COARSE-TO-FINE CONVOLUTIONAL NEURAL NETWORK-BASED MULTILABEL CLASS CLASSIFICATION METHOD AND APPARATUS
The present invention relates to a coarse-to-fine convolutional neural network-based multilabel class classification apparatus that generates a hierarchical structure-based plurality of group labels for a plurality of classes to be classified, by using a disjoint grouping method, predicts a class be...
Saved in:
Main Authors | , , , |
---|---|
Format | Patent |
Language | English French Korean |
Published |
11.05.2023
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | The present invention relates to a coarse-to-fine convolutional neural network-based multilabel class classification apparatus that generates a hierarchical structure-based plurality of group labels for a plurality of classes to be classified, by using a disjoint grouping method, predicts a class belonging to each of the plurality of group labels from among the plurality of classes by using a coarse-to-fine convolutional neural network including a main network and one or more sub-networks, completes training the coarse-to-fine convolutional neural network through prediction, and when an image including an object corresponding to a class is input, classifies the class included in the image as the main network of the coarse-to-fine convolutional neural network that has been trained receives an input of a feature map input from a last convolutional layer of the one or more sub-networks.
La présente invention concerne un appareil de classification de classes à plusieurs étiquettes basés sur un réseau neuronal convolutif grossier à fin qui génère une pluralité d'étiquettes de groupe basée sur une structure hiérarchique pour une pluralité de classes à classifier, en utilisant un procédé de regroupement disjoint, prédit une classe appartenant à chacune de la pluralité d'étiquettes de groupe parmi la pluralité de classes en utilisant un réseau neuronal convolutif grossier à fin comprenant un réseau principal et un ou plusieurs sous-réseaux, achève l'entraînement du réseau neuronal convolutif grossier à fin par prédiction, et lorsqu'une image comprenant un objet correspondant à une classe est entrée, classifie la classe incluse dans l'image à mesure que le réseau principal du réseau neuronal convolutif grossier à fin qui a été entraîné reçoit une entrée d'une entrée de carte de caractéristiques à partir d'une dernière couche de convolution du ou des sous-réseaux.
본 발명은 코스-투-파인 컨볼루션 뉴럴 네트워크(Coarse-to-Fine Convolutional Neural Network) 기반 다중 레이블 클래스 분류 장치로서, 디스조인트(disjoint) 그룹화 방법을 이용하여 분류 대상이 되는 복수의 클래스에 대한 계층적 구조 기반의 복수의 그룹 레이블을 생성하고, 메인 네트워크 및 하나 이상의 서브 네트워크를 포함하는 코스-투-파인 컨볼루션 뉴럴 네트워크를 이용하여 복수의 클래스 중 복수의 그룹 레이블 각각에 속하는 클래스를 예측하고, 예측을 통해 코스-투-파인 컨볼루션 뉴럴 네트워크의 학습을 완료하고, 클래스에 상응하는 객체를 포함하는 이미지가 입력되는 경우, 학습이 완료된 코스-투-파인 컨볼루션 뉴럴 네트워크의 메인 네트워크가 하나 이상의 서브 네트워크의 마지막 컨볼루션 레이어에서 입력되는 특징 맵을 입력 받아, 이미지에 포함되는 클래스를 분류하도록 한다. |
---|---|
Bibliography: | Application Number: WO2021KR17922 |