Surface Texture Classification Based on Transformer Network

An acceleration signal is generated when a person interacts with the surface of an object, which carries pertinent information about the surface material. This acceleration signal is unique to each surface and can be used to recognize the surface texture of an object. In this paper we developed a new transformer-based deep learning model for surface texture classification from haptic data. This approach leverages the self-attention process to learn the complex patterns and dynamics of time-series data. To the best of our knowledge this is the first time that the transformer or its variants are used for surface texture classification using tactile information. As a proof of concept, we collected data for 9 different textures and the evaluation experiments showed that the model achieved state-of-the-art classification accuracy.

Download paper here