EP3762866A1 - Procédé, dispositif, système et programme pour détecter une pièce à usiner et support de stockage - Google Patents

Procédé, dispositif, système et programme pour détecter une pièce à usiner et support de stockage

Info

Publication number
EP3762866A1
EP3762866A1 EP18722199.9A EP18722199A EP3762866A1 EP 3762866 A1 EP3762866 A1 EP 3762866A1 EP 18722199 A EP18722199 A EP 18722199A EP 3762866 A1 EP3762866 A1 EP 3762866A1
Authority
EP
European Patent Office
Prior art keywords
training data
category
threshold value
predetermined
classified
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
EP18722199.9A
Other languages
German (de)
English (en)
Inventor
Shuichi Misumi
Satoru Uchida
Jun Sasaki
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.)
Omron Corp
Original Assignee
Omron Corp
Omron Tateisi Electronics Co
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 Omron Corp, Omron Tateisi Electronics Co filed Critical Omron Corp
Publication of EP3762866A1 publication Critical patent/EP3762866A1/fr
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection

Definitions

  • the present disclosure relates to the field of workpiece detection, and specifically, relates to a method, a device, a system and a program for detecting a workpiece, and a storage medium.
  • a machine learning model can be used to detect the quality of a workpiece.
  • a method, a device, a system and a program for detecting a workpiece, and a storage medium are provided in the present disclosure, so as to solve the problem in the related art that the accuracy of a determination result cannot be effectively improved when a machine learning model is used to detect the quality of a workpiece.
  • a method for detecting a workpiece comprises: acquiring original training data related to the workpiece, the original training data comprising training data which is pre-classified into a first category and training data which is pre-classified into a second category; acquiring a first training data satisfying a predetermined condition from the original training data; and processing part or all of the content of the acquired first training data, to generate new training data which is classified into the first category or a second category; using the original training data and the new training data to train a predetermined machine learning model so as to obtained a machine learning model which has been trained; and using the machine learning model which has been trained to detect the workpiece.
  • the total amount of the training data is increased, and the predetermined machine learning model is trained by using the original training data and the new training data, effectively improving the accuracy of the determination result of the machine learning model, thereby improving the accuracy of detecting the quality of a workpiece.
  • acquiring the first training data satisfying the predetermined condition from the original training data comprises: determining the possibility of the training data in the first category being classified into the second category; and selecting the training data of which the possibility exceeds a predetermined threshold value as the first training data.
  • acquiring the first training data satisfying the predetermined condition from the original training data comprises: determining the possibility of the training data in the second category being classified into the first category; and selecting the training data of which the possibility exceeds a predetermined threshold value as the first training data.
  • the possibility of the training data in the first category being classified into the second category exceeding the predetermined threshold value or the possibility of the training data in the second category being classified into the first category exceeding the predetermined threshold value indicates that the possibility of the training data locating in a border region between the first category and the second category is large, and selecting such training data which is in the border region as the first training data to be processed can effectively increase the total amount of the training data, thereby improving the accuracy of the determination result of the machine learning model.
  • each training data is classified into a category in the plurality of categories according to the relationship between its own eigenvalue and at least one threshold value or a threshold value range; and the at least one threshold value or the threshold value range is the basis for distinguishing the borders of different categories, wherein, determining the possibility of the training data in the first category being classified into the second category comprises: determining, according to the relationship between the eigenvalue and the at least one threshold value or the threshold value range, the possibility of the training data in the first category being classified into the second category.
  • each training data is classified into a category in the plurality of categories according to the relationship between its own eigenvalue and at least one threshold value or a threshold value range; and the at least one threshold value or the threshold value range is the basis for distinguishing the borders of different categories, wherein, determining the possibility of the training data in the second category being classified into the first category comprises: determining, according to the relationship between the eigenvalue and the at least one threshold value or the threshold value range, the possibility of the training data in the second category being classified into the first category.
  • the relationship between the eigenvalue of the training data and the threshold value or the threshold value range it can be easily to determine whether the training data is located within the border region between the first category and the second category.
  • acquiring the first training data satisfying the predetermined condition from the original training data comprises: extracting predetermined training data, of which the eigenvalue is in a predetermined range, from the training data in the first category as the first training data, wherein, the predetermined training data is the training data for which at least N training data among M training data of which the eigenvalues are in the same predetermined range as the eigenvalue of the predetermined training data belongs to the second category, wherein, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 1.
