CN109919907A - The method and apparatus of LED epitaxial slice defects detection - Google Patents
The method and apparatus of LED epitaxial slice defects detection Download PDFInfo
- Publication number
- CN109919907A CN109919907A CN201910064372.3A CN201910064372A CN109919907A CN 109919907 A CN109919907 A CN 109919907A CN 201910064372 A CN201910064372 A CN 201910064372A CN 109919907 A CN109919907 A CN 109919907A
- Authority
- CN
- China
- Prior art keywords
- defect
- image
- epitaxial wafer
- type
- convolutional neural
- 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.)
- Pending
Links
Landscapes
- Testing Or Measuring Of Semiconductors Or The Like (AREA)
- Image Analysis (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The invention discloses a kind of method and apparatus of LED epitaxial slice defects detection, belong to technical field of semiconductors.The described method includes: obtaining the image of epitaxial wafer to be detected；The image of the epitaxial wafer to be detected is compared with the image of zero defect epitaxial wafer, the image of defect part is intercepted from the image of the epitaxial wafer to be detected；The image of the defect part is inputted into convolutional neural networks, obtains the defect type of the defect part, the parameter of the convolutional neural networks is trained to obtain by using the defect image of multiple defective types of calibration.The present invention passes through after the image for obtaining epitaxial wafer to be detected, first it is compared with the image of zero defect epitaxial wafer, therefrom intercept out the image of defect part, the convolutional neural networks for recycling the defect image of multiple defective types of calibration to train, detection efficiency can be greatly improved, and the accuracy of testing result can guarantee, especially meet industrial demand.
Description
Technical field
The present invention relates to technical field of semiconductors, in particular to a kind of method of LED epitaxial slice defects detection and
Device.
Background technique
Light emitting diode (English: Light Emitting Diode, referred to as: LED) it is that one kind can be converted to electric energy
The semiconductor diode of luminous energy.The core component of LED is chip, and chip includes epitaxial wafer and the electrode that extension on piece is arranged in.
Epitaxial wafer may generate various defects in process, such as hexagonal, micro- thick, scuffing, particle, atomization, green point
Deng, therefore after epitaxial wafer processing, defects detection can be generally carried out to epitaxial wafer, and according to testing result (including defect
Type, size, quantity etc.) divide epitaxial wafer grade.
The defect inspection method of most original is manually by micro- sem observation epitaxial wafer to be detected, detection efficiency and identification
Accuracy rate is all unable to satisfy industrial production demand.It is available to arrive extension to be detected later with the development of optical detection apparatus
The image of piece can determine that be checked by comparing the image of epitaxial wafer to be detected and the image of zero defect epitaxial wafer
Surveying epitaxial wafer whether there is defect, and detection efficiency and recognition accuracy have all obtained great promotion, but detectability is limited,
It cannot achieve the grade classification of epitaxial wafer.It is now based on being substantially improved for computer process ability, it can be by epitaxial wafer to be detected
Image of the image successively with various defect epitaxial wafers compare, the image of the highest defect epitaxial wafer of similarity is corresponding
Defect type of the defect type as epitaxial wafer to be detected, and then according to the defect type divided rank of epitaxial wafer to be detected.
In the implementation of the present invention, the inventor finds that the existing technology has at least the following problems:
The form of expression of same type defect is varied on epitaxial wafer, such as hexagonal crystal system, cubic system and orthorhombic system
The shape of upper planar defect is different.But the image of all representation defects can not be substantially provided in specific implementation
It is compared with the image of epitaxial wafer to be detected, the accuracy of testing result is not high, and the image of epitaxial wafer to be detected is divided
Image not with the various representations of same type defect compares, and compares substantial amounts, and detection efficiency is lower.And if only
A kind of image of form of expression defect is selected to compare, then testing result is likely to inaccuracy, is still unable to satisfy industrial life
Production demand.
Summary of the invention
The embodiment of the invention provides a kind of method and apparatus of LED epitaxial slice defects detection, it is able to solve existing
The problem of accuracy for having technological deficiency to detect is unable to satisfy industrial production demand.The technical solution is as follows:
On the one hand, the embodiment of the invention provides a kind of method of LED epitaxial slice defects detection, the methods
Include:
Obtain the image of epitaxial wafer to be detected；
The image of the epitaxial wafer to be detected is compared with the image of zero defect epitaxial wafer, from the extension to be detected
The image of defect part is intercepted in the image of piece；
The image of the defect part is inputted into convolutional neural networks, obtains the defect type of the defect part, it is described
The parameter of convolutional neural networks is trained to obtain by using the defect image of multiple defective types of calibration.
