Processing

Please wait...

Settings

Settings

Goto Application

1. WO2020025245 - METHOD AND DEVICE FOR ESTABLISHING AN EXPLANATION MAP

Publication Number WO/2020/025245
Publication Date 06.02.2020
International Application No. PCT/EP2019/067841
International Filing Date 03.07.2019
IPC
G06K 9/46 2006.1
GPHYSICS
06COMPUTING; CALCULATING OR COUNTING
KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
9Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
36Image preprocessing, i.e. processing the image information without deciding about the identity of the image
46Extraction of features or characteristics of the image
G06K 9/62 2006.1
GPHYSICS
06COMPUTING; CALCULATING OR COUNTING
KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
9Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
62Methods or arrangements for recognition using electronic means
CPC
G06K 9/6232
GPHYSICS
06COMPUTING; CALCULATING; COUNTING
KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
9Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
62Methods or arrangements for recognition using electronic means
6217Design or setup of recognition systems and techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
6232Extracting features by transforming the feature space, e.g. multidimensional scaling; Mappings, e.g. subspace methods
G06K 9/6271
GPHYSICS
06COMPUTING; CALCULATING; COUNTING
KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
9Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
62Methods or arrangements for recognition using electronic means
6267Classification techniques
6268relating to the classification paradigm, e.g. parametric or non-parametric approaches
627based on distances between the pattern to be recognised and training or reference patterns
6271based on distances to prototypes
G06V 10/454
G06V 10/764
G06V 10/82
G06V 2201/03
Applicants
  • ROBERT BOSCH GMBH [DE]/[DE]
Inventors
  • WAGNER, Joerg
  • KOEHLER, Jan Mathias
  • GINDELE, Tobias
  • HETZEL, Leon
  • WIEDEMER, Jakob Thaddaeus
Priority Data
10 2018 213 056.603.08.2018DE
Publication Language German (de)
Filing Language German (DE)
Designated States
Title
(DE) VERFAHREN UND VORRICHTUNG ZUM ERMITTELN EINER ERKLÄRUNGSKARTE
(EN) METHOD AND DEVICE FOR ESTABLISHING AN EXPLANATION MAP
(FR) PROCÉDÉ ET DISPOSITIF POUR DÉTERMINER UNE CARTE DESCRIPTIVE
Abstract
(DE) Verfahren zum Ermitteln einer Erklärungs karte (sϰ*) eines Bildes (x), in der all jene Pixel des Bildes (x) verändert sind, die maßgeblich für eine mit einem tiefen neuronalen Netz (ƒ) ermittelte Klassifikation des Bildes (x) sind, wobei die Erklärungskarte (sϰ*) derart gewählt wird, dass eine möglichst kleine Untermenge der Pixel des Bildes (x) verändert werden und die Erklärungskarte (sϰ*) dann, wenn sie dem tiefen neuronalen Netz (ƒ) zur Klassifikation zugeführt wird, möglichst nicht zum gleichen Klassifikationsergebnis (ƒ(sϰ*|θ)) führt wie das Bild (x), wobei die Erklärungskarte (sϰ*) derart gewählt wird, dass in Merkmalskarten (hı) des tiefen neuronalen Netzes (ƒ) eine durch die Erklärungs karte (sϰ*) hervorgerufene Aktivierung (hı(sϰ*)) eine durch das Bild (x) hervorgerufene Aktivierung (hı(sϰ*)) dann im Wesentlichen nicht übersteigt.
(EN) The invention relates to a method for establishing an explanation map (sϰ*) of an image (x), in which all the pixels of the image (x) that are critical for a classification of the image (x) established by means of a deep neural network (ƒ) are altered, wherein the explanation map (sϰ*) is selected in such a way that the smallest possible subset of the pixels of the image (x) is altered and the explanation map (sϰ*), when supplied to the deep neural network (ƒ) for classification, as far as possible does not lead to the same classification result (ƒ(sϰ*|θ)) as the image (x), wherein the explanation map (sϰ*) is selected in such a way that in feature maps (hı) of the deep neural network (ƒ) an activation (hı(sϰ*)) brought about by the explanation map (sϰ*) substantially does not exceed an activation (hı(sϰ*)) brought about by the image (x).
(FR) Procédé permettant de déterminer une carte descriptive (sϰ*) d’une image (x) dans laquelle sont modifiés tous les pixels de l'image (x) qui sont déterminants pour une classification de l’image (x) déterminée avec un réseau neuronal profond (ƒ), la carte descriptive (sϰ*) étant sélectionnée de telle sorte qu’un sous-quantité aussi petite que possible des pixels de l’image (x) peut être modifiée, et la carte descriptive (sϰ*), lorsqu’elle est fournie au réseau neuronal profond (ƒ) à des fins de classification, conduit le moins possible au même résultat de classification (ƒ(sϰ*|θ)) que l’image (x), la carte descriptive (sϰ*) étant choisie de sorte que dans des cartes de caractéristiques (hı) du réseau neuronal profond (ƒ) une activation (hı(sϰ*)) provoquée par la carte descriptive (sϰ*) ne dépasse sensiblement pas une activation (hı(sϰ*)) provoquée par l’image (x).
Latest bibliographic data on file with the International Bureau