EP4639472A1 - An automated method for glass defect classification and a system thereof - Google Patents

An automated method for glass defect classification and a system thereof

Info

Publication number
EP4639472A1
EP4639472A1 EP23906280.5A EP23906280A EP4639472A1 EP 4639472 A1 EP4639472 A1 EP 4639472A1 EP 23906280 A EP23906280 A EP 23906280A EP 4639472 A1 EP4639472 A1 EP 4639472A1
Authority
EP
European Patent Office
Prior art keywords
defect
module
defects
image
classification
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
Application number
EP23906280.5A
Other languages
German (de)
French (fr)
Inventor
Vishal Ramaswamy CHITTOOR VENKATASUBRAMANIAN
Santhosh Kumar CHIDAMBARAM TRICHY
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Saint Gobain Glass France SAS
Compagnie de Saint Gobain SA
Original Assignee
Saint Gobain Glass France SAS
Compagnie de Saint Gobain SA
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Saint Gobain Glass France SAS, Compagnie de Saint Gobain SA filed Critical Saint Gobain Glass France SAS
Publication of EP4639472A1 publication Critical patent/EP4639472A1/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/38—Concrete; Lime; Mortar; Gypsum; Bricks; Ceramics; Glass
    • G01N33/386—Glass
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Definitions

  • the present disclosure relates to the field of defects classification, particularly, this disclosure relates to classification of glass defects, and more particularly to a method and system for artificial intelligence (Al) based defects classification.
  • Al artificial intelligence
  • Glass manufacturing involves floating molten glass on a bed of molten metal to produce glass sheets with a very flat surface and uniform thickness. It involves besides various raw materials and different specific process parameters (such as and not limited to temperature and pressure). The variations which may occur in the Furnace, raw materials, repair work or other related parameters can lead to formation of defects in glass. In a glass manufacturing line, if formation of defects in glass is not an isolated case, but a recurring event, it may be attributed to one of said parameters such as raw material, process parameters etc. Usually such defects in glass manufacturing process is tracked by automated units or through vertical inspection of glass. However, the defect type needs to be identified for taking corrective measures in the manufacturing line.
  • WO2014083371A1 discloses a method and system for identifying defects in glass.
  • the method includes identifying each of a plurality of sheets of glass with an identifier. It further includes generating a map of glass attributes for each of the plurality of sheets of glass and associating the map of attributes of each of the plurality of sheets of glass with the identifier of a corresponding each of the plurality of sheets of glass. Further, the map of glass attributes of each of the sheets of glass is stored in a database. A level of access to information in the database is provided to customer to allow the customer to retrieve at least part of the map of attributes of sheets of glass acquired by the customer. However, with this solution also nothing could be obtained for rectifying manufacturing line related parameters so as to prevent the formation of defects.
  • WO2021/124775 discloses a method of detecting a defect in a glass by a defect classification device based on the image data. It is able to identify the defect against class of defect mapping and provides a yes or no based response against each class. Again, such information would be insufficient to provide further actions in the manufacturing line.
  • An object of the present invention is to provide a solution for overcoming the drawbacks of the prior art.
  • Another object of the present invention is to provide a robust solution for classification of glass defects.
  • Yet another object of the present invention is to provide a solution for classification of glass defects for quicker rectification actions in a manufacturing line.
  • a method for classification of defects comprises generating, by a defect classifier, an artificial intelligence (Al) based trained module, wherein said trained Al module is trained by a set of pre-processed defects data, and obtaining, by the defect classifier, an input defect sample for classification into a defect category.
  • Said classification of defects is based on origin.
  • the method further comprises predicting, by the defect classifier, using said trained Al module one or more distinct classes of defect of the input defect sample with at least likelihood, origin details and linked actions for said categories.
  • the step of generating the Al based trained module comprises obtaining, by a preprocessing module, an initial defect data set in a standard format.
  • the method includes pre-processing, by the pre-processing module, the data for effective identification of classes and adding, by the pre-processing module, a classified input defect sample into a raw data set for training the Al module.
  • the data pre-processing of a raw image data further comprises eliminating, by the pre-processing module, one or more irrelevant features in an image, identifying, by the preprocessing module, signature features of a defect in the image, identifying, by the pre-processing module, the localised defect region in the image and converting, by the pre-processing module, said localised defect region into a standardized Al suitable form.
  • the pre-processing includes segregation of background, foreground, identifying lines, edges, extraction of signature features of the defect, standardization of image to provide a precise localised region of the defect in the image to the Al module.
  • the method further includes cross-training, by the defect classifier, the Al module with two or more unbiased defects data sets and determining, by the defect classifier, one class of defect by means of elimination, wherein said elimination is based on chemical analysis, and/or position of the defects.
  • a system for identification and classification of defects in glass comprises a defect inspector configured to inspect and detect a defect in a glass from a manufacturing unit, an image obtaining unit configured to obtain a 3 -Dimensional high-resolution image of the defect and a defect classifier configured to obtain said image of the defect and classify the defect into one or more probable classes.
  • the defect classifier is configured to store a repository of pre-processed data set of defects for training an Al module.
  • the system comprises a processing unit configured to train the Al module with a set of pre-processed defects data and generate a trained Al module.
  • the processing unit is coupled with trained Al module are configured to classify the defect into one or more probable classes, wherein the defects are classified based on the origin of the defect.
  • the system comprises a display unit configured to display the probable classes of the defect along with details on origin and linked actions and an interface configured to obtain an image for defect classification from an input unit.
  • an artificial intelligence-based defect classifier configured to obtain a sample glass defect image and classify it into one or more probable defect classes, wherein the defects are classified based on the origin of the defect and wherein the artificial intelligence-based defect classifier is trained with a library of pre-processed data set of defects.
  • the present invention provides an improved and effective means for classifying defects present in glass through artificial intelligence by uniquely standardised set of defects that have specific origin. It minimizes the difference in operator level (human error), minimizes requirement of expertise availability (at different sites) and reduces response time to classify the defect.
  • This classification is specific to the defect origin and hence linked to the corrective actions in the manufacturing line thereby enabling control of defects.
  • the classification is made based on subfamily of the defect and thereby providing clearer direction towards the corrective measure.
  • the Al model is cross trained using different rather unbiased data sets. This advantageously improves the training of the Al by subjecting it to varied information. It has been observed that with such cross-training means, the prediction accuracy is seen to have improved.
  • FIG 1 illustrates an overview of the solution according to an embodiment of the present invention.
  • FIG 2 illustrates a method for classification of defects according to an embodiment of the present invention.
  • FIG 3 illustrates a method of generating an Al based trained module according to an embodiment of the present invention.
  • FIG 4 illustrates a method of pre-processing of a raw image data according to an embodiment of the present invention.
  • FIG 5 illustrates an overview of the system according to an embodiment of the present invention.
  • FIG 6 illustrates an example of the defect image obtained according to an embodiment of the present invention.
  • the one or more embodiments of the present invention are directed at a solution to classify defects present in glass.
  • defect classification is done by means of a module that uses artificial intelligence.
  • the module includes uniquely standardized set of defects that have specific origin. Since this classification is specific to the defect origin, it is linked to the corrective actions in the manufacturing line, thereby enabling efficient and effective control of defects.
  • the disclosure provides a method for classification of defects in glass, a glass defects classifier, and a system for glass defect inspection and classification thereof.
  • the disclosed solution can minimize the difference in operator level, it minimizes requirement of expertise availability at the various manufacturing sites and reduces response time to classify the defect.
  • FIG. 1 provides a snapshot of the solution.
  • the method comprises generating (SI 03) an artificial intelligence (Al) based trained module, and obtaining (SI 04), an input defect sample for classification into a defect category. Said classification of defects is based on origin.
  • the method further comprises predicting (SI 05) one or more classes or categories of defects for the input defect sample with at least a likelihood or confidence measure, origin details and linked actions to be executed in the manufacturing line.
  • the Al module is trained by a set of pre-processed defects data.
  • the Al module may advantageously be cross-trained (S1015).
  • the defect classifier may cross-train (SI 015) the Al module using two or more unbiased defects data sets.