Optionally, the image by the defect part inputs convolutional neural networks, obtains lacking for the defect part
Fall into type, comprising:
The image of the defect part is normalized, the image of predetermined dimension is obtained；
The image of the predetermined dimension is inputted into the convolutional neural networks, obtains the defect type of the defect part.
Optionally, the method also includes:
Obtain multiple defect images；
Receive the defect type of each defect image calibration；
Using the defect type of multiple defect images and each defect image calibration, to the convolutional Neural net
The parameter of network is trained.
Further, described using multiple defect images and the defect type of each defect image calibration, it is right
The parameter of the convolutional neural networks is trained, comprising:
Multiple defect images are successively inputted into the convolutional neural networks, obtain the defect of each defect image
Type, and in the defect type difference of defect type and calibration that the defect image obtains, backpropagation adjusts the volume
The parameter of product neural network, until the defect type that multiple defect images obtain is identical as the defect type of calibration.
Optionally, the method also includes:
Count the quantity, size and defect class of the image of the defect part intercepted in the image of the epitaxial wafer to be detected
Type determines the credit rating of the epitaxial wafer to be detected.
On the other hand, the embodiment of the invention provides a kind of device of LED epitaxial slice defects detection, the dresses
It sets and includes:
Detection image obtains module, for obtaining the image of epitaxial wafer to be detected；
Interception module, for the image of the epitaxial wafer to be detected to be compared with the image of zero defect epitaxial wafer, from
The image of defect part is intercepted in the image of the epitaxial wafer to be detected；
Determination type module obtains the defective part for the image of the defect part to be inputted convolutional neural networks
The parameter of the defect type divided, the convolutional neural networks is instructed by using the defect images of multiple defective types of calibration
It gets.
Optionally, the determination type module includes:
Submodule is normalized, is normalized for the image to the defect part, obtains the image of predetermined dimension；
It determines submodule, for the image of the predetermined dimension to be inputted the convolutional neural networks, obtains the defect
Partial defect type.
Optionally, described device further include:
Training image obtains module, for obtaining multiple defect images；
Receiving module, for receiving the defect type of each defect image calibration；
Training module, it is right for the defect type using multiple defect images and each defect image calibration
The parameter of the convolutional neural networks is trained.
Further, the training module is used for,
Multiple defect images are successively inputted into the convolutional neural networks, obtain the defect of each defect image
Type, and in the defect type difference of defect type and calibration that the defect image obtains, backpropagation adjusts the volume
The parameter of product neural network, until the defect type that multiple defect images obtain is identical as the defect type of calibration.
Optionally, described device further include:
Level determination module, the number of the image of the defect part intercepted in the image for counting the epitaxial wafer to be detected
Amount, size and defect type, determine the credit rating of the epitaxial wafer to be detected.
Technical solution provided in an embodiment of the present invention has the benefit that
By first it being compared with the image of zero defect epitaxial wafer after the image for obtaining epitaxial wafer to be detected,
Therefrom intercept out the image of defect part, can for defect part carry out defect type judgement, exclude non-defective part or
The influence of the other defect parts of person, it is with strong points, be conducive to improve accuracy and reduce operand.Multiple calibration are recycled to have scarce
The convolutional neural networks that the defect image of type trains are fallen into, the defect type of defect part is obtained by the image of defect part,
There is no multiple the problem of comparing, and can greatly improve detection efficiency.And the parameter of convolutional neural networks is lacked using various
Sunken image train come, various defect types, defect the various forms of expression can all be related to, the accuracy of testing result can
To guarantee, especially meet industrial demand.
Detailed description of the invention
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
Fig. 1 is a kind of flow chart of the method for LED epitaxial slice defects detection provided in an embodiment of the present invention；
Fig. 2 is the flow chart of the method for another LED epitaxial slice defects detection provided in an embodiment of the present invention；
Fig. 3 is a kind of structural representation of the device of LED epitaxial slice defects detection provided in an embodiment of the present invention
Figure.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with attached drawing to embodiment party of the present invention
Formula is described in further detail.
The embodiment of the invention provides a kind of methods of LED epitaxial slice defects detection.Fig. 1 is that the present invention is implemented
A kind of flow chart of the method for LED epitaxial slice defects detection that example provides.Referring to Fig. 1, this method comprises:
Step 101: obtaining the image of epitaxial wafer to be detected.
