Writer verification using directional morphological features and Markov chains

Much of the information about the writer’s style lies in the directionality of the lines that make up the letters of a word. This information, in order to be usable, must be reliably extracted. Directional Morphological Structuring Elements (SE), capture the inherent line information of a word and represent the alternation of this information along the word using Markov chains models. Thus, running along a word, the orientation of the best fitting SE changes according to the orientation of the lines of the word. This SE represents a single specific strip of the word. The directional SEs are considered as being the states of a Markov chain model that change along the word. The transition matrices of the Markov chain models constitute the features for writer verification. Using 3-pixel and 5-pixel long SEs it turns out that the cluster of each writer is well separated in the feature space, while each cluster presents very small dispersion. This cluster distribution results in a clear separation of the writers even using only a short word with 100% success in writer identification. The method is language independent, and its performance assessment is carried out in case of long words as well as in case of short words. Two different handwriting databases are employed. The verification problem can easily be faced using a simple distance metric in the reduced feature space.