  • the disclosed classification method includes creating (SI 01) a repository of classified defects data and pre-processing the data.
  • the disclosed method is performed by a defect classifier.
  • the automated defect identification method minimizes analysis time and enables faster management decision.
  • the cross-training of the Al module may include having different data sets to train and test. For instance, data set A may be used to train the Al module and data set B may be used to test said Al module. Additionally, data set B may further be used to train the Al module and test it using data set A.
  • the cross-training is particularly advantageously for the Al module to have a holistic view of the training data. This consequently improves the accuracy of the classification of defects.
  • an image capturing and classification method In an embodiment of the present invention is disclosed an image capturing and classification method.
  • the image capturing is based on a suitable protocol.
  • the defect observation is carried out in multiple steps.
  • a clear 3 -Dimensional image of the defect is captured. It is preferred to capture visuals of the defects and its details in high resolution. The depths, the height, and the length of the features of a defect is to be captured in high resolution.
  • captured images of the defects are to be obtained in a standard format.
  • the images of the defects may be captured by means of and not limited to digital microscope, gemological microscope or stereo microscope. It would be appreciated by a person skilled that the images may be obtained from any apparatus or device that can provide the details of the defects without compromising the resolution of the image. It is desirous to obtain the images in a computer readable format as well.
  • the dynamic creation of a collection of defect images such as a defects’ library.
  • the defect images may be collected from different data sources along with the respective detailed analysis reports.
  • Said defects’ library may be stored in a database in a server coupled with a processing unit.
  • the defects are identified and labelled into different classes. Each time when a new defect is identified it is analysed, observed, labelled and stored in the library. The defects are classified based on the origin of the defect in the manufacturing line. Typically, once a defect is inspected and identified in a glass, it is analysed to observe the significant feature of the defect. Based on such observations, the significant feature is mapped to an origin of the defect.
  • the library is dynamic in nature.
  • each time when a new defect is identified it is inspected, analyzed, labelled, and added to the library.
  • the data in defects’ library or repository is thoroughly analyzed and labelled for improved classification by the Al module. For instance, when a type of solid defect is identified, it is further analysed to identify the next level detail of the defect, which may be referred as a family level identification (say for instance silica stone). Still further analysis is conducted to identify the next level feature i.e., a sub-family level detail of silica stone defects. Defects identification and labelling are done correlating it to the origin. Such classification and labelling of defects at a sub- family level reveals information about defect origin and consequently helps in quick corrective action on the manufacturing line.
  • the visual inspection of the glass samples produced at a manufacturing site is done for quality check and identification of defects.
  • quality test to identify defects may be performed using automation.
  • inspection and identification of defects reveals only broad classification of defects.
  • Manufacturing based defects may include formation of seeds, rubs, inclusions and coating voids and pinholes.
  • Based on the glass defects it may be broadly classified as solid, vitreous, gaseous or surface.
  • classification of glass defect data or images do not provide precise details on origin of such defects. As a result, the defect rectification actions in the manufacturing line gets delayed as well.
  • the defects in glass are advantageously classified into non-overlapping separate origin-based classes. This classification is strategically performed based on sub-family instead of family-based classification.
  • This repository of classified defect data is used for training the Al based module for classification of a sample defect image.
  • the repository of defect data includes 3D high quality and high-resolution images of the defects in a standard format. Each class may be mapped with respective details and corrective measures to be adopted in the manufacturing line.
  • an artificial intelligence (Al) module for classification of a defect.
  • the library of labelled defects maybe used for training said Al module.
  • the defect classes are separated into origin-based classification i.e., sub-family-based classification.
  • This sub-family-based classification of the defect advantageously provide a clearer direction towards the corrective measure.
  • the sub-family classification or labelling are based on the more precise details on the kind of defect and are linked to the origin as well.
  • the labelling is so made that a defect is labelled or classified into nonoverlapping classes and collected in said library.
  • the library of defect images is filtered to retain only the high-resolution 3D images of the defects in a computer readable format and other nonstandard defect images are removed. Further, duplicate images are also removed. Filtering of the 3D images in the repository may be done by visual inspection or by means of other automated filtering modules.
  • the pre-processing of the data set in the defects’ image library In an embodiment of the present invention is disclosed the pre-processing of the data set in the defects’ image library.
  • the pre-processing of the training data set predominantly is performed for elimination of irrelevant features from each of the images of the library. Using the raw image from the library may provide undesirable unreliable results.
  • the pre-processing module is configured to improve accuracy of the image classification from 40% to 80%.
  • the pre-processing module is configured to detect features in the image and ignore the noises.
  • the pre-processing steps may include segregation of background, foreground of the image. It further comprises identifying lines and edges.
  • the pre-processing module is further configured to extract of signature features of the defect, standardize the image, and convert it into Al module suitable form.
  • the Al module suitable form is pre-processed image in a format suitable for improved classification of the defects. It may include removing irrelevant features and converting the image into a single channel.
  • the pre-processed data set is used for training an Al based neural network to obtain a trained model.
  • the trained model is obtained from a second processing unit.
  • the trained module obtained from the Al module is deployed for image classification prediction in a manufacturing line or such applicable sites.
  • the trained model is used to test images and analyse defect class prediction.
  • the trained model is also tested and validated before deployment.
  • generating (SI 03) the Al based trained module wherein the defect data from the repository of glass defects is pre-processed by a pre-processing module for training the Al module.
  • the step of generating (SI 03) Al based trained module comprises obtaining (SI Oi l) an initial defect data set in a standard format from the repository of data stored in a suitable storage medium. This is followed by sorting (SI 012) said data according to multiple pre-defined defect categories.
  • This data set includes the pre-fed 3D high resolution images of defects stored in a database.
  • the data set is pre-processed (SI 013) for effective identification of classes.
  • the pre-processing module further adds (S1014) the classified input defect sample into a raw data set for training the Al module.
  • FIG. 3 shows a snapshot of the method of generating the Al trained module.
  • the pre-processing module may be a module in a processing unit capable of associating with data library, server other processing units to bring an image data fit for the Al module.
  • the pre-processing of a raw image data as shown in FIG. 4. It comprises eliminating (S201), by the pre-processing module, one or more irrelevant features on an image. Any feature of the image data that do not serve in identifying the signature features of the defect may be regarded as irrelevant features.
  • An already pre-fed repository of features aid in identifying the signature features and eliminating noise.
  • Such a prefed repository may be a suitable memory unit.
  • the pre-processing module is further configured for identifying (S202) signature features of a defect in the image, identifying (S203) the localised defect region in the image and convert (S204) said localised defect region into a standardized Al suitable form.
  • the pre-processing includes segregation of background, foreground, identifying lines, edges, extracting signature features of the defect, standardizing of the image to provide a precise localised region of the defect in the image to the Al module.
  • the pre-processing module is capable of identifying precisely the signature features of the defect in a sample image. It is by means of the identification of said signature feature and removal of irrelevant features from the image that the classification by the Al module is improved.
  • the Al module having the neural network is advantageously able to train and classify the precisely localised region of the defect in an image.
  • the Al module can predict one or more two defect categories for a sample irrespective of the shape, size and nature of the defect.
  • a system for classification of defects in glass comprising an interface configured to obtain an image for defect for classification from an input unit, a second processing unit configured to obtain training data and create a trained model and a first processing unit configured to obtain the input defect and image predict the results.
  • the system further comprises a pre-processing module configured to convert both training data and input data into a format suitable for the Al module for obtaining precise and improved results, i.e., prediction.
  • the system comprises a defect inspector (101) configured to inspect and detect a defect in a glass from a manufacturing site.
  • the defect inspector (101) may be and not limited microscopic devices or tablets or camera units. The inspection and detection of defects may be performed manually or through automated systems having camera units, however, not limited to these.
