Seite - 120 - in Document Image Processing

J. Imaging 2018,4, 43 Figure22.BinarizationofBalinesemanuscriptwith ICFHRG2method. 5.2. TextLineSegmentation Theexperimental results for text linesegmentationtaskarepresentedinTable6.Accordingto theseresults,bothmethodsperformsufficientlywell formostdatasets,exceptKhmer1(Figures23–25). This is becauseall images in this set areof lowqualitydue to the fact that theyaredigitized from microfilms. Nevertheless, the adaptivepathfindingmethodachievesbetter results than the seam carvingmethodon all datasets of palm leafmanuscripts in our experiment. Themaindifference between these twoapproaches is that insteadoffindinganoptimal separatingpathwithinanarea constrainedbymedial seamlocationsof twoadjacent lines (in theseamcarvingmethod), theadaptive pathfinding approach tries to findapath close to an estimated straight seam line section. These linesectionsalreadyrepresentwell theseambordersbetweentwoneighboring lines, so theycanbe consideredabetterguide forfindinggoodpaths,henceproducingbetter results. Onecommonerror thatweencounter forbothmethods is in themedialpositioncomputation stage. Detecting correctmedial positions of text lines is crucial for the path-finding stage of the methods. Inourexperiment,wenoticedthatsomeparametersplayanimportant role. For instance, thenumberofcolumns/slices rof theseamcarvingmethodandthehighandlowthresholdingvalues of theedgedetectionalgorithmintheadaptivepathfindingapproachare important. Inorder toselect theseparameters,avalidationsetconsistingoffiverandompages isused. Theoptimalvaluesof the parametersare thenempiricallyselectedbasedontheresults fromthisvalidationset. Table6.Experimental results for text linesegmentationtask: thecountofgroundtruthelements (N), andthecountof resultelements (M), theone-to-one(o2o)matchscore iscomputedforaregionpair basedon90%acceptance threshold,detectionrate (DR), recognitionaccuracy(RA),andperformance metric (FM). Methods Manuscripts N M o2o DR(%) RA(%) FM(%) Seamcarving[47] Balinese1 140 167 128 91.42 76.64 83.38 Bali-2.1 181 210 163 90.05 77.61 83.37 Bali-2.2 182 219 161 88.46 73.51 80.29 Khmer1 191 145 57 29.84 39.31 33.92 Khmer2 476 665 356 53.53 74.79 62.40 Khmer3 971 1046 845 87.02 80.78 83.78 Sundanese 1 46 43 36 78.26 83.72 80.89 Sundanese 2 242 257 218 90.08 84.82 87.37 AdaptivePathFinding[27] Balinese1 140 143 132 94.28 92.30 93.28 Bali-2.1 181 188 159 87.84 84.57 86.17 Bali-2.2 182 191 164 90.10 85.86 87.93 Khmer1 191 169 118 61.78 69.82 65.55 Khmer2 476 484 446 92.15 93.70 92.92 Khmer3 971 990 910 93.71 91.91 92.80 Sundanese 1 46 50 41 89.13 82.00 85.41 Sundanese 2 242 253 222 91.73 87.74 89.69 120