Thispaperfocusesontheproblemofautomaticimageclassification(AIC)byproposingaframeworkbasedonlatentsemanticanalysis(LSA)andimageregionpairs.Thenovelframeworkemploysrelativespatialarrangementsforregionpairsastheprimaryfeaturetocapturesemantics.Thesignificanceofthispaperistwofold.Firstly,tothebestourknowledge,thisisthefirststudyoftheinfluenceofregionpairsaswellastheirrelativespatialinformationinlatentsemanticanalysisasappliedtoautomaticimageclassification.Secondly,ourproposedmethodforusingtherelativespatialinformationofregionpairsshowgreatpromiseinimprovingimagesemanticclassificationcomparedwiththeclassicallatentsemanticanalysismethodand2Dstringrepresentationalgorithm.