  • the system further comprises an image obtaining unit (102) configured to obtain a 3 -Dimensional high-resolution image of the defect and a defect classifier (103) configured to take obtain said image of the defect and classify said defect into one or more probable classes.
  • the system provides 2-4 probable classes by means of selection and the one class is determined by means of elimination.
  • the defect classifier is capable of effectively classifying the defect irrespective of its size, colour or texture.
  • the defect classifier (103) is configured to store a repository of pre-processed data set of defects for training an Al module (1031).
  • the system may further comprise a processing unit (1032) configured to train the Al module with a set of pre-processed defects data and generate a trained Al module.
  • a pre-processing module (1033) is plugged in the system operably coupled with processing unit (1032).
  • Said pre-processing module (1033) is configured to convert the 3-Dimensional high- resolution image of the defect to a standardized Al suitable form.
  • the standardized Al suitable form refers to the image of a definite size and in desired colour tone.
  • the processing unit (1032) is coupled with trained Al module and are configured to classify the defect into one or more probable classes, wherein the defects are classified based on the origin of the defect.
  • the system may further include a display unit (1032) such as and not limited to an LED display panel to display the out from the classifier.
  • the output includes the probable classes of the defect along with details on origin and linked actions.
  • the system comprises an interface configured to obtain an image for defect for classification from an input unit.
  • the defects in glass are identified, and the defect region is localised for obtaining a 3D high resolution image of the defect.
  • the identification and localisation of the defects are done at the site of occurrence of the defect such as the manufacturing line during production.
  • These images are classified using the Al based defect classifier as disclosed herein and are parallelly stored to the repository or library of images.
  • the root cause, and corrective actions of it are obtained as well. Based on the corrective action suggested by the defect classifier and the parameters observed during the defect generation period, the yield loss on the manufacturing line is addressed.
  • the disclosed solution thus helps in identifying a root cause of a defect and take the respective corrective action fast.
  • the solution reduces the dependency on human intervention, logistics of material movement, availability of expertise and time involved in the entire process.
  • the library of raw defects data containing defect images are stored in storage medium such as and not limited to server storage medium accessed by different terminal in a network.
  • Said raw defects data is accessed by a first processing unit for pre-processing the raw data.
  • the pre-processing module coupled with said first processing unit is configured to detect objects, segregate and remove noise to improve accuracy of prediction from 40% to 80%.
  • This pre-processed data is used to train an Al module.
  • the Al module may include a convolutional neural network to be trained for defect image classification.
  • Said Al module can be trained using the first processing unit or may be trained using a second processing unit with the first and second processing unit being connected via a network.
  • Any defect sample may be input to the system having the defect classifier for being classified into a probable category or class of defects.
  • the pre-processing module converts the input sample to Al module suitable form, using which the second processing unit triggers the Al module to predict 2-4 probable classes to which the sample defect may belong. These probable classes are displayed in the display unit along with a confidence measure.
  • the first and second processing unit may be the same.
  • Said processing unit may be any of central processing unit, cloud processing, graphic processing unit and like or such system performing the similar functionalities.
  • the defect classifier may be and not limited to a computation system capable of image processing and configured to provide support to artificial intelligence module.
  • the repository of defects is classified and labelled for training the Al module.
  • the defects classes are separated into origin-based nonoverlapping classes for labelling.
  • the classification performed on the basis of a feature that may be mapped to the sub-family rather than family-based classification.
  • the pre-processing of each image includes segregation of background, foreground, identifying lines, and optimizing the image to provide a precise localized region of the defect in the image to the Al module. This advantageously assists in not only identifying the sub-family class of the defect, but also helps in taking respective corrective measures.
  • an artificial intelligence-based defect classifier configured to obtain a sample glass defect image and classify it into at least two probable defect classes, wherein the defects are classified based on the origin of the defect.
  • the artificial intelligence-based defect classifier is trained with a library of pre-processed data set of defects.
  • the system and the method is capable predicting top two classes along with the likelihood of belongingness of a defect to said class.
  • the classifier can mimic a defect expertise approach for predicting the classes.
  • the classifier is configured to apply a combinatorial approach of selection followed by elimination. Further, the classifier is configured to identify if a second choice of the class could be a potential class depending upon the confidence level of the first class chosen.
  • the classifier is further configured to verify the prediction of the classes by applying a check on the confidence level of the probable classes.
  • the classifier is configured to check if the confidence level of the suggested top two classes put together is less than a threshold or not. The classifier makes the prediction of classes if the confidence level is more than the threshold value, else it will display a suitable message on the output device.
  • the prediction approach is configured to identify the features of the defect not only by selection but also by elimination.
  • the Al module is configured to select the probable classes of defects along with the confidence percentage. This may be followed by elimination to determine the defect class. The elimination may be performed by a determination module which is configured to determine the defect class of a particular defect based on at least other information like chemical analysis, position of defect and the like. Such determination of precise classes may also be performed manually after the Al module predicts the probable classes.
  • the disclosed method may be extended in identifying defects of all types of glass (TV Glass, Fiber glass, bottle or container glass) because they involve similar defects generated in glass manufacturing line or float.
  • the solution with modifications could be extended to work for microscopic and phone images of defects as well.
  • the disclosed system may in an alternate embodiment be run on different platforms depending on requirement. For instance, it could be integrated on a mobile device platform or on a cloud computing platform, so as to easily identify defects, their root cause and helping in actions to reduce them.
  • Example 1 Reference is made to FIG. 6 that shows an example of a defect in a glass.
  • the figure shows a clear visual appearance of the defect captured in a standard format.
  • This image data is analysed to identify the type of the defect, its family and further its sub-family. Such analysis reveals the origin of the defect. Based on the analysis the image is labelled. With the image is thus linked the associated origin and actions required on manufacturing line.
  • Such data are collected and labelled to build the defects library. It would be appreciated by the skilled person that the image collection, analysis, and labelling may be manual, or automated or a combination thereof.
  • Such a defects’ repository may be stored in a suitable database.
  • the defects’ library data is pre- processed by a processing unit and used to train the Al module, which is operably coupled with a processing unit.
  • this system for identification and classification of defects in glass may include one or more processing units and terminal device i.e., both input and output devices operably connected to classify defects. Said one or more processing units may belong to the same network or may be servers connected to same or different network.
  • Example 2 Reference is made to an example of a system for identification and classification of defects in glass. The system includes a defect inspector such as a scanning tool or an automated alarm at the glass manufacturing site to identify the defect.
  • An image obtaining unit such as a high-resolution camera or a microscope obtains the high-resolution image of the defect.
  • the defect classifier may be a computer (PCI) with peripheral devices to take the image for defect for classification as input.
  • Said computer (PCI) or any storage device communicatively connected with said computer is configured to store a repository of pre-processed data set of defects for training the Al module. Said repository of datasets are selectively used to train and test the Al module.
  • PCI may be coupled with a server to facilitate the training and subsequent generation of the Al module.
  • PCI is further configured to obtain the trained Al module to predict the defect classes.
  • PCI is configured to convert the 3-Dimensional high-resolution sample image of the defect to a standardized Al suitable form.
  • PCI is further configured to select two probable classes with confidence of the sample image. Additionally, the PCI may be configured to eliminate one class and select a single probable class for the sample image based on factors such as chemical analysis, position of the defects and the like
  • This classification is specific to the defect origin and hence linked to the corrective actions in the manufacturing line thereby enabling control of defects.
  • the classification is made based on sub-family of the defect and thereby providing clearer direction towards the corrective measure.
  • the pre-processing step detect objects, and extracts features. These pre-processed images help in improved classification accuracy and thereby enabling appropriate actions on the manufacturing line.
  • SI 016 determining by means of elimination
  • processing unit 1032 processing unit

