Transfer learning or design a custom CNN for tactile object recognition

Abstract

Novel tactile sensors allow treating pressure lectures as standard images due to its highresolution. Therefore, computer vision algorithms such as Convolutional Neural Networks (CNNs) can be used to identify objects in contact. In this work, a high-resolution tactile sensor has been attached to a robotic end-effector to identify objects in contact. Moreover, two CNNs-based approaches have been tested in an experiment of classification of pressure images. These methods include a transfer learning approach using a pre-trained CNN on an RGB images dataset and a custom-made CNN trained from scratch with tactile information. A comparative study of performance between them has been carried out.