Title:
IMAGE CLASSIFICATION METHOD USING PARTIAL DIFFERENTIAL OPERATOR-BASED GENERAL-EQUIVARIANT CONVOLUTIONAL NEURAL NETWORK MODEL
Document Type and Number:
WIPO Patent Application WO/2022/062164
Kind Code:
A1
Abstract:
An image classification method using a partial differential operator (PDO)-based general-equivariant convolutional neural network (CNN) model. An equivariant CNN model PDO-sCNNs is constructed using group representations and PDOs, and constraints are imposed on the convolution kernel of the CNN model, so that the CNN has a set symmetry or equivariance, and is used for efficient image classification and recognition visual analysis. By means of the method in the present invention, all equivariant convolutional layers satisfying conditions are solved by using differential operators and group representations, and an equivariant CNN model can be constructed by using an equivariant convolution solved by a convolutional layer in any existing CNN model, and then the model is used for image classification and recognition; the effect is better.
More Like This:
Inventors:
LIN ZHOUCHEN (CN)
HE LINGSHEN (CN)
SHEN ZHENGYANG (CN)
XU DAPENG (CN)
HE LINGSHEN (CN)
SHEN ZHENGYANG (CN)
XU DAPENG (CN)
Application Number:
PCT/CN2020/132017
Publication Date:
March 31, 2022
Filing Date:
November 27, 2020
Export Citation:
Assignee:
UNIV BEIJING (CN)
International Classes:
G06K9/62
Domestic Patent References:
WO2020025191A1 | 2020-02-06 |
Foreign References:
CN111401452A | 2020-07-10 | |||
CN107368886A | 2017-11-21 | |||
CN111160436A | 2020-05-15 | |||
CN107766794A | 2018-03-06 |
Attorney, Agent or Firm:
BEIJING WANXIANGXINYUE INTELLECTUAL PROPERTY OFFICE (CN)
Download PDF: