Résumé: In-air handwriting is a rapidly emerging human-machine interactive paradigm that helps users to write and communicate naturally and intuitively in free space. In this paper, we develop a hybrid one-dimensional convolutional recurrent attention framework model for in-air handwritten Assamese word recognition (IAHAWR) which associates an encoder and a decoder framework for efficiently recognizing air-written words. The encoder is an assimilation of 1D convolutional neural network and bidirectional gated recurrent unit neural network for input trajectory feature sequence learning, while the decoder is an attention-based gated recurrent unit for predicting the target words. In contrast to conventional pen-based handwriting, in-air handwriting is intricate in the sense that the handwriting is finished in a single continuous stroke giving rise to many irrelevant motions called ligatures in between adjacent character strokes. So, we have imbibed a salient stroke extraction and a critical point detection scheme into the proposed system, which helps in removal of insignificant ligatures thus enhancing the recognition performance. Further, air-writing trajectories contain intermittent jitters and suffer wide variations in writing patterns due to unrestricted writing in free space. So, we incorporate a multistage word normalization methodology which generalizes the air-written patterns and aids in efficient recognition. We have assessed the performance of our proposed system on an air-written Assamese word dataset as well as some air-written Latin words. Experimental evaluation connotes that our proposed IAHAWR system can effectively procure characteristic information from air-writing sequences and provides comparable recognition accuracy and computational performance with that of other state-of-the-art recognition frameworks.

Langue: Anglais