MCMC Based Generative Adversarial Networks for Handwritten Numeral Augmentation

Abstract

In this paper, we propose a novel data augmentation framework for handwritten numerals by incorporating the probabilistic learning and the generative adversarial learning. First, we simply transform numeral images from spatial space into vector space. The Gaussian based Markov probabilistic model is then developed for simulating synthetic numeral vectors given limited handwritten samples. Next, the simulated data are used to pre-train the generative adversarial networks (GANs), which initializes their parameters to fit the general distribution of numeral features. Finally, we adopt the real handwritten numerals to fine-tune the GANs, which greatly increases the authenticity of generated numeral samples. In this case, the outputs of the GANs can be employed to augment original numeral datasets for training the follow-up inference models. Considering that all simulation and augmentation are operated in 1-D vector space, the proposed augmentation framework is more computationally efficient than those based on 2-D images. Extensive experimental results demonstrate that our proposed augmentation framework achieves improved recognition accuracy.