CN115424099B - Model training method, recognition method and device for recognizing silicon carbide dislocation - Google Patents
Model training method, recognition method and device for recognizing silicon carbide dislocation Download PDFInfo
- Publication number
- CN115424099B CN115424099B CN202211383066.4A CN202211383066A CN115424099B CN 115424099 B CN115424099 B CN 115424099B CN 202211383066 A CN202211383066 A CN 202211383066A CN 115424099 B CN115424099 B CN 115424099B
- Authority
- CN
- China
- Prior art keywords
- silicon carbide
- dislocation
- picture samples
- picture
- dislocations
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- HBMJWWWQQXIZIP-UHFFFAOYSA-N silicon carbide Chemical compound [Si+]#[C-] HBMJWWWQQXIZIP-UHFFFAOYSA-N 0.000 title claims abstract description 358
- 229910010271 silicon carbide Inorganic materials 0.000 title claims abstract description 358
- 238000012549 training Methods 0.000 title claims abstract description 68
- 238000000034 method Methods 0.000 title claims abstract description 67
- 238000001514 detection method Methods 0.000 claims abstract description 76
- 238000003062 neural network model Methods 0.000 claims abstract description 40
- 238000002372 labelling Methods 0.000 claims abstract description 39
- 235000012431 wafers Nutrition 0.000 claims description 59
- XUIMIQQOPSSXEZ-UHFFFAOYSA-N Silicon Chemical compound [Si] XUIMIQQOPSSXEZ-UHFFFAOYSA-N 0.000 claims description 3
- 229910052710 silicon Inorganic materials 0.000 claims description 3
- 239000010703 silicon Substances 0.000 claims description 3
- 238000012545 processing Methods 0.000 abstract description 7
- 230000007547 defect Effects 0.000 description 15
- 238000010586 diagram Methods 0.000 description 9
- 238000005498 polishing Methods 0.000 description 6
- 238000012876 topography Methods 0.000 description 5
- 239000013078 crystal Substances 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 239000002585 base Substances 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 239000003086 colorant Substances 0.000 description 2
- 230000007797 corrosion Effects 0.000 description 2
- 238000005260 corrosion Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 239000003513 alkali Substances 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 230000005684 electric field Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000704 physical effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000012805 post-processing Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 239000000758 substrate Substances 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/778—Active pattern-learning, e.g. online learning of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30148—Semiconductor; IC; Wafer
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of silicon carbide, and discloses a model training method, a recognition method and a device for recognizing silicon carbide dislocation.A first silicon carbide dislocation picture sample with clean background and no overlap of different dislocations is obtained and marked, and then a 1 st generation detection model is obtained by training through a neural network model; acquiring a plurality of second silicon carbide dislocation picture samples with clean backgrounds and dislocation overlapping and staggering, detecting through a 1 st generation detection model, then labeling, and training the plurality of first silicon carbide dislocation picture samples and the second silicon carbide dislocation picture samples after labeling through a neural network model to obtain a 2 nd generation detection model; and obtaining a plurality of third silicon carbide dislocation picture samples with unclean backgrounds, and obtaining a 3 rd generation detection model according to the steps. By adopting the invention, under the condition of ensuring the identification accuracy, the time and labor cost are greatly saved, more than 30% of wafer processing cost is saved, and the detection cost is further reduced.
Description
Technical Field
The invention relates to the technical field of silicon carbide, in particular to a model training method, a recognition method and a device for recognizing silicon carbide dislocation.
Background
Silicon carbide is used as a third-generation compound semiconductor material, and is widely applied to the fields of aviation, military affairs and new energy automobiles due to excellent physical properties of wide forbidden band, high breakdown electric field intensity, high thermal conductivity, high electron mobility and the like; the silicon carbide adopts a PVT method to carry out a crystal growth process, and defects are easily introduced in the growth process; the type of dislocation and its dislocation density determine the quality of the substrate sheet; and high defect density can have a fatal effect on subsequent device performance.
Generally speaking, in order to monitor the crystal growth process, each crystal growth furnace periodically cuts off the head, middle and tail of a silicon carbide ingot at regular intervals, removes surface scratch damage through a series of processes, then corrodes in molten alkali for a certain time, then cleans the silicon carbide ingot, and places the silicon carbide ingot under a microscope to count the defect density and the number of defects in unit area, and the current detection mode includes manual detection and pattern recognition software based on machine learning. The manual detection speed is low, and the method is only suitable for small-batch pictures and a small amount of sparse defects in the pictures; however, since the defects on a wafer have at least 200-1000 points, accurate models are required to be obtained by training by using pattern recognition software based on machine learning, at least 1000 pictures are generally required, one picture needs to be marked manually for one day, and a large amount of time and labor are consumed for completing the marking; moreover, if the method completely depends on manual annotation, the quality of the picture annotation can influence the training result of the model; moreover, the classification of different defect types and the marking of overlapping defect types are based on personal professional experience to a great extent; if the picture is marked outside, not only the marking cost is increased, but also a certain degree of leakage is caused. And the model training time of the traditional machine learning method is long, a large amount of hardware resources are needed, if pictures with sizes corresponding to 1024x1024 pixels are detected after deployment, the pictures with the sizes corresponding to the 1024x1024 pixels need to be input for training, and the larger the sizes of the training pictures are, more hardware resources and longer training time are needed.
