CN103733322B - The method of crack detection forecast model and the method for the crackle on detection semiconductor structure are provided - Google Patents
The method of crack detection forecast model and the method for the crackle on detection semiconductor structure are provided Download PDFInfo
- Publication number
- CN103733322B CN103733322B CN201280038885.3A CN201280038885A CN103733322B CN 103733322 B CN103733322 B CN 103733322B CN 201280038885 A CN201280038885 A CN 201280038885A CN 103733322 B CN103733322 B CN 103733322B
- Authority
- CN
- China
- Prior art keywords
- semiconductor structure
- measurement
- point
- crackle
- descriptor
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/62—Systems in which the material investigated is excited whereby it emits light or causes a change in wavelength of the incident light
- G01N21/63—Systems in which the material investigated is excited whereby it emits light or causes a change in wavelength of the incident light optically excited
- G01N21/64—Fluorescence; Phosphorescence
- G01N21/6489—Photoluminescence of semiconductors
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/93—Detection standards; Calibrating baseline adjustment, drift correction
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/95—Investigating the presence of flaws or contamination characterised by the material or shape of the object to be examined
- G01N21/9501—Semiconductor wafers
-
- 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
- G06T7/001—Industrial image inspection using an image reference approach
-
- H—ELECTRICITY
- H01—ELECTRIC ELEMENTS
- H01L—SEMICONDUCTOR DEVICES NOT COVERED BY CLASS H10
- H01L22/00—Testing or measuring during manufacture or treatment; Reliability measurements, i.e. testing of parts without further processing to modify the parts as such; Structural arrangements therefor
- H01L22/10—Measuring as part of the manufacturing process
- H01L22/12—Measuring as part of the manufacturing process for structural parameters, e.g. thickness, line width, refractive index, temperature, warp, bond strength, defects, optical inspection, electrical measurement of structural dimensions, metallurgic measurement of diffusions
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10048—Infrared image
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- 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)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Pathology (AREA)
- Immunology (AREA)
- Biochemistry (AREA)
- Analytical Chemistry (AREA)
- Chemical & Material Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Manufacturing & Machinery (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Power Engineering (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Computer Hardware Design (AREA)
- Quality & Reliability (AREA)
- Theoretical Computer Science (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Testing Or Measuring Of Semiconductors Or The Like (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The present invention relates to the method that method provides the forecast model for the crack detection on semiconductor structure, the semiconductor structure is the prime thing of the photovoltaic solar cell in photovoltaic solar cell, manufacturing process, in particular for manufacturing the semi-conducting material of solar cell, this method includes following methods step：A）Reference semiconductor structure with least one crackle is provided；B）There is provided crack data at least one crackle, crack data includes the relevant geometric position data in position with crackle on reference to semiconductor structure；C）The measurement of spatial discrimination formula multiple local measurement points of caused luminescence generated by light and/or is absorbed come the measurement of spatial resolution reference semiconductor structure, D by the IR of the measurement of spatial resolution semiconductor structure in the semiconductor structure）By establishing forecast model according to the measurement of spatial resolution data obtained in method and step C and according to the training of the crack data progress learning algorithm provided in method and step B, wherein, the training of learning algorithm includes following methods step：D1）At least one descriptor at least one partial descriptions symbol point is established by following manner, i.e., a test zone is set or determined for descriptor point, and establish the descriptor according to the measurement data in test zone, the descriptor is characteristic vector and/or feature distribution and/or feature histogram, and D2）Utilize the descriptor and crack data training learning algorithm.The invention further relates to the method and apparatus for crack detection.
Description
Technical field
The present invention relates to the method for providing the crack detection forecast model on semiconductor structure and detection semiconductor structure
On crackle method, the semiconductor structure is the prime of the photovoltaic solar cell in photovoltaic solar cell, manufacturing process
Thing or the semi-conducting material particularly for photovoltaic solar cell as manufacture.
Background technology
The photovoltaic solar cell being made up of semi-conducting material early has been used for electromagnetic radiation being converted into electric energy.In typical case
Photovoltaic solar cell in, semi-conducting material occupies the suitable share of solar cell total manufacturing cost.Thus use into
This highly beneficial material such as polysilicon a, in addition, purpose of solar cell research is to reduce photovoltaic solar electricity
The thickness of pond manufacture semiconductor wafer used, thus reduces material cost.
But thus occur following dangerous, i.e. crackle adversely affects semi-conducting material stability.Such crackle is for example
It may be formed, occurred because of the load in the manufacture of chip raw material, because of defect and chip because of the fault in material in silicon crystal block
Mechanical stress during sawing and occur, and occur mechanical stress when transporting and carrying operation.In addition, semi-conducting material
Mechanical loading and thermic load are born in the fabrication process.
Once semi-conducting material crushes in process, just, there is sky high cost, because must especially stop manufacturing.
In spite of crackle but this semi-conducting material of whole manufacturing process is also subject to it is also possible to being caused in subsequent application in light
Lie prostrate notable power loss when solar cell uses.
Therefore reason, there is an urgent need to the method for detecting the crackle in semiconductor structure, the semiconductor structure to be by people
Above-mentioned semiconductor source material such as the solar cell prime thing in the semiconductor wafer for manufacturing solar cell, manufacturing process
Or finished product solar cell.Because crackle typically has micron order extension size, therefore it is also referred to as micro-crack.
It is known to be shot using so-called IR printing opacities to detect crackle.Therefore, semiconductor structure is by the spoke in infra-red range
Penetrate and receive printing opacity detection and be measured by imaging method such as ccd video camera.Because the different absorption characteristics in the range of IR, therefore make
User can visually have found crackle in the image so obtained.The crack detection using luminescence generated by light is also recalled that：Therefore,
Luminescence generated by light is produced in semiconductor structure and carries out measurement of spatial resolution by imaging method such as ccd video camera.Photoluminescence measurement
Using priciple on be known and for example in the proceedings of the 22nd European photovoltaic solar meeting, (2007, Milan, meaning is big
Profit) in Trupke, be described in T " the luminescence imaging progress for being used for the sign of silicon wafer and solar cell ".But now
Generally crackle can not for example be recombinated into effective noise spot with other effects to distinguish.
In addition, being known using the crack detection of ultrasonic vibration of resonating (RUV) and for example in Monastyrskyi, A.
Et al. " the resonance ultrasonic vibration for being used for the online crack detection of silicon wafer and solar cell " (the 33rd PVSC proceedings,
2008) it is described in.
The content of the invention
The task of the present invention is, there is provided it is a kind of it is reliable, especially can commercial Application semiconductor structure crack detection side
Method, the semiconductor structure are photovoltaic solar cell prime things in photovoltaic solar cell, manufacturing process or for manufacturing light
Lie prostrate the raw material of solar cell.
The task is by providing the method for the forecast model of crack detection and detection semiconductor junction on semiconductor structure
The method of crackle on structure solves.The present invention for detecting the device of the crackle on semiconductor structure also by completing.
Following understanding of the invention based on applicant, i.e. learning algorithm is applied to establish crack detection on semiconductor structure
Forecast model, thus combine forecast model and carry out crack detection.Therefore, the device principle of method of the invention and the present invention
It is upper to be different from photovoltaic art crack detecting method used so far, because for the first time by by the engineering of learning algorithm
Commonly use in crack detection.
