紀要論文

視線停留時間を活用した顔魅力評定システムの効率的表現に関する研究

A STUDY ON EFFICIENT REPRESENTATION OF FACIAL ATTRACTIVENESS ASSESSEMENT SYSTEM UTILIZING GAZE FIXATION TIME

jpn

24368083

https://doi.org/10.15002/00030728

法政大学大学院理工学研究科

法政大学大学院紀要. 理工学研究科編

65

1

6

2024-03-24

Version of Record

In facial attractiveness assessment, there are significant individual differences in gaze behavior of evaluators. Thus, it is crucial to consider these individual differences in the machine learning models that assess facial attractiveness based on gaze fixation time. One of the simplest approach to construct this system is to construct individual softmax classifiers for each evaluator. However, this approach cannot utilize any insights from other evaluators for constructing individual softmax classifiers for new evaluators. This study aims to achieve a model that can utilize insights from other evaluators while considering individual differences among evaluators. To accomplish this, we propose an efficient representation of all parameter matrices of indivisual softmax classifiers and reveal the geometric features common to all parameter matrices. First, we propose a model assuming that all the parameters of the indivisual softmax classifiers are given by the linear combination of a common set of basis matrices. Next, we apply this proposed model to real facial attractiveness assessment data and confirm numerically that it achieves high accuracy in estimating attractiveness ratings, exceeding 90%, even when using fewer basis matrices than the number of evaluators. The parameter matrices of the indivisual softmax classifier generated by the proposed model are localized in a narrow subspace spanned by a common set of few basis matrices. Based on this geometric insight, we propose a transfer learning method that utilizes pre-trained basis matrices for constructing softmax classifiers for new evaluators. Finally, through numerical experiments, we confirm that this transfer learning method is highly effective in reducing the time required to construct softmax classifiers.

Article