  • acquiring the first training data satisfying the predetermined condition from the original training data comprises: extracting predetermined training data, of which the eigenvalue is in a predetermined range, from the training data in the second category as the first training data, wherein, the predetermined training data is the training data for which at least N training data among M training data of which the eigenvalues are in the same predetermined range as the eigenvalue of the predetermined training data belongs to the first category, wherein, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 1.
  • the described method can effectively extract the training data within the border region.
  • processing part or all of the content of the acquired first training data comprises: extracting training data from the training data belonging to the first category according to a predetermined rule, or calculating the average value or median of all the training data belonging to the first category, as reference training data; and processing, on the basis of the reference training data, part or all of the content of the acquired first training data.
  • processing part or all of the content of the acquired first training data comprises: extracting training data from the training data belonging to the second category according to a predetermined rule, or calculating the average value or median of all the training data belonging to the second category, as reference training data; and processing, on the basis of the reference training data, part or all of the content of the acquired first training data.
  • Taking the training data of the first category or the training data of the second category as reference data to process the first training data can make the first training data after being processed definitely belongs to the first category or the second category.
  • the original training data is image data
  • processing part or all of the content of the acquired first training data comprises: modifying, according to the image data belonging to the first category or the image data belonging to the second category, at least one characteristic region of the image data which is the acquired first training data.
  • At least one characteristic region of the first training data is modified according to the image data belonging to the first category or belonging to the second category, such that the processing of the first training data is more intuitive.
  • a device for detecting a workpiece comprises: an acquiring unit, acquiring original training data related to the workpiece, and acquiring a first training data satisfying a predetermined condition from the original training data, wherein, the original training data comprising training data which is pre-classified into a first category and training data which is pre-classified into a second category; and a processing unit, processing part or all of the content of the acquired first training data, to generate new training data which is classified into the first category or a second category; a training unit, using the original training data and the new training data to train a predetermined machine learning model so as to obtain a machine learning model which has been trained; and a detecting unit, using the machine learning model which has been trained to detect the workpiece.
  • the acquiring unit comprises: a determining unit, determining the possibility of the training data in the first category being classified into the second category; and a selecting unit, selecting the training data of which the possibility exceeds a predetermined threshold value as the first training data.
  • the acquiring unit comprises: a determining unit, determining the possibility of the training data in the second category being classified into the first category; and a selecting unit, selecting the training data of which the possibility exceeds a predetermined threshold value as the first training data.
  • each training data is classified into a category in the plurality of categories according to the relationship between its own eigenvalue and at least one threshold value or a threshold value range; and the at least one threshold value or the threshold value range is the basis for distinguishing the borders of different categories, wherein, the determining unit determines, according to the relationship between the eigenvalue and the at least one threshold value or the threshold value range, the possibility of the training data in the first category being classified into the second category.
  • each training data is classified into a category in the plurality of categories according to the relationship between its own eigenvalue and at least one threshold value or a threshold value range; and the at least one threshold value or the threshold value range is the basis for distinguishing the borders of different categories, wherein, the determining unit determines, according to the relationship between the eigenvalue and the at least one threshold value or the threshold value range, the possibility of the training data in the second category being classified into the first category.
  • the acquiring unit extracts predetermined training data, of which the eigenvalue is in a predetermined range, from the training data in the first category as the first training data, wherein, the predetermined training data is the training data for which at least N training data among M training data of which the eigenvalues are in the same predetermined range as the eigenvalue of the predetermined training data belongs to the second category, wherein, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 1.
  • the acquiring unit extracts predetermined training data, of which the eigenvalue is in a predetermined range, from the training data in the second category as the first training data, wherein, the predetermined training data is the training data for which at least N training data among M training data of which the eigenvalues are in the same predetermined range as the eigenvalue of the predetermined training data belongs to the first category, wherein, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 1.
  • the processing unit extracts training data from the training data belonging to the first category according to a predetermined rule, or calculates the average value or median of all the training data belonging to the first category, as reference training data; and processes, on the basis of the reference training data, part or all of the content of the acquired first training data.
  • the processing unit extracts training data from the training data belonging to the second category according to a predetermined rule, or calculates the average value or median of all the training data belonging to the second category, as reference training data; and processes, on the basis of the reference training data, part or all of the content of the acquired first training data.
