Deep Hybrid Real-Complex-Valued Convolutional Neural Networks for Image Classification

摘要

Shallow complex-valued convolutional neural networks (CVCNNs) have displayed better performance than their real-valued counterparts (RVCNNs). This paper presents a deep CVCNN architecture inspired by the well known VGG architecture. The different structure of learning in the complex domain compared with the real domain means that CVCNNs make systematically different errors than RVCNNs. This led to the idea of a hybrid real-complex-valued ensemble of the two types of networks, which combines the advantages of both. Experiments done on the SVHN, CIFAR-10, and CIFAR- 100 datasets show better results of the CVCNNs compared with RVCNNs, and significantly better results of the hybrid real-complex-valued ensemble compared with both types of networks.