According　to　the　complex　manifestation　of　human　facial　expression　in　realistic　environment,　occlusion　problem　has　become　a　new　challenge　and　a　hot　spot　in　the　field　of　expression　recognition.　To　make　facial　expression　recognition　applied　in　broader　way,　the　main　work　is　to　increase　the　accuracy　under　different　partial　occlusion　with　feasible　robust,　which　is　limited　by　the　information　missing　and　insufficient　training　with　fewer　samples.　Therefore,　an　algorithm　with　a　deep　structure　has　been　proposed　in　this　paper　dealing　with　four　types　of　frequently　occurred　occlusion.　As　a　classic　method,　the　Gabor　filter　is　used　for　feature　extraction　at　first.　Then,　multi-layers　network　is　used　to　pre-train　the　training　data　samples,　with　re-describing　the　input　Gabor　features　in　complex　way　and　fine-tuning　the　weights　to　refine　the　learning　model.　The　experimental　results　on　JAFFE　database　show　that　the　proposed　method　is　valid　to　achieve　better　recognition　rate　especially　for　partial　occlusion　on　eyes　and　mouth.