A Novel Multi-Type Image Coding Method Acting on Supervised Hierarchical Deep Spiking Convolutional Neural Networks for Image Classification

Spiking neural networks (SNNs) have gained significant momentum in recent times as they transmit information via discrete spikes, similar to neuromorphic low-power systems. However, existing spike coding methods are often limited to a single scale of time or rate, and typically suffer from drawbacks such as reduced accuracy or long classification latency. In this paper, we propose a pixel-based multi-type image coding (PMIC) method inspired by the functional organization of primate visual systems to address the issues at hand. The encoded information comprises both spatial and temporal details, represented by spiking firing time and intensity, respectively. Subsequently, we combine the spiking firing time and intensity as inputs of a hierarchical spiking convolutional neural network (SCNN) including several convolutional and pooling layers. During the training phase, we use error backpropagation to optimize parameters. Comparison of experimental results with some state-of-the-art approaches on MNIST dataset, Fashion-MNIST dataset as well as ETH-80 dataset of image classification demonstrates that SCNN using PMIC can achieve the best test accuracy, which is 99.13%, 90.31%, and 94.29%, respectively. The proposed PMIC utilizes multiple filters and coding strategies to extract multi-type information and is more beneficial to the performance of SNNs compared to methods that extract single-scale or single-type information.