Loading...
Please wait, while we are loading the content...
Learning Based Defect Classification
| Content Provider | The Lens |
|---|---|
| Abstract | Methods, apparatuses and systems for classifying defects for a defect inspection system are disclosed. The defect inspection system can be used to inspect and manage wafer or reticle defects. The method includes receiving a defect record based on an inspection of a target specimen, the defect record comprising a defect image associated with an unknown defect, selecting, by a computing device using a first processing unit, components ranked by significance from the defect image using a first learning technique, and determining, by the computing device using the first processing unit, whether the defect image is associated with a known defect type based on the components ranked by significance using a second learning technique. |
| Related Links | https://www.lens.org/lens/patent/009-953-833-634-092/frontpage |
| Language | English |
| Publisher Date | 2019-03-05 |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Patent |
| Jurisdiction | United States of America |
| Date Applied | 2016-10-05 |
| Agent | Young Basile Hanlon & Macfarlane, P.c. |
| Applicant | Dongfang Jingyuan Electron Ltd |
| Application No. | 201615285667 |
| Claim | A method of classifying defects for a defect inspection system, comprising: receiving an image based on an inspection of a target specimen, the image including a defect and the image comprising components; selecting, by a computing device using a first processing unit, most significant components of the components, wherein a support vector machine (SVM) learning system determined, during a training process, respective weights for at least some of the components, and the most significant components being the components having the highest determined weights, and a first number of most significant components is smaller than a second number of components; and inputting the most significant components to a neural-network based image classifier to obtain a defect type of the defect. The method of claim 1 , wherein the computing device comprises the first processing unit and a second processing unit, and wherein the first processing unit comprises a central processing unit (CPU), and the second processing unit comprises a graphics processing unit (GPU). The method of claim 2 , wherein the target specimen comprises a wafer, and the each component comprises a pixel of the image and a position of the pixel, and the pixel is associated with at least one of an edge in the image, a corner in the image, a segment in the image, and a region of the image. The method of claim 1 , further comprising: receiving a defect image associated with a known defect type during the training process; selecting, by the computing device, components ranked by significance from the defect image associated with the known defect type using the first learning technique; determining, by the computing device using the second learning technique, whether the defect image is correctly associated with the same known defect type based on the components ranked by significance selected from the defect image associated with the known defect type; determining whether an accuracy requirement is met based on whether the defect image is correctly associated with the same known defect type; and based on a determination that the accuracy requirement is not met, updating parameters associated with the second learning technique. The method of claim 4 , wherein the accuracy requirement comprises a threshold of accuracy rate. The method of claim 4 , further comprising: based on the determination that the accuracy requirement is not met, updating parameters associated with the first learning technique before updating the parameters associated with the second learning technique. The method of claim 1 , wherein a significance value is indicative of at least one of a probability of a defect occurring in the each component or a relevancy of the each component for defect classification. The method of claim 1 , wherein the first number of most significant components is 10. A non-transitory computer-readable medium storing instructions for defect classification of a defect inspection system, which when executed by a computer system using a first processing unit become operational with the first processing unit for classify defects for a defect inspection system, the non-transitory computer-readable medium comprising instructions to: receive an image from the defect inspection system based on an inspection of a target specimen, the image including a defect and the image comprising components; select N components of the components, wherein a first learning system determined, during a training process, respective significance values for at least some of the components, and the N components being the components having the highest determined significance values, wherein a significance value of each component is indicative of at least one of a probability of a defect occurring in the each component and relevancy of the each component for defect classification, and N is an integer smaller than a total number of the components; and determine whether the image comprises a defect of a known defect type based on the N components using a second learning system. The non-transitory computer-readable medium of claim 9 , wherein the target specimen comprises a wafer, and the each component comprises a pixel of the image and a position of the pixel, and the pixel is associated with at least one of an edge in the image, a corner in the image, a segment in the image, and a region