As described above, the crack detecting method of the present invention and the crack detection device of the present invention are related to photovoltaic art.Accordingly
Herein and hereinafter the raw material such as silicon wafer, outstanding for manufacturing solar cell is called using term " semiconductor structure " in ground
It is polycrystalline silicon wafer, also to call photovoltaic solar cell prime thing at any process in the fabrication process and into
Product solar cell.
Included according to the present invention, the forecast model that provides crack detection on semiconductor structure method with lower section
Method step：
There is provided in method and step A with reference to semiconductor structure, this has at least one crackle with reference to semiconductor structure.
In method and step B, there is provided on the crack data of at least one crackle, the crack data is included with crackle at this
With reference to the relevant geometric position data in the position on semiconductor structure.
The semiconductor structure is usually planar elements, so that typically can be on the front of semiconductor structure or the appearance at the back side
State describes crackle, that is, typically by two-dimensional geometry data.Now it is within the scope of the invention that, crack data includes
The multiple points and/or line of crack growth are described, and/or at least two lines of crack data including star fracture intersecting are herein split
Line center.
Also it is within the scope of the invention that, it is the semiconductor structure for having crackle with reference to semiconductor structure, the crackle utilizes
Another mensuration is measured on geometric position.As an alternative or supplement also it is possible that being advised by delineation or the like
Given on fixed geometric position with reference to semiconductor structure addition crackle.
The measurement of spatial resolution with reference to semiconductor structure is carried out in method and step C.Therefore, for multiple local measurement points
Carry out the measurement of spatial resolution of caused luminescence generated by light and/or semiconductor structure infrared absorption (IR absorptions) in semiconductor structure
Measurement of spatial resolution.To perform spatial discrimination imaging method, measurement apparatus known per se especially ccd video camera can be used for
Measurement of spatial resolution.It is important that carry out measurement of spatial resolution for multiple local measurement points.Measurement point covers semiconductor structure
At least one region for including crackle on surface.
Forecast model is established in method and step D.Therefore, it is incorporated in the measurement of spatial resolution number obtained in method and step C
According to the training that learning algorithm is carried out with the crack data provided in method and step B.
Different from known crack detecting method, corresponded to typically without based on physical model according to measurement signal to distinguish
The measurement point of crackle and other measurement points.Replace, with reference to measurement data and crack data i.e. on reference to semiconductor structure
It is known that existing crackle is described to carry out the training of learning algorithm (Trainieren).Therefore do not provide to show crackle and
Explanation that effective geometry distinguishes and standard are such as recombinated with other elements for measurement signal, and (that is, the description is clear and definite
Statement), but be made up of learning algorithm training.
The training of learning algorithm now comprises the following steps：
At least one descriptor at least one partial descriptions symbol point is established in method and step D1.Now, for
The descriptor point sets or determines a test zone, and establishes the descriptor according to the measurement data in the test zone, the description
Symbol is characteristic vector and/or feature distribution and/or feature histogram.
Thus, the test zone includes at least one subset of the measurement point according to method and step C.Also within the scope of the present invention
Be that the test zone is included according to method and step C whole measurement points.
On the other hand, descriptor point forms a partial points of the descriptor, so as to binding characteristic vector and/or feature distribution
And/or feature histogram and formed in the case where considering test zone for partial descriptions symbol point description.Characteristic vector and/or
Feature distribution and/or the feature of feature histogram are based on the measurement data obtained in method and step C.Now in the scope of the invention
Interior is that this feature is directly derived from measurement data.But it is particularly advantageous the further processing that this feature is multiple measurement data
And/or association especially structure describes.
The descriptor and crack data training learning algorithm are utilized in method and step D2.
The training can be carried out using learning algorithm known per se in a manner known per se.Also within the scope of the present invention
, only instructed using a descriptor point and corresponding descriptor and/or only using a crackle and corresponding geometric data
Practice the learning algorithm.It is advantageous that carry out the training using multiple descriptor points and the descriptor accordingly established.Or
And especially it is further advantageous that performing methods described using multiple crackles and each self-corresponding crack data.
The training of learning algorithm can be carried out in a manner known per se as described above.At this point it is important to based on pre-
Fixed crack data and know whether there is crackle on the descriptor point.The information in training by learning algorithm for being realized
Classification formed be vital.
It is within the scope of the invention that multiple be used to train the learning algorithm with reference to semiconductor structure.Now not necessarily
Need it is each have crackle with reference to semiconductor structure because the descriptor point and descriptor of no crackle to training and have
Meaning.But at least one of which must have at least one crackle as described above with reference to semiconductor structure.
Following methods step is included according to the method for being used to detect the crackle on semiconductor structure of the present invention：
Semiconductor structure is provided in method and step A.
Different from reference to semiconductor structure, not being known a priori by certainly in the semiconductor structure for performing crack detection
Semiconductor structure has the information of flawless actually in other words for crackle geometric position.
Forecast model is provided in method and step B, the forecast model is established by the training of learning algorithm.
By measurement of spatial resolution, the multiple of caused luminescence generated by light locally survey in the semiconductor structure in method and step C
Amount point and/or absorbed by the IR of the measurement of spatial resolution semiconductor structure come the measurement of spatial resolution semiconductor structure.Especially
Advantageously, identical metering system is performed when establishing forecast model and in crack detection, that is be all luminescence generated by light
Measurement of spatial resolution or be all IR absorb measurement of spatial resolution.
Determine whether there is crackle on the checkpoint at least one toposcopy point in method and step D.The part
Checkpoint is a point on semicon-ductor structure surface and preferably corresponds to one of them described measurement according to method and step C
Point.
The determination includes following methods step：
In method and step D1, at least one descriptor is established for the checkpoint, in the descriptor for checkpoint
Set and determine a test zone, and the measurement data combined in test zone establishes the descriptor, and the descriptor is characteristic vector
And/or feature distribution and/or feature histogram.
Therefore, to crack detection it is also important that, establish a descriptor for the checkpoint, the descriptor thus be in spy
Vector and/or the feature of feature distribution and/or feature histogram form are levied, so as to consider for the characteristic of the checkpoint in side
It is mapped in the case of the measurement data in test zone obtained in method step C in the descriptor.
Determine whether there is crackle on the checkpoint using the descriptor and forecast model in method and step D2.
Thus, it is that a forecast model need only be provided the characteristics of crack detecting method of the invention, then according to known per se
Mode the measurement according to method and step C is performed in the form of luminescence generated by light or IR absorptiometries, and descriptor shape can be passed through
Determine whether there is crackle on checkpoint into using forecast model.
Now it is within the scope of the invention that only carrying out the above-mentioned determination according to method and step D2 a checkpoint.It is best
The determination is carried out in multiple checkpoints.Thus it is within the scope of the invention that, perform the every of measurement herein in method and step C
The individual measurement point is successively elected to be as checkpoint in succession.
As a result, it ensure that using the crack detecting method of the present invention and determine whether semiconductor structure has crackle and determination to split
The geometric position of line and/or extension size, without being made a reservation for by assuming, making a reservation for or physical model derive according to method and step C
Measurement data and abstract crackle characteristic between relation.
