Page - 118 - in Document Image Processing

J. Imaging 2018,4, 43 4.4.2. EvaluationMethod Theerror rate isdefinedbyeditdistancesbetweengroundtruthandrecognizeroutputand is computedusingtheprovidedOCRopyfunctionocropus-errs (https://github.com/tmbdev/ocropy/ blob/master/ocropus-errs) [56]. 5. ExperimentalResultsandDiscussion In this section, the performance of eachmethod for theDIA tasks on palm leafmanuscript collections ispresented. 5.1. Binarization Theexperimental results for thebinarization taskarepresented inTable5. Theseresults show that theperformanceofallmethodsoneachdataset is still quite low.Mostof themethodsachieve less thana50%FMscore. Thismeans thatpalmleafmanuscriptsarestill anopenchallengefor the binarizationtask. Thedifferentparametervalues for the localadaptivebinarizationmethodsshow significant improvement inperformance,but still giveunsatisfactoryresults. In theseexperiments, the ICFHRG1methodwasevaluatedfor theKhmerandSundanesedatasetsusing thepre-trained Balinese trainingsetweightedmodel. Basedontheseexperiments,Niblack’smethodgives thehighest FMscore forSundanesemanuscripts (Figure20), ICFHRG1methodgives thehighestFMscore for Khmermanuscripts (Figure21),andICFHRG2gives thehighestFMscore forBalinesemanuscripts (Figure22).However,visually, therearestillmanybrokenandunrecognizablecharacters/glyphs,and noise isdetected in the images. Figure20.BinarizationofSundanesemanuscriptwithNiblack’smethod. Figure21.BinarizationofKhmermanuscriptwith ICFHRG1method. 118