In practical applications, can use automatic optics inspection (English: Automated Optical Inspection,
Referred to as: AOI) the equipment image that obtains epitaxial wafer to be detected.AOI equipment is to be detected based on optical principle to workpiece, defect
Equipment.When AOI equipment detects workpiece, camera scanning workpiece can be first passed through, the image of workpiece is obtained；Reprocessing figure
Picture checks the defect on workpiece.But the defect kind that AOI equipment can identify is reduced, therefore the present invention is just with AOI
Equipment gets the image of epitaxial wafer to be detected.
Step 102: the image of epitaxial wafer to be detected being compared with the image of zero defect epitaxial wafer, from extension to be detected
The image of defect part is intercepted in the image of piece.
In the present embodiment, the image of zero defect epitaxial wafer is the image for not having defective epitaxial wafer.Do not have defective outer
Prolonging piece can be determined by the way of artificial detection, and the image of epitaxial wafer can use the acquisition of AOI equipment.
In practical applications, epitaxial wafer is that the substrate base that proper temperature is heated at one piece (mainly has sapphire, carbonization
Silicon, silicon etc.) on the specific monocrystal thin films that are grown.Due to the thinner thickness of epitaxial wafer entirety, usually just by epitaxial wafer
Face, i.e., image of the image for the film surface that last growth is formed as epitaxial wafer.
If there is no defect on epitaxial wafer to be detected, the image of epitaxial wafer to be detected and the image one of zero defect epitaxial wafer
Sample；If existing defects on epitaxial wafer to be detected, the image of defect part can be with zero defect epitaxial wafer in epitaxial wafer to be detected
The image of middle corresponding portion is different, while the image of other parts and corresponding portion in zero defect epitaxial wafer in epitaxial wafer to be detected
Image it is the same.Therefore, directly the image of epitaxial wafer to be detected is compared with the image of zero defect epitaxial wafer, is found out to be checked
Part different from zero defect epitaxial wafer in epitaxial wafer is surveyed, the image of defect part in epitaxial wafer to be detected can be obtained.
In addition, due to the image size of each epitaxial wafer be it is the same, specific implementation when, can be successively more to be checked
Survey the characteristic value of the characteristic value of each pixel and same position pixel on the image of zero defect epitaxial wafer in the image of epitaxial wafer, choosing
It takes characteristic value different or difference is more than all pixels of setting range, and distance mutual in the pixel of selection is not surpassed
Two pixels for crossing setting value are classified as the image of the same defect part.
For example, the image size of each epitaxial wafer is a pixel * b pixel, a and b are positive integer, then more to be detected outer
The feature of the 1st column pixel of 1st row in the characteristic value for prolonging the 1st column pixel of the 1st row in the image of piece, with the image of zero defect epitaxial wafer
Whether identical or difference is more than setting range to value；The feature of the 2nd column pixel of 1st row in the image of epitaxial wafer more to be detected
Value, and whether the characteristic value of the 2nd column pixel of the 1st row in the image of zero defect epitaxial wafer identical or difference is more than setting model
It encloses；……；In the image of epitaxial wafer more to be detected in the characteristic value of 1st row b column pixel, with the image of zero defect epitaxial wafer
Whether the characteristic value of the 1st row b column pixel is identical or difference is more than setting range；In the image of epitaxial wafer more to be detected
In the characteristic value of the 1st column pixel of 2 row, with the image of zero defect epitaxial wafer the characteristic value of the 1st column pixel of the 2nd row it is whether identical or
Difference is more than setting range；……；The characteristic value of a row b column pixel in the image of epitaxial wafer more to be detected, with zero defect
Whether the characteristic value of a row b column pixel is identical in the image of epitaxial wafer or difference is more than setting range.
If in the image of epitaxial wafer to be detected in the characteristic value of i-th row jth column pixel and the image of zero defect epitaxial wafer
The characteristic value of i-th row jth column pixel is different or difference is more than setting range, while m row in the image of epitaxial wafer to be detected
The characteristic value of n-th column pixel is different from the characteristic value of m row the n-th column pixel in the image of zero defect epitaxial wafer or difference is more than
Setting range, i, j, m and n are positive integer, i < a, j < b, m < a, n < b.If the i-th row jth column pixel and the i-th row jth
The distance between column pixel is no more than setting value, then by the i-th row jth column pixel in the image of epitaxial wafer to be detected and the i-th row the
(j+1) column pixel is classified as the image of the same defect part；If between the i-th row jth column pixel and the i-th row jth column pixel
Distance is more than setting value, then is classified as the i-th row jth column pixel and i-th row (j+1) column pixel in the image of epitaxial wafer to be detected
The image of two different defect parts.
