Views
3 years ago

Download the pdf - Committee to Protect Journalists

Download the pdf - Committee to Protect Journalists

Download the pdf - Committee to Protect

2009 10th International Conference on Document Analysis and RecognitionA Set of Chain Code Based Features for Writer RecognitionImran Siddiqi, Nicole VincentParis Descartes University, Laboratoire CRIP5 – SIP45, rue des Saints-Pères, 75006 Paris, France{siddiqi, nicole.vincent}@math-info.univ-paris5.frAbstractThis communication presents an effective method forwriter recognition in handwritten documents. We haveintroduced a set of features that are extracted from thecontours of handwritten images at differentobservation levels. At the global level, we extract thehistograms of the chain code, the first and secondorder differential chain codes and, the histogram of thecurvature indices at each point of the contour ofhandwriting. At the local level, the handwritten text isdivided into a large number of small adaptive windowsand within each window the contribution of each of theeight directions (and their differentials) is counted inthe corresponding histograms. Two writings are thencompared by computing the distances between theirrespective histograms. The system trained and testedon two different data sets of 650 and 225 writersrespectively, exhibited promising results on writeridentification and verification.1. IntroductionThe need to recognize the writer of a handwrittendocument is a recurrent problem not only from theperspective of behavioral biometrics [2,10,14] but alsoin the context of handwriting recognition [9] exploitingthe principle of adaptation of the system to the type ofwriter. Writer recognition is generally distinguishedinto writer identification and verification. Writeridentification involves a one-to-many search where,given a document of an unknown authorship, theobjective is to find its author in a reference base withdocuments of known writers. Writer verification on theother hand is a one-to-one comparison where, giventwo handwriting samples, one would like to determinewhether the two samples have been written by the sameperson or not.The early research in writer recognition has mainlywitnessed the text-dependent methods where the twowriting samples to be compared require to contain thesame fixed text for example; signature verification. Afew relatively recent studies [14,15] also present textdependentwriter identification systems. The textindependentmethods on the other hand identify thewriter of a document independent of its semanticcontent thus they are less constrained and more usefulfor practical applications. Another traditionalclassification of writer recognition methods is intoglobal and local approaches. The global methods [3,11]are based on the overall look and feel of the writingwhereas the local techniques [2,4] identify the writerbased on localized features, which are inherent in theway a writer specifically writes characters. The latesttrend in writer recognition is to use a set of patterns towhich the actual writing is compared [2,13].Combining the global and local features is also knownto improve the writer recognition performance [5,14]and our research is inspired by the same idea.We present a system for offline writer recognitionusing very simple features as recognition of the authorcan be done by human very instinctively. Human ismostly sensitive to contours and changes so we workon the contours of handwritten text images. We startwith a global analysis of handwriting using theclassically known histograms of chain code and theirdifferential forms. We then propose their local variantsthat are calculated from small segments of handwrittentext. Finally we perform a comparative evaluation ofthe two and explore their various combinations. Themethod has been detailed in the sections to follow.2. Feature ExtractionIn this section we present the proposed features andtheir extraction methods. Based on the hypothesis thatthe contour of a handwritten sample encapsulates thewriting style of its author, we introduce a number offeatures that are based on the contour of thehandwritten text images. We have chosen contour978-0-7695-3725-2/09 $25.00 © 2009 IEEEDOI 10.1109/ICDAR.2009.136981

Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Download the pdf - Committee to Protect Journalists
Read the magazine online (PDF) - Committee to Protect Journalists
annual report - Committee to Protect Journalists
Attacks on the Press - Committee to Protect Journalists
annual report 2011 - Committee to Protect Journalists
international press freedom - Committee to Protect Journalists
Spring/Summer 2005 - Committee to Protect Journalists
CPJ's 2012 Impunity Index - Committee to Protect Journalists
10 Most Censored Countries - Committee to Protect Journalists
After the Black Spring - Committee to Protect Journalists
Turkey's Press Freedom Crisis - Committee to Protect Journalists
Attacks on the Press in 2010 - Committee to Protect Journalists
CPJ Annual Report 2012 - Committee to Protect Journalists
Getting Away With Murder - Committee to Protect Journalists
Turkey's Press Freedom Crisis - Committee to Protect Journalists
Download PDF - International Center for Journalists
Download PDF - International Center for Journalists
Download PDF - International Center for Journalists
Download PDF - International Center for Journalists
Download PDF - International Center for Journalists
Download PDF - International Center for Journalists
PDF Download - Society of Environmental Journalists
Download PDF - International Center for Journalists