A Hierarchical Codebook Descriptor Approach for Online Writer Identification

This work proposes a technique to identify the authorship of an online handwritten document. The strategy focuses on encoding a set of feature vectors obtained from sample points of the online trace with a descriptor, by employing a pair of codebooks. The derived descriptors correspond to scoring each attribute of the feature vector on the basis of their proximity to the respective values of the assigned codevectors. The two codebooks consist of codevectors pre-learnt by a two level k-means clustering applied on the feature vectors derived from a subset of writers. The descriptors are constructed from features that are specified using a 'gap parameter', that capture the neighborhood information of the trace. Moreover, prior to classification with a SVM, we propose a weighting scheme for the descriptors corresponding to the codevectors generated after the second level of clustering. The weights, as such, are computed on the basis of entropy values obtained over a set of generated histograms. Experiments conducted on I AM Online Handwriting Database demonstrate the efficacy of the proposed descriptor.