Processing

Please wait...

Settings

Settings

Goto Application

1. WO2022025690 - NEURAL NETWORK MODEL TRAINING METHOD AND APPARATUS FOR COMPLEX CHARACTERISTIC CLASSIFICATION AND COMMON LOCALIZATION

Publication Number WO/2022/025690
Publication Date 03.02.2022
International Application No. PCT/KR2021/009939
International Filing Date 29.07.2021
IPC
G06N 3/08 2006.1
GPHYSICS
06COMPUTING; CALCULATING OR COUNTING
NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
3Computer systems based on biological models
02using neural network models
08Learning methods
G06N 3/04 2006.1
GPHYSICS
06COMPUTING; CALCULATING OR COUNTING
NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
3Computer systems based on biological models
02using neural network models
04Architecture, e.g. interconnection topology
CPC
G06N 3/04
GPHYSICS
06COMPUTING; CALCULATING; COUNTING
NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
3Computer systems based on biological models
02using neural network models
04Architectures, e.g. interconnection topology
G06N 3/08
GPHYSICS
06COMPUTING; CALCULATING; COUNTING
NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
3Computer systems based on biological models
02using neural network models
08Learning methods
Applicants
  • 주식회사 웨이센 WAYCEN INC. [KR]/[KR]
Inventors
  • 금지수 KEUM, Jisoo
  • 오상일 OH, Sangil
  • 김경남 KIM, Kyungnam
Agents
  • 이준성 LEE, Joon Sung
Priority Data
10-2020-009577331.07.2020KR
Publication Language Korean (ko)
Filing Language Korean (KO)
Designated States
Title
(EN) NEURAL NETWORK MODEL TRAINING METHOD AND APPARATUS FOR COMPLEX CHARACTERISTIC CLASSIFICATION AND COMMON LOCALIZATION
(FR) PROCÉDÉ ET APPAREIL D'APPRENTISSAGE DE MODÈLE DE RÉSEAU NEURONAL POUR UNE CLASSIFICATION DE CARACTÉRISTIQUES COMPLEXES ET UNE LOCALISATION COMMUNE
(KO) 복합 특성 분류와 공통 국소화를 위한 신경망 모델의 학습 방법 및 장치
Abstract
(EN) According to a neural network model training method for complex characteristic classification and common localization of an image, according to an embodiment of the present invention, a neural network model comprises: a convolution layer for performing a convolution operation on an input image by using a convolution filter; a pooling layer for performing pooling on an output of the convolution layer; and a plurality of fully connected layers for each class, respectively corresponding to a plurality of classes, in which complex characteristics are classified, and outputting a value obtained by multiplying an output of the pooling layer by a weight (wfc(Tt)) for each class, and the method comprises the steps of: (a) inputting the input image to the convolution layer; (b) calculating a plurality of observation maps for each class, on the basis of the output of the convolution layer; (c) calculating an observation loss (Lobs) common to the plurality of classes, on the basis of the plurality of observation maps for each class; and (d) back-propagating, to the neural network model, a loss based on the observation loss.
(FR) Selon un procédé d'apprentissage de modèle de réseau neuronal pour une classification de caractéristiques complexes et une localisation commune d'une image, selon un mode de réalisation de la présente invention, un modèle de réseau neuronal comprend : une couche de convolution pour effectuer une opération de convolution sur une image d'entrée à l'aide d'un filtre de convolution ; une couche de regroupement pour effectuer un regroupement sur une sortie de la couche de convolution ; et une pluralité de couches entièrement connectées pour chaque classe, correspondant respectivement à une pluralité de classes, dans laquelle des caractéristiques complexes sont classées, et délivrant en sortie une valeur obtenue par multiplication d'une sortie de la couche de regroupement par un poids (wfc(Tt)) pour chaque classe, et le procédé comprend les étapes suivantes : (a) l'entrée de l'image d'entrée dans la couche de convolution ; (b) le calcul d'une pluralité de cartes d'observation pour chaque classe, sur la base de la sortie de la couche de convolution ; (c) le calcul d'une perte d'observation (Lobs) commune à la pluralité de classes, sur la base de la pluralité de cartes d'observation pour chaque classe ; et (d) la rétropropagation, vers le modèle de réseau neuronal, d'une perte basée sur la perte d'observation.
(KO) 본 발명의 실시예에 따른 영상의 복합 특성 분류 및 공통 국소화를 위한 신경망 모델의 학습 방법에서, 상기 신경망 모델은, 입력 영상에 대해 컨볼루션 필터를 이용하여 컨볼루션 연산을 수행하는 컨볼루션층과, 상기 컨볼루션층의 출력에 대해 풀링(pooling)을 수행하기 위한 풀링층과, 복합 특성이 분류되는 복수의 클래스에 각각 대응하며, 상기 풀링층의 출력에 대해 클래스별 가중치(wfc(Tt))를 곱한 값을 출력하는 복수의 클래스별 완전결합층(fully connected layer)을 포함하고, 상기 방법은, (a) 입력 영상을 상기 컨볼루션층에 입력하는 단계; (b) 상기 컨볼루션층의 출력에 기초하여 복수의 클래스별 관찰 지도를 연산하는 단계; (c) 상기 복수의 클래스별 관찰 지도에 기초하여, 복수의 클래스에 공통되는 관찰 손실(Lobs)을 연산하는 단계; 및 (d) 상기 관찰 손실에 기초한 손실을 상기 신경망 모델에 역전파하는 단계를 포함한다.
Related patent documents
Latest bibliographic data on file with the International Bureau