Moreover, the model formed according to the prior art has poor background resistance and scratch resistance, is very easy to cause misjudgment, and cannot be identified or wrongly identified; for example, small-size defects are not easy to identify, the scratch resistance is poor, scratches are identified into BPD defects, so that the BPD count is larger, the background difference between the high-purity semi-insulating sheet and the conducting sheet after corrosion is different in darkness, different corrosion processes cause different sizes of the defects, and the depths of the scratches introduced by different processing processes such as polishing sheets and fine polishing sheets are different; the detection cost of the fine polishing piece is high, the traditional training method is adopted, in order to improve the identification accuracy as much as possible, a dislocation picture with a clean background and no scratch needs to be corroded by the fine polishing piece, the fine polishing piece needs a CMP (chemical mechanical polishing) process, and the cost of the CMP process accounts for one third of the cost of the whole wafer processing process; a wafer processing procedure: cutting-thinning-grinding-lapping-CMP (33% cost); the detection time of a single picture is long, for a large-size high-pixel picture, the identification time of each picture is more than 0.5 second, and at least 2 hours are needed after the test of one wafer; moreover, if a new process is introduced, and a new process picture has a large difference from a previous training picture, the recognition rate of the model to the new process picture is poor, a picture with a low recognition rate needs to be selected, picture labeling is performed again, the newly labeled picture and the previously labeled picture are input into the neural model together, retraining is performed, and a complete training process needs at least 24 hours, even multiple working days.
Disclosure of Invention
The invention aims to solve the problems of complicated training process, long time and poor detection effect of the existing neural network model for identifying silicon carbide dislocation, and provides a model training method, an identification method and a device for identifying silicon carbide dislocation.
In order to achieve the above object, the present invention provides a neural network model training method for identifying silicon carbide dislocations, comprising the steps of:
obtaining a plurality of first silicon carbide dislocation picture samples, wherein the background in the first silicon carbide dislocation picture samples is clean and different dislocations are not overlapped; marking dislocations in each first silicon carbide dislocation picture sample to obtain a plurality of first silicon carbide dislocation picture samples after marking; training the marked first silicon carbide dislocation picture samples through a neural network model to obtain a 1 st generation detection model;
acquiring a plurality of second silicon carbide dislocation picture samples, wherein the backgrounds in the second silicon carbide dislocation picture samples are clean and dislocation overlapping staggering exists; detecting the dislocation in each second silicon carbide dislocation picture sample through the 1 st generation detection model, and labeling the dislocation according to the detection result to obtain a plurality of labeled second silicon carbide dislocation picture samples; training the marked first silicon carbide dislocation picture samples and the marked second silicon carbide dislocation picture samples through a neural network model to obtain a 2 nd generation detection model;
acquiring a plurality of third silicon carbide dislocation picture samples, wherein the backgrounds in the third silicon carbide dislocation picture samples are unclean; detecting the dislocation in each third silicon carbide dislocation picture sample through the 2 nd generation detection model, and labeling the dislocation according to the detection result to obtain a plurality of labeled third silicon carbide dislocation picture samples; and training the marked first silicon carbide dislocation picture samples, the marked second silicon carbide dislocation picture samples and the marked third silicon carbide dislocation picture samples through a neural network model to obtain a 3 rd generation detection model.
As an implementation manner, after obtaining the 3 rd generation detection model, the method further includes: acquiring a plurality of fourth silicon carbide dislocation picture samples, wherein the fourth silicon carbide dislocation picture samples have influence factors further causing dislocation identification errors; and according to the steps, continuing further training to obtain a 4 th generation detection model.
As an implementation manner, the step of respectively obtaining a plurality of first silicon carbide dislocation picture samples, a plurality of second silicon carbide dislocation picture samples, and a plurality of third silicon carbide dislocation picture samples specifically includes:
acquiring a plurality of silicon carbide dislocation pictures, and cutting each silicon carbide dislocation picture to obtain a plurality of silicon carbide dislocation picture samples; and respectively selecting a plurality of first silicon carbide dislocation picture samples, a plurality of second silicon carbide dislocation picture samples and a plurality of third silicon carbide dislocation picture samples from the plurality of Zhang Tanhua silicon dislocation picture samples.
As an implementable mode, the number of dislocations in each of the first silicon carbide dislocation picture sample, the second silicon carbide dislocation picture sample and the third silicon carbide dislocation picture sample is not more than 10; the specific number of the first, second and third silicon carbide dislocation picture samples is 200-500.
As an implementation manner, labeling the dislocations in each first silicon carbide dislocation picture sample, and obtaining the labeled first silicon carbide dislocation picture samples specifically includes:
and manually labeling different dislocations in the first silicon carbide dislocation picture samples based on labeling software to obtain the labeled first silicon carbide dislocation picture samples.
As an implementation manner, labeling the dislocations in each first silicon carbide dislocation picture sample, and obtaining the labeled first silicon carbide dislocation picture samples specifically includes:
manually marking different dislocations in the multiple first silicon carbide dislocation picture samples based on marking software, and training the manually marked first silicon carbide dislocation picture samples through a neural network model to obtain a 0 th generation model; detecting a plurality of first silicon carbide dislocation picture samples one by using the 0 th generation model to obtain dislocation types and coordinate positions in each first silicon carbide dislocation picture sample; according to the dislocation types and the coordinate positions of the detected first silicon carbide dislocation picture samples, different dislocations in the first silicon carbide dislocation picture samples are automatically marked through marking software and are adjusted in an auxiliary mode through manual marking, and the marked first silicon carbide dislocation picture samples are obtained.
As an implementation manner, the step of detecting the dislocations in each second silicon carbide dislocation picture sample by using the generation 1 detection model, labeling the dislocations according to the detection result, and obtaining a plurality of labeled second silicon carbide dislocation picture samples specifically includes:
detecting a plurality of second silicon carbide dislocation picture samples by using the model of generation 1 to obtain dislocation types and coordinate positions in each second silicon carbide dislocation picture sample;
and according to the dislocation types and coordinate positions of different dislocations in the detected second silicon carbide dislocation picture samples, automatically marking the different dislocations in the second silicon carbide dislocation picture samples through marking software and performing auxiliary adjustment through manual marking to obtain the marked second silicon carbide dislocation picture samples.