The research of applicant shows, the formation of forecast model and using outstanding splitting suitable for detection semiconductor structure
Line.Significantly improved according to the crack detecting method of the present invention thus compared with known crack detecting method.Especially using
Associative learning Algorithm for Training is provided in crackle and other geometry knots in respective method and step C in the case of photoluminescence measurement method
Structure for example recombinates the high differentiation of accuracy at target between resulting structure, here, not necessarily prespecified differentiation crack structtire and its
The standard of its structure, but only formed by the training of learning algorithm.
The inventive method also has the advantage that, can be divided into the foundation of forecast model and examine forecast model for crackle
Survey：
Thus, for example can be by manufacturer by accurately preparing with reference to semiconductor structure and/or combining many crackles and crackle number
According to training accurately train learning algorithm and be supplied to user to use the forecast model being thusly-formed.
Therefore, it may not be necessary for user side also carries out learning algorithm training.User will directly can also be trained pre- by manufacturer
Survey model and be used for crack detection.
In the crack detecting method of the present invention, crackle reconstruction is carried out preferably in method and step E, in crackle reconstruction really
Determine the geometric data of crackle presentation.Thus rebuild for this based on a checkpoint for corresponding to crackle, at least to determine
The part of crackle and preferably whole geometrical representations.It is particularly advantageous that for an office around checkpoint in method and step E
Portion's reconstruction regions for each measurement point in reconstruction area determine orientation and by means of pattern identification means and star
The similarity system design of pattern and/or linearity pattern determines the measurement point corresponding to crackle.Above-mentioned preferred embodiment especially exists
It is during the use of photoluminescence measurement method and especially favourable in the use of polysilicon semiconductor structure.
Following understanding of this preferred embodiment of the inventive method based on applicant, i.e. splitting in semiconductor structure
Line typically has mulle.It is possible thereby to identification and reconstruction using mulle and/or linearity pattern known per se
The combination of method carries out crackle reconstruction.
As described above, it is within the scope of the invention that, in method and step D when establishing forecast model and/or in crackle
Successively now it can for example be used during detection in succession according to multiple partial points as descriptor point or checkpoint in various method steps
All local measurement points being measured in C.
But advantageously method is determined using key point to limit the quantity of descriptor point and/or checkpoint.Such side
Method is also referred to as point of interest (POI) and determines method.It is particularly advantageous that used to determine descriptor point and/or checkpoint
Key point search method, preferably use according to SIFT methods, GLOH methods and/or SURF methods or known to for feature detection
It is at least one of in method such as Tuscany oscillograph or Harris's oscillograph or similar, derived or similar method
The key point of individual method determines.The above method is known per se and is for example described below：
-GLOH：Krystian Mikolajczyk and Cordelia Schmid " A performance evulation
of local descriptors”(IEEE Transactions on Pattern Analysis and Machine
Intelligence, 10,27, the 1615-1630 pages, 2005).
-SIFT：Lowe, David G. (1999) " Object recognition from local scale-
Invariant features " (the 1150-1157 pages of international computer prospect conference proceedings 2., doi:10.1109/
ICCV.1999.790410.http://doi.ieeecs.org/10.1109/ICCV.1999.790410)。
-SURF：Bay, H., Tuytelaars, T., Gool, L.V. " SURF:Speeded Up Robust
Features " (proceedings of the 9th prospective capabilities in hardware European Convention, 2006.05).
-Harris：C.Harris and M.Stephens " A combined corner and edge detector "
(the 4th Alvey Vision Conference proceedings, 1988, the 147-151 pages).
The research of applicant shows, key point is determined to be particularly advantageous using the controllable wave filter in orientation to determine
State at least one key point.The wave filter that the derivative being advantageously employed particularly for calculating measurement data is preferably second dervative comes
Carry out the determination.The research of applicant shows that the bar chart wave filter being especially made up of Gaussian filter second dervative is fitted
Determined for key point.
In order to select key point, determine the key point in whole region in the scope of the invention especially since all measurement points
It is interior.Also it is within the scope of the invention that, prespecified finite geometry region on semicon-ductor structure surface and/or walked in method
The subset of the measurement point measured in rapid C, determines key point wherein.
It is determined that during key point, the different qualities and feature of regional area or measurement point can be used for determining associated picture office
Portion, such as the distribution of the presentation of measurement signal intensity or measurement signal intensity level.This feature can for example pass through factor such as side
The determination of difference, average value, symmetry or other factors known per se is depicted and is used for determining the key point.
Also mark sheet can be used on geometrical representation as Ru Jiao, side, line or circular configuration.Such mark sheet is as example may be used
By using during wave filter wave filter respond determine and/or by multiple filter banks and perhaps the pretreatment of spinoff and
Further handle to determine.The example of available filters is Suo Baier wave filters, the special wave filter of Puli's dimension, Gaussian filter, height
This difference filter, Laplacian.Using tensor such as Harris's oscillograph, hessian determinant also in the scope of the invention
It is interior, or using controllable filter, using small echo such as Gabor and Haar small echos also within the scope of the present invention.Moreover, use Soviet Union
Coral Corner Detection device or canny edge detection device are also within the scope of the present invention.
The research of applicant shows that the filtering using anisotropic bar chart wave filter is highly advantageous, the wave filter
It is the second dervative of Gaussian filter and for example in J.Malik " Contour and Texture Analysis for
It is described in Image Segmentation " (prospective capabilities in hardware International Periodicals 43 (1), the 7-27 pages, 2001).
Now it is particularly advantageous to use the wave filter with different spaces orientation.Now, positive wave filter response exists
It is associated in the case of the different spaces orientation of wave filter, it is preferably cumulative.It is thereby achieved that on rotating set SO
(2) Haar integration shapes, as example in H.Schulz-Mirbach " Anwendung von
Invarianzprinzipien zur Merkmalgewinnung in der Mustererkennung Dissertation”
Described in (hamburger port technology university, 1995.02.10, Nr.372, VDI are published).
The research of applicant shows that thus obtained value especially increases and still rotated constant in Crack Center.Cause
And it is possible thereby to which plain mode determines Crack Center according to rotating, i.e. it is determined that crackle on each measurement figure it is actual walk
To.Thus following advantageous effects are obtained, i.e. unrelated with the rotation of crackle or semiconductor structure during measurement to obtain one in analysis
The value of sample size, thus equally detected.
In addition, using the wave filter with different scales also within the scope of the present invention.It is derived from relative to crack size
Change insensitive result.
Now it is particularly advantageous that the interpretation of result in different scale ranks is considered for estimating the big of considered region
It is small：Thus corresponding high result refers to high scale rank, thus indicates larger crackle, so as to be in result compared with low scale level
Result when other is compared, and advantageously determines that key point, crack detection and/or crackle are rebuild according to larger image section.
Generally determine to obtain advantages below by key point, the quantity in potential crack region can be restricted to obvious less
Key point, thus the speed-raising of this method can be obtained.
Also it is within the scope of the invention that, key point is determined first with other measuring methods.In the case, such as
The testing result of IR measurements is used for determining key point in a conventional manner.Different measuring methods can also be mutually combined：Thus
Such as determined within the scope of the present invention according to the key point of IR measurements, then, for originally measuring the key point obtained with IR
Photoluminescence measurement is analyzed for crack detection.