Step 103: the image of defect part being inputted into convolutional neural networks, obtains the defect type of defect part, convolution
The parameter of neural network is trained to obtain by using the defect image of multiple defective types of calibration.
Convolutional neural networks (Convolutional Neural Network, abbreviation CNN) are the one of artificial neural network
Kind, it has also become the research hotspot of current speech analysis and field of image recognition.Its weight shares network structure and is allowed to more similar
In biological neural network, the complexity of network model is reduced, reduces the quantity of weight.The advantage is more in the input of network
What is showed when tieing up image becomes apparent, and allows image directly as the input of network, avoids complicated in tional identification algorithm
Feature extraction and data reconstruction processes.Convolutional neural networks are one Multilayer Perception of special designing for identification two-dimensional shapes
Device, this network structure have height invariance to the deformation of translation, inclination or other forms.
Input is set as the image of defect part, output on the basis of general convolutional neural networks by the present embodiment
It is set as the defect type of defect part, and using the defect image of multiple defective types of calibration in convolutional neural networks
Parameter is trained.
The embodiment of the present invention is by after the image for obtaining epitaxial wafer to be detected, first by the figure of itself and zero defect epitaxial wafer
As being compared, the image of defect part is therefrom intercepted out, the judgement of defect type can be carried out for defect part, is excluded non-
The influence of defect part or other defect parts, it is with strong points, be conducive to improve accuracy and reduce operand.It recycles more
The convolutional neural networks that the defect image of a defective type of calibration trains, obtain defect part by the image of defect part
Defect type can greatly improve detection efficiency there is no repeatedly comparing.And the parameter of convolutional neural networks is to adopt
With various defect images train come, various defect types, defect the various forms of expression can all be related to, testing result
Accuracy can guarantee, especially meet industrial demand.
The embodiment of the invention provides the methods of another LED epitaxial slice defects detection, are hairs shown in FIG. 1
A kind of specific implementation of the method for optical diode epitaxial wafer defects detection.Fig. 2 is provided in an embodiment of the present invention another luminous
The flow chart of the method for diode epitaxial slice defects detection.Referring to fig. 2, this method comprises:
Step 201: obtaining multiple defect images.
Illustratively, which may include:
Obtain the image of multiple epitaxial wafers；
Successively the image of each epitaxial wafer is compared with the image of zero defect epitaxial wafer, and will be in the image of epitaxial wafer
The parts different from the image of zero defect epitaxial wafer are as defect image.
In practical applications, the image of the epitaxial wafer can use the acquisition of AOI equipment.
In addition, the training effect (output of such as convolutional neural networks is accurate) of the parameter in order to ensure convolutional neural networks,
It needs to obtain a large amount of defect image to be trained, such as the quantity of the defect image of various defect types is at 1000 or more.
Optionally, after step 201, this method can also include:
Defect image is normalized, the defect image of predetermined dimension is obtained.
Wherein, predetermined dimension may include the size or format of setting.
In practical applications, in order to guarantee that convolutional neural networks export the accuracy of result, all inputs are subjected to normalizing
Change processing, for example it is normalized to the image of 64 pixel *, 64 pixel, convolution can be caused refreshing to avoid due to the inconsistent of input picture
Error is generated through network output result.
Step 202: receiving the defect type of each defect image calibration.
In practical applications, in order to be adjusted when convolutional neural networks export incorrect to the parameter of convolutional neural networks
Whole, while getting defect image, the defect type of the also correspondingly received each defect image artificially demarcated is (as different in used
Number respectively indicate different defect types), to be carried out according to the parameter of the defect type artificially demarcated to convolutional neural networks
Adjustment, the defect type for exporting convolutional neural networks are identical as received defect type.
Step 203: the defect type demarcated using multiple defect images and each defect image, to convolutional neural networks
Parameter is trained.
Since people is from part to the overall situation to extraneous cognition, and the space relationship of image is also that local pixel is more tight
It is close, it is then weaker apart from farther away pixel interdependence.Therefore, each neuron of neural network is not necessarily in fact to global image
It is perceived, it is only necessary to part be perceived, then get up the informix of part just to have obtained the overall situation in higher
Information.
Convolutional neural networks are using the convolutional layer (alternating being generally arranged at close to network inputs end position
Convolutional layer) it realizes to locally perceiving, being connected using being generally arranged at close to the complete of network output position
Layer is connect to realize the informix of part.
