Under　the　condition　of　multi-databases,　a　novel　algorithm　of　facial　expression　recognition　was　proposed　to　improve　the　robustness　of　traditional　semi-supervised　methods　dealing　with　individual　differences　in　facial　expression　recognition.　First,　the　regions　of　interest　of　facial　expression　images　were　determined　by　face　detection　and　facial　expression　features　were　extracted　using　Linear　Discriminant　Analysis.　Then　Transfer　Learning　Adaptive　Boosting　(TrAdaBoost)　algorithm　was　improved　as　semi-supervised　learning　method　for　multi-classification.　The　results　show　that　the　proposed　method　has　stronger　robustness　than　the　traditional　methods,　and　improves　the　facial　expression　recognition　rate　from　multiple　databases.　©　Springer　International　Publishing　2013.