A　new　face　detection　method　based　on　learning　is　proposed　in　this　paper,　it　has　three　properties:　First,　it　uses　not　only　the　local　facial　feature　but　also　the　global　facial　feature　to　design　weak　classifiers,　a　new　kind　of　global　facial　feature　called　as　the　unified　average　face　feature　(UAFF)　is　proposed;　Second,　it　uses　two　kinds　of　rectangle　feature　as　the　local　feature,　different　from　other　methods,　these　local　features　are　selected　and　calculated　only　in　the　partial　regions　of　face;　Third,　these　weak　classifiers　corresponding　to　the　global　facial　features　and　the　local　facial　features　are　combined　and　trained　by　our　novel　cascade　classifier　training　algorithm　to　construct　a　cascade　face　detector.　Because　of　these　properties,　our　face　detector　is　robust　and　generalizes　well.　Experimental　results　show　that,　with　a　small　number　of　features,　it　can　reach　higher　detection　rate　while　maintain　lower　false　alarm　rate.　Moreover,　it　can　detect　faces　with　partial　occlusion.　©　2004　IEEE.