixel deviation interval;

Traversing the M defect characteristic pixel sets to calculate pixel value deviations based on the comparison characteristic pixels, and obtaining M pixel value deviations;

and acquiring the proportionality coefficient of the M pixel value deviations belonging to the pixel deviation interval, setting the ith growing uncovered area as the associated defect type corresponding to the plurality of groups of local sample images when the proportionality coefficient is larger than or equal to a proportionality coefficient threshold value, and adding the associated defect type identification result of the pixel segmentation.

8. The chip surface defect intelligent recognition system based on the 5G technology is characterized by being applied to the chip surface defect intelligent recognition system, wherein the system is based on the 5G technology and industrial Internet communication connection and is used for implementing the chip surface defect intelligent recognition method based on the 5G technology as claimed in any one of claims 1 to 7, and comprises the following steps:

the correlation analysis module is used for transmitting the chip processing technological parameters and the chip model information to the industrial Internet to perform chip defect correlation analysis and obtain the correlation defect type;

the negative sampling module is used for transmitting the associated defect type and the chip model information to the industrial Internet for negative sampling to obtain a plurality of defect sample images, wherein the plurality of defect sample images are sample images with detected defects, and any one defect sample image is uniquely associated with one associated defect type;

The pixel characteristic analysis module is used for traversing the plurality of defect sample images to perform pixel characteristic analysis and obtaining a plurality of defect area pixel gradients, wherein the defect area pixel gradients represent pixel value deviations of a defect area and a chip normal area to which the defect belongs;

the pixel gradient judging module is used for traversing the pixel gradients of the defect areas and judging whether the pixel gradient is larger than or equal to a pixel gradient threshold value or not;

a correlation type adding module for adding the correlation defect types with the pixel gradients of the plurality of defect areas being greater than or equal to the pixel gradient threshold into a pixel segmentation correlation defect type;

the sample image acquisition module is used for transmitting the pixel segmentation association defect type and the chip model information to the industrial Internet for negative sampling to acquire a plurality of defect sample image sets;

and the defect comparison module is used for comparing the defects of the chip images to be detected based on the plurality of defect sample image sets and obtaining a pixel segmentation associated defect type identification result.