In the concrete realization, convolutional layer does convolution using at least one convolution kernel on the image, extracts each office in image
The feature of portion sensing region, the characteristics of image that different types of convolution kernel extracts are different.Full articulamentum is establish one layer each
The connection of a neuron and next layer of all neurons.All inputs of full articulamentum and any one output meet following public
Formula:
H=f (∑iWi*xi+b)；
Wherein, h is the output of full articulamentum, and b is bias, xiFor each input of full articulamentum, WiFor full articulamentum
Each weight for being input to output, i indicate any one input of full articulamentum, ∑iIndicate all inputs to full articulamentum
Summation, f () representative function relationship, generally sigmoid function or tanh function.
The parameter of convolutional neural networks may include the convolution kernel that convolutional layer uses in the present embodiment, and full articulamentum is adopted
Weight and bias.In addition, convolutional neural networks can also include being generally arranged at volume in addition to convolutional layer and full articulamentum
The pond layer (pooling layer) of lamination output end, for reducing the dimension of characteristics of image.Spy of the pond layer to different location
Sign carries out aggregate statistics.If convolutional neural networks further include pond layer, the parameter of convolutional neural networks can also include pond
Change the unit that layer is divided.
In practical applications, the initial value of the parameter of convolutional neural networks can be randomly provided, and then successively be lacked each
It falls into image and inputs convolutional neural networks, and after each defect image inputs convolutional neural networks, obtain convolutional neural networks
The defect type of output is compared with the defect type of calibration, when the two does not adjust the parameter of convolutional neural networks simultaneously, with
Keep the two identical.I.e. the step 203 may include:
Multiple defect images are successively inputted into convolutional neural networks, obtain the defect type of each defect image, and lacking
When the defect type that sunken image obtains and the defect type difference of calibration, backpropagation adjusts the parameter of convolutional neural networks, directly
The defect type obtained to multiple defect images is identical as the defect type of calibration.
For example, the quantity of defect image is 1000.The 1st defect image is first inputted into convolutional neural networks, obtains the 1st
The defect type of a defect image, if the defect type that the 1st defect image obtains is protected as the defect type of calibration
Hold the parameter constant of convolutional neural networks；If the defect type that the 1st defect image obtains is different from the defect type of calibration,
Then according to the parameter of the sequence adjustment convolutional neural networks of setting, the defect type obtained until the 1st defect image and calibration
Defect type it is the same.The 2nd defect image is inputted into convolutional neural networks again, obtains the defect type of the 2nd defect image,
If the defect type that the 2nd defect image obtains keeps the parameter of convolutional neural networks as the defect type of calibration
It is constant；If the defect type that the 2nd defect image obtains is different from the defect type of calibration, adjusted according to the sequence of setting
The parameter of convolutional neural networks, most until the defect type that the 2nd defect image obtains is as the defect type of calibration ...
The 1000th defect image is inputted into convolutional neural networks afterwards, obtains the defect type of the 1000th defect image, if the
The defect type that 1000 defect images obtain then keeps the parameter of convolutional neural networks not as the defect type of calibration
Become；If the defect type that the 1000th defect image obtains is different from the defect type of calibration, according to the sequence tune of setting
The parameter of whole convolutional neural networks, until the defect type that the 1000th defect image obtains is as the defect type of calibration.
Then above-mentioned circulation is repeated, until being always maintained at the parameter constants of convolutional neural networks in a circulation, then convolutional neural networks
Training finishes.
Further, backpropagation adjusts the parameter of convolutional neural networks, may include:
Using the parameter of gradient descent method adjustment convolutional neural networks.
In practical applications, can gradually adjust and (be gradually increased or be gradually reduced) parameter of convolutional neural networks, such as after
The continuous parameter for increasing or reducing convolutional neural networks, until obtained defect type is as the defect type of calibration.
It should be noted that step 201- step 203 is optional step, may be implemented pair by step 201- step 203
The training of the parameter of convolutional neural networks.
Step 204: obtaining the image of epitaxial wafer to be detected.
Illustratively, which can be identical as step 101, and this will not be detailed here.
Step 205: the image of epitaxial wafer to be detected being compared with the image of zero defect epitaxial wafer, from extension to be detected
The image of defect part is intercepted in the image of piece.
Illustratively, which can be identical as step 102, and this will not be detailed here.
Step 206: the image of defect part being inputted into convolutional neural networks, obtains the defect type of defect part, convolution
The parameter of neural network is trained to obtain by using the defect image of multiple defective types of calibration.
Optionally, which may include:
The image of defect part is normalized, the image of predetermined dimension is obtained；
The image of predetermined dimension is inputted into convolutional neural networks, obtains the defect type of defect part.