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • Chemical & Material Sciences (AREA)
  • Ceramic Engineering (AREA)
  • Quality & Reliability (AREA)
  • Databases & Information Systems (AREA)
  • Food Science & Technology (AREA)
  • Medicinal Chemistry (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
  • Image Analysis (AREA)

Abstract

Disclosed herein is a method for classification of defects and a system thereof Said method comprises generating (S103) an artificial intelligence (AI) based trained module, wherein said AI module is trained by a set of pre-processed defects data, obtaining (S104) an input defect sample for classification into a defect category, wherein said classification of defects is based on origin and predicting (S105) at least two categories of defect for the input defect sample with at least likelihood, origin details and linked actions for said categories. The pre-processing step detect objects, ignores scale bar, masks measurements readings, and extracts features. These pre-processed images help in improved classification accuracy and thereby enabling appropriate actions on the manufacturing line.

Description

TITLE OF THE INVENTION
AN AUTOMATED METHOD FOR GLASS DEFECT CLASSIFICATION AND A SYSTEM THEREOF
TECHNICAL FIELD
The present disclosure relates to the field of defects classification, particularly, this disclosure relates to classification of glass defects, and more particularly to a method and system for artificial intelligence (Al) based defects classification.
BACKGROUND
Background description includes information that may be useful in understanding the present disclosure. It is not an admission that any of the information provided herein is prior art or relevant to the presently claimed disclosure, or that any publication specifically or implicitly referenced is prior art.
Glass manufacturing involves floating molten glass on a bed of molten metal to produce glass sheets with a very flat surface and uniform thickness. It involves besides various raw materials and different specific process parameters (such as and not limited to temperature and pressure). The variations which may occur in the Furnace, raw materials, repair work or other related parameters can lead to formation of defects in glass. In a glass manufacturing line, if formation of defects in glass is not an isolated case, but a recurring event, it may be attributed to one of said parameters such as raw material, process parameters etc. Usually such defects in glass manufacturing process is tracked by automated units or through vertical inspection of glass. However, the defect type needs to be identified for taking corrective measures in the manufacturing line. Usually, once an unknown defect is identified, the glass sample with the defect is sent for manual detailed analysis to find probable cause and accordingly take the corrective measures in the manufacturing line. However, this is a time-consuming solution. Intuitively, automated glass defect classification would be preferred as compared to the manual ones. Reference is made to CN108335283A that discloses a method for automatic classification of glass defects. The method involves an inspection software to divide an image of glass according to the segmentation threshold of setting and thus obtain several division objects. It further includes the inspection software handling each division object through a neural network classification algorithm, and the corresponding defect characteristics of said object is extracted and further classified. Such classification may broadly identify the defect class, however, an expert opinion is required for rectification measures in the manufacturing line.
Another reference is made to WO2014083371A1 discloses a method and system for identifying defects in glass. The method includes identifying each of a plurality of sheets of glass with an identifier. It further includes generating a map of glass attributes for each of the plurality of sheets of glass and associating the map of attributes of each of the plurality of sheets of glass with the identifier of a corresponding each of the plurality of sheets of glass. Further, the map of glass attributes of each of the sheets of glass is stored in a database. A level of access to information in the database is provided to customer to allow the customer to retrieve at least part of the map of attributes of sheets of glass acquired by the customer. However, with this solution also nothing could be obtained for rectifying manufacturing line related parameters so as to prevent the formation of defects.
Yet another reference is made to WO2021/124775 that discloses a method of detecting a defect in a glass by a defect classification device based on the image data. It is able to identify the defect against class of defect mapping and provides a yes or no based response against each class. Again, such information would be insufficient to provide further actions in the manufacturing line.
Therefore, in view of the solutions known hitherto, there is a dire need of a solution for classification of glass defects such that the classification can be linked to one or more corrective actions in a glass manufacturing line. SUMMARY OF THE DISCLOSURE
An object of the present invention is to provide a solution for overcoming the drawbacks of the prior art.
Another object of the present invention is to provide a robust solution for classification of glass defects.
Yet another object of the present invention is to provide a solution for classification of glass defects for quicker rectification actions in a manufacturing line.
In an aspect of the present invention is disclosed a method for classification of defects. Said method comprises generating, by a defect classifier, an artificial intelligence (Al) based trained module, wherein said trained Al module is trained by a set of pre-processed defects data, and obtaining, by the defect classifier, an input defect sample for classification into a defect category. Said classification of defects is based on origin. The method further comprises predicting, by the defect classifier, using said trained Al module one or more distinct classes of defect of the input defect sample with at least likelihood, origin details and linked actions for said categories. The step of generating the Al based trained module comprises obtaining, by a preprocessing module, an initial defect data set in a standard format. It further includes sorting, by the pre-processing module, said data according to multiple pre-defined defect classes, wherein said data set includes pre-fed 3-Dimensional high-resolution images of defects. The method includes pre-processing, by the pre-processing module, the data for effective identification of classes and adding, by the pre-processing module, a classified input defect sample into a raw data set for training the Al module. The data pre-processing of a raw image data further comprises eliminating, by the pre-processing module, one or more irrelevant features in an image, identifying, by the preprocessing module, signature features of a defect in the image, identifying, by the pre-processing module, the localised defect region in the image and converting, by the pre-processing module, said localised defect region into a standardized Al suitable form. The pre-processing includes segregation of background, foreground, identifying lines, edges, extraction of signature features of the defect, standardization of image to provide a precise localised region of the defect in the image to the Al module. The method further includes cross-training, by the defect classifier, the Al module with two or more unbiased defects data sets and determining, by the defect classifier, one class of defect by means of elimination, wherein said elimination is based on chemical analysis, and/or position of the defects.
In another aspect of the present invention is provided a system for identification and classification of defects in glass. The system comprises a defect inspector configured to inspect and detect a defect in a glass from a manufacturing unit, an image obtaining unit configured to obtain a 3 -Dimensional high-resolution image of the defect and a defect classifier configured to obtain said image of the defect and classify the defect into one or more probable classes. The defect classifier is configured to store a repository of pre-processed data set of defects for training an Al module. The system comprises a processing unit configured to train the Al module with a set of pre-processed defects data and generate a trained Al module. It further comprises a pre-processing module configured to convert the 3 -Dimensional high-resolution image of the defect to a standardized Al suitable form. The processing unit is coupled with trained Al module are configured to classify the defect into one or more probable classes, wherein the defects are classified based on the origin of the defect. The system comprises a display unit configured to display the probable classes of the defect along with details on origin and linked actions and an interface configured to obtain an image for defect classification from an input unit.
In yet another aspect of the present invention an artificial intelligence-based defect classifier configured to obtain a sample glass defect image and classify it into one or more probable defect classes, wherein the defects are classified based on the origin of the defect and wherein the artificial intelligence-based defect classifier is trained with a library of pre-processed data set of defects.
The present invention provides an improved and effective means for classifying defects present in glass through artificial intelligence by uniquely standardised set of defects that have specific origin. It minimizes the difference in operator level (human error), minimizes requirement of expertise availability (at different sites) and reduces response time to classify the defect. This classification is specific to the defect origin and hence linked to the corrective actions in the manufacturing line thereby enabling control of defects. The classification is made based on subfamily of the defect and thereby providing clearer direction towards the corrective measure. The Al model is cross trained using different rather unbiased data sets. This advantageously improves the training of the Al by subjecting it to varied information. It has been observed that with such cross-training means, the prediction accuracy is seen to have improved.
