To　accelerate　the　speed　of　face　detection　and　get　low　false　alarm　rate,　an　approach　of　knowledge-based　face　detection　is　proposed.　It　integrates　the　skin-color　model　and　the　gravity-center　template.　In　the　process　of　rough　detection,　the　skin-color　model　is　used　to　segment　the　face　like　regions　from　any　input　image.　The　face-like　regions　are　further　checked　out　by　matching　the　gravity-center　template.　As　the　physical　structure　of　human　face　being　taken　into　careful　consideration,　a　dynamic　three　subsections　distribution　model　of　face,　is　proposed　and　used　to　establish　a　face　knowledge　base　by　analyzing　large　numbers　of　face　images　under　different　conditions.　All　the　face-like　regions　are　verified　if　they　are　genuine　human　faces　based　on　the　knowledge　base　of　faces.　The　experimental　results　show　that　this　approach　is　robust　for　human　face　images　under　complex　background,　different　sizes　and　certain　degree　of　rotation.