As an implementation manner, the step of acquiring multiple silicon carbide dislocation pictures specifically includes:
providing a silicon carbide wafer, placing the silicon carbide wafer on a chuck of a microscope, pulling the chuck to move according to a set step length through a stepping motor, and scanning and shooting the silicon carbide wafer after the microscope moves one step length at a time to generate a plurality of silicon carbide dislocation pictures; the silicon carbide dislocation picture shot after moving by one step length is named by the coordinate corresponding to the number of the horizontal and vertical moving step lengths; and when the diameter of the silicon carbide wafer is D, the step length set by the stepping motor is D/X1 for each movement in the X direction, and D/Y1 for each movement in the Y direction, the width of the scanned area shot in the X direction is Y1 for each movement of the chuck in the X direction, and the width of the scanned area shot in the Y direction is X1 for each movement of the chuck in the Y direction, so that the silicon carbide wafer images corresponding to the silicon carbide wafers formed after the silicon carbide images obtained by shooting are spliced keep a circular shape.
Correspondingly, the invention also provides a silicon carbide wafer image dislocation recognition method based on the neural network model training method, which comprises the following steps:
acquiring the 3 rd generation detection model and a plurality of silicon carbide dislocation pictures shot based on the silicon carbide wafer, and detecting the plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain dislocation types and coordinate positions in each silicon carbide dislocation picture;
according to the dislocation types and the coordinate positions of different dislocations in the detected multiple silicon carbide dislocation pictures, automatically marking the different dislocations in the multiple silicon carbide dislocation picture samples through marking software and performing auxiliary adjustment through manual marking to obtain multiple marked silicon carbide dislocation pictures;
and splicing the marked silicon carbide dislocation pictures according to the coordinates to obtain the marked silicon carbide wafer image.
As an implementation manner, the step of splicing the marked multiple silicon carbide dislocation pictures according to the coordinates specifically includes:
when a plurality of silicon carbide dislocation pictures are obtained after a stepping motor scans X1 columns and Y1 rows, setting the length-width ratio of each silicon carbide dislocation picture in the X direction and the Y direction as Y1: and X1, splicing each processed silicon carbide dislocation picture according to coordinates, so that after the shot silicon carbide pictures are spliced, the length and the width of the formed silicon carbide wafer images in the XY direction are consistent.
Correspondingly, the invention also provides a silicon carbide wafer image dislocation recognition device based on the neural network model training method, which comprises the following modules:
the acquisition module is used for acquiring the 3 rd generation detection model and a plurality of silicon carbide dislocation pictures shot based on the silicon carbide wafer, and detecting the plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain dislocation types and coordinate positions in each silicon carbide dislocation picture;
the marking module is used for automatically marking different dislocations in the samples of the multiple silicon carbide dislocation pictures through marking software according to the detected dislocation types and coordinate positions of the different dislocations in the multiple silicon carbide dislocation pictures and performing auxiliary adjustment through manual marking to obtain the marked multiple silicon carbide dislocation pictures;
and the splicing module is used for splicing the marked multiple silicon carbide dislocation pictures to obtain the marked silicon carbide wafer image.
The invention has the beneficial effects that: the invention provides a model training method, a recognition method and a device for recognizing silicon carbide dislocation.A first silicon carbide dislocation picture sample with clean background and non-overlapping different dislocations is obtained and marked, and then a 1 st generation detection model is obtained by training through a neural network model; acquiring a plurality of second silicon carbide dislocation picture samples with clean backgrounds and dislocation overlapping and staggering, detecting through a 1 st generation detection model, then labeling, and training the plurality of first silicon carbide dislocation picture samples and the second silicon carbide dislocation picture samples after labeling through a neural network model to obtain a 2 nd generation detection model; and obtaining a plurality of third silicon carbide dislocation picture samples with unclean backgrounds, and obtaining a 3 rd generation detection model according to the steps. By adopting the invention, under the condition of ensuring the identification accuracy, the time and labor cost are greatly saved, more than 30% of wafer processing cost is saved, and the detection cost is further reduced.
The invention can directly use the fine grinding sheet before the CMP process without using the fine grinding sheet obtained by the CMP process, and can save the wafer processing cost by more than 30 percent, thereby reducing the detection cost and greatly reducing the identification time of a single picture, wherein the time is less than 0.01 second in a GPU environment and less than 0.1 second in a CPU environment.
According to the invention, under the condition of ensuring the identification accuracy, small-size picture samples are obtained by cutting the silicon carbide dislocation pictures, the number of dislocations on each picture sample is small, and the silicon carbide dislocation picture samples with different influence factors are selected for training in a grading manner, so that a plurality of large pictures with a large number of dislocations do not need to be marked and put into a neural network model for training at one time as in the prior art, the training time is greatly reduced, and the accuracy of the model obtained by training is ensured.
Drawings
FIG. 1 is a schematic diagram of the steps of a neural network model training method for identifying silicon carbide dislocations according to an embodiment of the present invention;
fig. 2 is a dislocation labeling schematic diagram of a first silicon carbide dislocation picture sample labeled in the neural network model training method for identifying silicon carbide dislocations according to the embodiment of the present invention;
fig. 3 is a dislocation labeling schematic diagram of a second silicon carbide dislocation picture sample labeled in the neural network model training method for identifying silicon carbide dislocations according to the embodiment of the present invention;
4-6 are schematic dislocation labeling diagrams of a third silicon carbide dislocation picture sample labeled in the neural network model training method for identifying silicon carbide dislocations according to the embodiment of the present invention;
fig. 7 is a schematic dislocation labeling diagram of a large-size image labeled in the neural network model training method for identifying silicon carbide dislocations in the embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, the present embodiment provides a technical solution: a neural network model training method for identifying silicon carbide dislocations, comprising the steps of:
step S100, acquiring a plurality of first silicon carbide dislocation picture samples, wherein the backgrounds in the first silicon carbide dislocation picture samples are clean and different dislocations are not overlapped; marking dislocations in each first silicon carbide dislocation picture sample to obtain a plurality of first silicon carbide dislocation picture samples after marking; training the marked first silicon carbide dislocation picture samples through a neural network model to obtain a 1 st generation detection model;
step S200, a plurality of second silicon carbide dislocation picture samples are obtained, wherein the backgrounds of the second silicon carbide dislocation picture samples are clean and dislocation overlapping and staggering exist; detecting the dislocation in each second silicon carbide dislocation picture sample through the 1 st generation detection model, and labeling the dislocation according to the detection result to obtain a plurality of labeled second silicon carbide dislocation picture samples; training the marked first silicon carbide dislocation picture samples and the marked second silicon carbide dislocation picture samples through a neural network model to obtain a 2 nd generation detection model;
step S300, obtaining a plurality of third silicon carbide dislocation picture samples, wherein the backgrounds in the third silicon carbide dislocation picture samples are not clean; detecting the dislocation in each third silicon carbide dislocation picture sample through the 2 nd generation detection model, and labeling the dislocation according to the detection result to obtain a plurality of labeled third silicon carbide dislocation picture samples; and training the marked first silicon carbide dislocation picture samples, the marked second silicon carbide dislocation picture samples and the marked third silicon carbide dislocation picture samples through a neural network model to obtain a 3 rd generation detection model.
