A connected component-based approach for text-independent writer identification

Writer identification from off-line images of handwriting is a challenging task. In this work we assess the performance of textural descriptors for writer identification on different writing styles. The proposed method is different from the existing texture based methods: the earlier methods extract texture information at page, paragraph, fragment level to get a document descriptor, while the proposed method exploits the texture at a connected-component level. Specifically, an improvement of the textural features, the contour direction, angle, and length distributions are explored. Using these texture descriptors, the occurrence histograms are calculated in order to determine the similarities between different images. For the evaluation, the IFN/ENIT (411 writers) and IAM (657 writers) datasets were used. In our characterization of the individual and combination of textural features of connected-components we show that the proposed method outperforms the state-of-the-art algorithms and archives the best performance.