A Deep Learning Framework with Histogram Features for Online Writer Identification

This paper proposes a technique to identify the authorship of an online handwritten text using descriptors inspired by the Pattern of Local Gravitational Force (PLGF) from the area of computer vision. The strategy strives to leverage the spatial and temporal information in an online handwritten trace. A noteworthy aspect is that the spatial descriptors are based on histograms that are generated by incorporating the pressure and velocity information encapsulated in the sub-strokes of a writer. Thereafter, we utilize sequence modelling techniques to capture the temporal information present in the sequence of descriptors obtained from consecutive sub-strokes. By considering a sub-stroke as a basic temporal unit of a writer's trace, we propose a Long Short Term Memory (LSTM) autoencoder to compress the sequence data. The compressed representation is then used to train a Support Vector Machine (SVM) to obtain the final authorship of the handwritten text. The efficacy of our proposed approach is tested on the BIT CASIA Database. We achieve state-of-the art results at the text-line level.