ARTIFICIAL INTELLIGENCE DEVICE FOR DETECTING DEFECTIVE PRODUCT ON BASIS OF PRODUCT IMAGE, AND METHOD THEREFOR
An artificial intelligence device according to an embodiment of the present disclosure comprises: a memory for storing at least one normal product image belonging to a normal classification and at least one abnormal product image belonging to an abnormal classification; a learning processor for perf...
Gespeichert in:
Hauptverfasser: | , , , , |
---|---|
Format: | Patent |
Sprache: | eng ; fre ; kor |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | An artificial intelligence device according to an embodiment of the present disclosure comprises: a memory for storing at least one normal product image belonging to a normal classification and at least one abnormal product image belonging to an abnormal classification; a learning processor for performing contrastive learning on a feature extraction model so that representation vectors of product images belonging to the same classification move closer together and representation vectors of product images belonging to different classifications move farther away from each other; and a processor which inputs at least one normal product image belonging to the normal classification into the feature extraction model on which the contrastive learning has been performed, so as to acquire an embedding vector for each patch unit of at least one normal product image, and which acquires a normal distribution of the acquired embedding vector for each patch unit.
Un dispositif d'intelligence artificielle selon un mode de réalisation de la présente divulgation comprend : une mémoire pour stocker au moins une image de produit normal appartenant à une classification normale et au moins une image de produit anormal appartenant à une classification anormale ; un processeur d'apprentissage pour effectuer un apprentissage contrastif sur un modèle d'extraction de caractéristiques de telle sorte que des vecteurs de représentation d'images de produit appartenant à la même classification se rapprochent et des vecteurs de représentation d'images de produit appartenant à différentes classifications s'éloignent les uns des autres ; et un processeur qui entre au moins une image de produit normal appartenant à la classification normale dans le modèle d'extraction de caractéristiques sur lequel l'apprentissage contrastif a été effectué, de façon à acquérir un vecteur d'incorporation pour chaque unité de correctif d'au moins une image de produit normal, et qui acquiert une distribution normale du vecteur d'incorporation acquis pour chaque unité de correctif.
본 개시의 실시 예에 따른 인 공 지능 장치는 정상 분류에 속하는 적어도 하나의 정상 제품 이미지 및 비정상 분류에 속하는 적어도 하나의 비정상 제품 이미지를 저장하는 메모리, 동일 분류에 속하는 제품 이미지의 표현 벡터(Representation Vector)가 가까워지도록 하고, 서로 다른 분류에 속하는 제품 이미지의의 표현 벡터가 서로 멀어지도록 특징 추출 모델에 대한 대조 학습(Contrastive Leaning)을 시키는 러닝 프로세서 및 상기 대조 학습된 특징 추출 모델에 상기 정상 분류에 속하는 적어도 하나의 정상 제품 이미지를 입력하여 적어도 하나의 정상 제품 이미지의 패치 단위별 임베딩 벡터(embedding vector)를 획득하고, 상기 획득한 패치 단위별 임베딩 벡터의 정규 분포를 획득하는 프로세서를 포함한다. |
---|