  • the original training data is image data
  • the processing unit modifies, according to the image data belonging to the first category or the image data belonging to the second category, at least one characteristic region of the image data which is the acquired first training data.
  • a method for generating training data comprising: acquiring a first training data satisfying a predetermined condition; and processing part or all of the content of the acquired first training data, to generate new training data which is classified into a first category or a second category.
  • the total amount of the training data is increased, and then such training data is used to train the machine learning model, such that the accuracy of the determination result of the machine learning model can be improved.
  • a device for generating training data comprises: an acquiring unit, acquiring a first training data satisfying a predetermined condition; and a processing unit, processing part or all of the content of the acquired first training data, to generate new training data which is classified into the first category or a second category.
  • a system for detecting a workpiece comprising: a processing unit, executing the method according to the first aspect of the present disclosure; and an output unit, outputting a detection result of the workpiece.
  • a program for detecting a workpiece is provided, the program, when being executed, performing the method according to the first aspect of the present disclosure.
  • a storage medium which is stored thereon with a program, the program, when being executed, performing the method according to the first aspect of the present disclosure.
  • the accuracy of the determination result of the machine learning model which has been trained is effectively improved, thereby improving the accuracy of detecting the quality of a workpiece.
  • Figure 1 is a mode diagram showing the hardware structure of an information processing system according to an embodiment of the present disclosure.
  • Figure 2 shows a distribution diagram of training data in a case of determining whether a product is qualified according to an image
  • Figure 3 shows a flow chart of detecting a workpiece according to the present disclosure
  • Figure 4 shows an example of processing the training data within a border region according to the present disclosure.
  • Figure 5 is a block diagram of a device for detecting a workpiece according to the present disclosure.
  • the present disclosure increases the total amount of the training data, by means of generating new training data from the original training data, effectively improving the accuracy of a determination result of the machine learning model, thereby improving the accuracy of detecting the quality of a workpiece.
  • original training data related to a workpiece is acquired, and then the original training data is inputted into a machine learning model which has been trained, to obtain the values indicating the possibilities of the training data belonging to a first category and a second category, taking the training data for which the difference between the value of the first category and the value of the second category is below a predetermined value as the training data to be processed.
  • the eigenvalue or characteristic region of the training data to be processed is modified using the training data definitely belonging to the first category or the training data definitely belonging to the second category, to form new training data.
  • a machine learning model is trained using the original training data and the new training data, so as to be used to detect a workpiece, thereby improving the accuracy of the machine learning model detecting the quality of a workpiece.
  • FIG. 1 is a mode diagram showing the hardware structure of an information processing system according to an embodiment of the present disclosure.
  • an information processing system 100 can be achieved by a general purpose computer of general purpose computer architecture.
  • the information processing system 100 can include a processor 110, a main memory 112, a memory 114, an input interface 116, a display interface 118 and a communication interface 120. These parts, for example, can communicate with each other by an internal bus 122.
  • the processor 110 develops a program stored in the memory 114 on the main memory 112 so as to be executed, thereby achieving the functions and processing described later.
  • the main memory 112 can be configured to be a volatile memory, acting as a work memory required by the processor 110 to execute the program.
  • the input interface 116 can be connected to input units, for example, a mouse and a keyboard, for receiving the instructions inputted by operating the input units by an operator.
  • the display interface 118 can be connected to a display, and a variety of processing results generated by the processor 110 executing the program can be outputted to the display.
  • the communication interface 120 is used for communicating with a PLC and a database device and so on by a network 200.
  • the memory 114 can be stored with a program which can allow a computer to function as the information processing system 100, for example, an information processing program, an OS (operating system), and so on.
  • a program which can allow a computer to function as the information processing system 100, for example, an information processing program, an OS (operating system), and so on.
  • the information processing program stored in the memory 114 can be installed into the information processing system 100 by an optical recording medium, for example, a Digital Versatile Disc ( DVD ), or a semiconductor recording medium, for example, a Universal Serial Bus ( USB ) memory. Or the information processing program can also be downloaded from a server device on the network and so on.
  • an optical recording medium for example, a Digital Versatile Disc ( DVD ), or a semiconductor recording medium, for example, a Universal Serial Bus ( USB ) memory.
  • the information processing program can also be downloaded from a server device on the network and so on.
  • the information processing program according to the present embodiment can also be provided in the manner of combining with other programs.
  • the information processing program itself does not include the modules included by the other programs of such combination described above, but performs processing in cooperation with the other programs.
  • the information processing program according to the present embodiment can also be in the form of combining with the other programs.
  • Figure 1 shows an example of using a general purpose computer to achieve the information processing system 100, but the present disclosure is not limited thereto, and all or part of functions thereof can be achieved by dedicated circuits, for example, an Application Specific Integrated Circuit ( ASIC ) or a Field-Programmable Gate Array ( FPGA ) and so on.
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • part of the processing of the information processing system 100 can also be performed by an external device connected by the network.
  • detecting whether a product is a qualified product is taken as an example to describe the basic principles of generating new training data in the process of using the machine learning model to detect the quality of a workpiece
  • the training data used for training a machine learning model is the training data used for training a machine learning model which can determine whether a product is qualified according to a product image
  • the number of the classification category of the training data is two.
  • the training data can be training data used for any other purpose, the number of the classification category of the training data may be three or more.
  • training data is classified into two categories, in which, one category represents qualified products, which are represented by circles ( O :positive), and the other category represents unqualified products, which are represented by cross marks (c negative), as shown in figure 2.
  • Line L1 can represent a threshold value in relevant to an eigenvalue included in the training data.
  • the training data including the eigenvalue represents an unqualified product
  • the training data including the eigenvalue represents a qualified product.
  • the training data within the described border region is acquired, and then, after the characteristic thereof is modified, classified into a first category or a second category, so as to acquire new training data, and meanwhile, the training data within the border region is retained.
  • the amount of the training data is increased, and on the other hand, the new training data is acquired from the border region of the original training data, such that the accuracy of a determination result of the machine learning model can be effectively improved, thereby improving the accuracy of detecting the quality of a product.
  • Figure 3 shows a flow chart of detecting a workpiece according to the present disclosure, the workpiece here is a specific product, e.g. a coil and so on, and the specific details will be described in conjunction with figure 3 in the following.
  • step S301 original training data related to the workpiece is acquired.
  • the original training data here can be any type of training data, including but not being limited to image data, vector data and so on.
  • the number of the classification category of the original training data is two or more. For simplicity, it is assumed herein that the number of the classification category of the acquired original training data is two.
  • the classification category of the original training data can be“qualified product” and“unqualified product”. But the present disclosure is not limited thereto, the original training data is not limited to the training data for determining whether a product is qualified, and number of the classification category of the original training data can be three or more.
  • the way for acquiring original training data can be varied, and this is known in the art. In order to avoid unnecessarily obscuring the present disclosure, the way for acquiring original training data will be not redundantly described here.
  • step S303 training data satisfying a predetermined condition is acquired from the original training data.
  • each training data it can consist of one or more eigenvalues and a label data.
  • the eigenvalues of each training data can include: the shape of a product (round, oval), the color of a product (red, blue), the coil connectivity of a product and the like.
  • the label data thereof can be set to“1”, representing that the training data is pre-classified into“qualified product”, hereinafter referred to as “first category”.
  • the label data thereof can be set to“0”, representing that the training data is pre-classified into “unqualified product”, hereinafter referred to as“second category”.
  • the relationship between the eigenvalue of the training data and the threshold value or threshold value range corresponding to the corresponding eigenvalue distinguishes “first category” from“second category”.
  • the training data which can be definitely classified into a first category or a second category there are the training data which can be definitely classified into a first category or a second category, and the training data which falls within a border region.
  • the difference between the likelihood of it being classified into the first category and the likelihood of it being classified into the second category is calculated, so as to determine the possibility of the training data in the first category being classified into the second category.
  • the training data is a training data representing an qualified product or a unqualified product
  • the value obtained from the training data through the operation of the machine learning model is [a, b], wherein a represents the likelihood of the training data being classified into qualified products, and b represents the likelihood of the training data being classified into unqualified products.
  • training data A of which the value obtained by operation is [0.7, 0.1 ] it is definitely classified into qualified products, and the possibility of it being classified into unqualified products is relatively small, and in this case, the difference between a and b is 0.6.
  • training data B of which the value is [0.6, 0.3] it is also classified into qualified products, but the possibility of it being classified into unqualified products is larger than that of training data A, and in this case, the difference between a and b is 0.3. That is, the larger the difference between a and b is, the smaller the possibility of the training data in a qualified product category being classified into unqualified products is; and the smaller the difference between a and b is, the larger the possibility of the training data in a qualified product category being classified into unqualified products is.
  • a predetermined threshold value for a border region can be set, and when the described possibility exceeds the predetermined threshold value, it shows that the training data is in the border region between the first category and the second category, the training data like this is extracted as the training data to be processed.
  • a similar method as the method described above can be used to determine the possibility of the training data in the second category being classified into the first category.
  • the method for determining the possibility of the training data in the first category being classified into the second category or the method for determining the possibility of the training data in the second category being classified into the first category is not limited to the described method.
  • the plurality of eigenvalues are, for example, used for representing the shape feature, the color feature and the like of a workpiece.
  • Each eigenvalue has at least one threshold value or threshold value range corresponding to it and whether the feature of the product represented by the eigenvalue is quanlified or unqualified is determined according to the threshold value or threshold value range. That is, the at least one threshold value or threshold value range can be taken as the basis for distinguishing the borders of the training data of different categories.
  • each training data is classified into the first category or the second category according to the relationship between its own eigenvalue and at least one threshold value or threshold value range, for example, the qualified product category and the unqualified product category.
  • the threshold value or threshold value range for example, the qualified product category and the unqualified product category.
  • the training data in the qualified product category if the difference between its eigenvalue and the threshold value or threshold value range is relatively small, it shows that the training data is close to the border and the possibility of the training data being classified into unqualified products is relatively large; while for the training data in the qualified product category, if the difference between its eigenvalue and the threshold value or threshold value range is relatively large, it shows that the training data is far away from the border and thus the possibility of the training data being classified into unqualified products is relatively small.
  • the training data in the unqualified product category if the difference between its eigenvalue and the threshold value or threshold value range is relatively small, it shows that the training data is close to the border and the possibility of the training data being classified into qualified products is relatively large; while for the training data in the unqualified product category, if the difference between its eigenvalue and the threshold value or threshold value range is relatively large, it shows that the training data is far away from the border and thus the possibility of the training data being classified into qualified products is relatively small.
  • acquiring the training data belonging to the border region can also be achieved by the following way.
  • border region L2L3 there are both training data representing qualified products (first category) and the training data representing unqualified products (second category). That is, in border region L2L3, there is a plurality of training data representing unqualified products around the training data representing qualified products. With the distance far away from the border region, for example, in the right direction as shown in (c) of figure 1 , the training data representing unqualified products around the training data representing qualified products becomes fewer. Therefore, the number of the training data representing unqualified products around the training data representing qualified products can be used to characterize whether the training data representing qualified products is within the border region.
  • the training data D is considered to be in the border region and can be extracted as the training data to be processed, wherein, M is an integer larger than or equal to 2, N is an integer larger than or equal to 1 , and the predetermined range can be set according to actual needs.
  • the training data in the border region can also be extracted from the training data representing unqualified products so as to take these training data as the training data to be processed.
  • step S305 the acquired training data is processed to generate new training data.
  • the training data can be processed in various ways so as to allow the processed training data to be classified into the first category or the second category, thereby forming new training data.
  • training data which is extracted from the training data in the first category according to a predetermined rule is taken as reference training data
  • the predetermined rule for example, can be the likelihood generated from the training data in the first category through the operation of the machine learning model meet the following requirement, i.e. the difference of the likelihood of the training data belonging to the first category and the likelihood of the training data belonging to the second category is larger than a predetermined value.
  • the predetermined rule for example, can be that the difference between the eigenvalue of the training data in the first category and the corresponding threshold value or threshold value range is larger than a predetermined value.
  • the training data extracted from the training data in the first category, which definitely belongs to the first category is taken as reference training data.
  • Part or all of the content of the training data to be processed which is acquired from the border region is processed according to the reference training data.
  • the eigenvalue of the training data to be processed is modified according to the threshold value or threshold value range corresponding to one or more eigenvalues of the reference training data, such that the processed training data is classified into the first category, thereby forming new training data.
  • training data is extracted from the training data in the second category in a similar way as described above as reference training data.
  • the eigenvalue of the training data to be processed is modified according to the threshold value or threshold value range corresponding to one or more eigenvalues of the reference training data, such that the processed training data is classified into the second category, thereby forming new training data.
  • the average value or median of all the training data belonging to the first category is calculated as reference training data.
  • the same eigenvalues corresponding to all the training data in the first category are averaged or the median thereof is calculated, and the training data of which the corresponding eigenvalue is the average value or median is formed as reference training data.
