US 12,586,357 B2Granted

Collecting method for training data

AssigneeHonda Motor Co., Ltd.

InventorToshikazu Karube

Life of this patent

Estimating the term…

Loading patent…

Abstract

The present invention relates to a collecting method for training data for collecting defective product data as training data for learning a classification model that classifies the inspected object as a normal product or an abnormal product, including: collecting many pieces of defective product data (step 1 of FIG. 3); extracting a plurality of feature quantities respectively from many pieces of the defective product data (step 2); calculating a state sum for every feature quantity of the plurality of feature quantities that have been extracted, for the many pieces of the defective product data (step 3); calculating, as an index value, a logarithmic sum of a plurality of state sums that have been calculated (step 4); and ending collecting the defective product data, in a case where the calculated index value is equal to or greater than a predetermined target value (step 5).

A collecting method for training data for collecting defective product data that is image data obtained by imaging an appearance of an inspected object that has been determined to be an abnormal product by an inspection operator, as the training data for learning a classification model that classifies the inspected object as a normal product or the abnormal product, the collecting method comprising:

collecting many pieces of the defective product data;
extracting a plurality of feature quantities respectively from the many pieces of the defective product data that have been collected;
calculating, in an expression (1), a state sum Zk for every feature quantity of the plurality of feature quantities that have been extracted, for the many pieces of the defective product data, wherein the extracting and calculating are performed on multi-dimensional feature vectors derived from image data of the inspected object;
calculating, in an expression (2), as an index value HEG, a logarithmic sum of a plurality of state sums Zk that have been calculated;
and
ending collecting the defective product data, in a case where the index value HEG that has been calculated is equal to or greater than a predetermined target value
[
Expression
⁢
1
]

Z
k
=
a
*
∑
i
=
1
N
exp
⁡
(
-
(
x
i
-
mu
N
)
2
σ
N
2
)
(
1
)
Zk
:
STATE
⁢
SUM
⁢
FOR
⁢
EVERY
⁢
FEATURE
⁢
QUANTITY
⁢
OF
⁢
ALL
⁢
PIECES
⁢
OF
⁢
DEFECTIVE
⁢
PRODUCT
⁢
DATA
⁢
N
:
THE
⁢
NUMBER
⁢
OF
⁢
PIECES
⁢
OF
⁢
DEFECTIVE
⁢
PRODUCT
⁢
DATA
⁢
xi
:
VALUE
⁢
OF
⁢
FEATURE
⁢
QUANTITY
⁢
OF
⁢
EACH
⁢
PIECE
⁢
OF
⁢
DEFECTIVE
⁢
PRODUCT
⁢
DATA
⁢
muN
:
AVERAGE
⁢
VALUE
⁢
FOR
⁢
EVERY
⁢
FEATURE
⁢
QUANTITY
⁢
OF
⁢
ALL
⁢
PIECES
⁢
OF
⁢
DEFECTIVE
⁢
PRODUCT
⁢
DATA
⁢
σ
⁢
N
:
STANDARD
⁢
DEVIATION
⁢
a
:
NORMALIZATION
⁢
CONSTANT
[
Expression
⁢
2
]

HEG
=
∑
k
=
1
F
log
⁢
Z
k
(
2
)
HEG
:
INDEX
⁢
VALUE
⁢
F
:
DIMENSION
⁢
NUMBER
⁢
OF
⁢
FEATURE
⁢
QUANTITY

186 more elements

Facts

Application

18/191,256Filed 28 Mar 2023

Priority

JP 2022-05761730 Mar 2022

Published

US 2023/0316725 A15 Oct 2023

Granted

24 Mar 2026

Term ends

Not estimatedThe source records don’t support an estimate.

Classification15 codes

G06V 10/774
Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or…:…
G06F 18/2433
Relating to the number of classes: single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
G06N 3/09
Learning methods: supervised learning
G06T 7/0004
Inspection of images, e.g. flaw detection: industrial image inspection

Abstract

Claims1