The clear background means that no influence factors such as background scratches, background noise, interference spots and the like exist, so that dislocation characteristics can be accurately identified; the unclean background means that there are background scratches, background noise, interference spots, and the like, which may cause a dislocation recognition error.
Fig. 2 is a schematic diagram showing dislocation marks of a first silicon carbide dislocation picture sample marked with a clean background and no overlapping of different dislocations, which includes three kinds of dislocations, namely, a base plane dislocation BPD, a threading screw dislocation TSD and a threading edge dislocation TED, fig. 3 is a schematic diagram showing dislocation marks of a second silicon carbide dislocation picture sample marked with a clean background and overlapping and interleaving of dislocations, which includes a threading edge dislocation TED, which is shown in fig. 4-6 are schematic diagrams showing dislocations marked with a third silicon carbide dislocation picture sample marked with an unclean background, which includes three kinds of dislocations, namely, a base plane dislocation BPD, a threading screw dislocation TSD and a threading edge dislocation TED, which are shown in fig. 4, and fig. 5 and 6 both include two kinds of dislocations, namely, a threading edge dislocation TED and a threading screw dislocation TSD.
In this embodiment, a user may customize the defect deletion/addition type according to needs, in a special case, the user may adjust to focus on only a certain dislocation type, define three kinds of dislocations, i.e., a base plane dislocation BPD, a threading dislocation TSD, and a threading edge dislocation TED, and may further subdivide the three kinds of typical dislocations, i.e., BPD/TED/TSD/TED/BPD/others, into TMD/MPD/TSD/TED/BPD/others, which is not limited in this embodiment.
Further, after obtaining the 3 rd generation detection model, the method further includes: acquiring a plurality of fourth silicon carbide dislocation picture samples, wherein the fourth silicon carbide dislocation picture samples have influence factors further causing dislocation identification errors; and according to the steps, continuing further training to obtain a 4 th generation detection model.
In this embodiment, when a new process is introduced, there are influence factors that further cause dislocation recognition errors, and because a new process picture is greatly different from a previous training picture, the recognition rate of the new process picture by the model is poor, therefore, in this embodiment, the picture can be cut, small pictures with low recognition rate can be selected, the small pictures are directly input to a 3 rd generation model, the pictures are recognized through the 3 rd generation model to obtain detection results corresponding to dislocations, then the pictures are labeled by using labeling software according to the detection results and manually adjusted, the labeled pictures and the previously labeled pictures are input to a neural model together, and a 4 th generation model is obtained by retraining, so that when the new process is adopted, manual labeling of the new pictures is not required again, and a large amount of time and labor are saved.
In this embodiment, the step of respectively obtaining a plurality of first silicon carbide dislocation picture samples, a plurality of second silicon carbide dislocation picture samples, and a plurality of third silicon carbide dislocation picture samples specifically includes:
acquiring a plurality of silicon carbide dislocation pictures, and cutting each silicon carbide dislocation picture to obtain a plurality of silicon carbide dislocation picture samples; and respectively selecting a plurality of first silicon carbide dislocation picture samples, a plurality of second silicon carbide dislocation picture samples and a plurality of third silicon carbide dislocation picture samples from the plurality of Zhang Tanhua silicon dislocation picture samples.
The dislocation number in each first silicon carbide dislocation picture sample, each second silicon carbide dislocation picture sample and each third silicon carbide dislocation picture sample is not more than 10, so that the manual labeling time is saved; optionally, the number of the first silicon carbide dislocation picture sample, the second silicon carbide dislocation picture sample, and the third silicon carbide dislocation picture sample is 200-500, and may also be adjusted.
In the embodiment, after a large-size picture is cut into a plurality of small-size pictures, the number of defects is greatly reduced, the pictures are favorably marked manually, the marking correctness is favorably confirmed manually, in addition, the training of the small-size pictures is adopted, the training can be completed within 2 to 12 hours in the same hardware resources according to the model structure and the size of training data, and the training time can be shortened by more than 50 percent.
In this embodiment, labeling dislocations in each first silicon carbide dislocation picture sample, and obtaining a plurality of labeled first silicon carbide dislocation picture samples specifically includes:
and manually marking different dislocations in the multiple first silicon carbide dislocation picture samples based on marking software to obtain the marked multiple first silicon carbide dislocation picture samples.
It should be noted that in this embodiment, the professional uses the labeling software to label the silicon carbide dislocation picture sample with a non-color box, for example: the red frame is TSD, the blue frame is TED, and the purple frame is BPD.
Further, in the manual labeling, since the manual labeling largely depends on the experience of the labeling person, and whether the labeling frame is completely attached is not easy to be checked, as another embodiment, the step of labeling the dislocations in each first silicon carbide dislocation picture sample to obtain the labeled first silicon carbide dislocation picture samples specifically includes:
manually marking different dislocations in the multiple first silicon carbide dislocation picture samples based on marking software, and training the manually marked first silicon carbide dislocation picture samples through a neural network model to obtain a 0 th generation model; detecting a plurality of first silicon carbide dislocation picture samples one by using the 0 th generation model to obtain dislocation types and coordinate positions in each first silicon carbide dislocation picture sample; according to the dislocation types and the coordinate positions of the detected first silicon carbide dislocation picture samples, different dislocations in the first silicon carbide dislocation picture samples are automatically marked through marking software and are adjusted in an auxiliary mode through manual marking, and the marked first silicon carbide dislocation picture samples are obtained.
