Title:
TARGET IMAGE EXTRACTING METHOD, NEURAL NETWORK TRAINING METHOD, AND DEVICE
Document Type and Number:
WIPO Patent Application WO/2022/042120
Kind Code:
A1
Abstract:
A target image extracting method, a neural network training method, and a device, related to the technical field of image processing. The method comprises: performing a first feature extraction with respect to an input image to produce a reference feature map (S310); performing a second feature extraction with respect to the reference feature map to produce target feature maps of a number identical to the number of apexes of a target image in the input image (S320); normalizing eigenvalues of the target feature maps to produce probability feature maps corresponding one-to-one to the target feature maps (S330); acquiring, on the basis of the probability feature maps, coordinate feature maps corresponding to the probability feature maps, and utilizing the probability feature maps and the corresponding coordinate feature maps to determine target apex coordinates of each apex of the target image (S340); and completing the extraction of the target image on the basis of the target apex coordinates (S350). The precision of target image extraction is increased.
More Like This:
Inventors:
YAN BO (CN)
Application Number:
PCT/CN2021/106582
Publication Date:
March 03, 2022
Filing Date:
July 15, 2021
Export Citation:
Assignee:
GUANGDONG OPPO MOBILE TELECOMMUNICATIONS CORP LTD (CN)
International Classes:
G06K9/00
Foreign References:
| CN111950570A | 2020-11-17 | |||
| CN109670503A | 2019-04-23 | |||
| CN110059680A | 2019-07-26 | |||
| US20120099790A1 | 2012-04-26 |
Attorney, Agent or Firm:
SHENZHEN LUNGTIN LIANDING INTELLECTUAL PROPERTY AGENT LTD. (CN)
Download PDF: