WO2018169103A1 - Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant - Google Patents

Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant Download PDF

Info

Publication number
WO2018169103A1
WO2018169103A1 PCT/KR2017/002784 KR2017002784W WO2018169103A1 WO 2018169103 A1 WO2018169103 A1 WO 2018169103A1 KR 2017002784 W KR2017002784 W KR 2017002784W WO 2018169103 A1 WO2018169103 A1 WO 2018169103A1
Authority
WO
WIPO (PCT)
Prior art keywords
images
image
learning
data
labeled
Prior art date
Application number
PCT/KR2017/002784
Other languages
English (en)
Korean (ko)
Inventor
김성수
Original Assignee
(주)넥셀
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 (주)넥셀 filed Critical (주)넥셀
Priority to PCT/KR2017/002784 priority Critical patent/WO2018169103A1/fr
Publication of WO2018169103A1 publication Critical patent/WO2018169103A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N99/00—Subject matter not provided for in other groups of this subclass
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion

Definitions

  • the present invention relates to a method and apparatus for generating learning data that can automatically generate labeled learning data required in a machine learning process, and a self learning apparatus and method using the same.
  • Machine learning or machine learning is a field of artificial intelligence, which is a series of processes that train computer-prepared learning data to find an appropriate answer for new input based on the trained knowledge.
  • machine learning or machine learning is a field of artificial intelligence, which is a series of processes that train computer-prepared learning data to find an appropriate answer for new input based on the trained knowledge.
  • training data for training a computer is given both a training input and a training output, the training data is labeled.
  • Object detection is a problem of distinguishing the location and type of a specific object in an image or video.
  • machine learning In order to use machine learning for object detection, a large number of labeled learning data are required.
  • the important thing together with the feature extraction and the learning algorithm is the collection of the labeled learning data, and the more the labeled learning data is provided, the more effective the learning can be performed.
  • an object of the present invention is a self-learning device capable of self-learning using a learning data generation method and apparatus that can automatically generate labeled learning data for machine learning and the labeled learning data generated through the same. And providing a method.
  • Learning data generation method for achieving the above object, receiving a plurality of images having continuity, detecting and tracking the object in the plurality of images, the plurality of images and the detection in each image And storing the recognized object and the recognition rate by tracking, and if there is an image in which the recognition rate of the specific object is greater than or equal to a predetermined threshold value in the plurality of images, using the specific object recognized in the image as the final recognition object. Labeling to generate labeled data.
  • the method may further include detecting a continuous image from an input image and providing the image as the plurality of images, and providing the input image with images collected from a plurality of devices through a network.
  • Self-learning method for achieving the above object, receiving a plurality of images having a continuity, detecting and tracking the object in the plurality of images, the plurality of images and the detection and Storing the recognized object and the recognition rate by tracking; if there is an image in which the recognition rate of a specific object is greater than or equal to a predetermined threshold value in the plurality of images, labeling the plurality of images as a final recognition object by using the specific object recognized in the image Generating and storing labeled data, and machine learning using the labeled data as training data.
  • the apparatus for generating learning data includes an image input unit for receiving a plurality of images having continuity, detecting and tracking objects in the plurality of images, and recognizing objects and recognition rates recognized in each image.
  • An object detection unit for outputting information about the plurality of images, a short-term experience database unit for storing the plurality of images and information about the recognition rate and the object recognized in each image, and the short-term experience database unit to query the plurality of If there is an image in which the recognition rate of the specific object in the image is greater than or equal to a predetermined threshold, the learning data generation unit generates the labeled data by labeling the plurality of images with the specific object recognized in the image as the final recognition object.
  • a self-learning device capable of machine learning using the labeled data generated by using the learning data generation method as training data.
  • labeled learning data for machine learning can be automatically generated.
  • machine learning requiring labeled learning data can be effectively performed, and a self-learning device capable of self-learning using the labeled learning data generated through this can be provided.
  • the image data may be received from various devices distributed and distributed in various places to generate large-scale labeled learning data.
