Object Tactile Classification Based on Convolutional Spiking Neural Network with Attention Mechanism

Tactile sensation is an important way for humans to perceive information about the surrounding environment. For robots, the addition of tactile sensing provides more diverse information, which helps robots to realize the functions such as object grasping and recognition. Powerful and efficient tactile classification ability helps the robot to better complete tasks. However, current tactile classification algorithms are limited in accuracy and classification speed. Inspired by the brain-inspired computing, we use a tactile pressure classification algorithm based on the convolutional spiking neural network. Our method includes the convolutional structure and the attention mechanism module, extracting tactile pressure features and balancing the importance of different features respectively. Our method uses all-or-none spikes to encode pressure information in convolutional layers and fully connected layers. We collect tactile pressure data of the 35 objects using flexible thin film piezo-resistive sensors. The classification results show that the classification accuracy approaches 99%, and the speed of classifying a sample is about 23ms. The results of comparisons demonstrate the excellent comprehensive performance of our method. Our work provides a high-speed, accurate identification algorithm for the classification task based on tactile pressure and contributes to the investigation of haptic perception and neurorobotic framework during object classification task.