Title:
NEURAL NETWORK TRAINING METHOD, ELECTRONIC DEVICE, AND COMPUTER STORAGE MEDIUM
Document Type and Number:
WIPO Patent Application WO/2023/077809
Kind Code:
A1
Abstract:
A neural network training method, an electronic device, and a computer storage medium. The method comprises the following steps: obtaining a preset training sample image (S100); obtaining a prediction feature map (S200); and on the basis of a preset boundary aware depth loss function, training a neural network model by using the preset training sample image as an input and the prediction feature map as an output to obtain a trained neural network model (S300). By setting the boundary aware depth loss function, the attention level of the neural network model to a boundary area during training can be effectively increased, and it is ensured that the depth and the depth gradient of the boundary area have good correctness. Thus, a smoothing phenomenon is effectively suppressed, the problem that the boundary area is fuzzy is avoided, and finally, the boundary prediction precision can be effectively improved.
More Like This:
Inventors:
CUI YAN (CN)
CHANG QINGLING (CN)
YANG XIN (CN)
LIAO HUANHAO (CN)
WANG YUHAN (CN)
CHANG QINGLING (CN)
YANG XIN (CN)
LIAO HUANHAO (CN)
WANG YUHAN (CN)
Application Number:
PCT/CN2022/098767
Publication Date:
May 11, 2023
Filing Date:
June 14, 2022
Export Citation:
Assignee:
UNIV WUYI (CN)
GUANGDONG SIWEIKAN INTELLIGENT EQUIPMENT CO LTD (CN)
CHINA GERMANY ZHUHAI ARTIFICIAL INTELLIGENCE RES INSTITUTE CO LTD (CN)
4DAGE CO LTD (CN)
GUANGDONG SIWEIKAN INTELLIGENT EQUIPMENT CO LTD (CN)
CHINA GERMANY ZHUHAI ARTIFICIAL INTELLIGENCE RES INSTITUTE CO LTD (CN)
4DAGE CO LTD (CN)
International Classes:
G06V10/46
Foreign References:
CN109410261A | 2019-03-01 | |||
CN111784628A | 2020-10-16 | |||
CN111445432A | 2020-07-24 | |||
US20210089807A1 | 2021-03-25 | |||
CN114170438A | 2022-03-11 |
Attorney, Agent or Firm:
JIAQUAN IP LAW (CN)
Download PDF: