ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper Methods for Extracting Robust Features using Large Network Simulation for Face Recognition
Cited - time in scopus Share share facebook twitter linkedin kakaostory
Authors
Jaeyoon Jang, Hosub Yoon, Jaehong Kim, Minsu Jang
Issue Date
2019-06
Citation
International Conference on Ubiquitous Robots (UR) 2019, pp.1-2
Publisher
IEEE
Language
English
Type
Conference Paper
Abstract
In case of inputting a face image smaller than the input size of the recognition model due to factors such as distance in identity recognition problem, recognition performance is deteriorated because virtual information generated using interpolation is used. To solve this problem, we propose a new loss to maintain the quality of the feature vector. We design a network that uses a smaller size image than the existing recognition model, and use the feature quality loss (F.Q. loss) together with the general cross entropy loss in the learning stage. F.Q. loss is a loss designed to take advantage of the difference between a computed feature vector and a computed feature vector in a model receiving small inputs from a prelearned recognition model, which helps the model receiving a small input simulate the discernment of the model receiving a large input. We could experimentally confirm that applying F.Q. loss would maintain perceived performance even though the input size was halved compared to a decrease in perceived performance if only the input size was reduced without using the proposed loss.
KSP Keywords
Cross entropy, Entropy loss, Face image, Feature Vector, Identity Recognition, Large network, Learning Stage, Network Simulation, Quality Loss, Recognition model, Recognition performance