Open AccessFace recognition with compressed Fisher vector on multiscale convolutional features
Author(s) -
Deng Weihong,
Wang Hongjun
Publication year - 2018
Publication title -
iet biometrics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.434
H-Index - 28
eISSN - 2047-4946
pISSN - 2047-4938
DOI - 10.1049/iet-bmt.2017.0194
Subject(s) - discriminative model , computer science , pattern recognition (psychology) , artificial intelligence , dimension (graph theory) , representation (politics) , convolutional neural network , facial recognition system , face (sociological concept) , subspace topology , speech recognition , mathematics , social science , politics , sociology , political science , pure mathematics , law
Representations generated by Fisher vector (FV) have shown decent performances on many facial image datasets. However, discriminative information could be masked by noise if the authors directly sum all local responses with respect to the learned dictionary. Further, the high dimension of FV prohibits its practical use. To mitigate these problems, the authors propose a new framework called joint compressed Fisher vector (JCFV), which generate task‐specific data representation by jointly encoding multiscale deep convolutional activations. Firstly, they feed into the deep network facial images cropped with cascaded sub‐windows and resized into various scales. Next, they select discriminative convolutional features to form a dictionary. Then, they aggregate multiscale features with respect to the dictionary by calculating a re‐weighted first‐order statistics. JCFV halves the dimension of FV, and they could further compress the dimension with several combinations of subspace methods. They prove the effectiveness of their JCFV descriptor with comprehensive experiments on FERET, AR, LFW and FRGC 2.0 Experiment 4.