A compact representation of handwriting movements with mixtures of primitives

This paper concerns the issue of movement primitive extraction of handwriting, which is crucial to many applications with handwriting analysis. In particular, a novel analysis-by-synthesis paradigm is developed for extracting movement primitives from observed handwriting data. The logic behind our method is that complex human movements are the outcomes of the combinations of simple movement primitives. The contribution of this work is twofold. First, a convolutive mixture model with time-delay parameters of primitives is proposed to describe handwriting movements. Second, the problem of extracting handwriting motor primitives is formulated as an unsupervised learning task of nonnegative primitive decompositions with unknown time-delay parameters. A remarkable new feature of our unsupervised learning approach is to identify unknown time-delay parameters while seeking temporal nonnegative primitive decompositions. In doing this, the correlative maximization process for estimating time-delay parameters is combined with the temporal nonnegative primitive decomposition technique in an alternative iteration fashion. The experimental studies are conducted to validate the proposed method.