Multiscale Gaussian Markov Random Fields for writer identification

Writer identification recently has been considerably studied due to its various applications in forensic and commercial sections. Because offline, text-independent writer identification has limited requirements in writing sample collection, it has wider applications and meanwhile more difficult to handle. By considering handwriting images as visually distinctive textures, we propose a new method for offline, text-independent writer identification based on multiscale version of Gaussian Markov Random Fields (GMRF) model. The handwriting features are extracted in wavelet domain of handwriting textures in which global texture feature (such as directional information) from handwriting can be detected. In addition, GMRF is investigated to capture different local spatial structures of graphemes (character-shape) written by different people. The experimental results demonstrate that the proposed method outperforms both 2-D Gabor model and wavelet-based GGD method.