In　this　paper,　we　address　the　problem　of　learning　multi-scale　sparse　representations　of　natural　images　using　structured　dictionaries.　Dictionaries　learned　by　traditional　algorithms　have　two　major　limitations:　lack　of　structure　and　fixed　size.　These　methods　treat　atoms　independently　from　each　other,　and　do　not　exploit　possible　relationships　between　them.　Fixed　size　of　atoms　restricts　the　flexibility　of　representing　natural　images,　which　usually　consist　of　complicated　structure　and　texture.　We　put　forward　a　novel　approach　to　learn　a　dictionary　by　performing　structured　sparse　coding　under　a　multi-scale　binary　tree　model　of　patches.　Atoms　of　different　sizes　are　laid　out　in　a　grouped　or　hierarchical　fashion,　which　can　be　fully　exploited　by　structured　sparsity　regularization　techniques.　Experiments　show　that　both　quantitative　and　qualitative　improvements　are　achieved　for　restoration　tasks.　It　is　worth　noting　that　our　approach　can　be　easily　integrated　into　existing　sparse　representation-based　applications　in　image　processing.　©　2013　CSREA　Press.　All　rights　reserved.