In image recognition, a variety of elasticity, rotation and shift of image affects recognition performance seriously. To solve this problem, several methods based on 2D-HMM have been proposed. However, since the structure of 2D-HMM is more complex than that of 1D-HMM, high computational cost is required to calculate the likelihood and estimate the model parameters. In this paper, we define separable 2D-HNINI which consists of two state sequences: vertical and horizontal sequences, and propose a training algorithm for separable 2D-HMM based on variational approximation. The proposed method can reduce the computational complexity as compared to N-best approximation and the lower bound of log-likelihood is guaranteed to increase at each iteration of the training algorithm.
展开▼