Hermite Convolutional Networks

摘要

Convolutional Neuronal Networks (CNNs) have become a fundamental methodology in Computer Vision, specifically in image classification and object detection tasks. Artificial Intelligence has focused much of its efforts in the different research areas of CNN. Recent research has demonstrated that providing CNNs with a priori knowledge helps them improve their performance while reduce the number of parameters and computing time. On the other hand, the Hermite transform is a useful mathematical tool that extracts relevant image features useful for classification task. This paper presents a novel approach to combine CNNs with the Hermite transform, namely, Hermite Convolutional Networks (HCN). Furthermore, the proposed HCNs keep the advantages of CNN while leading to a more compact deep learning model without losing a high feature representation capacity.