  • the eigenvalue of the training data to be processed is modified according to the eigenvalue of the reference training data, such that the processed training data is classified into the first category, thereby forming new training data.
  • the average value or median of all the training data belonging to the second category is calculated as reference training data.
  • the same eigenvalues corresponding to all the training data in the second category are averaged or the median thereof is calculated, and the training data of which the corresponding eigenvalue is the average value or median is formed as reference training data.
  • the eigenvalue of the training data to be processed is modified according to the eigenvalue of the reference training data, such that the processed training data is classified into the second category, thereby forming new training data.
  • step S307 the original training data and the new training data is used to train a predetermined machine learning model, so as to obtain a machine learning model which has been trained.
  • the new training data is obtained after the training data acquired from the border region is processed, without replacing the original training data in the border region, namely, the new training data is generated while retaining the original training data.
  • the overall amount of the training data is increased, and as the new training data is generated by processing the training data in the border region between the first category and the second category, the new training data is definitely classified into the first category or the second category, thereby effectively improving the accuracy of a determination result of the machine learning model.
  • Step 309 the machine learning model which has been trained is used to detect the quality of the workpiece.
  • the detection data of the workpiece e.g. image data is inputted into the machine learning model which has been trained, and then whether the product is a qualified product or an unqualified product can be determined according to an output result of the machine learning model.
  • Figure 4 shows an example of processing the training data within a border region according to the present disclosure.
  • the training data is the image data of coils.
  • the examples here are merely exemplary examples, and the present disclosure is not limited thereto. The following description is made in conjunction with figure 4.
  • original image data of a coil is inputted into a machine learning model which has been trained to determine whether the coil is a qualified product, and whether the image data represents a qualified coil or an unqualified can be determined according to an output result of the machine learning model.
  • the output result being obtained is [c, d], wherein, c is the likelihood of the image data representing that the coil is a qualified product, and d is the likelihood of the image data representing that the coil is an unqualified product.
  • the coil represented by the image data is classified into qualified products
  • the coil represented by the image data is classified into unqualified products.
  • the likelihoods generated from the training data through the operation of the machine learning model are shown in the bottom of (a) of figure 4, which are the likelihoods of the training data belonging to“unqualified products” or“qualified products”.
  • the likelihood of the training data belonging to“unqualified products” is 0.31
  • the likelihood of the training data belonging to“qualified products” is 0.14. Namely, the likelihood of the training data belonging to“unqualified products” is higher than the likelihood of the training data belonging to“qualified products”.
  • the coil is disconnected, and the two joints of the disconnected part are staggered in the up and down direction and thud are in different heights.
  • the coil with such a disconnection will be easily to be determined as“unqualified products”.
  • the likelihood of the training data belonging to“qualified products” increases.
  • the likelihood of the training data belonging to “unqualified products” is 0.15
  • the likelihood of belonging to“qualified products” is 0.28.
  • the difference (0.13) between the likelihood of the latter belonging to “unqualified products” and the likelihood of the latter belonging to“qualified products” is less than the difference (0.17) between the likelihood of the former belonging to“unqualified products” and the likelihood of the former belonging to“qualified products”.
  • the training data of which the difference in likelihood is less than or equal to a predetermined threshold can be processed.
  • the predetermined threshold can be set to 0.15
  • the training data in (b) of figure 4 the difference (0.13) between the likelihood of it belonging to “unqualified products” and the likelihood of it belonging to “qualified products” is less than the predetermined threshold 0.15
  • the training can be processed to increase the likelihood of it belonging to “qualified products” or the likelihood of it belonging to “unqualified products”.
  • the difference (0.17) between the likelihood of it belonging to “unqualified products” and the likelihood of it belonging to“qualified products” is greater than the predetermined threshold 0.15, the training will not be processed.
  • the training data in (b) of figure 4 can be extracted and image processing is performed on the extracted training data.
  • the disconnected part is connected together, so as to increase the likelihood of it belonging to “qualified products”, and then the modified training data can be classified into“qualified product” category, thereby obtaining new training data that represents“qualified products”.
  • the training data in (b) of figure 4 can be extracted and image processing is performed on the extracted training data.
  • the two joints of the disconnected part is made to be at different heights, or the training data is made to have several disconnected parts which are the same as shown in (b) of figure 4, improving the likelihood of it belonging to“unqualified products”, and then the modified training data is definitely classified into“unqualified product” category, thereby obtaining new training data that represents“unqualified products”.
  • the disconnected parts in (a) and (b) of figure 4 can be detected by known methods such as edge detecting. Or the disconnected part can be confirmed by the human eyes.
  • Figure 4 shows that the training data (images in (b) of figure 4) in the border region between a first category (qualified products) and a second category (unqualified products) is extracted and the training data is modified so that the disconnected parts of the coils are connected together or the characteristics of the disconnected parts of the coils become more obvious, such that the modified image data is definitely classified into the first category or the second category, thereby forming new image data.