It is corresponding with step 201, image is normalized and inputs convolutional neural networks again, it can be to avoid due to input
The inconsistent of image causes convolutional neural networks output result to generate error.
Step 207: counting the quantity, size and defect of the image of the defect part intercepted in the image of epitaxial wafer to be detected
Type determines the credit rating of epitaxial wafer to be detected.
In practical applications, can according in the target level of product quality of industry universal to defects count, size and type etc.
Epitaxial wafer to be detected is divided into corresponding grade by the demand of aspect according to statistical result.
It should be noted that step 207 is optional step, the screening of epitaxial wafer may be implemented, by step 207 so as to right
Epitaxial wafer carries out different processing.For example, for the epitaxial wafer of (such as defective proportion be lower than 10%) and up-to-standard of having good quality
The epitaxial wafer of (such as defective proportion is between 20%~30%) carries out different packing, for (such as defective proportion off quality
40% or more) epitaxial wafer re-work.
The embodiment of the invention provides a kind of device of LED epitaxial slice defects detection, be adapted to carry out Fig. 1 or
The method of person's LED epitaxial slice defects detection shown in Fig. 2.Fig. 3 is a kind of light-emitting diodes provided in an embodiment of the present invention
The structural schematic diagram of the device of pipe epitaxial wafer defects detection.Referring to Fig. 3, which includes:
Detection image obtains module 301, for obtaining the image of epitaxial wafer to be detected；
Interception module 302, for the image of epitaxial wafer to be detected to be compared with the image of zero defect epitaxial wafer, to
Detect the image that defect part is intercepted in the image of epitaxial wafer；
Determination type module 303 obtains lacking for defect part for the image of defect part to be inputted convolutional neural networks
Type is fallen into, the parameter of convolutional neural networks is trained to obtain by using the defect image of multiple defective types of calibration.
Optionally, determination type module 303 may include:
Submodule is normalized, is normalized for the image to defect part, obtains the image of predetermined dimension；
It determines submodule, for the image of predetermined dimension to be inputted convolutional neural networks, obtains the defect class of defect part
Type.
Optionally, which can also include:
Training image obtains module, for obtaining multiple defect images；
Receiving module, for receiving the defect type of each defect image calibration；
Training module, the defect type for being demarcated using multiple defect images and each defect image, to convolutional Neural
The parameter of network is trained.
Further, training module can be used for,
Multiple defect images are successively inputted into convolutional neural networks, obtain the defect type of each defect image, and lacking
When the defect type that sunken image obtains and the defect type difference of calibration, backpropagation adjusts the parameter of convolutional neural networks, directly
The defect type obtained to multiple defect images is identical as the defect type of calibration.
Optionally, which can also include:
Level determination module, the quantity of the image of the defect part intercepted in the image for counting epitaxial wafer to be detected,
Size and defect type determine the credit rating of epitaxial wafer to be detected.
It should be understood that the device of LED epitaxial slice defects detection provided by the above embodiment shines in detection
It, only the example of the division of the above functional modules, can be according to need in practical application when diode epitaxial slice defect
It wants and is completed by different functional modules above-mentioned function distribution, i.e., the internal structure of device is divided into different function moulds
Block, to complete all or part of the functions described above.In addition, LED epitaxial slice defect provided by the above embodiment
The device of detection and the embodiment of the method for LED epitaxial slice defects detection belong to same design, and specific implementation process is detailed
See embodiment of the method, which is not described herein again.
Those of ordinary skill in the art will appreciate that realizing that all or part of the steps of above-described embodiment can pass through hardware
It completes, relevant hardware can also be instructed to complete by program, the program can store in a kind of computer-readable
In storage medium, storage medium mentioned above can be read-only memory, disk or CD etc..
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of method of LED epitaxial slice defects detection, which is characterized in that the described method includes:
Obtain the image of epitaxial wafer to be detected；
The image of the epitaxial wafer to be detected is compared with the image of zero defect epitaxial wafer, from the epitaxial wafer to be detected
The image of defect part is intercepted in image；
The image of the defect part is inputted into convolutional neural networks, obtains the defect type of the defect part, the convolution
The parameter of neural network is trained to obtain by using the defect image of multiple defective types of calibration.
2. the method according to claim 1, wherein the image by the defect part inputs convolutional Neural
Network obtains the defect type of the defect part, comprising:
The image of the defect part is normalized, the image of predetermined dimension is obtained；
The image of the predetermined dimension is inputted into the convolutional neural networks, obtains the defect type of the defect part.