  • FIG. 1 is a block diagram for explaining the configuration of a learning data generating apparatus according to an embodiment of the present invention
  • Figure 2 is a flow chart provided in the description of the learning data collection method according to an embodiment of the present invention.
  • FIG. 4 is a block diagram for explaining a configuration of an apparatus for generating training data according to another embodiment of the present invention.
  • FIG. 5 is a diagram illustrating an example of a self-learning apparatus using the method of generating learning data according to the present invention.
  • a component when referred to as being "connected” or “connected” to another component, a component may be directly connected to or connected to another component, but other components in between It should be understood that the element may exist.
  • Other expressions describing relationships between components, such as “between” or “neighboring”, and the like, such that a component “transmits” a signal to another component, should be interpreted as well.
  • FIG. 1 is a block diagram illustrating a configuration of an apparatus for generating learning data according to an embodiment of the present invention.
  • the apparatus 100 for generating training data may include an image input unit 110, an object detector 120, a short-term experience database unit 150, and a training data generator 160.
  • the object detector 120 may include an object detector 130 and an object tracker 140.
  • Such components may be configured by combining two or more components into one component, or by dividing one or more components into two or more components as necessary when implemented in an actual application.
  • the image input unit 110 may receive image data, and may include a device capable of receiving a camera or other image data.
  • the image input unit 110 may process an image frame such as a still image or a moving image obtained by an image sensor or the like, and output a temporally continuous image including the same object.
  • the object detector 120 may include an object detector 130 and an object tracker 140.
  • the object detector 120 may extract feature points from an image output from the image inputter 110 to determine whether a specific object exists or not. If present, the location and type, and the position of the object in the continuous image can be tracked. As a result, the object detector 120 may output information about the object recognized in each image and the recognition rate of the recognized object.
  • Various methods and algorithms may be used to recognize an object in an image and to calculate recognition rate information of the recognized object.
  • the image output from the image input unit 110, and information about the object and the recognition rate recognized in each image output from the object detector 120 is stored in a database.
  • the training data generation unit 160 inquires the short-term experience database unit 150, and if there is an image having a recognition rate equal to or greater than a preset threshold of a specific object in a plurality of recently stored images, finalize the specific object recognized in the image. Labeled data is generated by labeling a number of recently stored images as recognition objects.
  • FIG. 2 is a flowchart provided to explain an automatic learning data collection method according to an embodiment of the present invention.
  • the object detector 120 executes an object detection and tracking process (S210).
  • the object detection unit 120 detects and tracks the object recognized in each image and the recognition rate for the recognized object in a short-term experience database unit 150 is stored in a database with a plurality of images (S220).
  • the learning data generation unit 160 queries the short-term experience database unit 150 and, when there is an image in which the recognition rate of a specific object is greater than or equal to a preset threshold value in a plurality of images, the specific object recognized in the image as the final recognition object.
  • the recently stored continuity image is labeled to generate labeled data (S240).
  • FIG. 3 is a view referred to for describing the learning data generation method according to the present invention.
  • a portion denoted by reference numeral 300 is a short-term experience with respect to the first image 301, the second image 303, and the third image 305 that are continuous in time including the same object.
  • the object recognized in the first image 301 and the first image and its recognition rate 302 the object recognized in the second image 303 and the second image and its recognition rate 304
  • the object recognized in the third image 305 and the third image and the recognition rate 306 are stored.
  • the first image 301 and the second image 303 may not know clearly which of two objects to be recognized.
  • the portion denoted by reference numeral 320 is determined by the learning data generation unit 160.
  • the first image 301, the second image 303, and the third image 305 are labeled with the final recognition object as an object in which the recognition rate exceeds the threshold value in the three images 305 to generate labeled training data. Indicates the state.
  • the learning data when an object is recognized as a value greater than or equal to a threshold value in one image among consecutive images including the same object, the learning data may be generated by labeling the remaining object as the recognized object. Accordingly, the learning data can be generated based on the image taken from another angle even for an image that is clear but has not been recognized because it was not previously learned.
