The　exemplar-based　image　inpainting　algorithm　consists　of　two　procedures:　patch　selection　and　patch　estimation.　In　this　paper,　we　focus　on　the　latter.　Conventional　exemplar-based　approaches　cannot　fully　exploit　the　information　existing　in　the　image,　which　makes　the　patch　estimation　unstable　and　produces　false　colors　and　artifacts.　In　this　paper,　we　propose　a　robust　patch　estimation　algorithm　for　exemplar-based　image　inpainting.　A　set　of　most　relevant　patches　is　extracted　as　reference　patches　from　the　known　region　of　the　image.　A　dictionary　is　learned　both　globally　and　adoptively　to　best　represent　the　target　patch,　which　is　then　approximated　under　the　assumption　that　similar　patches　admit　similar　decompositions.　Experiments　show　that　both　quantitative　and　qualitative　improvements　are　achieved　by　our　method.