3. method according to claim 1 or 2, which is characterized in that the method also includes:
Obtain multiple defect images；
Receive the defect type of each defect image calibration；
Using the defect type of multiple defect images and each defect image calibration, to the convolutional neural networks
Parameter is trained.
4. according to the method described in claim 3, it is characterized in that, described using multiple defect images and each described scarce
The defect type for falling into image calibration, is trained the parameter of the convolutional neural networks, comprising:
Multiple defect images are successively inputted into the convolutional neural networks, obtain the defect class of each defect image
Type, and in the defect type difference of defect type and calibration that the defect image obtains, backpropagation adjusts the convolution
The parameter of neural network, until the defect type that multiple defect images obtain is identical as the defect type of calibration.
5. method according to claim 1 or 2, which is characterized in that the method also includes:
Quantity, size and the defect type of the image of the defect part intercepted in the image of the epitaxial wafer to be detected are counted, really
The credit rating of the fixed epitaxial wafer to be detected.
6. a kind of device of LED epitaxial slice defects detection, which is characterized in that described device includes:
Detection image obtains module, for obtaining the image of epitaxial wafer to be detected；
Interception module, for the image of the epitaxial wafer to be detected to be compared with the image of zero defect epitaxial wafer, from described
The image of defect part is intercepted in the image of epitaxial wafer to be detected；
Determination type module obtains the defect part for the image of the defect part to be inputted convolutional neural networks
The parameter of defect type, the convolutional neural networks is trained by using the defect image of multiple defective types of calibration
It arrives.
7. device according to claim 6, which is characterized in that the determination type module includes:
Submodule is normalized, is normalized for the image to the defect part, obtains the image of predetermined dimension；
It determines submodule, for the image of the predetermined dimension to be inputted the convolutional neural networks, obtains the defect part
Defect type.
8. device according to claim 6 or 7, which is characterized in that described device further include:
Training image obtains module, for obtaining multiple defect images；
Receiving module, for receiving the defect type of each defect image calibration；
Training module, for the defect type using multiple defect images and each defect image calibration, to described
The parameter of convolutional neural networks is trained.
9. device according to claim 8, which is characterized in that the training module is used for,
Multiple defect images are successively inputted into the convolutional neural networks, obtain the defect class of each defect image
Type, and in the defect type difference of defect type and calibration that the defect image obtains, backpropagation adjusts the convolution
The parameter of neural network, until the defect type that multiple defect images obtain is identical as the defect type of calibration.
10. device according to claim 6 or 7, which is characterized in that described device further include:
Level determination module, the quantity of the image of the defect part intercepted in the image for counting the epitaxial wafer to be detected,
Size and defect type determine the credit rating of the epitaxial wafer to be detected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910064372.3A CN109919907A (en) | 2019-01-23 | 2019-01-23 | The method and apparatus of LED epitaxial slice defects detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910064372.3A CN109919907A (en) | 2019-01-23 | 2019-01-23 | The method and apparatus of LED epitaxial slice defects detection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109919907A true CN109919907A (en) | 2019-06-21 |
Family
ID=66960632
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910064372.3A Pending CN109919907A (en) | 2019-01-23 | 2019-01-23 | The method and apparatus of LED epitaxial slice defects detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109919907A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112599438A (en) * | 2021-03-03 | 2021-04-02 | 惠州高视科技有限公司 | High-precision detection system and detection method for MiniLED wafer defects |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105891215A (en) * | 2016-03-31 | 2016-08-24 | 浙江工业大学 | Welding visual detection method and device based on convolutional neural network |
| CN106204618A (en) * | 2016-07-20 | 2016-12-07 | 南京文采科技有限责任公司 | Product surface of package defects detection based on machine vision and sorting technique |
| CN106290378A (en) * | 2016-08-23 | 2017-01-04 | 东方晶源微电子科技（北京）有限公司 | Defect classification method and defect inspecting system |
| CN106504238A (en) * | 2016-10-31 | 2017-03-15 | 成都交大光芒科技股份有限公司 | Railway contact line defect inspection method based on image procossing and convolutional neural networks |
