Spiking neural network for road signs classification

Abstract

In recent years, with the rapid development of big data technology and parallel computing technology, deep networks play an active role in image tasks. But the neurons of this deep network are based on real-valued computation, and backpropagation is used for network learning. As the network depth increases, additional computational cost and storage resources are required, limiting the practical application of truly deep networks. As a new type of modern artificial neural network, Spike neural networks is inspired by brain science, deeply simulates the dynamic characteristics of biological neurons, and is currently the most biologically reliable neural network. At present, the network structure of most SNN models is relatively simple, and the network depth is not large enough to handle more complex tasks. In response to this problem, this project proposes a deep SNN model using the DenseNet structure and the time-involved batch normalization method. In order to study the application of the SNN model to the field of autonomous vehicles, this project evaluates the classification ability of the model for different road signs. The results show that without any network pre-training, the established SNN model can be directly trained and achieve significant classification results. Moreover, this project proves that the proposed SNN model has good generality by means of transfer learning.