What Should We Be Comparing for Writer Identification?

Certain approaches to writer identification encode handwriting as texture, producing a single histogram of visual features, ignoring any information about the lexical content of the passage. In contrast, other approaches first segment elements of the text, such as characters or big rams, so that they can be compared like-for-like with other instances of the same element. The difference between the two types of methods can therefore be viewed in terms of what is being compared. In this work we assess the performance of the different approaches as well as exploring the performance of a novel method that lies in between the two. We show that this novel method outperforms the other schemes, achieving an error rate of 0.4% when discriminating samples of 20 characters among 251 writers. This suggests that optimal performance may be obtained by making comparisons between groups of characters, or sub textures, within a passage.