| CN107607554A (en) * | 2017-09-26 | 2018-01-19 | 天津工业大学 | A kind of Defect Detection and sorting technique of the zinc-plated stamping parts based on full convolutional neural networks |
| CN107611046A (en) * | 2017-07-28 | 2018-01-19 | 华灿光电（浙江）有限公司 | Inner layer defect detection device and method |
| WO2018106827A1 (en) * | 2016-12-07 | 2018-06-14 | Kla-Tencor Corporation | Data augmentation for convolutional neural network-based defect inspection |
| WO2018192662A1 (en) * | 2017-04-20 | 2018-10-25 | Hp Indigo B.V. | Defect classification in an image or printed output |
| CN109064454A (en) * | 2018-07-12 | 2018-12-21 | 上海蝶鱼智能科技有限公司 | Product defects detection method and system |
-
2019
- 2019-01-23 CN CN201910064372.3A patent/CN109919907A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105891215A (en) * | 2016-03-31 | 2016-08-24 | 浙江工业大学 | Welding visual detection method and device based on convolutional neural network |
| CN106204618A (en) * | 2016-07-20 | 2016-12-07 | 南京文采科技有限责任公司 | Product surface of package defects detection based on machine vision and sorting technique |
| CN106290378A (en) * | 2016-08-23 | 2017-01-04 | 东方晶源微电子科技（北京）有限公司 | Defect classification method and defect inspecting system |
| CN106504238A (en) * | 2016-10-31 | 2017-03-15 | 成都交大光芒科技股份有限公司 | Railway contact line defect inspection method based on image procossing and convolutional neural networks |
| WO2018106827A1 (en) * | 2016-12-07 | 2018-06-14 | Kla-Tencor Corporation | Data augmentation for convolutional neural network-based defect inspection |
| WO2018192662A1 (en) * | 2017-04-20 | 2018-10-25 | Hp Indigo B.V. | Defect classification in an image or printed output |
| CN107611046A (en) * | 2017-07-28 | 2018-01-19 | 华灿光电（浙江）有限公司 | Inner layer defect detection device and method |
| CN107607554A (en) * | 2017-09-26 | 2018-01-19 | 天津工业大学 | A kind of Defect Detection and sorting technique of the zinc-plated stamping parts based on full convolutional neural networks |
| CN109064454A (en) * | 2018-07-12 | 2018-12-21 | 上海蝶鱼智能科技有限公司 | Product defects detection method and system |
Non-Patent Citations (1)
| Title |
|---|
| 姚锡凡 等: "《制造物联网技术》", 31 December 2018, 华中科技大学出版社 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112599438A (en) * | 2021-03-03 | 2021-04-02 | 惠州高视科技有限公司 | High-precision detection system and detection method for MiniLED wafer defects |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109919908A (en) | The method and apparatus of light-emitting diode chip for backlight unit defects detection | |
| CN101981683B (en) | Method for classifying defects and device for classifying defects | |
| CN109977808A (en) | A kind of wafer surface defects mode detection and analysis method | |
| US8965116B2 (en) | Computer-aided assignment of ratings to digital samples of a manufactured web product | |
| US20130322733A1 (en) | System for detection of non-uniformities in web-based materials | |
| US20210334946A1 (en) | Method and system for classifying defects in wafer using wafer-defect images, based on deep learning | |
| CN113192040A (en) | Fabric flaw detection method based on YOLO v4 improved algorithm | |
| CN112633382B (en) | Method and system for classifying few sample images based on mutual neighbor | |
| CN109583324A (en) | A kind of pointer meters reading automatic identifying method based on the more box detectors of single-point | |
| CN109684906B (en) | Method for detecting red fat bark beetles based on deep learning | |
| CN113920107A (en) | Insulator damage detection method based on improved yolov5 algorithm | |
| CN109215015A (en) | A kind of online visible detection method of silk cocoon based on convolutional neural networks | |
| CN111652273B (en) | Deep learning-based RGB-D image classification method | |
| CN113222982A (en) | Wafer surface defect detection method and system based on improved YOLO network | |
| CN116091505B (en) | Automatic defect detection and classification method and system for sapphire substrate | |
| CN114359245A (en) | Method for detecting surface defects of products in industrial scene | |
| CN114494250A (en) | Floor defect detection method based on neural network model | |
| CN109919907A (en) | The method and apparatus of LED epitaxial slice defects detection | |
| CN114549391A (en) | Circuit board surface defect detection method based on polarization prior | |
| CN116843615B (en) | Lead frame intelligent total inspection method based on flexible light path | |
| CN115830302A (en) | Multi-scale feature extraction and fusion power distribution network equipment positioning identification method | |
| Nafi et al. | High accuracy swin transformers for image-based wafer map defect detection | |
| CN113256576B (en) | Automatic optical element detection system and method based on polarization imaging and machine learning | |
| CN113469984B (en) | Method for detecting appearance of display panel based on YOLO structure | |
| CN115082726A (en) | Ceramic biscuit product classification method for toilet based on PointNet optimization |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190621 |
|
| RJ01 | Rejection of invention patent application after publication |