Home // International Journal On Advances in Networks and Services, volume 13, numbers 3 and 4, 2020 // View article

Machine Learning-based Classification and Generation of Vibrotactile Information

Authors:
Satoshi Saga
Shotaro Agatsuma
Simona Vasilache
Shin Takahashi

Keywords: Tactile Information; Machine Learning; Convolutional Neural Network; Generative Adversarial Network

Abstract:
In the field of tactile displays, many researchers are developing systems that employ recorded tactile information as an input signal for tactile display. Various kinds of tactile information have been recorded from real textures and used to present high- quality tactile sensations via the displays. However, collecting, classifying and generating large amounts of tactile information data under many different conditions with complicated sensors are difficult to realize. Thus, we developed a method of collecting accelerations in haptic behaviors using wireless microcomput- ers and implemented a Convolutional Neural Network-based classification method of tactile information. We had succeeded in classifying 30 types of data with an accuracy of 88.9%. Furthermore, we proposed to generate unrecorded data under various conditions from recorded data. We construct a data gener- ation model using a Generative Adversarial Network. The model generates unrecorded three-axis, time-series acceleration data from recorded acceleration data obtained by stroking real objects. To evaluate the quality of the data generated, we presented generated vibrotactile information to users via a tactile display. We revealed that the generated data were indistinguishable from real data. Besides, by mixing and generating data of two or more classes, we generated unrecorded data that have mixed features of the original classes.

Pages: 115 to 124

Copyright: Copyright (c) to authors, 2020. Used with permission.

Publication date: December 30, 2020

Published in: journal

ISSN: 1942-2644