A learning-based method to detect and segment text from scene images

(整期优先)网络出版时间:2007-04-14
/ 1
Thispaperproposesalearning-basedmethodfortextdetectionandtextsegmentationinnaturalsceneimages.First,theinputimageisdecomposedintomultipleconnected-components(CCs)byNiblackclusteringalgorithm.ThenalltheCCsincludingtextCCsandnon-textCCsareverifiedontheirtextfeaturesbya2-stageclassificationmodule,wheremostnon-textCCsarediscardedbyanattentionalcascadeclassifierandremainingCCsarefurtherverifiedbyanSVM.AlltheacceptedCCsareoutputtoresultintextonlybinaryimage.Experimentswithmanyimagesindifferentscenesshowedsatisfactoryperformanceofourproposedmethod.