Handwriting Text/Non-Text Classification on Mobile Device

This paper is dedicated to classification of handwritten/drawn input made on screen of mobile devices into two classes: Text and Non-Text. A deep-learning solution using gated recurrent and feed-forward artificial neural networks has been proposed. Two approaches have been compared: a real-time approach, designed to process data at input time with preliminary strokes grouping, and a batch processing approach, designed for analysis of completed handwriting documents having access to document contexts and performing text line grouping after classification. The presented solutions have been validated using the benchmark IAMonDo dataset [1] and specially collected Samsung Mobile HandWriting Document (MHWD) dataset, containing about 10 000 free-form documents combining unconstrained handwriting in seven different languages and different heterogeneous elements. The obtained precision by text class is 98.09% and recall by text class is 99.07% for the proposed batch processing approach. The results of the research have become the basis for development of Document Structure Analysis Engine focused on mobile platform and included in Samsung Handwriting Recognition Solution.