Descriptor in method and step D1 preferably based on SIFT methods, GLOH methods, HOG methods, LESH methods and/or
SURF methods and/or at least one of method modification of the feature to produce descriptor of test zone method is combined to determine.Make
For substitute or supplement and advantageously, in order to establish descriptor, using adjustable and the preferably above-mentioned bar shaped of rotatable wave filter
The maximal filter response of figure wave filter.LESH methods are also known in itself and for example in Sarfraz, S., Hellwich,
" being described in Head Pose Estimation in Face Recognition across Pose Scenarios " O.
(meetings of VISAPP 2008, prospective capabilities in hardware theory and application message conference, Portugal Madeira, the 235-242 pages,
2008.01 (optimal student papers prizes)).
It is particularly advantageous that establish descriptor in the way of comprising following methods step：
For each partial points of test zone in method and step D1.1, preferably by determining for each of the test zone
The parameter of measurement point determines amplitude and orientation, the parameter be further preferred that for the measured value of each measurement point gradient or
Maximal filter response and orientation for the controllable filter of each measurement point.
At least selected key point for test zone in method and step D1.2 establishes feature histogram.
Now it is particularly advantageous that also determining the main orientation of crackle and carrying out repairing for feature histogram according to the main orientation of crackle
Just, constant feature histogram is rotated to obtain.
The description for the test zone being thus combined when descriptor is established around the descriptor point or checkpoint of feature.It is special
Sign description can be for example made up of the isometric chart of one or more features, feature distribution or its parameter or some presentations of feature.
The descriptor of basis known per se in algorithm SIFT, GLOH, HOG, LESH or SURF known per se can now be used
Feature statement.Establish the measurement data that the basis of descriptor always determines in method and step C.
In above-mentioned preferred embodiment, determined advantageously between method and step D1.1 and D1.2 the main orientation of crackle and
Carry out establishing feature histogram by using the measurement point in scanning algorithm scanography area in method and step D1.2.Adopt herein
It is especially advantageous with according to the scanning algorithm known per se of GLON and/or SIFT or its variant.
In order to establish descriptor, therefore certain scanning and/or division of image local can be used.This can especially be used
The known scanning according to algorithm known of body.
For establish descriptor carry out feature selecting, just as known in SIFT, GLOH, HOG, LESH or SURF algorithm that
Sample.
The research of applicant shows, uses the maximal filter of anisotropic bar chart wave filter as described above herein
Response is also especially favourable.Instead of the use in the gradient orientation and gradient amplitude such as known from SIFT and GLOH, thus have
Sharply using the orientation and amplitude of the bar chart wave filter responded with maximal filter.
It is within the scope of the invention that pretreatment or the description of after-treatment feature when descriptor is established.Root can especially be passed through
Realize that rotation is constant as described above according to the main orientation amendment descriptor of identified crackle.And the scaling of descriptor can be passed through
And/or leveling is obtained relative to changes in contrast or the stability of noise.
Preferably in method and step D1 the test zone is determined comprising following methods step：
First, the determination of at least one local key point is realized as before.
It is then determined that the measurement point in the test zone, in the measurement point regulation geometry extension size for test zone and
Shape.The rectangle, ellipse or border circular areas that its central point is key point are set preferably for test zone.
In an Advantageous embodiments of the inventive method, the geometric position of crystal boundary is as the crystal boundary in semiconductor structure
Data set and/or are determined by measurement of spatial resolution method, and are weighted according to crystal boundary data, wherein,
- it is determined that during key point, as described above, the partial points on crystal boundary are not selected as key point, or it is determined that
With the weight lower than remaining partial points during key point, and/or
When establishing descriptor in the-method and step D when establishing forecast model and/or in crack detection, do not consider
Partial points on crystal boundary, or the partial points on crystal boundary are considered with the weight lower than remaining partial points, and/or
- in crackle reconstruction, as described above, not considering partial points on the crystal boundary or with lower than remaining partial points
Weight considers the partial points on the crystal boundary.
Thus especially in application the inventive method when for providing the preferred embodiment during polycrystalline silicon wafer.As long as crystal boundary
Geometric position is known or can simply determined based on available measuring apparatus that then error-prone property during crack detection thus can quilt
Reduce again, its way is as described above with the methods of the invention by grain boundary sites information.
Can be learning algorithm known per se for establishing forecast model or performing the learning algorithm of crack detection.Especially
It is that the use of neuron net or Bayes classifier or core engine is within the scope of the present invention.The research of applicant shows, uses branch
It is especially advantageous to hold vector machine algorithm, and its foundation is that the SVMs determines to be used to divide to complete classification task
The optimal hyperplane that the feature of " cracking " class and " not ftractureing " class describes.Thus remarkable advantage is obtained, because dividing surface preferably exists
Between class.Feature description not in class circle not adversely affects to the alignment for the boundary line.Appoint in order to complete classification by dividing surface
Business, data are mapped to a high-dimensional space by so-called " core skill ".The way is known per se and for example existed
BernhardAlex Smola " Learning with Kernels:Support Vector
(" adaptive computation and machine learning ", MIT goes out Machines, Regularization, Optimization, and Beyond "
Version society, Cambridge, MA, 2002, ISBN0-262-19475-9) it is described in.
As described above, by learning algorithm, stated in descriptor feature and whether had crackle in some partial points
Know that the relation between information is used to form the forecast model.Then, can combine forecast model will be unknown in crack detection
Image section or the descriptor formed for this are categorized as crackle or flawless.
Advantageously crackle is performed in the crack detecting method of the present invention in method and step E as described above to rebuild.This
When, crack structtire can be rebuild based on the architectural characteristic of the corresponding checkpoint peripheral region in test zone of interest.Category
In crackle partial points architectural characteristic now be not belonging to crackle architectural characteristic demarcate open.Architectural characteristic in the case may be used
Be position, density value, gradient and wave filter response.
The research of applicant is it has been shown that advantageously it is assumed that carrying out pattern knowledge in the case of the star of crackle or linear structure
Not.
It is particularly advantageous that reconstruction is so designed that in the case of linear or star fracture, i.e. preferably closed first as more than
A Crack Center is determined as described in when key point search describes, and limits the central point as test zone.
Then, the architectural characteristic of the test zone measurement point is extracted.The description of measurement point around key point with according to itself
The description of one corresponding points of the mulle of known pattern identification means is compared.Show the region phase with mulle
As structure such as orientation structure therefore high probability belong to crackle and thus can be assessed as belonging to crackle.In order to calculate structure
The similitude of characteristic, it can be advantageous to realize additional weight, such as pass through each partial points density value, gradient or corresponding wave filter
Response.Judge whether partial points belong to crackle in the case of closely similar.The judgement can for example utilize base known per se
Completed in the sluggish method of Crack Center.
The research of applicant shows, the orientation and density of structure are particularly suitable for use in description scheme characteristic.
Herein, for computer azimuth, filtering of the above-mentioned anisotropic bar chart wave filter in different azimuth is utilized
Also it is proved to be favourable.Now advantageously, using maximal filter response and its orientation as structural information.