It can be seen that, in this embodiment, after carrying out manual marking to first silicon carbide dislocation picture sample earlier, input neural network model training obtains 0 th generation detection model, uses 0 th generation detection model to detect first silicon carbide dislocation picture sample again, carries out automatic marking and adjustment to the testing result through the marking software again for the mark of the many first silicon carbide dislocation picture samples that obtain is more accurate, the mark frame is also laminated more.
Wherein, carry out automatic marking through labeling software also the system adopts not marking with the color frame, for example: the red frame is TSD, the blue frame is TED, and the purple frame is BPD.
In this embodiment, the step of detecting dislocations in each second silicon carbide dislocation picture sample by using the generation 1 detection model, and labeling the dislocations according to the detection result to obtain a plurality of second silicon carbide dislocation picture samples after labeling specifically includes:
detecting a plurality of second silicon carbide dislocation picture samples by using the generation-1 model to obtain dislocation types and coordinate positions in each second silicon carbide dislocation picture sample;
according to the dislocation types and the coordinate positions of different dislocations in the detected second silicon carbide dislocation picture samples, automatically marking the different dislocations in the second silicon carbide dislocation picture samples through marking software and carrying out auxiliary adjustment through manual marking to obtain the marked second silicon carbide dislocation picture samples.
Detecting the dislocation in each third silicon carbide dislocation picture sample through the 2 nd generation detection model, labeling the dislocation according to the detection result, and obtaining a plurality of labeled third silicon carbide dislocation picture samples specifically comprises the following steps:
detecting a plurality of third silicon carbide dislocation picture samples by using the 2 nd generation model to obtain dislocation types and coordinate positions of different dislocations in each third silicon carbide dislocation picture sample;
and automatically labeling different dislocations in the third silicon carbide dislocation picture samples through labeling software, and performing auxiliary adjustment through manual labeling to obtain the labeled third silicon carbide dislocation picture samples.
In this embodiment, the auxiliary adjustment is performed through manual labeling mainly for adjusting labeling frames and the like which may occur frequently, the neural network model includes a plurality of convolutional layers/pooling layers/nonlinear active layers/full-link linear layers and the like, the learning rate lr =0.01 to 0.02, the training round number epoch is greater than 200, the model indexes are mAP@0.5>90% mAP@0.5-0.9 >90%, P >90%, and R >90%.
Specifically, in the prior art, since each scratch in the silicon carbide dislocation picture corresponds to two parallel lines, when a small black dot appears between the parallel lines, the detection model may erroneously identify the black dot and the adjacent two parallel lines as a BPD because the shape of the BPD is that the two lines converge to a point;
therefore, in order to solve the above problems, the detection model of this embodiment introduces hough transform, firstly detects the lines in the silicon carbide dislocation picture through hough transform, and then after hough transform, extracts the features of dislocations in the line range through a series of convolutions, and identifies whether dislocations fall around the lines and are right on the lines, so that the line parts without dislocations can be excluded, thereby excluding the interference of scratches.
In the present embodiment, three kinds of dislocations are distinguished by extracting the contours, dislocation areas, dislocation gradations, dislocation irregularities, and the like of TSD, TED, and BPD, which are three kinds of dislocations, by arranging a series of convolution kernels.
After the dislocation is identified, the silicon carbide dislocation picture sample is detected through the detection model, the coordinates of the upper, lower, left and right vertexes of the corresponding dislocation in the silicon carbide dislocation picture sample can be automatically identified, the coordinates of the vertexes are sent to the labeling software, and the labeling software draws a rectangular frame according to the coordinates of the upper, lower, left and right vertexes of the corresponding dislocation; common open source labeling software includes: offline labeling software labelme, online labeling makesense.
And when the manual marking is carried out by using marking software, the picture frame is directly and manually carried out on the corresponding dislocation.
In the embodiment, machine learning is introduced to assist in marking the picture in the picture marking stage, so that the efficiency is higher, the marking is more accurate, the verification of the marked picture is facilitated, the marking cost is reduced, and the leakage of key information is prevented.
In this embodiment, the step of obtaining a plurality of silicon carbide dislocation pictures includes:
providing a silicon carbide wafer, placing the silicon carbide wafer on a chuck of a microscope, driving the silicon carbide wafer to move under the traction of a stepping motor to the chuck, scanning X1 columns and Y1 rows to obtain X1X Y1 silicon carbide dislocation pictures corresponding to the silicon carbide wafer, and naming the shot silicon carbide dislocation pictures by coordinates consisting of the corresponding row number and column number; e.g., 1 u 4.Jpeg, representing the picture at the row 1, column 4 position.
After the 3 rd generation detection model is obtained, the embodiment detects a plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain the dislocation types and the coordinate positions in each silicon carbide dislocation picture; according to the dislocation types and the coordinate positions of different dislocations in the detected multiple silicon carbide dislocation pictures, automatically marking the different dislocations in the multiple silicon carbide dislocation picture samples through marking software and performing auxiliary adjustment through manual marking to obtain multiple marked silicon carbide dislocation pictures; and splicing the marked silicon carbide dislocation pictures according to the coordinates to obtain the marked silicon carbide wafer image.