These and other objects of the invention are achieved by the following aspects of the invention. The following disclosure presents a simplified summary of the invention in order to provide a basic understanding of some aspects of the invention. This presents some concept of the invention in a simplified form to a more detailed description of the invention presented later. It is a comprehensive summary of the disclosure, and it is not an extensive overview of the present invention. The intend of this summary is to provide a fundamental understanding of some of the aspects of the present invention.
The significant features of the present invention and the advantages of the same will be apparent to a person skilled in the art from the detailed description that follows in conjunction with the annexed drawings.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS
The following briefly describes the accompanying drawings, illustrating the technical solution of the embodiments of the present invention or the prior art, for assisting the understanding of a person skilled in the art to comprehend the invention. It would be apparent that the accompanying drawings in the following description merely show some embodiments of the present invention, and persons skilled in the art can derive other drawings from the accompanying drawings without deviating from the scope of the disclosure.
FIG 1 illustrates an overview of the solution according to an embodiment of the present invention. FIG 2 illustrates a method for classification of defects according to an embodiment of the present invention.
FIG 3 illustrates a method of generating an Al based trained module according to an embodiment of the present invention.
FIG 4 illustrates a method of pre-processing of a raw image data according to an embodiment of the present invention.
FIG 5 illustrates an overview of the system according to an embodiment of the present invention.
FIG 6 illustrates an example of the defect image obtained according to an embodiment of the present invention.
Persons skilled in the art will appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions of some of the elements in the figures may be exaggerated relative to other elements to help to improve understanding of embodiments of the disclosure.
DETAILED DESCRIPTION
The present disclosure is now discussed in more detail referring to the drawings that accompany the present application. It would be appreciated by a skilled person that this description to assist the understanding of the invention, but these are to be regarded as merely exemplary.
The terms and words used in the following description are not limited to the bibliographical meanings and the same are used to enable a clear and consistent understanding of the invention. Accordingly, the terms/phrases are to be read in the context of the disclosure and not in isolation. Additionally, descriptions of well-known functions and constructions are omitted for clarity and conciseness. The one or more embodiments of the present invention are directed at a solution to classify defects present in glass. Such defect classification is done by means of a module that uses artificial intelligence. The module includes uniquely standardized set of defects that have specific origin. Since this classification is specific to the defect origin, it is linked to the corrective actions in the manufacturing line, thereby enabling efficient and effective control of defects. The disclosure provides a method for classification of defects in glass, a glass defects classifier, and a system for glass defect inspection and classification thereof. The disclosed solution can minimize the difference in operator level, it minimizes requirement of expertise availability at the various manufacturing sites and reduces response time to classify the defect. FIG. 1 provides a snapshot of the solution.
In an embodiment of the of present invention is disclosed a method for classification of defects as shown in FIG. 2. The method comprises generating (SI 03) an artificial intelligence (Al) based trained module, and obtaining (SI 04), an input defect sample for classification into a defect category. Said classification of defects is based on origin. The method further comprises predicting (SI 05) one or more classes or categories of defects for the input defect sample with at least a likelihood or confidence measure, origin details and linked actions to be executed in the manufacturing line. The Al module is trained by a set of pre-processed defects data. The Al module may advantageously be cross-trained (S1015). The defect classifier may cross-train (SI 015) the Al module using two or more unbiased defects data sets. For a given input image, such crosstraining enables the Al module to select the correct class of defect with improved accuracy. The disclosed classification method includes creating (SI 01) a repository of classified defects data and pre-processing the data. The disclosed method is performed by a defect classifier. The automated defect identification method minimizes analysis time and enables faster management decision.
In an embodiment of the present invention, the cross-training of the Al module may include having different data sets to train and test. For instance, data set A may be used to train the Al module and data set B may be used to test said Al module. Additionally, data set B may further be used to train the Al module and test it using data set A. The cross-training is particularly advantageously for the Al module to have a holistic view of the training data. This consequently improves the accuracy of the classification of defects.
In an embodiment of the present invention is disclosed an image capturing and classification method. The image capturing is based on a suitable protocol. In glass manufacturing site, once the defects are identified, the defect observation is carried out in multiple steps. A clear 3 -Dimensional image of the defect is captured. It is preferred to capture visuals of the defects and its details in high resolution. The depths, the height, and the length of the features of a defect is to be captured in high resolution. Thus, captured images of the defects are to be obtained in a standard format. The images of the defects may be captured by means of and not limited to digital microscope, gemological microscope or stereo microscope. It would be appreciated by a person skilled that the images may be obtained from any apparatus or device that can provide the details of the defects without compromising the resolution of the image. It is desirous to obtain the images in a computer readable format as well.
In an implementation of the present invention is disclosed the dynamic creation of a collection of defect images such as a defects’ library. The defect images may be collected from different data sources along with the respective detailed analysis reports. Said defects’ library may be stored in a database in a server coupled with a processing unit. The defects are identified and labelled into different classes. Each time when a new defect is identified it is analysed, observed, labelled and stored in the library. The defects are classified based on the origin of the defect in the manufacturing line. Typically, once a defect is inspected and identified in a glass, it is analysed to observe the significant feature of the defect. Based on such observations, the significant feature is mapped to an origin of the defect. The library is dynamic in nature. This follows that each time when a new defect is identified, it is inspected, analyzed, labelled, and added to the library. The data in defects’ library or repository is thoroughly analyzed and labelled for improved classification by the Al module. For instance, when a type of solid defect is identified, it is further analysed to identify the next level detail of the defect, which may be referred as a family level identification (say for instance silica stone). Still further analysis is conducted to identify the next level feature i.e., a sub-family level detail of silica stone defects. Defects identification and labelling are done correlating it to the origin. Such classification and labelling of defects at a sub- family level reveals information about defect origin and consequently helps in quick corrective action on the manufacturing line.
The visual inspection of the glass samples produced at a manufacturing site is done for quality check and identification of defects. Such quality test to identify defects may be performed using automation. However, such inspection and identification of defects reveals only broad classification of defects. Manufacturing based defects may include formation of seeds, rubs, inclusions and coating voids and pinholes. Based on the glass defects, it may be broadly classified as solid, vitreous, gaseous or surface. However, such classification of glass defect data or images do not provide precise details on origin of such defects. As a result, the defect rectification actions in the manufacturing line gets delayed as well. In an implementation of the present invention, the defects in glass are advantageously classified into non-overlapping separate origin-based classes. This classification is strategically performed based on sub-family instead of family-based classification. Consequently, the classes of the defects could thus be linked to origin of the defect, thereby providing ease in taking corrective measures in the manufacturing line. This repository of classified defect data is used for training the Al based module for classification of a sample defect image. The repository of defect data includes 3D high quality and high-resolution images of the defects in a standard format. Each class may be mapped with respective details and corrective measures to be adopted in the manufacturing line.