  • FIG. 4 is a block diagram of an apparatus for generating training data according to another exemplary embodiment.
  • an object detector 420 including an image input unit 410, an object detector 430, and an object tracker 440, a short-term experience database unit 450, and learning data
  • the configuration and function of the generation unit 460 is basically the same as described in the above-described embodiment.
  • the continuity detection unit 470 is further included to detect a part of the continuity in time in which the same object exists in the image input by scene change detection or the like. With such a configuration, it is possible to provide a temporally continuous image in which the same object exists from the input image, and thus, it is possible to generate a labeled learning data by receiving an image from a TV video or various other input sources.
  • an image providing device that can deliver the image data from a plurality of devices, such as a plurality of vehicles or CCTV, distributed in various places through the network, it is also possible to generate large-scale labeled learning data.
  • FIG. 5 shows an example of a self-learning apparatus using the method of generating learning data according to the present invention.
  • an object detector 520 including an image input unit 510, an object detector 530, and an object tracker 540, and a short-term experience database.
  • the configuration and function of the unit 550, the learning data generator 560, and the continuity detector 570 are basically the same as those described in the above-described embodiment.
  • the object detection device 580 is configured to self-learn by using the labeled learning data generated by the learning data generator 560 as training data.
  • the object detection algorithm 580 may update the object recognition algorithm used in the object search unit 520 according to a machine learning object recognition algorithm using the labeled learning data, and may continuously improve performance.
  • the automatically generated labeling training data can be applied to various machine learning.
  • the learning data generation method and apparatus and the self-learning apparatus and method using the same according to the present invention is not limited to the configuration of the embodiments described as described above, the above embodiments can be made in various modifications All or part of each of the embodiments may be configured to be selectively combined so that.
  • the subject matter of the present invention is not limited to the use of hardware or software, and can be applied to any other computing or processing environment. It may be implemented in hardware, software or a combination of hardware software described in the present invention.
  • the present invention can be implemented using circuits. That is, one or more programmable logic circuits, i.e., application specific integrated circuits (ASICs) or logic circuits (AND, OR, NAND gates) or processing devices (e.g., microprocessors, controllers) may be implemented.
  • ASICs application specific integrated circuits
  • AND, OR, NAND gates logic circuits
  • processing devices e.g., microprocessors, controllers
  • the present invention can be implemented as a computer program on a programmable computer.
  • a computer may include a processor, a storage device, an input device, and an output device.
  • Program code may be input by a mouse or a keyboard input device to implement the contents described in the present invention.
  • These programs can be implemented in high-level or object-oriented languages. It can also be implemented as a computer system implemented in assembly or machine code.
  • the present invention can be used to automatically generate labeled learning data for machine learning, and can also be used in a self-learning apparatus and method capable of self-learning using the generated labeled learning data.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention concerne un procédé et un dispositif de génération de données d'apprentissage, et un dispositif d'apprentissage auto-orienté et un procédé l'utilisant. Le procédé de génération de données d'apprentissage selon la présente invention comprend les étapes consistant : à recevoir de multiples images ayant une continuité ; à détecter et à suivre un objet dans les multiples images ; à stocker les multiples images, un objet reconnu par la détection et le suivi d'objet dans chacune des images, et le taux de reconnaissance de l'objet ; et lorsque, parmi les multiples images, il y a une image dans laquelle le taux de reconnaissance d'un objet particulier est supérieur ou égal à une valeur seuil prédéterminée, à étiqueter les multiples images en utilisant l'objet particulier reconnu dans l'image correspondante en tant qu'objet reconnu final, de façon à générer des données étiquetées. Selon la présente invention, des données d'apprentissage étiquetées requises pour un apprentissage automatique peuvent être automatiquement générées.
PCT/KR2017/002784 2017-03-15 2017-03-15 Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant WO2018169103A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/KR2017/002784 WO2018169103A1 (fr) 2017-03-15 2017-03-15 Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/KR2017/002784 WO2018169103A1 (fr) 2017-03-15 2017-03-15 Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant

Publications (1)

Publication Number Publication Date
WO2018169103A1 true WO2018169103A1 (fr) 2018-09-20

Family

ID=63522501

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/KR2017/002784 WO2018169103A1 (fr) 2017-03-15 2017-03-15 Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant

Country Status (1)

Country Link
WO (1) WO2018169103A1 (fr)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130132311A1 (en) * 2011-11-18 2013-05-23 Honeywell International Inc. Score fusion and training data recycling for video classification
US20130142418A1 (en) * 2011-12-06 2013-06-06 Roelof van Zwol Ranking and selecting representative video images
US8762299B1 (en) * 2011-06-27 2014-06-24 Google Inc. Customized predictive analytical model training
US20140279739A1 (en) * 2013-03-15 2014-09-18 InsideSales.com, Inc. Resolving and merging duplicate records using machine learning
US20150033362A1 (en) * 2012-02-03 2015-01-29 See-Out Pty Ltd. Notification and Privacy Management of Online Photos and Videos

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8762299B1 (en) * 2011-06-27 2014-06-24 Google Inc. Customized predictive analytical model training
US20130132311A1 (en) * 2011-11-18 2013-05-23 Honeywell International Inc. Score fusion and training data recycling for video classification
US20130142418A1 (en) * 2011-12-06 2013-06-06 Roelof van Zwol Ranking and selecting representative video images
US20150033362A1 (en) * 2012-02-03 2015-01-29 See-Out Pty Ltd. Notification and Privacy Management of Online Photos and Videos
US20140279739A1 (en) * 2013-03-15 2014-09-18 InsideSales.com, Inc. Resolving and merging duplicate records using machine learning

Similar Documents

Publication Publication Date Title
EP3673417B1 (fr) Système et procédé d'apprentissage distributif et de distribution de poids dans un réseau neuronal
CN102457680B (zh) 图像处理装置及图像处理方法
WO2017164478A1 (fr) Procédé et appareil de reconnaissance de micro-expressions au moyen d'une analyse d'apprentissage profond d'une dynamique micro-faciale
CN104732187B (zh) 一种图像跟踪处理的方法及设备
CN107886048A (zh) 目标跟踪方法及系统、存储介质及电子终端
WO2019040213A1 (fr) Réseau de détection et d'examen en temps réel adaptatif (arden)
WO2015178540A1 (fr) Appareil et procédé de suivi de cible à l'aide d'un transfert intercellulaire entre des caméras
CN102087746A (zh) 图像处理装置、图像处理方法和程序
WO2016013885A1 (fr) Procédé d'extraction d'image et dispositif électronique associé
WO2021107650A1 (fr) Apprentissage conjoint de mouvement visuel et de confiance à partir de correctifs locaux dans des caméras d'événements
US20140226857A1 (en) Information processing system, information processing method and program
WO2020246655A1 (fr) Procédé de reconnaissance de situation et dispositif permettant de le mettre en œuvre
WO2021075772A1 (fr) Procédé et dispositif de détection d'objet au moyen d'une détection de plusieurs zones
WO2021020866A1 (fr) Système et procédé d'analyse d'images pour surveillance à distance
EP3765995B1 (fr) Système et procédés de reconnaissance entre caméras d'individus et de leurs propriétés
WO2018111011A1 (fr) Système et procédé de détection d'objets en mouvement
WO2021101045A1 (fr) Appareil électronique et procédé de commande associé
WO2021100919A1 (fr) Procédé, programme et système pour déterminer si un comportement anormal se produit, sur la base d'une séquence de comportement
WO2021235682A1 (fr) Procédé et dispositif de réalisation d'une prédiction de comportement à l'aide d'une attention auto-focalisée explicable
CN114359976A (zh) 一种基于人物识别的智能安防方法与装置
WO2022197136A1 (fr) Système et procédé permettant d'améliorer un modèle d'apprentissage machine destiné à une compréhension audio/vidéo au moyen d'une attention suscitée à multiples niveaux et d'une formation temporelle par antagonisme
WO2024090786A1 (fr) Procédé d'entraînement de modèle de détection de chute basé sur des données radar
WO2018169103A1 (fr) Procédé et dispositif de génération de données d'apprentissage automatique et dispositif d'apprentissage auto-orienté et procédé l'utilisant
CN114140751B (zh) 一种考场监控方法及系统
US11722763B2 (en) System and method for audio tagging of an object of interest

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17901126

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 030120)

122 Ep: pct application non-entry in european phase

Ref document number: 17901126

Country of ref document: EP

Kind code of ref document: A1