Using a convolutional neural network to construct a pen-type tactile sensor system for roughness recognition

Abstract

The tactile texture of a product is one of many important factors that influences an impression of a product. Numerical evaluation methods for recognizing tactile texture measures, such as roughness, are required because there are indeed individual differences in how users experience tactile textures. In this paper, we propose a tactile sensor to recognize roughness using a convolutional neural network (CNN). Our sensor system consists of a pressure sensor and a six-axis acceleration sensor for detecting time-series data. The sensor measures time-series data, which are the pressure, speed, and posture of the sensor when the sensor is touching an object and is moved by a user. The surface roughness is calculated from these time-series data using a CNN. Here, our system configuration is simple and therefore easy and inexpensive to construct. To evaluate our approach, we constructed a prototype sensor and measured the roughness of six objects. The average correct recognition rate proved to be 71% for the experimental data acquired by one user, which are categorized into learning data and evaluation data. Further, the total average recognition rate for evaluation data by our five users for considering each individual using the sensor system was 42%. While the problem of roughness recognition by each individual user remains, we were able to show the possibility of roughness recognition via our approach. We conclude that our proposed sensor system is useful as a functional and useful device.