Page - 78 - in Document Image Processing

J. Imaging 2018,4, 37 ..... Query Retrieved Results Fast DTW with all database images Mean vector for query Query specific principal alignments Class1 Class2 Class3 ClassN Mean vectors of frequent classes Best Matching Cut Portions Fast DTW Matching using cut−specific principal alignments ..... Figure3.OverallSchemeforNNDQC. Inanofflinephase, themeanvectors for the frequentword classesare computedand their cut-specificprincipal alignmentsare computed. Toprocessaquery word image, it isdivided intocutportionsandFastDTWmatching isused toget thebestmatching cut-portions fromthefrequentclassmeanvectorswith thecut-portionsof thequery image. Thesebest matchingcut-portionsareusedtoconstruct themeanvector for thequeryclassandthequeryspecific principal alignments. FastDTW [20]matchingbetween thequery imageand thedatabase images is doneusingthequeryspecificprincipalalignments. )UHTXHQW &ODVV )UHTXHQW &ODVV L )UHTXHQW &ODVV M )UHTXHQW &ODVV Q &XW VSHFLILF 3ULQFLSDO $OLJQPHQWV 4XHU\ 6SHFLILF 3ULQFLSDO $OLJQPHQWV (a) (b) Figure 4. Synthesis of query specific principal alignments. (a) Cut specific principal alignments corresponding to “ground” and “leather” are joined to form the principal alignments for “great”. Note that theappropriatecutportionsareautomatically found. (b) Inageneral setting,queryspecific principalalignmentsgets formedfrommultipleconstituentcutspecificprincipalalignmentscomputed for frequentclasses. Toensurewiderapplicabilityofourapproach,weconsider that thealignments trainedonone datasetmaynotworkwellonanotherdataset. This ismainlydue to theprintandstylevariations. Foradaptingtodifferentstyles,weusequeryexpansion (QE),apopularapproach in the information retrievaldomain inwhich thequery is reformulated to further improve the retrievalperformance. An index is built over thegiven samplevectors fromthedatabase andusingapproximatenearest neighbor search, the top10 similarvectors to thegivenqueryare computed. These top10 similar vectorsare thenaveragedtoget thenewreformulatedquery. This reformulatedquery isexpectedto bettercapture thevariations in thequeryclass. Inourexperiments, this further improves theretrieval performance.ApproximatenearestneighborsareobtainedusingFLANN[29]. 5.ResultsandDiscussions In this section, we validate the DQC classifier using query specific Fast DTW distance for efficient indexing onmultipleword image collections and also demonstrate its quantitative and qualitativeperformance. 78