Learn-to-Segment by Operational Object Segmentation Networks 15.9


Project Start Date: Jul 1, 2015
Funding: Member Funded
Project Tags: ,

Project Summary

Object  extraction  in  an  image  is  a  challenging  problem  since  segmenting  an  object  needs  a  combination  of  various  visual  cues  such  as  texture   information,  distribution  of  the  pixel  intensities,  object  boundaries  and  dissimilarity  from  other  objects/regions.  Furthermore,  the  visual  scenery  may  bear   more  than  one  object,  or  partially  occluded  objects.  Unfortunately,  image  segmentation  that  can  be  defined  as  extracting  meaningful  regions  from  an  image  is   an  ill-­‐posed  problem  since  one  has  to  first  define  and  fix  a  priori  what  a  “meaningful  region”  is.  As  a  result,  current  automatic  object  segmentation  algorithms   suffer  from  various  problems  that  make  them  unreliable  and  deficient  for  the  purpose  of  (salient)  object  extraction.  Learn-­‐to-­‐segment  (L2S)  proposes  a  novel   operational  neural  network  (ONN)    in  order  to  improve  the  traditional  deep  learning  neural  networks,  in  which  “convolution”  is  the  only  operator  used.    We   believe  that  extending  the  set  of  operations,  both  linear  and  non-­‐linear,  can  lead  to  improving  the  generalization  capability  of  the  neural  network.  However   ONNs,  like  their  predecessor  CNNs,  will  be  classifiers  that  are  proven  to  localize  the  object(s)  in  an  image  and  can  (learn  to)  extract  the  features  for  the   purpose  of  image  (object)  classification.    Therefore,  we  shall  then  modify  the  ONN  topology,  its  training  and  outcome  so  as  to  serve  to  the  L2S  purpose  to  yield   Operational  (Object)  Segmentation  Networks  (OSNs)  where  the  output  of  a  OSN  will  be  the  saliency  map  of  the  image  from  which  salient  object(s)  can  be   segmented.