Writer Identification Using a Statistical and Model Based Approach

The state-of-the-art writer identification systems use a variety of different features and techniques in order to identify the writer of the handwritten text. In this paper several statistical and model based features are presented. Specifically, an improvement of a statistical feature, the edge hinge distribution, is attempted. Furthermore, the combination of this feature with a model-based feature is explored, that is based on a codebook of graphemes. For the evaluation, the Fire maker DB was used, which consists of 250 writers, including 4 pages per writer. The best result for the statistical suggested approach, the skeleton hinge distribution, achieved accuracy of 90.8%, while the combination of this method with the codebook of graphemes reached 96%.