In an embodiment of the present invention is disclosed an artificial intelligence (Al) module for classification of a defect. The library of labelled defects maybe used for training said Al module. In this implementation, the defect classes are separated into origin-based classification i.e., sub-family-based classification. This sub-family-based classification of the defect advantageously provide a clearer direction towards the corrective measure. The sub-family classification or labelling are based on the more precise details on the kind of defect and are linked to the origin as well. The labelling is so made that a defect is labelled or classified into nonoverlapping classes and collected in said library. The library of defect images is filtered to retain only the high-resolution 3D images of the defects in a computer readable format and other nonstandard defect images are removed. Further, duplicate images are also removed. Filtering of the 3D images in the repository may be done by visual inspection or by means of other automated filtering modules.
In an embodiment of the present invention is disclosed the pre-processing of the data set in the defects’ image library. The pre-processing of the training data set predominantly is performed for elimination of irrelevant features from each of the images of the library. Using the raw image from the library may provide undesirable unreliable results. The pre-processing module is configured to improve accuracy of the image classification from 40% to 80%. The pre-processing module is configured to detect features in the image and ignore the noises. The pre-processing steps may include segregation of background, foreground of the image. It further comprises identifying lines and edges. The pre-processing module is further configured to extract of signature features of the defect, standardize the image, and convert it into Al module suitable form. The Al module suitable form is pre-processed image in a format suitable for improved classification of the defects. It may include removing irrelevant features and converting the image into a single channel. The pre-processed data set is used for training an Al based neural network to obtain a trained model. In an implementation of the present invention, the trained model is obtained from a second processing unit. In an implementation of the present invention, the trained module obtained from the Al module is deployed for image classification prediction in a manufacturing line or such applicable sites. The trained model is used to test images and analyse defect class prediction. The trained model is also tested and validated before deployment.
In an embodiment of the present invention is disclosed generating (SI 03) the Al based trained module, wherein the defect data from the repository of glass defects is pre-processed by a pre-processing module for training the Al module. The step of generating (SI 03) Al based trained module comprises obtaining (SI Oi l) an initial defect data set in a standard format from the repository of data stored in a suitable storage medium. This is followed by sorting (SI 012) said data according to multiple pre-defined defect categories. This data set includes the pre-fed 3D high resolution images of defects stored in a database. The data set is pre-processed (SI 013) for effective identification of classes. The pre-processing module further adds (S1014) the classified input defect sample into a raw data set for training the Al module. FIG. 3 shows a snapshot of the method of generating the Al trained module. In an implementation of the present invention the pre-processing module may be a module in a processing unit capable of associating with data library, server other processing units to bring an image data fit for the Al module.
In an embodiment of the present invention is disclosed the pre-processing of a raw image data as shown in FIG. 4. It comprises eliminating (S201), by the pre-processing module, one or more irrelevant features on an image. Any feature of the image data that do not serve in identifying the signature features of the defect may be regarded as irrelevant features. An already pre-fed repository of features aid in identifying the signature features and eliminating noise. Such a prefed repository may be a suitable memory unit. The pre-processing module is further configured for identifying (S202) signature features of a defect in the image, identifying (S203) the localised defect region in the image and convert (S204) said localised defect region into a standardized Al suitable form. The pre-processing includes segregation of background, foreground, identifying lines, edges, extracting signature features of the defect, standardizing of the image to provide a precise localised region of the defect in the image to the Al module.
In an implementation of the present invention is disclosed removal of irrelevant features from an image by the pre-processing module. It may further include segregation of background, foreground and segmenting to identify and remove noise. The pre-processing module is capable of identifying precisely the signature features of the defect in a sample image. It is by means of the identification of said signature feature and removal of irrelevant features from the image that the classification by the Al module is improved. As a result, the Al module having the neural network is advantageously able to train and classify the precisely localised region of the defect in an image. The Al module can predict one or more two defect categories for a sample irrespective of the shape, size and nature of the defect.
In an embodiment of the present invention is disclosed a system for classification of defects in glass. The system comprises an interface configured to obtain an image for defect for classification from an input unit, a second processing unit configured to obtain training data and create a trained model and a first processing unit configured to obtain the input defect and image predict the results. The system further comprises a pre-processing module configured to convert both training data and input data into a format suitable for the Al module for obtaining precise and improved results, i.e., prediction.
In an embodiment of the present invention is disclosed a system for identification and classification of defects in glass as shown in FIG. 5. The system comprises a defect inspector (101) configured to inspect and detect a defect in a glass from a manufacturing site. The defect inspector (101) may be and not limited microscopic devices or tablets or camera units. The inspection and detection of defects may be performed manually or through automated systems having camera units, however, not limited to these. The system further comprises an image obtaining unit (102) configured to obtain a 3 -Dimensional high-resolution image of the defect and a defect classifier (103) configured to take obtain said image of the defect and classify said defect into one or more probable classes. In a preferred embodiment the system provides 2-4 probable classes by means of selection and the one class is determined by means of elimination. The defect classifier is capable of effectively classifying the defect irrespective of its size, colour or texture. The defect classifier (103) is configured to store a repository of pre-processed data set of defects for training an Al module (1031). The system may further comprise a processing unit (1032) configured to train the Al module with a set of pre-processed defects data and generate a trained Al module. A pre-processing module (1033) is plugged in the system operably coupled with processing unit (1032). Said pre-processing module (1033) is configured to convert the 3-Dimensional high- resolution image of the defect to a standardized Al suitable form. The standardized Al suitable form refers to the image of a definite size and in desired colour tone. Said instances of the Al suitable form are provided as an example and not as limitation. The processing unit (1032) is coupled with trained Al module and are configured to classify the defect into one or more probable classes, wherein the defects are classified based on the origin of the defect. The system may further include a display unit (1032) such as and not limited to an LED display panel to display the out from the classifier. The output includes the probable classes of the defect along with details on origin and linked actions. The system comprises an interface configured to obtain an image for defect for classification from an input unit.
In an implementation of the present invention, the defects in glass are identified, and the defect region is localised for obtaining a 3D high resolution image of the defect. The identification and localisation of the defects are done at the site of occurrence of the defect such as the manufacturing line during production. These images are classified using the Al based defect classifier as disclosed herein and are parallelly stored to the repository or library of images. Once a defect identified from the manufacturing line is classified into a class, the root cause, and corrective actions of it are obtained as well. Based on the corrective action suggested by the defect classifier and the parameters observed during the defect generation period, the yield loss on the manufacturing line is addressed. Overall, the disclosed solution thus helps in identifying a root cause of a defect and take the respective corrective action fast. The solution reduces the dependency on human intervention, logistics of material movement, availability of expertise and time involved in the entire process.
