Porn　image　detection　has　been　treated　as　a　binary　classification　task　in　previous　learning　based　studies.　In　binary　classification　methods,　the　training　samples　cannot　be　collected　broadly　enough,　which　will　limit　the　generalization　ability　of　the　classifiers.　This　paper　seeks　to　solve　this　problem　with　one-class　method.　First,　BoW　model　mixed　with　skin　color　detection　is　used　for　feature　extraction　and　image　representation.　Then　we　train　a　one-class　visual　dictionary　from　pornographic　image　training　set.　Third,　to　make　BoW　model　suitable　for　one-class　classification,　we　employ　random　forests　to　select　important　variables　and　optimize　the　original　BoW　vectors.　At　last,　we　train　a　classifier　using　one-class　SVM　with　unlabeled　data　(optimized　image　representation　vectors)　of　pornographic　images　training　set.　As　far　as　we　are　aware,　it　is　the　first　attempt　to　treat　pornographic　image　detection　as　a　one-class　classification　task.　Experimental　results　on　different　kinds　of　testing　images　demonstrate　that　the　proposed　one-class　method　can　achieve　a　good　performance　on　both　normal　images　and　pornographic　images　with　lower　time　consumption.　©　2012　ACM.