Feature's Selection-Based Shape Complexity for Writer Identification Task

Writer Identification task has attracted a lot of research interests due to its wide variety of applications. Different approaches based on various features exist in the literature. However, all these approaches use all the information available in the handwritten sample to identify the writer (relevant or irrelevant). In this paper, we propose an original approach based on a double feature selection process, where the features are represented by graphemes resulting from a segmentation process. These features are analyzed based on their shape complexity, using the Fourier Elliptic transform, and the complexity score is assigned to each grapheme (FECS). The second phase of feature selection is to eliminate the redundancy among the resulting using a sequential clustering algorithm. Two similarity measures are proposed to evaluate the proposed system on 100 writers of the IAM dataset. We obtained a good identification rate of 96% using only 25 graphemes, which is equivalent to 3--4 words.