In an implementation of the present invention is disclosed the training of the Al module. The library of raw defects data containing defect images are stored in storage medium such as and not limited to server storage medium accessed by different terminal in a network. Said raw defects data is accessed by a first processing unit for pre-processing the raw data. The pre-processing module coupled with said first processing unit is configured to detect objects, segregate and remove noise to improve accuracy of prediction from 40% to 80%. This pre-processed data is used to train an Al module. The Al module may include a convolutional neural network to be trained for defect image classification. Said Al module can be trained using the first processing unit or may be trained using a second processing unit with the first and second processing unit being connected via a network. Any defect sample may be input to the system having the defect classifier for being classified into a probable category or class of defects. The pre-processing module converts the input sample to Al module suitable form, using which the second processing unit triggers the Al module to predict 2-4 probable classes to which the sample defect may belong. These probable classes are displayed in the display unit along with a confidence measure. In an alternative embodiment, the first and second processing unit may be the same. Said processing unit may be any of central processing unit, cloud processing, graphic processing unit and like or such system performing the similar functionalities. The defect classifier may be and not limited to a computation system capable of image processing and configured to provide support to artificial intelligence module. In an embodiment of the present invention, the repository of defects is classified and labelled for training the Al module. The defects classes are separated into origin-based nonoverlapping classes for labelling. The classification performed on the basis of a feature that may be mapped to the sub-family rather than family-based classification. Further, the pre-processing of each image includes segregation of background, foreground, identifying lines, and optimizing the image to provide a precise localized region of the defect in the image to the Al module. This advantageously assists in not only identifying the sub-family class of the defect, but also helps in taking respective corrective measures.
In an embodiment of the present invention is disclosed an artificial intelligence-based defect classifier configured to obtain a sample glass defect image and classify it into at least two probable defect classes, wherein the defects are classified based on the origin of the defect. The artificial intelligence-based defect classifier is trained with a library of pre-processed data set of defects.
In an implementation of the present invention is disclosed that the system and the method is capable predicting top two classes along with the likelihood of belongingness of a defect to said class. The classifier can mimic a defect expertise approach for predicting the classes. The classifier is configured to apply a combinatorial approach of selection followed by elimination. Further, the classifier is configured to identify if a second choice of the class could be a potential class depending upon the confidence level of the first class chosen. The classifier is further configured to verify the prediction of the classes by applying a check on the confidence level of the probable classes. The classifier is configured to check if the confidence level of the suggested top two classes put together is less than a threshold or not. The classifier makes the prediction of classes if the confidence level is more than the threshold value, else it will display a suitable message on the output device.
In an implementation, the prediction approach is configured to identify the features of the defect not only by selection but also by elimination. In an implementation, the Al module is configured to select the probable classes of defects along with the confidence percentage. This may be followed by elimination to determine the defect class. The elimination may be performed by a determination module which is configured to determine the defect class of a particular defect based on at least other information like chemical analysis, position of defect and the like. Such determination of precise classes may also be performed manually after the Al module predicts the probable classes.
The disclosed method may be extended in identifying defects of all types of glass (TV Glass, Fiber glass, bottle or container glass) because they involve similar defects generated in glass manufacturing line or float. The solution with modifications could be extended to work for microscopic and phone images of defects as well. The disclosed system may in an alternate embodiment be run on different platforms depending on requirement. For instance, it could be integrated on a mobile device platform or on a cloud computing platform, so as to easily identify defects, their root cause and helping in actions to reduce them.
Example 1 : Reference is made to FIG. 6 that shows an example of a defect in a glass. The figure shows a clear visual appearance of the defect captured in a standard format. This image data is analysed to identify the type of the defect, its family and further its sub-family. Such analysis reveals the origin of the defect. Based on the analysis the image is labelled. With the image is thus linked the associated origin and actions required on manufacturing line. Such data are collected and labelled to build the defects library. It would be appreciated by the skilled person that the image collection, analysis, and labelling may be manual, or automated or a combination thereof. Such a defects’ repository may be stored in a suitable database. The defects’ library data is pre- processed by a processing unit and used to train the Al module, which is operably coupled with a processing unit. When an unclassified sample defect captured as an image is standard format is attained, it may be fed to the defect classifier by means of an interface and an input unit. Said sample defect is pre-processed for the Al module to precisely identify and predict at least two probable classes to which the defect may belong along with the likelihood as well. In an implementation, this system for identification and classification of defects in glass may include one or more processing units and terminal device i.e., both input and output devices operably connected to classify defects. Said one or more processing units may belong to the same network or may be servers connected to same or different network. Example 2: Reference is made to an example of a system for identification and classification of defects in glass. The system includes a defect inspector such as a scanning tool or an automated alarm at the glass manufacturing site to identify the defect. An image obtaining unit such as a high-resolution camera or a microscope obtains the high-resolution image of the defect. The defect classifier may be a computer (PCI) with peripheral devices to take the image for defect for classification as input. Said computer (PCI) or any storage device communicatively connected with said computer is configured to store a repository of pre-processed data set of defects for training the Al module. Said repository of datasets are selectively used to train and test the Al module. PCI may be coupled with a server to facilitate the training and subsequent generation of the Al module. PCI is further configured to obtain the trained Al module to predict the defect classes. PCI is configured to convert the 3-Dimensional high-resolution sample image of the defect to a standardized Al suitable form. PCI is further configured to select two probable classes with confidence of the sample image. Additionally, the PCI may be configured to eliminate one class and select a single probable class for the sample image based on factors such as chemical analysis, position of the defects and the like.
Industrial Application: With the solution of the disclosed invention automation of defect identification process is achieved. In manufacturing lines usually, when defect occurrence is noted, the glass samples produced are inspected to manually or tracked by defected defect tracking systems. Said glass samples are further manually analysed to identify the defect and its origin thereof, further to which the corrective actions are taken in the manufacturing line. The entire method after tracking of defects in glass to the corrective actions takes about 10 days. But if such conventional method is replaced with the herein disclosed system, it minimizes analysis time and enables quicker decision on the manufacturing line. Consequently, corrective measures may be taken on the manufacturing line in about 2-3 days, thereby bringing about significant economic impact. With the present invention, at least 20% economic benefit is obtained in the glass manufacturing line.
Some advantages of the present invention are enlisted in the following:
• It provides an improved and effective means for classifying defects present in glass through artificial intelligence by uniquely standardised set of defects that have specific origin. • It minimizes the difference in operator level (human error), minimizes requirement of expertise availability (at different sites) and reduces response time to classify the defect.
• This classification is specific to the defect origin and hence linked to the corrective actions in the manufacturing line thereby enabling control of defects. The classification is made based on sub-family of the defect and thereby providing clearer direction towards the corrective measure.
• The pre-processing step detect objects, and extracts features. These pre-processed images help in improved classification accuracy and thereby enabling appropriate actions on the manufacturing line.
• The solution does not require the size of the defect for classification which makes it more robust. This enables faster decision and reduce yield loss.
• The cross-training of the Al module improves the accuracy of the classification process.
Features that are described and/or illustrated with respect to one embodiment may be used in the same way or in a similar way in one or more other embodiments and/or in combination with or instead of the features of the other embodiments.
List of reference numerals appearing in the accompanying drawings and the corresponding features:
S1O1-S1O5: steps for classification of a defect
S1011-S1014: steps for generating Al trained module
SI 015: cross-training
SI 016: determining by means of elimination
S201-S204: pre-processing steps
101 : defect inspector
102: an image obtaining unit
103: a defect classifier
1031: Al module
1032: processing unit
1033: pre-processing unit

