Tactile Sensing with Contextually Guided CNNs: A Semisupervised Approach for Texture Classification

Abstract

Texture classification plays a crucial role in applications ranging from object recognition and product design to surface exploration. Utilizing deep learning methods with sensors, such as accelerometers, offers a way to identify key surface features without the need to precisely replicate human touch. A Contextually Guided Convolutional Neural Network (CG-CNN) employs contextual guidance by developing auxiliary tasks during its training. These tasks offer implicit, yet rigorous, internal supervision signals. When trained with these subtasks, CG-CNN learns to represent the innate structure and patterns within the data, resulting in robust, transferrable, and local/contextual-neighborhood-preserving domain representations. This paper extends the CG-CNN framework for texture classification by integrating semisupervised learning. Empirical evaluations on the VibTac-12 texture dataset reveal that CG-CNN effectively generalizes to novel and unfamiliar textures, even when trained with scarce labeled examples. By harnessing vast amounts of unlabeled, contextually relevant data alongside the labeled samples, CG-CNN ensures robust and precise texture classification. Such advancements hold promise for applications in robotics, prosthetics, and haptic interfaces.