Classification method, classification module and computer program product using the same

Patent metadata

Field	Value
Publication number	US-10489687-B2
Application number	US-201715589423-A
Country	US
Kind code	B2
Filing date	May 8, 2017
Priority date	Nov 23, 2016
Publication date	Nov 26, 2019
Grant date	Nov 26, 2019

Abstract

A classification method includes the following steps. Firstly, a classification module including a deep neural network (DNN) is provided. Then, to-be-classified sample is obtained. Then, the DNN automatically extracts a feature response of the to-be-classified sample. Then, whether the feature response of the to-be-classified sample falls within a boundary scope of several training samples is determined; wherein the training samples are classified into several categories. Then, if the feature response of the to-be-classified sample falls within the boundary scope, the DNN determines that to-be-classified sample belongs to which one of the categories according to the training samples.

First claim

What is claimed is:

1. A classification method, comprising:
providing a classification module comprising a Deep Neural Network (DNN);
obtaining a to-be-classified sample of a product, wherein the to-be-classified sample is an image captured by an Automated Optical Inspection (AOI);
automatically extracting a feature response of the to-be-classified sample by the Deep Neural Network;
determining whether the feature response of the to-be-classified sample is within a boundary scope of a plurality of training samples, wherein the training samples belong to a plurality of categories; and
if the feature response of the to-be-classified sample is within the boundary scope of the training samples, determining that the to-be-classified sample belongs to which one of the categories;
wherein the categories comprise a defect category and a non-defect category; the step of determining that the to-be-classified sample belongs to which one of the categories comprises:
determining whether the to-be-classified sample belongs to the defect category; and

wherein the classification method further comprises:
if the to-be-classified sample belongs to the defect category, reconfirming whether the to-be-classified sample belongs to the defect category or the non-defect category; and
if the to-be-classified sample belongs to the defect category after being reconfirmed, adding the to-be-classified sample to become one of the training samples.

2. The classification method according to claim 1 , further comprising:
if the feature response of the to-be-classified sample is not within the boundary scope of the training samples, classifying the to-be-classified sample.

3. The classification method according to claim 1 , further comprising:
if the feature response of the to-be-classified sample is outside the boundary scope of the training samples, manually classifying the to-be-classified sample.

4. The classification method according to claim 2 , wherein after the step of classifying the to-be-classified sample, the classification method further comprising:
adding the to-be-classified sample to become one of the training samples.

5. The classification method according to claim 1 , further comprising:
if the to-be-classified sample does not belong to the defect category after being reconfirmed, adding the to-be-classified sample to become one of the training samples.

6. The classification method according to claim 1 , wherein the training samples belong to the defect category and the non-defect category, the number of the training samples belonging to the defect category and the number of the training samples belonging to the non-defect category are equal.

7. The classification method according to claim 1 , wherein before the step of determining whether the feature response of the to-be-classified sample is within the boundary scope of the training samples, the classification method further comprises a learning process comprising:
obtaining a plurality of the training samples;
learning the feature response of each training sample by the Deep Neural Network;
determining the boundary scope of the training samples.

8. A classification module, comprising:
a Deep Neural Network configured to automatically extract a feature response of a to-be-classified classified sample of a product, wherein the to-be-classified sample is an image captured by an Automated Optical Inspection; and
a determination unit configured to determine whether the feature response of the to-be-classified sample is within a boundary scope of a plurality of training samples, wherein the training samples belong to a plurality of categories;
wherein if the feature response of the to-be-classified sample is within the boundary scope of the training samples, the Deep Neural Network determines that the to-be-classified sample belongs to which one of the categories;
wherein the categories comprise a defect category and a non-defect category; the Deep Neural Network is further configured to determine whether the to-be-classified sample belongs to the defect category; and
wherein the classification module further comprises:
a boundary learning unit configured to: if the to-be-classified sample belongs to the defect category after reconfirming whether the to-be-classified sample belongs to the defect category or the non-defect category, add the to-be-classified sample to become one of the training samples.

9. The classification module according to claim 8 , wherein the training samples belong to the defect category and the non-defect category, the number of the training samples belonging to the defect category and the number of the training samples belonging to the non-defect category are equal.

10. The classification module according to claim 8 , wherein the Deep Neural Network is further configured to learn the feature response of each training sample; the boundary learning unit is configured to determine the boundary scope of the training samples.

11. A non-transitory computer readable medium storing a program causing a classification module to execute a classification method, the classification method comprising:
automatically extracting a feature response of a to-be-classified sample of a product, wherein the to-be-classified sample is an image captured by an Automated Optical Inspection;
determining whether the feature response of the to-be-classified sample is within a boundary scope of a plurality of training samples, wherein the training samples belong to a plurality of categories; and
if the feature response of the to-be-classified sample is within the boundary scope of the training samples, determining that the to-be-classified sample belongs to which one of the categories;
wherein the categories comprise a defect category and a non-defect category; the step of determining that the to-be-classified sample belongs to which one of the categories comprises:
determining whether the to-be-classified sample belongs to the defect category; and

wherein the classification method further comprises:
if the to-be-classified sample belongs to the defect category, reconfirming whether the to-be-classified sample belongs to the defect category or the non-defect category; and
if the to-be-classified sample belongs to the defect category after being reconfirmed, adding the to-be-classified sample to become one of the training samples.

Assignees

• Ind Tech Res Inst

Inventors

• Tsai Dong-Chen
• Tang Tsann-Tay

Classifications

• G06V10/82 Primary

  using neural networks · CPC title

• G06F18/24 Primary

  Classification techniques · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F18/2411

  based on the proximity to a decision surface, e.g. support vector machines · CPC title