Claims

1. A method for classification of defects, comprising: generating (SI 03), by a defect classifier, an artificial intelligence (Al) based trained module, wherein said trained Al module is trained by a set of pre-processed defects data; obtaining (SI 04), by the defect classifier, an input defect sample for classification into a defect category, wherein said classification of defects is based on origin; and predicting (SI 05), by the defect classifier, using said trained Al module one or more distinct classes of defect of the input defect sample with at least likelihood, origin details and linked actions for said categories.
2. The method as claimed in claim 1, wherein generating (SI 03) the Al based trained module comprises: obtaining (SI 011), by a pre-processing module, an initial defect data set in a standard format; sorting (SI 012), by the pre-processing module, said data according to multiple pre-defined defect classes, wherein said data set includes pre-fed 3 -Dimensional high-resolution images of defects; and pre-processing (SI 013), by the pre-processing module, the data for effective identification of classes.
3. The method as claimed in any one of the preceding claims, comprising adding (S1014), by the pre-processing module, a classified input defect sample into a raw data set for training the Al module.
4. The method as claimed in any one of the preceding claims, wherein the data pre-processing (SI 03) of a raw image data further comprises: eliminating (S201), by the pre-processing module, one or more irrelevant features in an image; identifying (S202), by the pre-processing module, signature features of a defect in the image; identifying (S203), by the pre-processing module, the localised defect region in the image; and converting (S204), by the pre-processing module, said localised defect region into a standardized Al suitable form.
5. The method as claimed in any one of the preceding claims, wherein the pre-processing (SI 013) includes segregation of background, foreground, identifying lines, edges, extraction of signature features of the defect, standardization of image to provide a precise localised region of the defect in the image to the Al module.
6. The method as claimed in any one of the preceding claims, comprising cross-training (SI 015), by the defect classifier, the Al module with two or more unbiased defects data sets.
7. The method as claimed in any one of the preceding claims, comprising determining (SI 016), by the defect classifier, one class of defect by means of elimination, wherein said elimination is based on chemical analysis and/or position of the defect.
8. A system for identification and classification of defects in glass, wherein the system comprises: a defect inspector (101) configured to inspect and detect a defect in a glass from a manufacturing unit; an image obtaining unit (102) configured to obtain a 3 -Dimensional high-resolution image of the defect; and a defect classifier (103) configured to obtain said image of the defect and classify the defect into one or more probable classes.
9. The system as claimed in claim 8, wherein the defect classifier (103) is configured to store a repository of pre-processed data set of defects for training an Al module (1031).
10. The system as claimed in claim 8 or claim 9, comprising a processing unit (1032) configured to train the Al module with a set of pre-processed defects data and generate a trained Al module.
11. The system as claimed in any one of the preceding claims 8-10, comprising a preprocessing module (1033) configured to convert the 3 -Dimensional high-resolution image of the defect to a standardized Al suitable form.
12. The system as claimed in any one of the preceding claims 8-11, wherein the processing unit (1032) coupled with trained Al module are configured to classify the defect into one or more probable classes, wherein the defects are classified based on the origin of the defect.
13. The system as claimed in any one of the preceding claims 8-12, comprising a display unit (1034) configured to display the probable classes of the defect along with details on origin and linked actions.
14. The system as claimed in any one of the preceding claims 8-13, comprising an interface (1035) configured to obtain an image for defect classification from an input unit.
15. An artificial intelligence-based defect classifier configured to obtain a sample glass defect image and classify it into one or more probable defect classes, wherein the defects are classified based on the origin of the defect; and wherein the artificial intelligence-based defect classifier is trained with a library of pre- processed data set of defects.
EP23906280.5A 2022-12-23 2023-10-30 An automated method for glass defect classification and a system thereof Pending EP4639472A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
IN202241074767 2022-12-23
PCT/IN2023/050993 WO2024134664A1 (en) 2022-12-23 2023-10-30 An automated method for glass defect classification and a system thereof

Publications (1)

Publication Number Publication Date
EP4639472A1 true EP4639472A1 (en) 2025-10-29

Family

ID=91587975

Family Applications (1)

Application Number Title Priority Date Filing Date
EP23906280.5A Pending EP4639472A1 (en) 2022-12-23 2023-10-30 An automated method for glass defect classification and a system thereof

Country Status (5)

Country Link
EP (1) EP4639472A1 (en)
CN (1) CN120435725A (en)
CO (1) CO2025009823A2 (en)
MX (1) MX2025007276A (en)
WO (1) WO2024134664A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118817720A (en) * 2024-09-14 2024-10-22 四川省华兴宇电子科技有限公司 A printed circuit board online automatic testing method and system

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110473170B (en) * 2019-07-10 2023-04-07 苏州卓融新能源科技有限公司 Artificial intelligence detection method suitable for judging true and false defects of PCB
CN113658108A (en) * 2021-07-22 2021-11-16 西南财经大学 Glass defect detection method based on deep learning

Also Published As

Publication number Publication date
CO2025009823A2 (en) 2025-07-28
WO2024134664A1 (en) 2024-06-27
CN120435725A (en) 2025-08-05
MX2025007276A (en) 2025-08-01

Similar Documents

Publication Publication Date Title
KR102749767B1 (en) Detecting defects in semiconductor specimens using weak labeling
US11568531B2 (en) Method of deep learning-based examination of a semiconductor specimen and system thereof
EP3791336B1 (en) System and method for detecting defects on imaged items
US20220067523A1 (en) Method of deep learining-based examination of a semiconductor specimen and system thereof
CN114494780B (en) Semi-supervised industrial defect detection method and system based on feature comparison
KR102956982B1 (en) Machine learning-based classification of defects in semiconductor specimens
US20010016061A1 (en) Method for analyzing circuit pattern defects and a system thereof
CN110648310A (en) Weakly Supervised Casting Defect Recognition Method Based on Attention Mechanism
CN120580226B (en) Automatic detection and classification method of intelligent manufacturing defects based on machine vision
CN118038454A (en) Wafer defect root cause analysis method and device based on proprietary large model
JP2022512292A (en) Classification of defects in semiconductor samples
CN118761973A (en) Method and system for identifying silicon wafer surface defects
WO2024134664A1 (en) An automated method for glass defect classification and a system thereof
CN114119562A (en) A method and system for detection of outer surface defects of brake disc based on deep learning
CN116228679A (en) Infrared non-destructive testing method and system for composite material defects based on U-Net neural network
Sun et al. Visual-semantic alignment for automatic structural defect detection and diagnosis of prestressed concrete bridges
CN120641938A (en) Method for detecting and characterizing defects in microscopic photographs of objects using artificial intelligence
CN116843608A (en) Clothing defect detection methods, devices and computer equipment based on machine vision
KR20230068445A (en) Automatic quality inspection device and method of hole image based on artificial intelligence
JP2023035643A (en) Defect inspection device, defect inspection method and prediction model generation method
TWM631676U (en) Equipment for detecting fasteners with artificial intelligence module
CN118961756B (en) A method for detecting sealing failure of biological products based on image recognition
CN118761989B (en) Industrial product quality detection method and system for artificial intelligence
Quantrill et al. Applying Deep Learning for automated visual verification of manual bracket installations
CN119399754A (en) Intelligent identification method of microplastic quantity and type based on machine vision learning

Legal Events

Date Code Title Description
STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE

PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE

17P Request for examination filed

Effective date: 20250723

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC ME MK MT NL NO PL PT RO RS SE SI SK SM TR

DAV Request for validation of the european patent (deleted)
DAX Request for extension of the european patent (deleted)