Image Recognition Based on Convolutional Neural Networks Using Features Generated from Separable Lattice Hidden Markov Models
Takayuki Kasugai, Yoshinari Tsuzuki, Kei Sawada, Kei Hashimoto, Keiichiro Oura, Yoshihiko Nankaku, Keiichi Tokuda · 2018
An image recognition method based on convolutional neural networks (CNNs) using features generated from separable lattice hidden Markov models (SLHMMs) is proposed. A major problem in image recognition is that the recognition performance is degraded by geometric variations in the size and position of the object to be recognized. To solve this problem, SLHMMs have been proposed as an extension of HMMs with size and locational invariances based on state transitions. Although SLHMMs are generative models that can represent the generation processes of observations well, there is a possibility that they are not specialized for discrimination compared to discriminative models. Our method integrates SLHMMs that extract features invariant to geometric variations with CNNs that build an accurate classifier based on discriminative models with the extracted features. Face recognition experiments showed that the proposed method improves recognition performance.