Brief description of the drawings
Below with reference to the further feature and preferred embodiment of embodiment and the accompanying drawing description present invention, in figure：
Fig. 1 shows the luminescence generated by light picture for the polycrystalline silicon wafer for being substantially centered crackle in component a), wherein, dividing
The partial enlarged drawing of crackle peripheral region is drawn out in scheming b)；
Fig. 2 be by bar chart wave filter to contain crackle region when maximal filter respond view；
After Fig. 3 shows orientation histogram around ungraded crackle in component a) and shows leveling in component b)
Orientation histogram；
Fig. 4 shows crackle reconstructed results.
Embodiment
The offer according to the present invention is described in conjunction with the accompanying for the method for the forecast model of crack detection and according to this hair
The embodiment of bright crack detecting method.
Polycrystalline silicon wafer is used as with reference to semiconductor structure, and its is substantially square, and it has the about 15.6cm length of sides and about 180 μm
Thickness.This polycrystalline silicon wafer is the typical raw material for manufacturing photovoltaic solar cell.
In order to form the semiconductor structure as with reference to semiconductor structure, using automatic percussion device substantially centrally by one
Individual metal tip is that so-called punching point is highly dropped on semiconductor structure from regulation drop, so as to form crackle and also be aware of
Geometric position is the X-coordinate and Y-coordinate for the Crack Center that the rum point with metal tip on chip overlaps.
The reference semiconductor structure so prepared is provided in method and step A.In addition, Crack Center is aware of as described above
Location coordinates and thus provided as crack data in the method and step B.
In method and step C, the measurement of spatial resolution with reference to semiconductor structure is carried out.Now, using use known per se
In the equipment of photoluminescence measurement：Excited by using the semiconductor optical of electromagnetic irradiation to produce electron-hole pair.Because restructuring
Process, by sending luminescence generated by light with reference to semiconductor structure, it is by silicon ccd video camera measurement of spatial resolution.Therefore, walked as method
Rapid C result, it is in the netted measurement point being scattered in reference to semicon-ductor structure surface for many, is respectively present a measured value,
The measured value corresponds to the intensity or at least associated with intensity of the luminescence generated by light calculated by these measurement points.
Measurement result is as shown in Figure 1a.It is clearly seen the high spatial resolution of (1024 × 1024) individual point.Initially substantially occupy
In see star fracture structure.Crack structtire peripheral region is indicated with black rectangle.Fig. 1 b show the partial enlarged drawing of the rectangle,
The crack structtire is high-visible (in drawn black circle) in rectangle.
But Fig. 1 a and 1b also show that a number of other structures have similar measured value, thus to for being incited somebody to action with reference to measurement signal
It is proposed to be strict with corresponding to the measurement point and the method that remaining measurement point distinguishes of crackle.
Thus, in the embodiment of the inventive method, determined in method and step D by being incorporated in method and step C
Measurement of spatial resolution data and the learning algorithm of the crack data provided in stepb train and establish a forecast model：
Now it is possible that establishing a descriptor respectively for each measurement point as described above and then training this
Learning algorithm.
But in embodiment described herein, the detection at potential crack center is carried out first to search key point, so as to significantly
Processing duration and processing cost are limited, because only carrying out subsequent step to key point.
Key point (so-called point of interest POI) can be searched using the above method in principle in the case.But applicant's
Research shows that because crackle has star structure, therefore it is especially advantageous to search key point using anisotropy bar chart wave filter
's.The second dervative of Anisotropic fractals device is used in the present embodiment.The wave filter is orientated according to different azimuth to build
It is vertical, it is according to such orientation herein：
After measured value screening, the positive wave filter response to these bar chart wave filters is cumulatively added.Thus especially splitting
Line center is that high wave filter response is obtained in crackle crosspoint or region.In order to further limit potential Crack Center, take out
Take such measurement point, i.e. its cumulative wave filter response is higher by defined threshold, preferably based on equation 1 below：
Herein, A (x, y) is when having different azimuth orientation for the measurement point with coordinate (x, y), utilizes bar chart
The positive wave filter response of the image of wave filter is amounted to.A binary values are distributed to each measurement point, it elects 1 as, only
The wave filter response for being used for the measurement point is more than or equal to the threshold value for the maximal filter response for being multiplied by all measured values.Remaining
In the case of distribution 0 value.Therefore sensitivity being adjusted by selecting threshold value c, c selects bigger, and obtained key point is fewer, or instead
It.Threshold value c obtains by experience, preferably so obtains, includes all Crack Centers.It turns out that Crack Center typically has
There is very big value.
Fig. 2 shows Measurement and Data Processing, is shown in which each cumulative wave filter response for each pixel.Clearly see
There is very big value to Crack Center.
Then, associated structure is preferably reduced to a partial points.This is in a manner known per se by form
Computing realizes, wherein, target perforations are contracted to a little.As a result, i.e., remaining point is key point therewith, wherein, each pass
Key point is potentially Crack Center.
To each key point, descriptor is established in method and step D1 respectively.Thus each key point is successively in succession selected
Symbol point is described, is intended for each test zone of descriptor point respectively and establishes descriptor.
It will be described below two modifications of descriptor foundation：One is established by the descriptor of gradient, the second is by
The descriptor of bar chart wave filter is established.
It is similar with GLOH methods known per se when establishing descriptor by gradient in the flexible program of the present embodiment first
Ground is established.First, gradient orientation is considered.But because the noise in measured value significantly impacts gradient, therefore first by adding
Enter Gaussian filter and measure value leveling.This way is known per se.By using Gaussian filter derivative in x directions and y
Convolution measurement data on direction, measurement data are leveled, while obtain the gradient on x directions and y directions.Thus obtain for the phase
Hope the orientation and amplitude of the gradient vector of position (x, y).The orientation usable levels [0,2 π] of two-dimensional space represent.
Therefore, in the method and step D1 of the present embodiment, a test zone is limited first for each key point.This is at this
It is circle (measurement point) of the center of circle using 40 pixels as radius to be defined as in example using key point.Subsequently, in test zone
Each measurement point, gradient orientation and gradient amplitude are obtained in method and step B1.1 as described above.
In order to calibrate gradient orientation according to key point surrounding structure, the orientation of key point surrounding structure is identified below first：
The gradient orientation obtained in advance in test zone is drawn in orientation histogram, its for example with 36 bars,
That is, a uniform azimuthal bin is covered in 36 independent sortings, each classification.Fig. 3 a show this division result.Now may
There is situations below, there is a bar with smaller value between two bars with maximum.It is preferably this that there is smaller value
Orientation be not considered for key point orientation calculating.It is therefore advantageous that flatten the histogram with method known per se.Figure
3b shows this leveling result.
In order to obtain rotation constant result now, i.e. the descriptor constant for the rotation of key point, make test zone with it
The key point orientation of preceding determination rotates.Descriptor is established can be indirectly by the gradient orientation according to key point orientation and descriptor
Correct to realize.
Based on this, key point descriptor is used in particular according to the measurement data of the measurement point in test zone the key that determines
Point.
Therefore, such key point is determined so far, its orientation is calculated and makes inspection according to key point orientation
Area rotates.This way is known per se and for example in (Lowe, D.G. " Distinctive Image features
From scale-invariant keypoints " (prospective capabilities in hardware International Periodicals, 2004:The 91-110 pages)) in be described.
Then, a descriptor is established now for each key point：
This can similarly be realized with GLOH methods known per se.
