MCMC based Generative Adversarial Networks for Handwritten Numeral Augmentation
Institute of Electrical and Electronics Engineers (IEEE)
© Copyright 2017 IEEE
Reason for embargo
Currently under an indefinite embargo pending publication by IEEE. No embargo required on publication. Add DOI on publication
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 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.
This work was supported by grants from the Chinese Scholarship Council (CSC) program.
This is the author accepted manuscript.
6th International Conference on Communications, Signal Processing, and Systems (CSPS), 14-16 July 2017, Harbin, China