  • Figure 5 is a block diagram of a device for detecting a workpiece according to the present disclosure.
  • the device for detecting comprises an acquiring unit 502 and a processing unit 504, a training unit 506, and a detecting unit 508, wherein, the acquiring unit 502 is configured to acquiring original training data related to the workpiece, and acquiring the training data in a border region from the original training data; and a processing unit 504 is configured to process part or all of the content of the acquired training data in the border region, so as to generate new training data which is classified into a first category or a second category; a training unit 506, using the original training data and the new training data to train a predetermined machine learning unit so as to obtain a machine learning unit which has been trained; and a detecting unit 508, using the machine learning unit which has been trained to detect the workpiece.
  • the acquiring unit 502 further comprises a determination unit and a selection unit (not shown).
  • the determination unit can determine the possibility of the training data in the first category being classified into the second category or the possibility of the training data in the second category being classified into the first category, and the selection unit can select the training data of which the possibility exceeds a predetermined threshold value as the training data in the border region.
  • the new training data is obtained after the training data acquired from the border region is processed and classified, without replacing the existing training data in the border region, namely, the new training data is generated while retaining the existing training data.
  • the overall amount of the training data is increased, and the increased training data is generated by processing and classifying the training data in the border region between the first category and the second category, thereby effectively improving the accuracy of a determination result of the machine learning model, thereby improving the accuracy of detecting the quality of a workpiece.
  • a method for generating training data comprises: acquiring a first training data satisfying a predetermined condition; and processing part or all of the content of the acquired first training data, to generate new training data which is classified into the first category or a second category.
  • a device for generating training data comprises: an acquiring unit, acquiring a first training data satisfying a predetermined condition; and a processing unit, processing part or all of the content of the acquired first training data, to generate new training data which is classified into the first category or a second category.
  • a variety of implementation methods for acquiring the first training data of the border region and for processing the to-be-processed first training data acquired from the border region have been described above in detail. These implementation methods can respectively performed by the acquiring unit and the processing unit. Therefore, for the avoidance of repetition, the acquiring unit acquiring the first training data in the border region and the processing unit processing the to-be-processed training data acquired from the border region will not be described here redundantly.
  • a system for detecting a workpiece comprising a processing unit and an output unit, wherein, the processing unit performs the method for detecting a workpiece described herein, and the output unit outputs a detection result of the workpiece.
  • a program for detecting a workpiece is provided, the program, when being executed, performing the method for detecting a workpiece described herein.
  • a storage medium which is stored thereon with a program, the program, when being executed, performing the method described in embodiment 1.
  • the present disclosure is not limited thereto, and the present disclosure can be applied to the determination of the state or posture of human and so on.
  • image data representing an ambiguous posture such image data can be extracted according the method described in the present disclosure, and the extracted image data can be processed, such that the image data can be classified into a category representing a definite posture.
  • the disclosed technical content may be implemented in other ways.
  • the described device embodiments are merely exemplary.
  • the unit division can be logical function division and may be other division in actual implementation.
  • a plurality of units or components may be merged or integrated into another system, or some features may be ignored or not performed.
  • the displayed or discussed mutual couplings or direct couplings or communication connections may be implemented through some interfaces.
  • the indirect couplings or communication connections between the units or modules may be implemented in electronic or other forms.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention concerne un procédé, un dispositif, un système et un programme pour détecter une pièce à usiner et un support de stockage, le procédé pour détecter une pièce à usiner consistant à : acquérir des données d'apprentissage d'origine concernant la pièce à usiner ; acquérir des premières données d'apprentissage satisfaisant une condition prédéterminée à partir des données d'apprentissage d'origine ; et traiter une partie ou la totalité du contenu des premières données d'apprentissage pour générer de nouvelles données d'apprentissage qui sont classifiées dans la première catégorie ou la seconde catégorie ; utiliser les données d'apprentissage d'origine et les nouvelles données d'apprentissage pour entraîner un modèle d'apprentissage machine prédéterminé, de façon à obtenir un modèle d'apprentissage machine qui a été entraîné, et utiliser le modèle d'apprentissage machine qui a été entraîné pour détecter la pièce à usiner. Les nouvelles données d'apprentissage sont obtenues par traitement des données d'apprentissage d'origine de telle sorte que la précision d'un résultat de détermination de détection de pièce à usiner est efficacement améliorée.
EP18722199.9A 2018-03-05 2018-03-05 Procédé, dispositif, système et programme pour détecter une pièce à usiner et support de stockage Pending EP3762866A1 (fr)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/IB2018/051394 WO2019171122A1 (fr) 2018-03-05 2018-03-05 Procédé, dispositif, système et programme pour détecter une pièce à usiner et support de stockage