The simply possible mode that descriptor is established is, for example, foundation gradient orientation to establish characteristic vector.Each orientation
Or azimuthal bin indicates a feature.From the gradient amplitude with the orientation it is cumulative in obtain certain orientation feature presentation.
In addition to foregoing description symbol is established, sampling known per se can also be used.Pole sampling herein is especially advantageous
Because crackle to be sought typically have with further from measurement point compared with higher close Crack Center gradient amplitude.
The pole sampling that thus its center overlaps with key point is especially advantageous.Another advantage is the above-mentioned inspection according to key point orientation
Area, which rotates, to be polar coordinates by simple conversion.
Further improvement and the GLOH that descriptor is established by inspection similarly by dividing into different zones to realize.For
Per sub-regions, a characteristic vector is established.Then, descriptor is obtained by the combination of the characteristic vector of different subregions.Thus
The additional information on regional structure is obtained, the additional information is highly beneficial to difference crackle and other structures.
The research of applicant shows, is established above-mentioned using gradient in a modification of descriptor, advantageously by above-mentioned bar
The wave filter of shape figure wave filter is responded for calculating descriptor.Therefore, the side of bar chart wave filter is selected instead of gradient orientation
Position, it produces maximal filter response.
Respective strengths similar to above-mentioned way, instead of gradient amplitude selecting maximal filter to respond.The modification is with adopting
Magnitude matrix is especially shown with the difference of modification during gradient：In the modification using gradient, fracture edges have maximum,
And in the modification responded using the wave filter of bar chart wave filter, crackle has maximum in itself.
In principle, the selection for differing only in input of this descriptor is formed modification and above-mentioned modification：Instead of key point
The Grad of the test zone of surrounding, responded now using the wave filter of anisotropic bar chart wave filter.Remaining route is similar
Carry out.
It is unrelated with selected flexible program, established descriptor is then utilized in method and step D2 and in method and step
The crack data that is there is provided in B trains learning algorithm.
The research of applicant shows especially can reliably use SVMs.
The result of this embodiment of the inventive method is thus the prediction mould obtained by the training using SVMs
Type.
In a second embodiment, one embodiment of the crack detecting method of the present invention will now be described.
Above-mentioned forecast model is used herein.
Semiconductor structure is provided in method and step A, and above-mentioned forecast model is provided in method and step B.
Then, the measurement of spatial resolution of semiconductor structure is carried out in method and step C.This preferably with first embodiment
Described in, similarly carried out according to the method and step C spatial discrimination photoluminescence measurement described there.
As a result, there is measurement data in multiple local measurement points for semiconductor structure, the measurement data with it is produced
The current local strength of luminescence generated by light be associated.
Determine whether there is crackle on checkpoint at least one toposcopy point in method and step D.
Now in principle also it is possible that for example performing such determine for each measurement point.
But particularly advantageously, also determine in crack detection key point first as described above and just for identified
Key point performs above-mentioned inspection.The determination of key point can now be carried out as described on first embodiment.
For each checkpoint, descriptor is established in method and step D1, in the descriptor for checkpoint it is predetermined or
Determine test zone.The foundation of descriptor is realized in a manner of previously in the identical described in first embodiment.
Different from first embodiment, the forecast model is so used in method and step D2 in crack detection, i.e. by
Descriptor and forecast model determine whether there is crackle on the checkpoint.It can thus be classified by the forecast model,
Corresponding key point is divided into the point for belonging to crackle or the point for being not belonging to crackle in the classification.
In the second embodiment of the inventive method described herein, then it is directed to and is classified as in method and step D
The crackle for belonging to the key point of crackle is rebuild.
Crackle is rebuild to be realized using sluggish method known per se：Start in the key point of usually Crack Center, one by one
Rebuild to measurement point the original shapes of crackle.Reconstruction is based on following hypothesis, i.e. crackle is extended in star and determined based on this
Around crackle with the similitude of mulle.If measurement point position is in the bar structure expanded from Crack Center, then in crackle
Therefore a pericentral measurement point belongs to the crackle high probability,.It is therefore advantageous that in order to each partial points reconstruction and
Orientation and the wave filter response of bar chart wave filter are considered as described above.
By the measurement point orientation around Crack Center compared with mulle orientation.Now it can preferably be used as similar
Degree weights come the amplitude for calculating the cosine between differential seat angle and being responded with each wave filter.
Preferably also examine the morphological character of image occurring because sluggish, obtaining.It is particularly advantageous that the face of crackle
Product, orientation and length are considered for associated with chip steadiness.
Fig. 4 shows the result that this crackle is rebuild.
Claims (28)
1. a kind of method for being used to provide the forecast model for the crack detection being used on semiconductor structure, the semiconductor structure
It is the prime thing of the photovoltaic solar cell in photovoltaic solar cell or manufacturing process,
This method includes following methods step：
A. provide with reference to semiconductor structure, this has at least one crackle with reference to semiconductor structure；
B. the crack data at least one crackle is provided, the crack data is included with crackle in the reference semiconductor junction
The relevant geometric position data in position on structure；
C. multiple local measurement points of caused luminescence generated by light and/or passed through in the semiconductor structure by measurement of spatial resolution
The IR of the measurement of spatial resolution semiconductor structure, which absorbs, carrys out the measurement of spatial resolution reference semiconductor structure, and
D. provided by the measurement data according to the spatial discrimination obtained in method and step C and foundation in method and step B
Crack data carry out the training of learning algorithm and establish forecast model, wherein, the training of the learning algorithm includes following methods
Step：
D1. at least one descriptor that point is accorded with least one partial descriptions is established by following manner, i.e. for the part
Descriptor point is set or determines a test zone, and establishes the descriptor according to the measurement data in the test zone, and this is retouched
It is characteristic vector and/or feature distribution and/or feature histogram to state symbol, and
D2. the learning algorithm is trained using the descriptor and the crack data；
Wherein, the geometric position of crystal boundary is set and/or by measurement of spatial resolution as the crystal boundary data in semiconductor structure
Method determines, and
So it is weighted according to crystal boundary data,
When-the descriptor in method and step D is established, the partial points on crystal boundary are not considered, or with than remaining partial points
Low weight considers the partial points on crystal boundary.
2. according to the method for claim 1, it is characterised in that the semiconductor structure is by for manufacturing solar cell
Semi-conducting material formed.
3. according to the method for claim 1, it is characterised in that at least one partial descriptions symbol point is searched by key point
Method is sought to determine, the crucial point searching method is at least one method in SIFT methods, GLOH methods and SURF methods.
4. a kind of method for being used to detect the crackle on semiconductor structure, the semiconductor structure is photovoltaic solar cell or system
Photovoltaic solar cell prime thing during making, this method include following methods step：
A., semiconductor structure is provided；
B. the forecast model established according to claim 1 is provided, the forecast model is established by the training of learning algorithm；
C. multiple local measurement points of caused luminescence generated by light and/or passed through in the semiconductor structure by measurement of spatial resolution
The IR of the measurement of spatial resolution semiconductor structure, which absorbs, carrys out the measurement of spatial resolution semiconductor structure；
D. determine whether there is crackle on the toposcopy point at least one toposcopy point, including following method step
Suddenly：
D1. at least one descriptor of the toposcopy point is established by following manner, i.e. for the toposcopy point setting or
Determine a test zone and combine the test zone in measurement data establish the descriptor, the descriptor be characteristic vector and/
Or feature distribution and/or feature histogram, and
D2. determine whether there is crackle in the toposcopy point using the descriptor and forecast model.
