The　existence　of　mixed　pixels　pose　difficulty　for　many　hyperspectral　image　applications.　In　this　paper,　we　propose　an　endmember　dictionary　based　algorithm　for　hyperspectral　image　unmixing.　A　material　class　based　endmember　dictionary　pre-trained　on　a　standard　spectral　library　is　used　as　the　set　of　endmembers　for　unmixing.　A　K-SVD　based　sparse　decomposition　algorithm　is　adopted　to　capture　the　abundances　of　the　endmembers.　By　training　the　endmembers　from　standard　spectral　library,　typicality　of　the　endmembers　are　improved　and　correlations　between　the　endmembers　are　reduced.　Experimental　results　show　that　the　proposed　algorithm　improves　performances　for　both　simulated　and　real　data,　especially　in　low　SNR　cases.