of the image. The non-transitory computer-readable medium of claim 9 , wherein the first learning system comprises a support vector machine (SVM) technique, and the second learning system comprises a deep learning (DL) technique. The non-transitory computer-readable medium of claim 11 , wherein the instructions to determine whether the image comprises the defect of the known defect type further comprise instructions which when executed by a CPU and a GPU become operational with the CPU and the GPU to: in accordance with a rank of weighted significance values of the total number of components, select the N components as components having top-N weighted significance values in the rank; input the N components to a neural-network based image classifier, wherein the neural-network based image classifier outputs data indicative of known defect types; and determine whether the image comprises the defect of the known defect type based on the data. The non-transitory computer-readable medium of claim 9 , further comprising instructions to: receive a defect image associated with a known defect type during the training process; select components ranked by significance from the defect image associated with the known defect type using the first learning system; determine, using the second learning system, whether the defect image is correctly associated with the same known defect type based on the components ranked by significance selected from the defect image associated with the known defect type; determine whether an accuracy requirement is met based on whether the defect image is correctly associated with the same known defect type; and based on a determination that the accuracy requirement is not met, update parameters associated with the second learning system. The non-transitory computer-readable medium of claim 13 , further comprising instructions to: based on the determination that the accuracy requirement is not met, update parameters associated with the first learning system before updating the parameters associated with the second learning system. A defect inspection system, comprising: a first processing unit; a second processing unit; and a memory coupled to the first and the second processing units, the memory configured to store a set of instructions which when executed by the first and the second processing units become operational with the first and the second processing units to: receive, using the first processing unit, a defect record based on an inspection of a target specimen, the defect record comprising a defect image associated with an unknown defect; select, using the first processing unit, components ranked by significance from the defect image using a support vector machine (SVM)), wherein the components ranked by significance are determined using both the first processing unit and the second processing unit; and determine, using both the first processing unit and the second processing unit, whether the defect image is associated with a known defect type based on the components ranked by significance using a deep learning system. The system of claim 15 , wherein the first processing unit comprises a central processing unit (CPU), and the second processing unit comprises a graphics processing unit (GPU). The system of claim 15 , wherein the memory is further configured to store a set of instructions which when executed by the first and the second processing units become operational with the first and the second processing units to: receive a defect image associated with a known defect type during a training process; select components ranked by significance from the defect image associated with the known defect type using the first learning technique; determine, using the second learning technique, whether the defect image is correctly associated with the same known defect type based on the components ranked by significance selected from the defect image associated with the known defect type; determine whether an accuracy requirement is met based on whether the defect image is correctly associated with the same known defect type; and based on a determination that the accuracy requirement is not met, update parameters associated with the second learning technique. The system of claim 17 , wherein the memory is further configured to store a set of instructions which when executed by the first and the second processing units become operational with the first and the second processing units to: based on the determination that the accuracy requirement is not met, update parameters associated with the first learning technique before updating the parameters associated with the second learning technique. The apparatus of claim 17 , wherein the accuracy requirement comprises a threshold ratio of times of the CNN correctly classifying the defect type. |
| CPC Classification | IMAGE DATA PROCESSING OR GENERATION; IN GENERAL IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING ELECTRIC DIGITAL DATA PROCESSING |
| Examiner | Phuoc Tran |
| Extended Family | 090-822-339-353-84X 158-241-085-961-318 051-685-108-354-359 010-500-326-852-005 009-953-833-634-092 |
| Patent ID | 10223615 |
| Inventor/Author | Ma Weimin Zhang Jian Zhang Zhaoli |
| IPC | G06T7/00 G06V10/764 |
| Status | Active |
| Owner | Dongfang Jingyuan Electron Limited |
| Simple Family | 051-685-108-354-359 009-953-833-634-092 |
| CPC (with Group) | G06T7/0004 G06T2207/30148 G06V10/764 G06V10/82 G06V2201/06 G06F18/211 G06F18/2411 |
| Issuing Authority | United States Patent and Trademark Office (USPTO) |
| Kind | Patent/New European patent specification (amended specification after opposition procedure) |