Furthermore, due to the weak post-processing function of the pictures, after a plurality of marked silicon carbide dislocation pictures are obtained, the silicon carbide dislocation pictures obtained by shooting with different row and column numbers cannot be reconstructed at will, for example, the scanning parameters of the silicon carbide wafer are as follows; the diameter of the silicon carbide wafer is D, the silicon carbide wafer needs to be shot in 40 rows and 40 columns under the traction of a stepping motor, namely, 40X 40 pictures need to be shot, wherein a circular shape map can be pieced up due to the fact that the number of rows and columns of the silicon carbide wafer moving in the XY direction is equal, and the length and the width of the formed silicon carbide wafer images in the XY direction are consistent, but if the number of rows and columns of the silicon carbide wafer moving in the XY direction is not equal, as 80 rows and 60 columns need to be shot, gaps may exist between the pictures shot in each step length of the stepping motor and the pictures shot at the last time, so after scanning of different rows and columns in the X direction and the Y direction, the obtained silicon carbide dislocation pictures can be pieced into an oval dislocation distribution map, and the length and the width of the formed silicon carbide wafer images in the XY direction are inconsistent; the present embodiment thus provides a solution, specifically as follows:
when a plurality of silicon carbide dislocation pictures are obtained after a stepping motor scans X1 columns and Y1 rows, setting the length-width ratio of each silicon carbide dislocation picture in the X direction and the Y direction as Y1: x1, splicing each processed silicon carbide dislocation picture according to coordinates, so that after the shot silicon carbide pictures are spliced, the length and the width of the formed silicon carbide wafer image in the XY direction are consistent, and the silicon carbide wafer in the silicon carbide wafer image is circular; for example, when the diameter of the silicon carbide wafer is D, the length and width of each silicon carbide dislocation picture in the X direction and the Y direction may be set to D/X1 and D/Y1, respectively.
Furthermore, the embodiment adopts a self-adaptive algorithm, and can reconstruct the dislocation density topography with different scanning steps at will, for example, when scanning a wafer with a diameter D, a circular topography map can be pieced up by scanning and photographing 40 rows and 40 columns, and if the scanning steps in the XY direction are not consistent, a circular topography map can still be pieced up by scanning and photographing 80 rows and 60 columns; the algorithm is as follows:
clustering the dislocation densities of all the shot silicon carbide dislocation pictures, dividing the dislocation densities into 6 grades, and expressing the numerical values of each grade by different colors to generate corresponding legends; generating a large-size blank picture, enabling each silicon carbide dislocation picture to correspond to one area on the blank picture based on the coordinate of the silicon carbide dislocation picture, representing the dislocation density of different areas by using different colors based on a legend, and setting the length ratio of each silicon carbide picture in the X direction and the Y direction as Y1: and X1, the length and the width of the spliced silicon carbide wafer images in the XY direction are consistent, so that the silicon carbide wafers in the obtained silicon carbide wafer images are always kept in a circular shape after a plurality of silicon carbide images are shot and spliced, and a large color image with dislocation density is obtained.
In this embodiment, the names of the single silicon carbide dislocation pictures include the position information of the single silicon carbide dislocation pictures in the whole silicon carbide wafer, and the defect density of the single silicon carbide dislocation pictures is combined, so that the dislocation topography distribution of the whole silicon carbide dislocation picture can be spliced through software, and the adjustment of the crystal growth thermal field is facilitated according to the wafer topography distribution map, and the ingot quality is improved.
However, in another embodiment, the microscopically scanned pictures may be pieced together into one giant picture, the single picture may be sequentially read, the tensor representation may be converted, the tensors of the pictures may be pieced together to generate one large tensor, and the large tensor may be transmitted to the model detection, and the model may detect all defects at once by using the strong calculation power of the GPU.
In this embodiment, after training a model through multiple iterations, the model can be used for detecting a large-size picture, and as shown in fig. 7, a schematic diagram of a dislocation detection result of the large-size picture shows that two kinds of dislocations, namely threading edge dislocation TED and threading dislocation TSD, are included in the detected result.
It should be noted that, in this embodiment, the magnification of fig. 2 to 7 is 100 times, and the scale is 1.2 μm/pixel.
In the invention, by using the training method provided by the invention, the precision polished wafer obtained by the CMP process is not needed, but the precision polished wafer before the CMP process is directly used, so that the wafer processing cost can be saved by more than 30 percent under the condition of ensuring the identification accuracy, and further, the detection cost is reduced, the identification time of a single picture is less than 0.01 second under the GPU environment, and is less than 0.1 second under the CPU environment.
Based on the same inventive concept, the invention also provides a silicon carbide wafer image dislocation recognition method based on the neural network model training method, which comprises the following steps:
acquiring the 3 rd generation detection model and a plurality of silicon carbide dislocation pictures shot based on the silicon carbide wafer, and detecting the plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain dislocation types and coordinate positions in each silicon carbide dislocation picture;
according to the dislocation types and the coordinate positions of different dislocations in the detected multiple silicon carbide dislocation pictures, automatically marking the different dislocations in the multiple silicon carbide dislocation picture samples through marking software and performing auxiliary adjustment through manual marking to obtain multiple marked silicon carbide dislocation pictures;
and splicing the marked multiple silicon carbide dislocation pictures according to the coordinates to obtain the marked silicon carbide wafer image.
Based on the same invention concept, the invention also provides a silicon carbide wafer image dislocation recognition device based on the neural network model training method, which comprises the following modules:
the acquisition module is used for acquiring the 3 rd generation detection model and a plurality of silicon carbide dislocation pictures shot based on the silicon carbide wafer, and detecting the plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain dislocation types and coordinate positions in each silicon carbide dislocation picture;
the marking module is used for automatically marking different dislocations in the samples of the multiple silicon carbide dislocation pictures through marking software according to the detected dislocation types and coordinate positions of the different dislocations in the multiple silicon carbide dislocation pictures and performing auxiliary adjustment through manual marking to obtain the marked multiple silicon carbide dislocation pictures;
and the splicing module is used for splicing the marked multiple silicon carbide dislocation pictures to obtain the marked silicon carbide wafer image.
The step of splicing the marked multiple silicon carbide dislocation pictures according to the coordinates specifically comprises the following steps:
when a plurality of silicon carbide dislocation pictures are obtained after a stepping motor scans X1 columns and Y1 rows, setting the length-width ratio of each silicon carbide dislocation picture in the X direction and the Y direction as Y1: and X1, splicing each processed silicon carbide dislocation picture according to coordinates, so that after the shot silicon carbide pictures are spliced, the length and the width of the formed silicon carbide wafer images in the XY direction are consistent.
