The following tutorials will take place on Monday, 1st September 2014.
Handwritten Text Recognition: Word-Graphs, Keyword Spotting and Computer Assisted Transcription
Moisés Pastor, Verónica Romero, Joan Andreu Sánchez, Alejandro H. Toselli and Enrique Vidal
Current technology for Handwriting Text Recognition (HTR) is based on fully automatically trainable models such as Hidden Markov Models and N-gram Language Models. This technology typically yields good (but not perfect) transcription results. For applications where this is not enough, HTR techniques may still be useful as a front-end. This is particularly the case for two important applications; namely, Keyword Spotting (KWS) and Computer Assisted Transcription of Text Images (CATTI). Both applications can be advantageously dealt with using by-products of the HTR process known as “Word Graphs” (WG).
In KWS, WGs are used to eﬃciently compute very accurate scores, which can are useful to index and search large image collections under the precision-recall tradeoﬀ model. On the other hand, CATTI is useful when perfect transcripts are needed. This often requires heavy human intervention to “post-edit” the automatic transcripts. CATTI uses WGs to implement an Interactive-Predictive model, where human corrective feedback is used to anticipate further corrections, thereby reducing the editing user eﬀort.
This one-day tutorial will ﬁrst present basic HTR concepts and methods, followed by hand-on practical exercises using well known HTR software. Finally, applications to KWS and CATTI will be presented including demonstration of real systems. All the supporting software and documentation will be open-source and freely available for the participants.
Further information about this tutorial, including slides, practical, exercises, etc. can be found here.
Automatic signature verification: state of the art and recent trends
Angelo Marcelli, Guiseppe Pirlo, Marcus Liwicki, Michael Blumenstein
Signature verification is possibly the field of handwriting recognition that, in the short time, may lead to several applications in different domains, such as office automation and forensic examination of documents, as well as biometric identification. Furthermore, the widespread adoption of devices for on-line signature acquisition resulting from both national and international Governmental efforts demands online handwriting technologies, including signature verification, that can be implemented on hand held device.
The tutorial aims at introducing the state of the art and the open issues in the field of automatic signature verification for both on-line and offline data. PhD students and researchers working in the academia will to get a comprehensive state of the art and suggestions on the most promising approach to address open issues, while people working with company that are already dealing with or are considering to introduce such a technology will have better understanding of the factors determining the impact of their introduction and their expected and actual performance. The ultimate aim of the tutorial will be that of providing the audience a careful and updated review of the state of the art, outlining future direction and trends for encouraging new research in the field, as well as raising in the audience interested in considering the deployment of such a technology within an organization the relevance of the performance evaluation topic.
There are no particular requirements for attending the tutorials, since all the concepts and terms used will be defined. Of course, a background in computer science, and particularly pattern recognition, will allow to get the most out of the presentations.
Statistical Models for Handwriting Recognition and Retrieval
Gernot A. Fink
In this tutorial the foundations of two important pattern recognition techniques, namely Hidden Markov Models (HMMs) and of Bag-of-Features (BoF) representations based on local image descriptors, and their application to handwriting recognition and retrieval will be presented.
Starting from the general architecture of an HMM-based handwriting recognition system, the theoretical concepts behind HMMs as well as the important algorithms necessary for applying this technique in practice will be presented. In order to give the audience a better understanding of how these techniques are applied to real-world problems, it will be shown how HMM-based methods are embedded in real handwriting recognition frameworks.
In the second part of the tutorial, the principle of word spotting will be introduced as a tool for analyzing documents without transcription. It will be shown, how the bag-of-words principle known from statistical text modeling can be transferred to the domain of image processing leading to Bag-of-Features representations. Afterwards, segmentation-free word spotting techniques will be covered which heavily rely on advanced BoF models. As both BoF models and HMMs are statistical models of patterns, both techniques can be combined and it will be shown how these so-called BoF-HMMs can be used to build state-of-the-art word spotting systems for historical manuscripts.
The tutorial slides can be downloaded from here.