Non-supervised determination of allograph sub-classes for on-line omni-scriptor handwriting recognition

We present a novel clustering algorithm dedicated to the determination of character allographs. The "problem of the allographs" specific of the dynamic handwriting in omni-scriptor context renders the implementation of "classical" clustering algorithms particularly delicate because it introduces the notion of heterogeneous classes characterized by strongly variable example densities. We propose an hybrid clustering algorithm that combines a prototype placement stage and an adaptation stage. The first realizes an under-optimal determination of kernels in the different clusters composing the classes. It is followed by a kernel adaptation stage driving to an optimization of their position. The process drastically reduces the number of references to examine during a k-nn classification while preserving to the classifier a high level of performances. The experience has been driven on an extensive alphabet comprising 80 classes (upper- and lower-case letters, digits and mathematical symbols). Recognition rate, evaluated on near 35000 examples from the UNIPEN database show the sturdiness of the modelization.