Although the present invention has been described with reference to the preferred embodiments, it is not intended to limit the present invention, and those skilled in the art can make possible variations and modifications of the present invention using the method and the technical contents disclosed above without departing from the spirit and scope of the present invention.
Claims (10)
1. A neural network model training method for identifying silicon carbide dislocations, comprising the steps of:
obtaining a plurality of first silicon carbide dislocation picture samples, wherein the background in the first silicon carbide dislocation picture samples is clean and different dislocations are not overlapped; marking dislocations in each first silicon carbide dislocation picture sample to obtain a plurality of first silicon carbide dislocation picture samples after marking; training the marked first silicon carbide dislocation picture samples through a neural network model to obtain a 1 st generation detection model;
acquiring a plurality of second silicon carbide dislocation picture samples, wherein the backgrounds of the second silicon carbide dislocation picture samples are clean and dislocation overlapping and staggering exist; detecting the dislocation in each second silicon carbide dislocation picture sample through the 1 st generation detection model, and labeling the dislocation according to the detection result to obtain a plurality of labeled second silicon carbide dislocation picture samples; training the marked first silicon carbide dislocation picture samples and the marked second silicon carbide dislocation picture samples through a neural network model to obtain a 2 nd generation detection model;
acquiring a plurality of third silicon carbide dislocation picture samples, wherein the backgrounds in the third silicon carbide dislocation picture samples are unclean; detecting the dislocation in each third silicon carbide dislocation picture sample through the 2 nd generation detection model, and labeling the dislocation according to the detection result to obtain a plurality of labeled third silicon carbide dislocation picture samples; and training the marked first silicon carbide dislocation picture samples, the marked second silicon carbide dislocation picture samples and the marked third silicon carbide dislocation picture samples through a neural network model to obtain a 3 rd generation detection model.
2. The method of neural network model training for identifying silicon carbide dislocations of claim 1, further comprising, after deriving the 3 rd generation detection model: acquiring a plurality of fourth silicon carbide dislocation picture samples, wherein the fourth silicon carbide dislocation picture samples have influence factors further causing dislocation identification errors; and according to the steps, continuing further training to obtain a 4 th generation detection model.
3. The neural network model training method for identifying silicon carbide dislocations according to claim 1, wherein the step of respectively obtaining a plurality of first silicon carbide dislocation picture samples, a plurality of second silicon carbide dislocation picture samples and a plurality of third silicon carbide dislocation picture samples specifically comprises:
obtaining a plurality of silicon carbide dislocation pictures, and cutting each silicon carbide dislocation picture to obtain a plurality of silicon carbide dislocation picture samples; and respectively selecting a plurality of first silicon carbide dislocation picture samples, a plurality of second silicon carbide dislocation picture samples and a plurality of third silicon carbide dislocation picture samples from the plurality of Zhang Tanhua silicon dislocation picture samples.
4. The method of neural network model training for identifying silicon carbide dislocations of claim 1, wherein the number of dislocations in each of the first picture sample of silicon carbide dislocations, the second picture sample of silicon carbide dislocations, and the third picture sample of silicon carbide dislocations is no more than 10; the number of the first, second and third silicon carbide dislocation picture samples ranges from 200 to 500.
5. The method for training the neural network model for identifying silicon carbide dislocations according to claim 1, wherein the step of labeling dislocations in each first silicon carbide dislocation picture sample to obtain a plurality of labeled first silicon carbide dislocation picture samples specifically comprises:
and manually labeling different dislocations in the first silicon carbide dislocation picture samples based on labeling software to obtain the labeled first silicon carbide dislocation picture samples.
6. The method of training a neural network model for identifying silicon carbide dislocations as claimed in claim 1, wherein the step of labeling dislocations in each of the first silicon carbide dislocation picture samples to obtain a plurality of labeled first silicon carbide dislocation picture samples comprises:
manually marking different dislocations in the multiple first silicon carbide dislocation picture samples based on marking software, and training the manually marked first silicon carbide dislocation picture samples through a neural network model to obtain a 0 th generation model; detecting a plurality of first silicon carbide dislocation picture samples one by using the 0 th generation model to obtain dislocation types and coordinate positions in each first silicon carbide dislocation picture sample; according to the dislocation types and the coordinate positions of the detected first silicon carbide dislocation picture samples, different dislocations in the first silicon carbide dislocation picture samples are automatically marked through marking software and are adjusted in an auxiliary mode through manual marking, and the marked first silicon carbide dislocation picture samples are obtained.
7. The method for training the neural network model for identifying silicon carbide dislocations according to claim 1, wherein the step of detecting dislocations in each second silicon carbide dislocation picture sample through the 1 st generation detection model, labeling dislocations according to detection results, and obtaining a plurality of labeled second silicon carbide dislocation picture samples specifically comprises:
detecting a plurality of second silicon carbide dislocation picture samples by using the generation-1 model to obtain dislocation types and coordinate positions in each second silicon carbide dislocation picture sample;
and according to the dislocation types and coordinate positions of different dislocations in the detected second silicon carbide dislocation picture samples, automatically marking the different dislocations in the second silicon carbide dislocation picture samples through marking software and performing auxiliary adjustment through manual marking to obtain the marked second silicon carbide dislocation picture samples.
8. An image dislocation recognition method of silicon carbide wafer based on the neural network model training method for recognizing silicon carbide dislocation of any one of claims 1 to 7, comprising the steps of:
acquiring the 3 rd generation detection model and a plurality of silicon carbide dislocation pictures shot based on the silicon carbide wafer, and detecting the plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain dislocation types and coordinate positions in each silicon carbide dislocation picture;
according to the dislocation types and the coordinate positions of different dislocations in the detected multiple silicon carbide dislocation pictures, automatically marking the different dislocations in the multiple silicon carbide dislocation picture samples through marking software and performing auxiliary adjustment through manual marking to obtain multiple marked silicon carbide dislocation pictures;
and splicing the marked multiple silicon carbide dislocation pictures according to the coordinates to obtain the marked silicon carbide wafer image.