Publications (1)

Publication Number Publication Date
EP3762866A1 true EP3762866A1 (fr) 2021-01-13

Family

ID=62111128

Family Applications (1)

Application Number Title Priority Date Filing Date
EP18722199.9A Pending EP3762866A1 (fr) 2018-03-05 2018-03-05 Procédé, dispositif, système et programme pour détecter une pièce à usiner et support de stockage

Country Status (3)

Country Link
EP (1) EP3762866A1 (fr)
JP (1) JP7103421B2 (fr)
WO (1) WO2019171122A1 (fr)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7334801B2 (ja) * 2019-12-25 2023-08-29 日本電気株式会社 学習装置、学習方法および学習プログラム

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140247978A1 (en) * 2013-03-04 2014-09-04 Xerox Corporation Pre-screening training data for classifiers

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6298351B1 (en) * 1997-04-11 2001-10-02 International Business Machines Corporation Modifying an unreliable training set for supervised classification
US7426497B2 (en) * 2004-08-31 2008-09-16 Microsoft Corporation Method and apparatus for analysis and decomposition of classifier data anomalies
JP2011192032A (ja) * 2010-03-15 2011-09-29 Nippon Steel Corp 疵学習装置、疵学習方法、及びコンピュータプログラム
US8903167B2 (en) * 2011-05-12 2014-12-02 Microsoft Corporation Synthesizing training samples for object recognition
JP6118752B2 (ja) * 2014-03-28 2017-04-19 セコム株式会社 学習データ生成装置
US10878286B2 (en) * 2016-02-24 2020-12-29 Nec Corporation Learning device, learning method, and recording medium
WO2018005489A1 (fr) * 2016-06-27 2018-01-04 Purepredictive, Inc. Détection et compensation de qualité de données pour un apprentissage machine

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140247978A1 (en) * 2013-03-04 2014-09-04 Xerox Corporation Pre-screening training data for classifiers

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
BI JINBO ET AL: "Learning classifiers from dual annotation ambiguity via a min-max frame", NEUROCOMPUTING, ELSEVIER, AMSTERDAM, NL, vol. 151, 14 October 2014 (2014-10-14), pages 891 - 904, XP029105339, ISSN: 0925-2312, DOI: 10.1016/J.NEUCOM.2014.10.005 *
See also references of WO2019171122A1 *

Also Published As

Publication number Publication date
JP2021516386A (ja) 2021-07-01
JP7103421B2 (ja) 2022-07-20
WO2019171122A1 (fr) 2019-09-12

Similar Documents

Publication Publication Date Title
US9965719B2 (en) Subcategory-aware convolutional neural networks for object detection
CN112381775B (zh) 一种图像篡改检测方法、终端设备及存储介质
CN109325538B (zh) 目标检测方法、装置和计算机可读存储介质
CN106897746B (zh) 数据分类模型训练方法和装置
US20160026900A1 (en) Image processing device, information storage device, and image processing method
JP2018200685A (ja) 完全教師あり学習用のデータセットの形成
KR20190063839A (ko) 제조 공정에서 딥러닝을 활용한 머신 비전 기반 품질검사 방법 및 시스템
US20150186726A1 (en) Scene Recognition Method and Apparatus
US20180089818A1 (en) Image inspection device, image inspection method, and image inspection program
US11403560B2 (en) Training apparatus, image recognition apparatus, training method, and program
JP4745207B2 (ja) 顔特徴点検出装置及びその方法
US10169673B2 (en) Region-of-interest detection apparatus, region-of-interest detection method, and recording medium
US11741685B2 (en) Commodity identification device, non-transitory computer-readable storage medium, and learning method
CN105117729A (zh) 一种识别翻拍图像的方法和装置
US20170220896A1 (en) Fault-Aware Matched Filter and Optical Flow
JP2015041164A (ja) 画像処理装置、画像処理方法およびプログラム
CN108875474A (zh) 评估人脸识别算法的方法、装置及计算机存储介质
US11023714B2 (en) Suspiciousness degree estimation model generation device
KR20190018274A (ko) 이미지에 포함된 특징 포인트의 시간 또는 공간의 움직임에 기초하여 이미지에 존재하는 피사체를 인식하는 장치 및 방법
WO2019171122A1 (fr) Procédé, dispositif, système et programme pour détecter une pièce à usiner et support de stockage
EP4030343A1 (fr) Procédé et appareil de détection de la peau du visage
Choksuriwong et al. Comparison of invariant descriptors for object recognition
US20240020918A1 (en) Information processing apparatus and information processing method
CN111144438A (zh) 一种广告单中商品检测方法及装置
CN105678321A (zh) 一种基于融合模型的人体姿态估计方法

Legal Events

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

Free format text: STATUS: UNKNOWN

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: 20200907

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 MK MT NL NO PL PT RO RS SE SI SK SM TR

AX Request for extension of the european patent

Extension state: BA ME

DAV Request for validation of the european patent (deleted)
DAX Request for extension of the european patent (deleted)
STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: EXAMINATION IS IN PROGRESS

17Q First examination report despatched

Effective date: 20220623