Publication
RECOGNITION OF WHITEBOARD NOTES - Online, Offline and Combination
Marcus Liwicki; Horst Bunke
Series in Machine Perception and Artificial Intelligence, Vol. 71, World Scientific, 2008.
Abstract
This book addresses the issue of processing online handwritten notes acquired from an electronic whiteboard. Notes written on a whiteboard is a new modality in handwriting recognition research that has received relatively little attention in the past. The main motivation for this book is smart meeting room applications, where not only speech and video data of a meeting are recorded, but also notes written on a whiteboard are captured. The aim of a smart meeting room is to automate standard tasks usually performed by humans in a meeting. In order to allow for retrieval of the meeting data by means of a browser, semantic information needs to be extracted from the raw sensory data. The main achievements of this book can be summarized as follows. A new online handwritten database has been compiled, and four handwriting recognition systems have been developed. These are an offline and an online recognition system, a system combining offline and online data, and a writer-dependent recognition system. The online recognition system includes novel preprocessing and normalization strategies which have been developed especially for whiteboard notes. A novel classification strategy based on bidirectional long short-term memory networks has been applied for the first time in the field of handwriting recognition. In the combination experiments both the offline and online system were integrated into a single recognizer. To the best of the authors' knowledge these are the first experiments in the field of online sentence recognition combining systems based on offline and online features. Furthermore, external recognition systems were included in the combination experiments. The experimental results on the test set show a highly significant improvement of the recognition performance over the individual systems. The optimal combination achieved a word level accuracy of more than 86,%, implying a relative error reduction of about 26,%, compared to the best individual classifier.