5. according to the method for claim 4, it is characterised in that the semiconductor structure is by for manufacturing solar cell
What semi-conducting material was formed.
6. according to the method for claim 4, it is characterised in that when detecting crackle according to method and step D, in method and step
Crackle reconstruction is carried out in the way of the geometric data for obtaining crackle characteristic in E.
7. according to the method for claim 6, it is characterised in that for one around toposcopy point in method and step E
Orientation is determined for each measurement point in the reconstruction area and by means of pattern identification means in individual partial reconstruction region
The measurement point corresponding to the crackle is determined with the similarity system design of mulle.
8. according to the method for claim 4, it is characterised in that at least one toposcopy point is to utilize key point search
Method determines.
9. according to the method for claim 8, it is characterised in that the crucial point searching method is SIFT methods, GLOH methods
With at least one method in SURF methods.
10. according to the method for claim 9, it is characterised in that the crucial point searching method is using being capable of control azimuth
Wave filter realize.
11. according to the method for claim 10, it is characterised in that the wave filter is bar chart wave filter.
12. according to the method for claim 11, it is characterised in that the bar chart wave filter is led using Gaussian filter
Number.
13. according to the method for claim 12, it is characterised in that the derivative is second dervative.
14. the method according to one of claim 4 to 9, it is characterised in that in the crack detecting method and the offer
In forecast model method, the descriptor in method and step D1 be according to SIFT methods, GLOH methods, HOG methods, LESH methods and
At least one method in SURF methods determines, and/or
In order to establish the descriptor, the maximal filter response of scalable and rotatable wave filter and orientation are used.
15. according to the method for claim 14, it is characterised in that the wave filter is bar chart wave filter.
16. according to the method for claim 15, it is characterised in that in the crack detecting method, according to comprising following
The mode of method and step establishes the descriptor：
D1.1 for the test zone each toposcopy point, by determine for each measurement point of the test zone parameter come
Determine amplitude and orientation；
D1.2 establishes feature histogram at least selected toposcopy point of the test zone.
17. according to the method for claim 16, it is characterised in that the parameter is further：
The gradient of the measured value of each measurement point；Or
The maximal filter response of the controllable filter of each measurement point and orientation.
18. according to the method for claim 17, it is characterised in that the main orientation of crackle is also determined, and according to crackle master
Orientation rotates constant feature histogram to correct this feature histogram to obtain.
19. according to the method for claim 18, it is characterised in that the main orientation of crackle is between method and step D1.1 and D1.2
To determine, and this feature histogram is by according to GLOH and/or SIFT and/or GLOH or SIFT in method and step D1.2
The scanning algorithm of flexible program scan the measurement point in the test zone to realize.
20. the method according to one of claim 9 to 13, it is characterised in that in the crack detecting method, according to bag
The mode for including following methods step determines the test zone in method and step D1：
At least one toposcopy point is determined,
The predetermined geometry for the test zone being incorporated in around the toposcopy point extends size to determine the measurement point in the test zone.
21. according to the method for claim 20, it is characterised in that in the crack detecting method, with the local inspection
The predetermined geometry of the test zone of rectangle, ellipse or the circle put centered on making an inventory of extends size to determine the survey in the test zone
Amount point.
22. the method according to claim 4 or 5, it is characterised in that in the crack detecting method, the geometry of crystal boundary
Position sets and/or determined by measurement of spatial resolution method as the crystal boundary data in semiconductor structure, and
So it is weighted according to crystal boundary data：
In the crack detecting method, when the descriptor in method and step D is established, the part on crystal boundary is not considered
Point, or the partial points on crystal boundary are considered with the weight lower than remaining partial points.
23. the method according to claim 6 or 7, it is characterised in that in the crack detecting method, the geometry of crystal boundary
Position sets and/or determined by measurement of spatial resolution method as the crystal boundary data in semiconductor structure, and
So it is weighted according to crystal boundary data：
In the crack detecting method, when the descriptor in method and step D is established, the part on crystal boundary is not considered
Point, or the partial points on crystal boundary are considered with the weight lower than remaining partial points, and/or
When crackle is rebuild, do not consider partial points on the crystal boundary or consider to be located at the weight lower than remaining partial points
Partial points on crystal boundary.
24. the method according to one of preceding claims 8 to 13, it is characterised in that brilliant in the crack detecting method
The geometric position on boundary sets and/or determined by measurement of spatial resolution method as the crystal boundary data in semiconductor structure,
And
So it is weighted according to crystal boundary data,
- in the crack detecting method and the offer forecast model method, when key point determines, the office on crystal boundary
Portion's point is not selected as key point, or it is determined that the partial points being pointed on crystal boundary during key point assign it is lower than remaining partial points
Weight, and/or
- in the crack detecting method, when the descriptor in method and step D is established, do not consider the part on crystal boundary
Point, or the partial points on crystal boundary are considered with the weight lower than remaining partial points.
25. the method according to one of preceding claims 4 to 13, it is characterised in that the learning algorithm be neuron net or
Bayes classifier or the learning algorithm based on core engine.
26. according to the method for claim 25, it is characterised in that the learning algorithm is algorithm of support vector machine.
27. it is a kind of be used on semiconductor structure detect crackle device, the semiconductor structure be photovoltaic solar cell or
The prime thing of photovoltaic solar cell in manufacturing process, comprising semiconductor source material,
Caused luminescence generated by light and/or measurement of spatial resolution should in the semiconductor structure including being used for measurement of spatial resolution for the device
Measuring unit that the IR of semiconductor structure absorbs and for reference to determined by the measuring unit measurement of spatial resolution data Lai
Detect the analytic unit of crackle, it is characterised in that the analytic unit is configured to be split according to one of claim 4 to 26
Line detects.