9. The silicon carbide wafer image dislocation identification method according to claim 8, wherein the step of splicing the plurality of marked silicon carbide dislocation pictures according to coordinates specifically comprises:
when a plurality of silicon carbide dislocation pictures are obtained after a stepping motor scans X1 columns and Y1 rows, setting the length-width ratio of each silicon carbide dislocation picture in the X direction and the Y direction as Y1: and X1, splicing each processed silicon carbide dislocation picture according to coordinates, so that after the shot silicon carbide pictures are spliced, the length and the width of the formed silicon carbide wafer images in the XY direction are consistent.
10. An image dislocation recognition apparatus for silicon carbide wafers based on the neural network model training method for recognizing silicon carbide dislocations as claimed in any one of claims 1 to 7, comprising the following modules:
the acquisition module is used for acquiring the 3 rd generation detection model and a plurality of silicon carbide dislocation pictures shot based on the silicon carbide wafer, and detecting the plurality of silicon carbide dislocation pictures through the 3 rd generation detection model to obtain dislocation types and coordinate positions in each silicon carbide dislocation picture;
the marking module is used for automatically marking different dislocations in the samples of the multiple silicon carbide dislocation pictures through marking software according to the detected dislocation types and coordinate positions of the different dislocations in the multiple silicon carbide dislocation pictures and performing auxiliary adjustment through manual marking to obtain the marked multiple silicon carbide dislocation pictures;
and the splicing module is used for splicing the marked multiple silicon carbide dislocation pictures to obtain the marked silicon carbide wafer image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211383066.4A CN115424099B (en) | 2022-11-07 | 2022-11-07 | Model training method, recognition method and device for recognizing silicon carbide dislocation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211383066.4A CN115424099B (en) | 2022-11-07 | 2022-11-07 | Model training method, recognition method and device for recognizing silicon carbide dislocation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115424099A CN115424099A (en) | 2022-12-02 |
| CN115424099B true CN115424099B (en) | 2023-04-18 |
Family
ID=84207611
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211383066.4A Active CN115424099B (en) | 2022-11-07 | 2022-11-07 | Model training method, recognition method and device for recognizing silicon carbide dislocation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115424099B (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021166161A1 (en) * | 2020-02-20 | 2021-08-26 | 株式会社日立ハイテク | Defect inspection system, defect inspection method, and method for creating teacher data |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200365685A1 (en) * | 2019-05-17 | 2020-11-19 | Cree, Inc. | Nondestructive characterization for crystalline wafers |
| US11361454B2 (en) * | 2020-02-28 | 2022-06-14 | Wolfspeed, Inc. | Alignment for wafer images |
| CN112557402B (en) * | 2020-12-04 | 2023-03-14 | 北京天科合达半导体股份有限公司 | Dislocation detection system |
| CN115219517A (en) * | 2021-04-19 | 2022-10-21 | 广州南砂晶圆半导体技术有限公司 | Rapid detection method for dislocation distribution in silicon carbide wafer |
| CN113255590A (en) * | 2021-06-25 | 2021-08-13 | 众芯汉创（北京）科技有限公司 | Defect detection model training method, defect detection method, device and system |
-
2022
- 2022-11-07 CN CN202211383066.4A patent/CN115424099B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021166161A1 (en) * | 2020-02-20 | 2021-08-26 | 株式会社日立ハイテク | Defect inspection system, defect inspection method, and method for creating teacher data |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115424099A (en) | 2022-12-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106952250B (en) | Metal plate strip surface defect detection method and device based on fast R-CNN network | |
| CN108562589B (en) | Method for detecting surface defects of magnetic circuit material | |
| CN111179225B (en) | Test paper surface texture defect detection method based on gray gradient clustering | |
| CN111080622B (en) | Neural network training method, workpiece surface defect classification and detection method and device | |
| CN109829903B (en) | Chip surface defect detection method based on convolution denoising autoencoder | |
| CN112233092A (en) | Deep learning method for intelligent defect detection of unmanned aerial vehicle power inspection | |
| CN113239930B (en) | Glass paper defect identification method, system, device and storage medium | |
| CN110390677B (en) | Defect positioning method and system based on sliding self-matching | |
| CN113393438B (en) | Resin lens defect detection method based on convolutional neural network | |
| CN114742818A (en) | Visual defect detection system based on industrial area array camera and detection method thereof | |
| CN116091505B (en) | Automatic defect detection and classification method and system for sapphire substrate | |
| KR20230030259A (en) | Deep learning-based data augmentation method for product defect detection learning | |
| CN112381175A (en) | Circuit board identification and analysis method based on image processing | |
| CN113409234A (en) | Minimum supervision Automatic Inspection (AI) of wafers supported by Convolutional Neural Network (CNN) algorithm | |
| CN114663382A (en) | Surface defect detection method for electronic component based on YOLOv5 convolutional neural network | |
| CN115424099B (en) | Model training method, recognition method and device for recognizing silicon carbide dislocation | |
| CN114820541A (en) | Defect detection method based on reconstructed network | |
| CN112784894A (en) | Automatic labeling method for rock slice microscopic image | |
| CN117132564A (en) | YOLOv 3-based sapphire substrate surface defect detection method and system | |
| CN110738698A (en) | Floating type seabed data measuring method and device and electronic equipment | |
| CN111598851A (en) | Solar cell fragment detection method based on morphological image processing | |
| CN113469984B (en) | Method for detecting appearance of display panel based on YOLO structure | |
| CN114841980A (en) | Insulator defect detection method and system based on line patrol aerial image | |
| CN114494142A (en) | Mobile terminal middle frame defect detection method and device based on deep learning | |
| CN114202508A (en) | Detection method and system suitable for highlighting surface defects of various ceramic tiles |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |