Seite - 36 - in Document Image Processing

J. Imaging 2018,4, 80 Themeasureusedtoevaluate the technique is theroot-mean-squareerror (RMSE) RMSE= √√√√√ N∑d=1 ( Slantdgr−Slantdes )2 N (6) whereSlantgr isthegroundtruthslantofdocumentdandSlantes theslantestimatedusingthetechnique. Thismeasuregives a comparative result that is independent of the right or left slantdirection. N refers to the amount of documents. Next, a short descriptionof thedatabases is given,while the setupontheparameters follows.As initialparametervalues, theparametersusedin[17]areusedin ourexperiments,andassoonas thebestparametervalue isestimated, it isusedfurtheron. Finally, experimental results for the fourdatabasesarepresented(seeSection3.9). 3.1. TrigraphSlantDB TheTrigraphSlant [18]database contains imagesofhandwritingproducedundernormaland forcedslantconditions. It includes190handwrittendocument images,writtenby47people. Foreach image, the slant has been estimated by two researchers (Axel and Rolland) from the average slant computed from10measurements on eachdocument image. In Figure 6, an example of the TrigraphSlantdatabaseafter theapplicationof theproposedtechnique isshown. Figure6.Applicationof the techniqueonasample fromtheTrigraphSlantdatabase (left); correctedby +28degrees (right). 3.2.GeorgeWashingtonDB This archive contains a set of 20 page images from theGeorgeWashington collection [19] at theLibraryofCongress in theUnitedStates. Aprocess similar to that used for theTrigraphSlant databasewas followed. Ten slantsweremeasuredby twohumansoneachpageand themeanof thesemeasurementswasconsideredtobethepageslant. Figure7showsanexampleof theGeorge WashingtonDBafter theapplicationof the technique. 36