28. a kind of method for being used to detect crackle on semiconductor structure, the semiconductor structure be photovoltaic solar cell or
Photovoltaic solar cell prime thing in manufacturing process, this method include following methods step：
A., semiconductor structure is provided；
B. forecast model is provided, the forecast model is established by the training of learning algorithm；
C. multiple local measurement points of caused luminescence generated by light and/or passed through in the semiconductor structure by measurement of spatial resolution
The IR of the measurement of spatial resolution semiconductor structure, which absorbs, carrys out the measurement of spatial resolution semiconductor structure；
D. determine whether there is crackle on the toposcopy point at least one toposcopy point, including following method step
Suddenly：
D1. at least one descriptor of the toposcopy point is established by following manner, i.e. for the toposcopy point setting or
Determine a test zone and combine the test zone in measurement data establish the descriptor, the descriptor be characteristic vector and/
Or feature distribution and/or feature histogram, and
D2. determine whether there is crackle in the toposcopy point using the descriptor and forecast model,
Wherein, when detecting crackle according to method and step D, according to the geometric data for obtaining crackle characteristic in method and step E
Mode carries out crackle reconstruction,
Wherein, the geometric position of crystal boundary is set and/or by measurement of spatial resolution as the crystal boundary data in semiconductor structure
Method determines, and
When-the descriptor in method and step D is established, the partial points on crystal boundary are not considered, or with than remaining partial points
Low weight considers the partial points on crystal boundary；And/or
- when crackle is rebuild, do not consider partial points on the crystal boundary or position is considered with the weight lower than remaining partial points
In the partial points on crystal boundary.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE102011105182.5 | 2011-06-17 | ||
| DE102011105182A DE102011105182A1 (en) | 2011-06-17 | 2011-06-17 | A method of providing a predictive model for crack detection and a method of crack detection on a semiconductor structure |
| PCT/EP2012/061496 WO2012172073A1 (en) | 2011-06-17 | 2012-06-15 | Method for providing a prediction model for crack detection and method for crack detection on a semiconductor structure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103733322A CN103733322A (en) | 2014-04-16 |
| CN103733322B true CN103733322B (en) | 2017-11-24 |
Family
ID=46384360
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201280038885.3A Expired - Fee Related CN103733322B (en) | 2011-06-17 | 2012-06-15 | The method of crack detection forecast model and the method for the crackle on detection semiconductor structure are provided |
Country Status (4)
| Country | Link |
|---|---|
| CN (1) | CN103733322B (en) |
| DE (2) | DE102011105182A1 (en) |
| HU (1) | HUP1400302A1 (en) |
| WO (1) | WO2012172073A1 (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102015119360A1 (en) | 2015-11-10 | 2017-05-11 | Albert-Ludwigs-Universität Freiburg | Method and device for testing the contacting quality of an electrical contact between a solar cell and a contacting unit |
| CN105869164A (en) * | 2016-03-28 | 2016-08-17 | 国网浙江省电力公司宁波供电公司 | Method and system for detecting on/off state of switch |
| CN106814086B (en) * | 2016-12-07 | 2019-12-27 | 青岛海尔股份有限公司 | Inner container cracking analysis method |
| EP3367166A1 (en) * | 2017-02-24 | 2018-08-29 | ASML Netherlands B.V. | Method of measuring variation, inspection system, computer program, and computer system |
| CN110431395A (en) * | 2017-03-13 | 2019-11-08 | 通用电气公司 | Fatigue crack growth prediction |
| CN107389697B (en) * | 2017-07-10 | 2019-08-30 | 北京交通大学 | A kind of crack detection method based on half interactive mode |
| EP3569147B1 (en) * | 2018-05-16 | 2021-07-21 | Siemens Healthcare GmbH | Method and device for determining a geometric calibration for an imaging device and method for determining allocation data for geometric calibration |
| CN108986086A (en) * | 2018-07-05 | 2018-12-11 | 福州大学 | The detection of typographical display panel inkjet printing picture element flaw and classification method and its device |
| EP3627568B1 (en) * | 2018-09-21 | 2024-03-27 | Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. | Method for processing of images of semiconductor structures and for process characterization and process optimisation by means of semantic data compression |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH07201946A (en) * | 1993-12-28 | 1995-08-04 | Hitachi Ltd | Manufacture of semiconductor device and apparatus for manufacture the same, testing of the same and testing apparatus |
| DE19914115A1 (en) * | 1998-04-20 | 1999-11-04 | Gfai Ges Zur Foerderung Angewa | Error analysis of polycrystalline wafer, solar cell, and solar module |
| US6539106B1 (en) * | 1999-01-08 | 2003-03-25 | Applied Materials, Inc. | Feature-based defect detection |
| US7096207B2 (en) * | 2002-03-22 | 2006-08-22 | Donglok Kim | Accelerated learning in machine vision using artificially implanted defects |
-
2011
- 2011-06-17 DE DE102011105182A patent/DE102011105182A1/en not_active Withdrawn
-
2012
- 2012-06-15 WO PCT/EP2012/061496 patent/WO2012172073A1/en active Application Filing
- 2012-06-15 DE DE112012002509.1T patent/DE112012002509A5/en active Pending
- 2012-06-15 HU HU1400302A patent/HUP1400302A1/en unknown
- 2012-06-15 CN CN201280038885.3A patent/CN103733322B/en not_active Expired - Fee Related
Non-Patent Citations (2)
| Title |
|---|
| Analysis of luminescence images applying pattern recognition techniques;Demant M,et al;《25th european photovoltaic solar energy conference and exhibition and 5th world conference on photovoltaic energy conversion》;20100906;第1079-1082页第2.3、3、4部分，附图4-5 * |
| Automatic inspection of surface crack in solar cell images;Li Bin,et al;《Chinese control and decision conference》;20110523;第993-997页第II、III部分，附图1-7 * |
Also Published As
| Publication number | Publication date |
|---|---|
| HUP1400302A1 (en) | 2014-10-28 |
| WO2012172073A1 (en) | 2012-12-20 |
| DE112012002509A5 (en) | 2014-12-04 |
| DE102011105182A1 (en) | 2012-12-20 |
| CN103733322A (en) | 2014-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103733322B (en) | The method of crack detection forecast model and the method for the crackle on detection semiconductor structure are provided | |
| CN104751178B (en) | Lung neoplasm detection means and method based on shape template matching combining classification device | |
| CN113592845A (en) | Defect detection method and device for battery coating and storage medium | |
| CN107230203B (en) | Casting defect identification method based on human eye visual attention mechanism | |
| CN108765389A (en) | A kind of microcosmic wafer surface defects image detecting method | |
| CN110210448B (en) | Intelligent face skin aging degree identification and evaluation method | |
| CN103940708B (en) | A kind of Quick Measurement, sophisticated category method of steel holotype state crystal grain | |
| CN107122737A (en) | A kind of road signs automatic detection recognition methods | |
| CN102621150A (en) | Airplane skin damage identification method based on gray level co-occurrence matrix and support vector machine | |
| Posilović et al. | Flaw detection from ultrasonic images using YOLO and SSD | |
| TW201220039A (en) | Inspection recipe generation and inspection based on an inspection recipe | |
| CN104880389A (en) | Mixed crystal degree automatic measurement and fine classification method for steel crystal grains, and system thereof | |
| Rodolà et al. | A simple and effective relevance-based point sampling for 3D shapes | |
| EP3785223A1 (en) | Systems and methods for segmentation and analysis of 3d images | |
| CN112488211A (en) | Fabric image flaw classification method | |
| Gui et al. | Object-based crack detection and attribute extraction from laser-scanning 3D profile data | |
| He et al. | Robust illumination invariant texture classification using gradient local binary patterns | |
| Lin et al. | Identification of broken rice kernels using image analysis techniques combined with velocity representation method | |
| Jin et al. | End Image Defect Detection of Float Glass Based on Faster Region-Based Convolutional Neural Network. | |
| Ghamisi et al. | Anomaly detection in automated fibre placement: learning with data limitations | |
| CN102722718B (en) | Method for classifying cells | |
| Dawood et al. | Texture image classification with improved weber local descriptor | |
| CN107478656A (en) | Paper pulp mixing effect method of determination and evaluation based on machine vision, device, system | |
| CN106327462B (en) | Print the positioning core extracting method and extraction element of image | |
| CN109636781B (en) | Defect extraction method based on feature mining and weighted Bayesian classifier |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20171124 |